Algorithm 1. General Steps of RDPN FCDAE Model.

Input: Bearing datasets consisting of rolling bearing vibration signal samples

Output: Diagnostic results, including classification accuracy, F-metrics, confusion matrix, feature segmentation, etc.

Step 1: data collection

1.1. Collect the vibration signal of rolling bearing through sensors, data acquisition system and host computer system.

Step 2: data preprocessing

2.1. Sort the collected original rolling bearing vibration signal datasets to obtain sample sets for each category, and add an additive white Gaussian noise with appropriate signal-to-noise ratio to each sample set signal to be closer to the actual working conditions.

2.2. Use CWT to convert the signal into a time-frequency image and express it in grayscale.

2.3. Normalize the obtained image, divide it into training set, verification set and test set according to a certain proportion, and do a good job of classification labeling.

Step 3: encoding-decoding training

3.1. Build the RDPN-FCDAE network structure, and randomly initialize the network weights and offsets in the form of Gaussian distribution.

3.2. Encoding and decoding training of the proposed RDPN-FCDAE model on the training datasets in an unsupervised manner.

Step 4: Fine-tune the parameters

4.1. Remove the decoding network of RDPN-FCDAE model.

4.2. Add GAP layer and Softmax classification layer on top of the coding network to form RDPFCN, and fine-tune the parameters through back propagation. This process is labeled.

Step 5: Verify algorithm performance

5.1. Verify the effectiveness of the algorithm on the test datasets and evaluate the diagnosis results.