Table 13.

Advantages and limitations of the artificial neural networks employed for mechanical fault detection and fault prognosis

Publication	Advantages	Limitations
[31]	Proposed approach: early detection of fault features; diagnosis of machine’s health status under time-varying operation.	Algorithm: performance decreases when training and test data don’t share the same distribution.
[40]	Proposed approach: faster convergence; improved prediction performance; less computational complexity.	Algorithm: slow convergence speed; low precision; falls easily into local minimum; number of hidden layers difficult to determine.
[57]	Algorithm: fault tolerance; learns complex nonlinear relationships; strong generalization abilities.	Algorithm: lack of interpretability.
[81]	Proposed approach: knowledge sharing; generation of vast amounts of data through simulation of the entire product life cycle.	Algorithm: performance decreases when training and test data don’t share the same distribution.; poor performance in case of insufficient or low-quality training data.
[78]	Proposed Approach: improved classification performance; maintains temporal information and learns time-invariant features.	Not identified
[77]	Algorithm: learns complex nonlinear relationships; uncovers patterns in raw time series data.	Not identified
[55]	Proposed approach: online learning capability; autonomous structural evolution; capable of adapting to drifts in the input data; capable of learning under finitely/infinitely delayed label scenarios.	Not identified
[73]	Algorithm: capable of processing time series data; superior feature extraction capability; better forecasting ability.	Algorithm: choice of suitable hyperparameters is complex and affects the performance of the network;
	Proposed approach: can be adjusted to different types of machines and labels.	Proposed approach: using multiple ANNs can be a difficult task; might not be able to identify neighbour states; requires labelled data.
[79]	Proposed approach: capable of generating large volumes of fault data; avoids mode collapse; improved accuracy and efficiency.	Not identified
[66]	Algorithm: denoising effect; automatic extraction of meaningful features; capable of learning complex probability distributions; well-understood and stable; can be used for conditional data generation.	Not identified
	Proposed approach: estimation of machine’s health status under time-varying operations; capable of handling sparse industrial data; product-specific health index can be used for scheduling maintenance and production.