Abnormal vibration signal generation method and system

Abstract

This disclosure discloses a method and system for generating abnormal vibration signals. The method includes the following steps: A training dataset includes multiple actual vibration signals, and the training dataset includes labeled categories for each actual vibration signal. A classifier for identifying vibration anomalies is pre-trained based on the training dataset. A generative adversarial network model including a generator, a discriminator, and a classifier is trained based on the training dataset to obtain a trained generator and a trained classifier. The discriminator and classifier are updated synchronously based on a total loss value. An abnormal vibration signal generation task is performed using the generator, or a vibration signal classification task is performed using the classifier.

Claims

1. A method for generating abnormal vibration signals, the method comprising: The process involves: acquiring a training dataset on the vibration status of a mechanical device, wherein the training dataset includes multiple actual vibration signals and labeled with category tags for each of the multiple actual vibration signals; pre-training a classifier for identifying vibration anomalies based on the training dataset; training a Generative Adversarial Network (GAN) model comprising a generator, a discriminator, and the classifier based on the training dataset to obtain the trained generator and the trained classifier, wherein the discriminator and the classifier are updated synchronously based on a total loss value; and using the generator to perform an abnormal vibration signal generation task, or using the classifier to perform a vibration signal classification task.

2. The abnormal vibration signal generation method as described in claim 1, wherein the classifier includes an auxiliary physical model and a classification model, the plurality of actual vibration signals includes a first actual vibration signal, and the step of pre-training the classifier based on the training dataset includes: The first actual vibration signal is input into the classification model, and the probability of multiple categories relative to multiple abnormal categories is obtained by the classification model. Based on the multiple anomaly feature detection rules in the auxiliary physical model corresponding to the multiple anomaly categories, multiple anomaly scores of the first actual vibration signal relative to the multiple anomaly categories are determined; based on the multiple category probabilities and the multiple anomaly scores, a final classification probability for each of the multiple anomaly categories is determined; and based on the final classification probability for each of the multiple anomaly categories, an anomaly classification result for the first actual vibration signal is determined; and the anomaly classification result is compared with the labeled category tag of the first actual vibration signal to update the classification model.

3. The abnormal vibration signal generation method as described in claim 2, wherein the step of determining the final classification probability of each of the plurality of abnormal categories based on the plurality of category probabilities and the plurality of abnormal scores includes: Calculate the weighted sum of the category probability of one of the plurality of anomaly categories and the anomaly score to obtain the final classification probability of one of the plurality of anomaly categories.

4. The abnormal vibration signal generation method as described in claim 2, wherein the step of determining the multiple abnormal scores of the first actual vibration signal relative to the multiple abnormal categories according to the multiple abnormal feature detection rules in the auxiliary physical model corresponding to the multiple abnormal categories includes: The degree of matching between the time-domain or frequency-domain features of the first actual vibration signal and the various abnormal feature detection rules is determined, so as to determine the scores of each of the various abnormalities based on the degree of matching.

5. The abnormal vibration signal generation method as described in claim 1, wherein the plurality of actual vibration signals includes a first actual vibration signal, and the step of training the generative adversarial network model including the generator, the discriminator, and the classifier based on the training data set to obtain the trained generator and the trained classifier includes: The first actual vibration signal is processed by time-frequency conversion to generate a frequency domain signal; The generator generates a fused feature vector based on the frequency domain signal and a target anomaly category; and inputs the fused feature vector into the generator's generation network model, which then generates a fake vibration signal.

6. The abnormal vibration signal generation method as described in claim 5, wherein the forged vibration signal is a time-domain signal, and the generation network model is a U-net model architecture with skip connections omitted.

7. The abnormal vibration signal generation method as described in claim 5, wherein the step of generating the fused feature vector based on the frequency domain signal and the target abnormality category includes: A frequency domain feature vector is generated by extracting features from the frequency domain signal; a category feature vector is generated based on the target anomaly category; And the frequency domain feature vector and the category feature vector are fused to obtain the fused feature vector.

8. The abnormal vibration signal generation method as described in claim 5, wherein the step of training the generative adversarial network model including the generator, the discriminator, and the classifier based on the training dataset to obtain the trained generator and the trained classifier further includes: The fake vibration signal is input to the discriminator model of the discriminator so that the discriminator model can identify whether the fake vibration signal is real or fake. The fake vibration signal is input to the classifier, which classifies the fake vibration signal into a predicted anomaly category; the total loss value is generated based on the discrimination result of the discriminator and the classification result of the classifier; and the discriminator and the classifier are updated using the total loss value.

9. The abnormal vibration signal generation method as described in claim 8, wherein the step of generating the total loss value based on the discrimination result of the discriminator and the classification result of the classifier includes: Calculate an adversarial error loss based on the discrimination result of the discriminator; calculate a cross-entropy loss based on the classification result of the classifier; The total loss value is obtained by summing the adversarial error loss and the cross-entropy loss.

10. The abnormal vibration signal generation method as described in claim 1, wherein the steps of performing the abnormal vibration signal generation task using the generator or performing the vibration signal classification task using the classifier include: Collect a vibration signal during the operation of a piece of on-site mechanical equipment; The vibration signal is input to the generator to generate an extended abnormal vibration signal corresponding to a specific abnormality category; and the vibration signal is input to the classifier to predict one of the abnormality classification results corresponding to the vibration signal.

11. An abnormal vibration signal generation system, comprising: Storage device that records multiple instructions; The processor, connected to the storage device, is configured to execute the instructions to: acquire a training dataset relating to the vibration status of a mechanical device, wherein the training dataset includes multiple actual vibration signals and the training dataset includes labeled categories for each of the multiple actual vibration signals; pre-train a classifier for identifying vibration anomalies based on the training dataset; train a Generative Adversarial Network (GAN) model including a generator, a discriminator, and the classifier based on the training dataset to obtain the trained generator and the trained classifier, wherein the discriminator and the classifier are updated synchronously based on a total loss value; and perform an abnormal vibration signal generation task using the generator, or perform a vibration signal classification task using the classifier.

Images