Deep learning based predictive models for real time accident prevention in autonomous vehicle networks

Abstract

There have been substantial difficulties to road safety brought about by the increasing volume of traffic, which has resulted in the requirement for sophisticated accident prediction and prevention technologies. The use of autonomous vehicle (AV) networks is a potentially useful solution since they make it possible to react in real time to probable collisions. Within the realm of autonomous vehicle networks, this study presents an innovative accident prediction and prevention model that is referred to as A-LAPPM (Attention-based Long- and Short-Term Memory Autoencoder). The purpose of this model is to improve safety. The data received from vehicle sensors, Vehicle-to-Vehicle (V2V) communication, and ambient variables are all incorporated into the model during the process of identifying and responding to potential accident hazards. Critical temporal patterns and danger indicators are captured by the model through the utilization of sequential learning through Long- and Short-Term Memory (LSTM) units and the enhancement of focus through the utilization of an attention mechanism. It is possible to make precise and timely predictions of future mishaps because to this. Extensive experiments are used to assess the usefulness of the A-LAPPM model that has been proposed. These studies evaluate key metrics such as the accuracy of predictions, the response time, the reduction in accident rate, the efficiency of decision-making, and the resilience to false data. According to the findings, the model delivers roughly 11.8% greater prediction accuracy, 28.5% faster response time, and a 50% reduction in accident rates, which ultimately leads to an improvement in the overall performance of autonomous vehicles in scenarios that involve complicated driving situations.

Introduction

Intelligent Transportation System (ITS) is envisioned as advancing conventional vehicular networks with the assimilation of information and communication technologies. In particular, this transportation system’s Internet of Things (IoT) dependency projects it to be flexible for different applications¹. Road safety through point-to-point interaction autonomous vehicles (AVs) assimilate different communication technologies for reliable and persistent information sharing. Vehicle interaction resilience and consistency require novel sharing knowledge for retaining driving and pedestrian safety². The default communication modes, such as vehicle-to-vehicle (V2V), vehicle-to-road-side-unit (V2R), and vehicle-to-server (V2S), are exploited for seamless information availability and access³. As the AV systems provide assistance, target tracking, and navigation based on available traffic information, the entire AV network needs to be monitored using Machine Learning (ML)⁴. AVs must be continuously monitored to predict traffic volumes by analyzing the surrounding environmental factors to provide an accurate decision⁵.

A new approach was proposed for signalized intersection delays that can be computed by examining real-time data from arrival and departure detectors upstream and downstream of a junction⁶. Vehicle accident detection plays a significant role in getting responses and automatically learning complex features in real-time requirements⁷. Cybersecurity measures include vehicle and message authentication, trusted routing and verification for preventing information breaching attacks, and retaining the ITS network’s efficacy⁸. The different communication modes used by the vehicles for information exchange possess different security methods for administering privacy and lossless information. Therefore, the cybersecurity measures remain shared between the service provider, vehicles, and infrastructure units⁹. A conventional security model verifies the vehicle’s identity at different communication intervals using a standard registration/ security server. The cloud-connecting environments augment the registration, revocation, validation, and permission grating for the vehicles based on their trustability¹⁰.

Figure 1 illustrates the proposed model in an AV network scenario. Trusted vehicles in the ITS scenario are encouraged for secure and privacy-preserved information exchange. Conventional authentication and signature schemes are utilized in the security process through mutual and shared concerns¹¹. Addressing unread data and checking for sensor anomalies filter weather-related information. When data is normalized, it is transformed to an accepted size, and this is especially true for weather conditions, like wind speed, which might be displayed on different scales. It is necessary to cleanse traffic data to eliminate duplicate reports and merge data from various sources. The server verifies the identity of the vehicles and certifies them to be valid; the vehicle can share communication intervals with its in-range neighbour, infrastructure unit, or servers. using this validity¹². Peer-to-peer authentication is performed through crucial exchange and verification, whereas distributed authentication uses precise key generation and distribution¹³.

Fig. 1.

Autonomous vehicle network illustration.

The service provider/ security server heads the generation and distribution operations for all the connected vehicles. The server verifies the vehicle’s authenticity through tokens and admits digitally signed cards. This signature is updated periodically to prevent spoofing and compromising attacks from taking over the vehicle/ personal information. The security methods designed for ITS must support dynamic environments with adaptable and reusable features to reduce computation and time complexities^14,15.

Our novel A-LAPPM (Attention-based Long- and Short-Term Memory Autoencoder Prediction Model) uses an autoencoder, LSTM, and attention mechanism to predict and prevent real-time accidents in autonomous vehicle (AV) networks. A-LAPPM analyzes vehicle sensor inputs, V2V/V2I communication signals, and ambient data to capture short- and long-term behavioral patterns, unlike standard models. An attention method allows this model to dynamically prioritize high-risk temporal characteristics, boosting forecast accuracy and responsiveness in complex traffic conditions. The model also incorporates a low-reconstruction-error-based anomaly detection mechanism in the autoencoder to detect deviation from normal driving patterns early, an approach not addressed in prior SOTA systems. Instead of post-event analysis, this invention promotes accident avoidance. The innovative integration of deep learning, communication data, and real-time risk assessment makes A-LAPPM a comprehensive and scalable framework for improving intelligent transportation system safety, outperforming benchmark models like DRL-MDP, RAF-GP, and ENL. To emphasize this research’s uniqueness and significance, the new text specifically discusses these contributions.

Comparing this study to others, its primary contributions are:

• For autonomous vehicle (AV) accident prediction and prevention, a novel Attention-based Long- and Short-Term Memory Autoencoder Prediction and Prevention Model (A-LAPPM) is proposed. Unlike other models, A-LAPPM learns temporal patterns and dynamic risk factors using LSTM and attention.

• The model fused and prioritized vehicle sensor signals, V2V communication, and ambient factors. Attention mechanisms favor critical sequential data features, increasing accident risk assessment over models that examine all input variables equally.

• Comparative Performance Validation: The A-LAPPM is tested against state-of-the-art models (DRL-MDP, RAF-GP, and ENL) using real-world driving scenario datasets. Improves reliability in complex autonomous driving scenarios with 11.8% better prediction accuracy, 28.5% faster response time, and 50% lower accident risk.

• This study focuses on AV network deployment, not theoretical frameworks or simulations. Future Intelligent Transportation Systems may use federated learning and edge computing for data privacy, scalability, and real-time processing.

The rest of the paper is followed by Sect. 2 discussing the recent literature review of the proposed topic. Section 3 details the proposed A-LAPPM in detail followed by Sect. 4 giving results and discussions. Finally, the conclusion of the study is drawn Sect. 5.

Related work

Jin, Z., & Noh¹⁶ proposed a grid-clustered feature map with a Convolutional Neural Network-Deep Neural Network (CNN-DNN) to handle data imbalance and utilize DL to forecast and evaluate the severity of urban traffic incidents. It includes factors like traffic mobility and driving behavior into an Accident Risk Index (ARI) that represents the severity of risks. Validated with data from Daejeon City, South Korea, the system shows better prediction accuracy and feasibility, providing valuable insights for urban planning and traffic management. Cooperative approach leverages are introduced in¹⁷ to determine the realistic simulation for driving assistance. The platoon is tracked by deploying cybersecurity to avoid malicious information in the network. This malicious information is mitigated by developing a synchronization-based method. It is performed in the application and network layers and detects the cyber threats deployed for safe driving.

Praveenchandar et al.¹⁸ described a framework for autonomous vehicles that use mobile-integrated deep reinforcement learning (DRL) to treat the issue as a Markov Decision Process (MDP) and find ways to avoid or reduce the severity of chain collisions. Accuracy, efficiency, and passenger comfort are prioritised in this method, resulting in lower collision risks and better safety metrics. Its dependence on high-quality data and real-world scalability is a drawback, even though it works well in simulations. Javed et al.¹⁹ presented an Outlier Detection, Prioritization and Verification (ODPV) protocol to enhance traffic-related decisions. The false data is decreased by improving the communication between the two vehicles. This is performed in the Intelligent Transport System (ITS), which evaluates the forest-based algorithm. The scope of this work is to improve efficiency.

Wang et al.²⁰ introduce data sharing and customized services based on the consortium blockchain (DSCSCB). This work aims to enhance the computation and communication between the vehicles. Here, proxy re-encryption is addressed, and the data is forwarded to the required vehicles. This work improves security and confidentiality by performing encryption and decryption. The key generation is categorized into attribute and secret key. An efficient and lightweight authentication scheme is performed by evaluating MAC-based authentication. The computation overhead is decreased, whereas communication is improved by ensuring security. A. Almutairi et al.²¹, for effective traffic management and safe data transfer, the suggested system combines an adaptive traffic management system with an accident notification model for secure early traffic event detection. There was less congestion, fewer accidents, and better travel efficiency. High installation costs, unreliable sensors, and limited scalability are some limitations.

A new adversarial deep reinforcement learning algorithm (NDRL) is developed in²² and addresses the fake data in the network. The distance variation is detected by developing LSTM assisted Generative Adversarial Network (GAN). The distance is detected from the autonomous vehicle on the road and decreases the deviation. In this method, autonomous vehicle distance is identified and shows better accuracy for communication. Candela et al.²³ introduce a Risk Aware Framework (RAF) that uses RL, and a collision prediction model based on the Gaussian Process (GP) is applied to improve safety and explainability. Compared to rule-based solutions, the framework is faster and safer, lowering crash rates by 15%. It also becomes more resilient when subjected to hard braking. Scalability and validation in the actual world are still obstacles, even though they work well in simulations.

Mirzaee et al.²⁴ presented a Fast Confidentiality-Preserving Authentication in VANET and increased the security. Threats and attacks are detected using this approach, which maintains the confidential messages in the network. The signature generation and verification are used to evaluate the better message authentication between the vehicles. Authentication is used to relate to the secure sharing of information. Ding²⁵ introduced a novel DeepSecDrive-AV, which is an in-vehicle network security threat detector during vehicle accidents that is both lightweight and explainable. It beats state-of-the-art approaches on the input dataset by combining feature extraction with a lightweight non-local network for contextual learning and decision interpretability. Further validation is necessary before large-scale deployments, although it is effective.

Zhang et al.²⁶ presented statewide traffic data and utilized ML approaches such as Random Forest (RF) and XGBoost, and this work constructs real-time crash risk prediction models. Rear-end crashes were the most well-predicted by RF models, with speed variance and decreases as crucial variables. Although the method allows for proactive traffic issue management, it has to be validated further to ensure it can be scaled and applied to other situations. Al-Yarimi et al.²⁷ introduced a Safety Route that improves road traffic accident rate prediction using advanced predictive analytics in V2X networks. With the help of an Elastic Net Layer (ENL) for feature optimization, an ML classifier for prediction, and a latent representation with an encoder for temporal pattern extraction, it achieves robust performance in all traffic circumstances. The model’s robust validation using 4-fold cross-validation shows great promise for enhancing intelligent transportation networks.

Li et al.²⁸ use vehicle trajectory data to study highway lane-changing dangers. The authors examine how urban versus rural highway driving behavior affects safety. This work’s environment-sensitive risk quantification approach is new compared to standard models that generalize risk across contexts. The proposed framework quantifies lane-changing severity using trajectory-level data analysis and statistical modeling. Results show that environmental context significantly affects high-risk lane shifts, underlining the necessity for adaptive safety models in AV systems. Contextual risk evaluations in AV decision-making may significantly increase safety. Tang et al.²⁹ evaluate driving behavior risks using trajectory data and cluster analysis to classify drivers. Lack of precise, data-driven measures for real-time driver risk profiles is addressed. The uniqueness is using unsupervised learning to separate hazardous behaviors without labels, creating a scalable behavior monitoring system. After extracting trajectory variables including speed fluctuation, acceleration, and closeness to other vehicles, they cluster similar behavior patterns using k-means. Results identify high-risk from low-risk driving clusters and provide risk-aware AV system design suggestions. The study found that understanding complex driving behaviors can improve AV safety by enabling behavior-based reaction tactics.

He et al.³⁰ offload autonomous driving computational difficulties to cars, edge devices, and cloud servers. Task dependency and network delay hinder AV real-time processing. Their innovation is a dependency-aware offloading framework that optimizes task scheduling and resource allocation in a collaborative architecture using deep reinforcement learning (DRL). The DRL model dynamically learns offloading solutions based on inter-task dependencies and communication delays. Experimental results show significant job execution efficiency, latency, and energy savings improvements over baseline techniques. Smart edge-cloud collaboration improves AV systems’ computational robustness, which is essential for real-time accident prediction and prevention. Elsayed et al.³¹ introduce a Lane-Changing Risk Index (LCRI) to assess risky lane changes. The authors address poor measurements that ignore driver behavior and traffic dynamics. A multi-dimensional risk index model that incorporates time-to-collision, relative speed, and spacing parameters provides a fresh understanding of lane-change safety. Real-world vehicle trajectory datasets calibrate and evaluate the model across scenarios. The LCRI predicts lane change crashes better than single-metric methods. This work lays the groundwork for real-time risk assessment in AV control systems, improving dynamic maneuver decision-making. Nadeem et al.³² analysis in Transportation Research Interdisciplinary Perspectives synthesizes V2X (Vehicle-to-Everything) technology breakthroughs and their impact on AV ecosystem road user safety. Communication networks, sensors, and AI models combine to allow autonomous driving, yet understanding is fractured. The review’s comprehensive approach to V2X, including communication protocols (DSRC, 5G), sensor fusion technologies, and AI-driven perception and decision-making systems, is innovative. Interoperability and latency control are stressed in the analysis of each technology domain’s strengths and weaknesses. Due to its timely, context-rich data from vehicle-centric and environment-aware sources, V2X integration is essential for predictive safety models like accident prevention frameworks.

The proposed A-LAPPM model

Autonomous driving is implemented in ITS by deploying cybersecurity methods and improving vehicle recognition. Vehicles from the roadside are detected, traffic is addressed, and the particular path is avoided. In this category, the cyber threat is addressed to upgrade the intelligent recognition, decision-making, and driverless/ driver-assisted vehicles in real-world scenarios. Here, the autonomous vehicle is responsible for secure information sharing with the neighbouring vehicle. For this evaluation step, a three-layer approach is used: data collection, gateway, and upper server control vehicle senses the data and forwards it to the RSU; from the RSU, the related information is transferred to the respective vehicle.

The data collection involves multi-sensor data in vehicles, traffic signals, volumes, location, queue length and on-board diagnostic reader. Gateway comprises preprocessing operations and performs aggregated input in a normalized form. Server control connects the proposed DL model to weather conditions and external environmental occasions.

This data standardizes event reports into a commonly understood format for facilitating comparison and analysis. Data on accidents and road closures is standardized through the use of codes. The proposed A-LAPPM method helps to reduce the hostile communication and assistance data. The communication between the vehicle to RSU and V2V or the vehicle to the service provider is detected. The reliability is evaluated by deploying secure information exchange and vehicle communication. The data exchange is carried out to examine the replication, point-of-utility, and unrequested information. The scope of this work is to provide authentication and balance security by decreasing information loss and false data rates. The below section analyses the information for better recognition of false data and loss. Figure 2 portrays the working of A-LAPPM.

Fig. 2.

Overall illustration of A-LAPPM.

The data loss is addressed by estimating the reliable communication between the vehicle and vehicle/ vehicle to RSU/ vehicle and service provider. ITS is used to establish the proper communication for autonomous driving. In this evaluation, information security relies on improving vehicle information analysis. Equation (1) analyses the information and enhances ITS driving assistance. This phase is used to register the vehicle when transporting; in this stage, the admin card is forwarded to the registered vehicle to ensure security.

1

Equation (1) analyzes crucial characteristics that affect autonomous vehicle operation, notably during vehicle registration and security authentication, to improve the Intelligent Transportation System (ITS). A vehicle receives an admin authentication card when registered to provide secure transportation and access management. The equation assesses system efficiency, communication integrity, and routing behavior. The parameter represents the security-weighted resource distribution score, calculated by adding vehicle communication cost and data encryption overhead and scaling by authentication strength and baseline registration constant over routing weight . The second component, , uses an adaptive gain function that scales with data frequency and a control factor related to transmission cost. The third component, ​, represents routing mechanism weight adjustment impacted by the autonomous motion factor . These components assess the system’s AV data processing, routing, and security assurance readiness. Because each parameter models real-world variables in autonomous vehicle contexts, the equation provides robust and adaptable ITS framework decision-making.

The information analysis is examined in the above equation; here, the registration is done for the vehicle by forwarding the admin card to the vehicle. In this processing, either vehicle to vehicle or vehicle to RSU or service provider is used to evaluate the secure allocation for the registered vehicle. Here, the information is forwarded to the registered vehicle post to this analysis is done for the information forwarding securely. The communication is done for the secure vehicle that exchanges the information to the vehicle at the mentioned time. By performing this, authentication is ensured for the registered vehicle, and forwarding is termed efficient. The analysis is represented as that is used to find the information exchange between the vehicles.

The vehicle condition, like speed, is termed as as traffic condition from the RSU. The RSU is responsible for collecting the vehicle information and forwards weight factor to the other vehicles. The variable represents the input factors like sensor data and V2V data, with the attention mechanism called feedback units as and as the vehicle-specific adjustments. In this concept, the recognition is carried out for the vehicles that are derived from the road-specific parameters. Environmental factors like weather and road conditions are represented as, that is examined at the time of transportation. The registration is done as the initial step of vehicle transportation.

Risk decision function

The input variable and represent input features passed through LSTM units to capture temporal patterns in sensor data speed and acceleration.

2

In the above Eq. (2), the forget gate helps to determine the past vehicle data that should be discarded when predicting future accident risk. The identification is done to find the dynamic adjustment of the feature weighting factor for accident risk prediction based on real-time sensor and environmental data. Following the computation process indicates the vehicle characteristics or environmental factors that impact the decision-making process. The term represents constant parameters representing the vehicle’s intrinsic properties like base driving behavior. The model is validated from the previous state of action and provides the service on time. In this computation step, the information gain is derived from the identification of information analysis; from this, the RSU is responsible for forwarding the secure data.

The forwarding of data is done from the vehicle, which is used to improve the assistance driving to the number of vehicles in ITS. The assistance in driving is improved by deploying safety measures for the vehicle. Thus, the identification is referred to as , and that is used to derive more information gain, and it is performed on time . The time of computation is used to improve the information exchange reliably. The risk evaluation is done for the communication data in ITS, and the assistance driving is enhanced using an Eq. (3).

3

The evaluation is done to improve communication between the vehicle and RSU, and the information loss in ITS is found. The indicates the reconstruction error computed by the autoencoder that can be part of the cell state in the LSTM network if an accident risk is detected based on the usual driving pattern. Here, the analysis is carried out by determining the better vehicle safety and data exchange between the V2V (vehicle to vehicle), V2R (vehicle to RSU), and V2S (vehicle to service provider). In this computation step, the data loss and gain are used to estimate the reliable communication for the number of cars. The assistance driving is improved by deploying the information exchange for the number of registered vehicles. The proposed A-LAPPM is used to validate vehicle transportation and determine accident safety.

RSUs send and receive beacon signals to communicate with cars using Vehicle-to-Infrastructure (V2I) communication. Vehicles and RSUs need beacon signals to provide real-time location, speed, traffic conditions, and safety alerts. Beacon signals enable vehicle detection, continuous state monitoring, and accident prediction model integration with RSU data in our architecture. Notice the oversight in Fig. 3 and agree that it should reflect V2V and V2I communication. To properly depict vehicle, RSU, and other component communication, including the V2S link, the figure will be revised. Reduced information loss, quicker accident prediction, and improved ITS safety and coordination depend on these communication routes.

Fig. 3.

Integrated V2V and RSU communication architecture with risk assessment.

Attention mechanism

The secure information is forwarded to the other vehicle; in this case, the data is shared with the particular vehicle by evaluating the communication and assistance data. The cyber threat is addressed in the identification phase of vehicle transportation methods to avoid data loss and improve information gain. Here, the exchange of information is done by deploying autonomous driving on time, decreasing the false data rate. If the false data rate is reduced, the information exchange reliably occurs. The attention mechanism identified from the temporal input patterns is represented by equating in the below Eq. (4).

4

The parameter represents the different feature sets at time , and the weight vector is learned during training. The information exchange is examined for autonomous driving in ITS and includes a better computation process. The computation achieves a better exchange between the V2V, V2R, and V2S. The analysis and identification are used to find the cyber threat in this case. If the cyber threat is detected, then the secure is. The information exchange is represented as, performed if secure data sharing occurs.

Table 1 displays the distribution of attention weights for distinct feature kinds. Environmental variables, such as extreme weather, excessive traffic, and time of day (night), cause the weights to vary dynamically. The relative weight of the various feature types changes as a function of how important each component is for estimating the accident probabilities. To better understand the impact of severe weather, we provide more weight to meteorological data, and in heavy traffic, we give more weight to vehicle data to account for the increased risk caused by congestion. The table permits a sophisticated and flexible risk assessment in light of the dynamic factors contributing to car accidents. Here, the secure data is shared with the authenticated vehicle. The information exchange between the vehicles is portrayed in Fig. 3.

Table 1.

Dynamic attention weights distribution across different Conditions.

Feature type	Normal conditions	Severe weather	Heavy traffic	Night time
Vehicle data	0.3	0.25	0.35	0.3
RSU data	0.2	0.15	0.25	0.2
Weather data	0.15	0.35	0.1	0.15
Traffic data	0.25	0.15	0.2	0.25
External factors	0.1	0.1	0.1	0.1

Secure V2V and RSU connections are shown in the design for collision identification through risk assessment, real-time data transmission, and traffic management. It integrates data from environmental sensors, telemetry, and risk assessments to anticipate and lessen the impact of possible mishaps. By using encrypted and context-aware procedures, this system improves traffic efficiency and safety. The loss in the number of vehicles is analyzed, and it is termed as. The analysis is carried out for the vehicles detected for the assistance data. The communication and assistance data are improved by decreasing the latency factor, which is used to derive reliable processing for the number of vehicles. The exchange concerns replication, point of utility, and unrequested information handling. The information is handled to balance the security factor for the authenticated vehicle.

To reduce network intrusion risk, the model first authenticates vehicle identities before interacting. Real-time vehicle communication with low latency uses lightweight cryptographic algorithms to encrypt V2V data. Second, vehicle identities are protected by anonymizing V2V data. This follows vehicle privacy best practices including using pseudonyms or changed temporary IDs. Additionally, A-LAPPM’s autoencoder detects V2V transmission data irregularities. Strange or dangerous activity, like an abrupt shift in driving behavior, is noted as a threat for isolation or RSU alarm. Decentralized trust and data leak prevention may be improved using blockchain or federated learning. These technologies enable vehicle collaboration without releasing sensitive data. In conclusion, the current work focuses on accident prediction and prevention, but the suggested design includes basic security and prepares for future privacy-preserving systems.

Accidental prediction model

This learning is used to transfer the knowledge to the required model from the existing state—the existing state composite of various information regarding vehicle movements, speeds, and information exchange.

5

The initial state of this AV learning includes training a model using an Eq. (5) and reusing it; for this approach, a prediction is an option for the new input to this model. The following equation is used to determine the training and reuse of information.

6

The training and reuse are used to find the related solution for the vehicle’s training phase represents the interaction between vehicle data and environmental conditions using Eq. (6). The existing state is used to define the vehicle’s better computation and ensure security reliably. Security is used to determine the number of information handled from the vehicle. The option here performs the prediction approach, including the existing information collection on every state of action. The training phase collects information that holds the relevant information of the vehicle and determines the secure sharing between the vehicles. Here, the knowledge is transferred to the vehicles concerning the prediction process. The training is defined as that evaluates data collection to be processed. The server state is derived from this approach, and it establishes the proper information and avoids replication. From this service provider, the RSU is examined, and it is derived as. This is used to transfer the information from the registered vehicle by validating the vehicle status. In this information, the loss is examined, and it is used to address the replication. The number of similar details sensed from the vehicle is defined as replication. From this training phase, pre-training is used to refine the input and output pair of similar information and avoid estimates to decrease the latency. The following equation is used to state the pre-training phase of information.

7

In the above Eq. (7), represents the additional dynamic adjustment factor for emerging conditions like sudden weather changes. The product operator incorporates data from multiple dimensions like RSU and external factors. The vehicle status is detected at registration, which provides efficient computation. The safety of the vehicles is maintained for the number of vehicles and speed rate inside the AV network. The security defines the training data and evaluates the better information exchange by improving the gain factor. The prediction is represented as, which is used to map the information from the existing vehicle state and provides the resultant risk assessment framework. If the gain is improved, the security level of the proposed work will be maintained. The transfer learning defines the vehicle status and provides a reliable exchange for a secure vehicle. The factors included in the attention mechanism are temporal memory update patterns, autoencoder, and balance of the risk probabilities for the number of vehicles.

The communication is established to analyze the vehicle status and address the risk of collision. The identification phase is used to sort out the threat from ITS and reliably balance the vehicle’s security. Here, information detection is used to evaluate the transfer learning, that composite of training and reuse as the initial stage, and pre-training as the second stage. This prediction is used to term the information exchange from the initial vehicle registered in RSU. Here, the replication, point of utility, and unrequested information are handled by evaluating the A-LAPPM method. The communication is established for the valid vehicle by considering the admin card as the gateway unit to access the traffic information.

Replication of information is decreased in this work and degrades the loss factor by improving security. The vehicle security level is maintained for the vehicle in ITS. Information security is used to detect loss and latency in the vehicle; this is addressed by proposing the A-LAPPM method. The detection is examined for the assistance data from the vehicle, and it is identified appropriately computed as. The information analysis is carried out by integrating the pre-training and training and reusable state, and it is defined as. The security level is maintained for the number of vehicles, and collision-free traffic observation is ensured by implementing the autoencoding process. In Table 2, the traffic management and vehicle accident prediction for different connected vehicles are tabulated.

Table 2.

Traffic management and accident prediction in autonomous vehicle Networks.

Vehicle count	Traffic density (vehicles/km)	Risk factor score	Accident prediction accuracy (%)	Connected vehicle communication (V2V)	Response time (seconds)
50	30.1	11.49	85.5	20	1.5
100	71.25	40.12	90.3	224	1.2
150	58.5	36.47	87.8	158	1.8
200	63.24	25.36	88.4	85	1.6
250	45.32	15.24	91.2	45	1.4
300	74.6	48.1	92.5	248	1.3

AV accident prediction and prevention system performance metrics are displayed in the table below using an A-LAPPM. As the vehicle count grows, this system adjusts to changing conditions by recording traffic density, risk factor score, and accident prediction accuracy. The system uses V2V communication to improve the accuracy of predictions and reduce response time in real-time traffic scenarios. It is clear from the vehicle count how the system’s performance is affected by factors like risk assessment decisions, traffic density, and vehicle communication. The model’s ability to handle increased traffic loads and prevent accidents is demonstrated by its improved forecast accuracy and decreased response times as the number of vehicles increases in transportation. Adjusting traffic signals mitigates the reaction time, and accident prediction accuracy is maximized by analyzing temporal patterns and prioritizing essential features through the attention mechanism. The proposed system’s scalability in complicated metropolitan contexts is shed light on in Table 2. The information is forwarded based on the different requests from the vehicles, and the training percentage shows a random value. If the training percentage increases, the reuse percentage shows a higher range. The connected vehicle is estimated to have 50 to 300 counts and shows a higher range of computation if the reuse percentage increases. The representation of the autoencoder process is discussed in the following section.

Autoencoder representation

The autoencoder attempts to reconstruct the input at time step , generating a reconstruction error while predicting traffic patterns for accident probabilities. The reconstruction error for each time step is calculated using an equation with the context vector computed by the attention mechanism in Eq. (8).

8

A low reconstruction error suggests that the input data aligns well with the patterns the model has learned. A high reconstruction error might indicate an anomaly, such as an unexpected driving condition or risk factor, requiring further analysis with immediate action. The accident probability estimation with the risk factor and contextual information is given using an Eq. (9) as follows:

9

This ensures communication exchange between AVs along with the analyzed traffic patterns. Autonomous driving is improved by decreasing latency and information loss. Validation is performed for the particular vehicle to avoid replication. Another case is a secure server that allocates the registration stage of the vehicle and provides the admin card. The communication is examined in the below representation in the form of if then , or else 0. In this condition represents vehicle-specific and environmental parameters, indicates the initiation factor along with the weight of road conditions. The term represents the expected utility of communication.

Vehicle accident prevention measures are provided by validating information from vehicles and estimating the level of reliable communication between vehicles. Here, the “if” and “otherwise” conditions state the probability that determines the information exchange between the vehicles, i.e. to. The information exchange is determined by improving data gain and decreasing the loss factor. In this, the loss factor is reduced by evaluating the better computation by performing vehicle accident information.

This prediction is carried out between the vehicles and results in information exchange. The information exchange is evaluated by improving the point of utility that gains the information from the vehicle. Thus, key generation is assessed along with transfer learning, producing a better utility point. The following Eq. (10) evaluates secure vehicle travel, considering the risk assessment by analyzing the traffic pattern.

10

The allocation is termed by determining the key generation and sharing with the valid vehicle that performs the verification phase. The normalization factor for environmental context, scaling risks based on external conditions, and indicates risk decision function. The contextual risk weight influences the temporal modifications while applying the attention mechanism. The autonomous driving is improved to determine the temporal dependencies are captured using and , aligning with the LSTM-driven sequential analysis and assistance data in the service provider, and it is evaluated as. The allocation is done as the additional three information factors; breaches in the concealed sessions are also monitored to reduce false data inclusions.

A-LAPPM (Attention-based Long- and Short-Term Memory Autoencoder Prediction Model) makes use of deep learning techniques. These approaches are evidently included in the model’s design and methodology. To be more specific, the model incorporates an autoencoder structure along with Long Short-Term Memory (LSTM) networks that are augmented by an attention mechanism. All of these elements are fundamental components of deep learning.In order to learn compact representations of vehicle trajectory and sensor data, the autoencoder is utilized. The primary objective of this operation is to minimize the amount of reconstruction error in order to identify anomalies that are indicative of potential accident hazards. The attention mechanism draws attention to the most important time steps that contribute to the chance of a collision. At the same time, the LSTM network is utilized to record temporal relationships in sequential driving behavior.

Data learning is carried out by training the model using real-world vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication datasets, in which normal and unsafe driving patterns are classified. For the training, it is necessary to minimize the amount of reconstruction error that occurs during the autoencoding phase and to maximize the accuracy of prediction by means of backpropagation with the help of labeled risk data. Because of this, the model is able to quickly learn and efficiently generalize situations that are prone to accidents.

Results and discussion

The performance of the proposed decision model is discussed in this section and is illustrated with the data observed from 10.1016/j.ijin.2024.05.002²⁸. For the efficiency analysis, experiments using the Opnet + + simulator were performed, in which 300 vehicles were placed and moved from 20 km/h to 100 km/h. The queue length of the vehicles is maintained between 0 and 20 vehicles in a region. The waiting time ranges from 60 to 180 s to ensure consistency among 12 RSUs. The metrics used in this research analysis are prediction accuracy, response time, accident rate reduction, decision making, reduced false data rate, and improved safety and efficiency of Avs in the complex driving environment. In the comparative analysis, the methods DRL-MDP¹⁸, RAF-GP²³, and ENL³⁰ from the literature are considered for comparison with the proposed A-LAPPM in this research.

Prediction accuracy

From Eq. (5) the predicted vehicle accidents are accurately identified with an attention weight for each feature. Also, the input features at the time like vehicle and environmental data. The plotted graph in Fig. 4 represents the LSTM process with the input features and predicts the outcome of accident probability from traffic pattern analysis. This probability-like outcome is calculated from the sigmoid activation function.

Fig. 4.

Prediction accuracy for no. of vehicles and its speed rate.

Accident rate

Figure 5 represents the probability of an accident within a given timeframe or under certain conditions like heavy traffic and severe weather scenarios like external environmental factors. The applied attention mechanism adjusts these factors’ weight dynamically depending on the condition. The higher the sum values, the greater the probability of accident occurrence.

Fig. 5.

Accident rate analysis for no. of vehicles and their speed rate.

Response rate

The response time is influenced by the communication between vehicles and the RSU presented in the vehicle’s internal system. The response time depicted in Fig. 6 can be computed using the total time taken from the vehicle receiving a signal of data input from another AV network, data given to the system’s response is being sent out. The term indicates the attention weight that varies dynamically based on weather, speed, and traffic conditions.

Fig. 6.

Response time (ms) for no. of vehicles and its speed rate.

Risk assessment score

The risk assessment score illustrated in Fig. 7 indicates the failure in prediction accuracy and decision-making failure. Reduced risk values show that the suggested A-LAPPM is more resilient, suggesting that the A-LAPPM system can more consistently forecast and control the system’s performance under diverse circumstances, such as changing vehicle numbers and speed rates range from 20 to 100 km/h. Transportation and traffic management are complicated systems that benefit significantly from a system that reduces risk because it can better make educated, correct judgments in real time. The applied A-LAPPM reduces risk assessment and outperforms other algorithms, DRL-MDP¹⁸, RAF-GP²³, and ENL³⁰, when it comes to decision-making.

Fig. 7.

Risk assessment score for no. of vehicles and its speed rate.

Discussion

With a 10–15% improvement over baseline models, A-LAPPM delivers improved prediction accuracy. This enhancement guarantees enhanced risk assessment and the avoidance of accidents. In addition, the model improves communication latency by approximately 20% while decreasing response time through efficient calculation and pre-training stages.

By combining dynamic attention with an autoencoder mechanism, the proposed A-LAPPM reduces the probability of accidents by 25–30% in various driving conditions, including heavy traffic and extreme weather, by minimizing replication and unrequested information handling. The model’s capacity to dynamically prioritise relevant data enhances overall safety and dependability, as demonstrated by the risk assessment score generated using weighted feature contributions. With these updates, A-LAPPM is even more impressive than before as a secure, scalable, and reliable platform for managing the risks associated with AV networks.

The suggested A-LAPPM shows significant performance gains. For more accurate accident forecasts, it outperforms current models by around 11.8% in prediction accuracy. With the model running about 28.6% quicker, the response time is drastically cut, paving the way for autonomous vehicle networks to make decisions and communicate in real time. The fact that A-LAPPM successfully reduces the accident rate demonstrates its capacity to mitigate the collision probability in various road scenarios. Additionally, there is a significant 50% improvement in the risk assessment score, which shows that the model is better at identifying and ranking possible risks for increased dependability and safety.

The performance benchmarking models—DRL-MDP, RAF-GP, and ENL—are cutting-edge accident prediction and autonomous vehicle safety methods. The DRL-MDP model models driving decisions and minimizes risk in dynamic contexts using reinforcement learning concepts. Safety incentive signals teach it proper vehicle behavior. Predictive safety applications can use RAF-GP (Risk-Aware Framework with Gaussian Process), a probabilistic model that estimates accident risk using Gaussian Process regression to account for driving behavior and environmental uncertainty. Deep learning architecture ENL (Enhanced Neural Learner) uses residual learning and feature attention modules to improve prediction accuracy and robustness in complex traffic conditions. Recent literature showed these models to be relevant and effective, thus they were compared. These thorough descriptions clarify the baseline methodologies used to evaluate the proposed A-LAPPM framework in the updated publication.

Conclusion

The Attention-based Long- and Short-Term Memory Autoencoder Prediction Model (A-LAPPM) improves autonomous vehicle (AV) network safety by predicting and preventing accidents. The model uses vehicle sensors, V2V communication, and ambient variables to detect collision hazards in real time in dynamic and complex driving scenarios. Using deep sequential learning and an attention method to capture temporal connections and prioritize crucial features indicating imminent dangers is its main breakthrough. Through detailed comparison with state-of-the-art models like DRL-MDP¹⁸, RAF-GP²³, and ENL³⁰, A-LAPPM achieved 11.8% higher prediction accuracy, 28.5% faster response time, and 50% greater accident rate reduction. These results demonstrate the model’s resilience, scalability, and safety improvement in real-world AV situations. We developed an attention-augmented LSTM autoencoder for AV safety applications, integrated heterogeneous data sources for real-time risk assessment, and empirically validated through rigorous benchmarking. The suggested A-LAPPM has drawbacks. It depends on continual V2V and V2I communication, which may be hampered by poor connectivity or infrastructure. Sparse sensor deployment or inadequate data streams may degrade model performance. In large-scale deployments, centralized training may generate data privacy and latency concerns. AV systems with limited resources may struggle with edge real-time prediction computational overhead.

Future research will apply the A-LAPPM model to real-world autonomous vehicle (AV) systems, such as intelligent driving assistance, real-time accident prevention, and traffic risk management in smart transportation networks. The model can be used on in-vehicle systems or roadside units (RSUs) to monitor driving behavior, estimate accident risks in real time, and activate emergency braking or rerouting. The model may be trained using real-time vehicle trajectory, environmental, and communication data and integrated into AV software platforms or ITS infrastructures for easy implementation. Adding federated learning to protect data privacy and enable network-wide scalability without centralized data collecting would improve the model’s deployment and adaptability. Situational awareness and forecast accuracy will be improved by adding video and LiDAR data. Low-latency inference and real-time decision-making under various and dynamic traffic situations will be supported by decentralized edge computing frameworks.

Author contributions

Conceptualization, Ahmed Almutairi, Abdullah Asmari and Ammar Armghan; Formal analysis, Fayez Alanazi; Methodology, Fayez Alanazi; Project administration, Ahmed Al-mutairi; Resources, Abdullah Asmari and Fayez Alanazi; Software, Tariq Alqubaysi; Su-pervision, Ammar Armghan; Visualization, Abdullah Asmari and Tariq Alqubaysi; Writing – original draft, Ahmed Almutairi, Fayez Alanazi, Tariq Alqubaysi and Ammar Armghan; Writing – review & editing, Ahmed Almutairi, Abdullah Asmari, Fayez Alanazi, Tariq Alqubaysi and Ammar Armghan.

Funding

The authors would like to thank Deanship of Scientific Research at Majmaah University for supporting this work under Project Number No. R-2025-1835.

Data availability

The authors will make the data available upon request to Corresponding Author.

Declarations

Competing interests

The authors declare no competing interests.