Transmission system gear fault diagnosis method based on incremental learning

Abstract

The invention provides a transmission system gear fault diagnosis method based on incremental learning, and the method comprises the steps: firstly collecting transmission system monitoring data through a sensor, constructing an incremental data set, and dividing the incremental data set into fault diagnosis tasks of different stages; constructing an initial model, and training the model based on the initial fault data; in the test stage, a generalized entropy index is defined to judge whether the fault is a new fault; when new fault type data appears, a small number of samples are randomly selected from the old type data and fused with the new type data to construct a training set, and in order to avoid forgetting of the model to the old type fault data, knowledge distillation is carried out on attention weight and old type prediction output during training; and when the generalized entropy index continues to judge that an unknown fault occurs, repeating the network training step to update the model, and finally realizing incremental learning of a fault mode. According to the method, continuous updating of the fault diagnosis model can be realized for streaming data, the state monitoring capability of the transmission system is improved, and a guarantee is provided for the service safety of a mechanical system.

Description

Technical Field

[0001] The present invention relates to the field of mechanical equipment fault diagnosis, and particularly to a gear fault diagnosis method for a transmission system based on incremental learning. Background Art

[0002] As a key component in the transmission system, gears are responsible for high-efficiency rotational transmission. The harsh service environment makes the monitored gear vibration signals non-stationary and non-linear. Diagnostic methods based on traditional signal processing techniques are difficult to accurately identify fault modes and provide maintenance guidance for maintenance personnel. In recent years, intelligent diagnostic methods represented by deep learning have been widely used to extract features due to their powerful non-linear feature representation capabilities. Existing intelligent diagnostic methods usually assume that the number of fault categories remains unchanged during the service process of mechanical systems during deployment. However, in the actual industrial environment, the operating conditions are complex and changeable, resulting in accidental faults occurring from time to time. Most existing intelligent diagnostic methods do not consider the problem of new faults, and may result in misjudgment of faults after online deployment, bringing difficulties to operation and maintenance, and thus are difficult to be applied in practice.

[0003] Incremental learning can perform progressive training for streaming data, update the diagnostic model in the case of new faults, and retain the memory ability for old-class faults while learning the features of new classes. Nevertheless, how to find effective monitoring indicators to detect new faults still needs to be studied. At the same time, in order to avoid the phenomenon of data explosion, incremental learning only selects some typical examples of old fault samples for training in the new stage, resulting in the occurrence of catastrophic forgetting phenomenon, that is, the model loses the diagnostic ability for early-stage faults during the progressive training process. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a gear fault diagnosis method for a transmission system based on incremental learning. By using the generalized entropy index to monitor new faults, when new faults occur, it enters the incremental training stage, and optimizes the model parameters by designing the attention distillation loss and the task distillation loss, alleviates the catastrophic forgetting phenomenon of incremental learning, and realizes the update of the diagnostic model for streaming data.

[0005] For the above purpose, the present invention proposes the following technical route: A gear fault diagnosis method for a transmission system based on incremental learning, the steps are as follows:

[0006] Step 1: Collect the monitoring data of the gears in the transmission system, construct a data set, and then divide it into fault diagnosis tasks at different stages. Each fault diagnosis task at a different stage includes fault category data at different stages;

[0007] Step 2: Construct an initial model for gear fault diagnosis of the transmission system, and train the initial model using the initial fault category data in the fault diagnosis task to obtain a trained model;

[0008] Step 3: Input the test data into the trained model for fault diagnosis to determine whether it belongs to a new fault; when a new fault appears, regard the test data at this time as new fault category data, regard the existing fault category data as old fault category data, randomly select a small number of samples from the old fault category data, fuse them with the new fault category data to update the training set, and perform incremental learning to obtain a new model. Otherwise, continue to use the trained model (old model) to diagnose faults; it should be noted that due to the appearance of new fault category data, the model performs incremental learning, so there is a "new model"; correspondingly, the model before incremental learning is the "old model".

[0009] Furthermore, in the incremental learning stage, use the updated training set to train the model. To alleviate the catastrophic forgetting phenomenon of the model for old fault category data, design an attention distillation loss to maintain the consistency of the feature responses of the new model and the old model to old fault categories, so as to achieve incremental learning of fault patterns. The feature response is the learned attention weight, indicating the attention of the model to the corresponding feature.

[0010] Preferably, the monitoring data in Step 1 is collected by a vibration sensor.

[0011] Preferably, the monitoring data in Step 1 includes: vibration signals of gears in normal state and vibration signals of gears in fault state.

[0012] Preferably, the gears in fault state include: tooth surface wear, half tooth breakage, uniform wear, crack, tooth missing, and foreign object protrusion.

[0013] Preferably, the data set in Step 1 includes: training set, validation set, and test set, and construct the training set, validation set, and test set according to the ratio of 6:2:2, and divide them into fault diagnosis tasks in different stages. Each fault diagnosis task in a different stage includes fault category data in different stages;

[0014] Preferably, a channel attention module and a spatial attention module are added after each residual block of the initial model to enhance the model's attention to key fault features.

[0015] Preferably, for the test data in Step 3, first, the data set in seven states is divided into a training set and a test set. A sequential fault occurrence setting is made for the training set, and the samples of the test set remain unchanged.

[0016] Preferably, step 3 further includes: defining a generalized entropy index to determine whether the fault is a new fault;

[0017] When a new sample is inferred by the old model, the generalized entropy index defines the confidence that the sample belongs to the old fault category. When the index is higher than the threshold, the system considers that the sample belongs to the old fault, otherwise it considers that the sample belongs to the new fault.

[0018] Preferably, the threshold is set to the confidence at the 90% position after sorting the confidences in the training set from large to small.

[0019] Further, the gear vibration signal described in step 1 includes 7 states: normal, tooth surface wear, half tooth breakage, uniform wear, crack, tooth missing, and foreign object bulge. Define T incremental diagnosis tasks and divide the corresponding data sets , where represents different diagnostic stages, represents the initial diagnostic stage when and other stages are incremental learning stages; represents the data set at the t-th stage, where and represent the sample and the label respectively, represents the number of samples in this stage; represents the number of fault categories at the t-th stage.

[0020] Further, the model feature extractor in step 2 consists of a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, three residual module layers, and a max pooling layer; the classifier consists of a single fully connected layer; channel and spatial attention modules are designed after each residual module layer. Further, the initial model parameters in step 2 are trained and updated by cross-entropy loss, which can be described as:

[0021]

[0022] where is the true label while is the predicted label.

[0023] Further, in step 3, the generalized entropy index is used to detect new faults, and this index can be described as:

[0024]

[0025] where is a hyperparameter, is the predicted output after SoftMax normalization for the j-th sample. When a new sample is inferred by the old model, the generalized entropy metric defines the confidence of the sample belonging to the old fault class. When this metric is higher than the threshold, the system considers the sample to belong to the old fault, and the old model is used to diagnose the fault. Otherwise, the sample is considered to belong to a new fault, and then a certain number of typical examples are randomly selected to construct a training set with the new fault samples for the new stage model training.

[0026] Furthermore, in the incremental training stage of step 4, the old samples in the training set constructed in step 3 are input into the old model for inference, and the network output results of all old classes are stored. For the old class samples in the training set, the new model aims to reproduce the predicted output (distillation loss) stored in the old model, while for the new class samples in the training set, the new model encourages the network to output the correct class metrics of the new class samples (classification loss). Therefore, the designed loss is:

[0027]

[0028] where the former term is the cross-entropy loss for the classification of new class data, and the latter term is the distillation loss for the prediction of old classes, is the predicted output of the previous step for old classes, represents the designed diagnostic model, and represent the samples and corresponding labels in the sample set.

[0029] To further alleviate the catastrophic forgetting phenomenon in incremental learning, an attention distillation loss is designed to require the attention responses of the new model and the old model to the old class data to be consistent, which can be achieved by minimizing the cosine similarity of the attention weights of the new and old models:

[0030]

[0031] where is the cosine distance between the attention maps of the j-th layer of the new model and the old model. represents the cosine similarity, represents the L2 norm, is the feature of the j-th layer, represents the attention map of the feature of the j-th layer, and the symbols t and t-1 represent the new model and the old model respectively. Reducing the cosine distance between the attention maps of the two models increases their similarity, and the attention distillation loss can be expressed as:

[0032]

[0033] Where the superscripts c and s represent spatial attention and channel attention respectively, and N represents the number of layers for distillation. The loss function for incremental training can be expressed as:

[0034]

[0035] where is the balance coefficient, is the classification loss.

[0036] In summary, the present invention constructs a gear fault diagnosis method for a transmission system based on incremental learning, designs a generalized entropy index to detect emerging faults in the model inference stage, and designs a task distillation loss and an attention distillation loss in the incremental training stage to maintain the model's memory ability for the feature responses and prediction outputs of old samples, alleviating the catastrophic forgetting phenomenon in incremental learning.

[0037] The present invention also provides a gear fault diagnosis system for a transmission system based on incremental learning, including:

[0038] An acquisition module for acquiring monitoring data of the gears of the transmission system;

[0039] A construction module for constructing an initial model;

[0040] A training module for training the initial model and updating the training set after emerging faults occur;

[0041] A judgment module for judging whether the test data belongs to emerging faults.

[0042] The present invention also provides a computer storage medium, wherein the storage medium includes computer instructions, and when it runs on a computer, it enables the computer to execute the method described in any one of the foregoing items.

[0043] The present invention also provides an electronic device, wherein the electronic device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the method described in any one of the foregoing items.

[0044] The present invention can realize continuous update of the fault diagnosis model for streaming data, improve the state monitoring ability of the transmission system, and provide guarantee for the service safety of the mechanical system.

[0045] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail. Description of the Drawings

[0046] Figure 1 is a flowchart of a gear fault diagnosis method for a transmission system based on incremental learning provided by an embodiment of the present disclosure;

[0047] Figure 2 It is a schematic diagram of a transmission system provided by an embodiment of the present disclosure;

[0048] Figure 3 It is a photo of gears in 6 fault states provided by an embodiment of the present disclosure;

[0049] Figure 4 It is a schematic diagram of vibration signals collected from 6 faulty gears in an embodiment of the present disclosure. These signals are used as real inputs to prove that the present invention can diagnose Figure 3 the 6 faults shown;

[0050] Figure 5 It is a schematic diagram of the structure of a network model in an embodiment of the present disclosure;

[0051] Figure 6 It is a schematic diagram of the process of incremental training after detecting a new fault provided by an embodiment of the present disclosure;

[0052] Figure 7 It is the result of new fault detection of the stage 1 model provided by an embodiment of the present disclosure;

[0053] Figure 8 It is a comparison of the accuracies of the baseline method and the proposed method at different training stages provided by an embodiment of the present disclosure. Detailed implementation manners

[0054] Next, specific embodiments of the present invention will be described in detail with reference to the appended Figures 1 to 8 Although specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be completely conveyed to those skilled in the art.

[0055] It should be noted that in the description of the specification and the claims, certain terms are used to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. The description of the specification and the claims does not use the difference in nouns as a way to distinguish components, but uses the difference in the functions of components as the criterion for distinction. As mentioned throughout the specification and the claims, "comprising" or "including" is an open-ended term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is for the purpose of implementing the preferred embodiments of the present invention, but the description is for the general purpose of the specification and is not used to limit the scope of the present invention. The protection scope of the present invention shall be defined by the appended claims.

[0056] For the convenience of understanding the embodiments of the present invention, the following will further explain and illustrate with specific embodiments in conjunction with the accompanying drawings, and the various drawings do not constitute a limitation on the embodiments of the present invention.

[0057] A gear fault diagnosis method for a transmission system based on incremental learning, as Figure 1 shown, the steps are as follows:

[0058] Step 1: Collect the monitoring data of the gears of the transmission system, construct a data set, and then divide it into fault diagnosis tasks at different stages. Each fault diagnosis task at a different stage includes fault category data at different stages;

[0059] Step 2: Construct an initial model for gear fault diagnosis of the transmission system, and use the initial fault category data in the fault diagnosis task to train the model;

[0060] Step 3: Input the test data into the trained model for fault diagnosis to determine whether it belongs to a new fault; when a new fault appears, regard the test data at this time as new fault category data, regard the existing fault category data as old fault category data, randomly select a small number of samples from the old fault category data, and fuse and update the training set with the new fault category data, and perform incremental learning to obtain a new model. Otherwise, continue to use the trained model (old model) to diagnose faults; it should be noted that due to the appearance of new fault category data, the model performs incremental learning, so there is a "new model"; relatively, the model before incremental learning is the "old model".

[0061] Furthermore, in the incremental learning stage, use the updated training set to train the model. To alleviate the catastrophic forgetting phenomenon of the model for old fault category data, design an attention distillation loss to maintain the consistency of the responses of the new model and the old model to the features of old fault categories, so as to achieve incremental learning of fault patterns.

[0062] In one embodiment, Step 1 further includes: using a vibration sensor to collect the monitoring data of the gears of the transmission system, constructing a training set, a validation set, and a test set according to a ratio of 6:2:2, and dividing them into fault diagnosis tasks at different stages. Each fault diagnosis task at a different stage includes fault category data at different stages;

[0063] In another embodiment, Step 2 further includes: in the initial training stage, design a network to train the model based on the initial fault category data; to enhance the model's attention to key fault features, add channel attention and spatial attention modules after each residual block;

[0064] In another embodiment, for the test data in step 3, first, a dataset of seven states is divided into a training set and a test set. A sequential fault setting is performed on the training set, and the samples in the test set remain unchanged.

[0065] Furthermore, in step 1, gear vibration signals of a total of 7 states including normal, tooth surface wear, half tooth breakage, uniform wear, crack, tooth missing, and foreign object protrusion are collected; the collected signals are used to generate a sample set and normalized; further, T incremental diagnosis tasks are defined and corresponding data sets are divided , where represents different diagnostic stages, represents the initial diagnostic stage when, and other stages are incremental learning stages; represents the data set in the t-th stage, where and represent samples and labels respectively, represents the number of samples in this stage; represents the number of fault categories in the t-th stage.

[0066] Furthermore, the model feature extractor in step 2 consists of a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, three residual module layers, and a max pooling layer; the classifier consists of a single fully connected layer; channel and spatial attention modules are designed after each residual module layer to enhance the generalization ability of the model.

[0067] Using the data set in the initial stage to train the initial diagnostic model, the loss function is categorical cross-entropy loss, which can be written as:

[0068]

[0069] where is the true label, is the predicted label.

[0070] Furthermore, in step 3, a generalized entropy index is defined to detect emerging faults, and this index can be described as:

[0071]

[0072] where is a hyperparameter, is the predicted output after SoftMax normalization for the j-th sample. When a new sample is inferred by the old model, the generalized entropy metric defines the confidence that the sample belongs to the old fault class. When this metric is higher than the threshold, the system considers that the sample belongs to the old fault, and then uses the old model to diagnose the fault. Otherwise, it is considered that the sample belongs to a new-born fault, and a certain number of typical examples are randomly selected to construct a training set with the new fault samples for the training of the new stage model.

[0073] Furthermore, the way to achieve incremental training in step 4 is as follows:

[0074] Input the old samples in the training set constructed in step 3 into the old model for inference, and store all the output results of the old model . For the old-class samples in the training set, the new model aims to reproduce the predicted output stored in the old model (distillation loss), while for the new-class samples in the training set, the new model encourages the network to output the correct class metrics for the new-class samples (classification loss). Therefore, the designed loss is:

[0075]

[0076] where the former term is the cross-entropy loss for the classification of new-class data, and the latter term is the distillation loss for the prediction of old classes, is the output result of the old model, represents the designed new diagnostic model, and represent the samples and corresponding labels in the sample set. It should be noted that when classifying by inputting new and old data into the new diagnostic model together, for new-class data, the cross-entropy loss is normally used, and for old-class data, its output is the same as the output when the old-class data is input into the old model (achieved through the first loss term).

[0077] To alleviate the catastrophic forgetting phenomenon in incremental learning, the attention distillation loss is designed to require the attention responses of the new model and the old model to the old-class data to be consistent, which can be achieved by minimizing the cosine similarity of the attention weights of the new and old models:

[0078]

[0079] where is the cosine distance between the attention maps of the j-th layer of the new model and the old model. represents the cosine similarity, represents the L2 norm, is the feature of the j-th layer, represents the attention map of the feature of the j-th layer. The symbols t and t - 1 represent the new model and the old model respectively. Reducing the cosine distance between the attention maps of the two models increases their similarity, and the attention distillation loss can be expressed as:

[0080]

[0081] Among them, the superscripts c and s represent spatial attention and channel attention respectively, and N represents the number of layers for distillation. The loss function for incremental training can be expressed as:

[0082]

[0083] where is the balance coefficient, is the classification loss.

[0084] In summary, the present invention constructs a gear fault diagnosis method for a transmission system based on incremental learning, designs a generalized entropy index to detect emerging faults in the model inference stage, and designs a task distillation loss and an attention distillation loss in the incremental training stage to maintain the model's memory ability for the feature responses and prediction outputs of old samples, alleviating the catastrophic forgetting phenomenon in incremental learning.

[0085] In another embodiment, the gear fault diagnosis method for a transmission system based on incremental learning includes the following steps:

[0086] In the first step, a three-channel vibration sensor is used to collect the vibration signals of gears in normal and faulty states in the transmission system. The types of gear faults include: normal state, tooth surface wear, half-tooth breakage, uniform wear, crack, tooth missing, and foreign object protrusion. The signals collected each time of movement are used as a sample and normalized. There are 1040 samples for each category, the number of channels is 3, and the sample length is divided into a training set, a validation set, and a test set according to the ratio of 6:2:2. Define 1 initial diagnosis task and 5 incremental diagnosis tasks and divide the corresponding data sets where represents different diagnosis stages, represents the initial diagnosis stage when and other stages are incremental learning stages; represents the data set at the t-th stage, where and represent the sample and the label respectively, represents the number of samples in this stage; represents the number of fault categories at the t-th stage, where in the initial stage includes two types of faults, normal and tooth surface modes, and one more fault category data is added in each subsequent incremental stage. Table 1 defines the fault categories included in each diagnosis task.

[0087] Table 1 Incremental fault diagnosis tasks designed based on the gear fault data set of the transmission system

[0088]

[0089] Table 2 Backbone Network Parameters

[0090]

[0091] In the second step, a convolutional network is used as the backbone network, and the channel attention and spatial attention mechanisms are fused to encourage the network to focus on fault-related features and suppress fault-unrelated features. The detailed structural parameters of the network are shown in Table 2. Specifically, as Figure 5 shown, the model feature extractor consists of a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation function, three residual module layers, and a max pooling layer; the classifier consists of a single fully connected layer; the channel and spatial attention modules are designed after each residual module layer and are not listed because they do not cause changes in the feature dimensions. The symbol C represents the number of classes, which varies according to different task settings. The initial model parameters are trained and updated using the cross-entropy loss, which can be described as:

[0092]

[0093] where is the true label and is the predicted label.

[0094] In the third step, when the tested streaming data is input into the initial model for fault diagnosis, a generalized entropy index is defined to determine whether it belongs to a new fault. The principle of the generalized entropy index is to amplify the small deviation between the predicted distribution and the ideal label encoding, and it can distinguish probability vectors near the boundary better than the Shannon entropy, judging whether the test sample belongs to the categories in the training set. The generalized entropy index can be described as:

[0095]

[0096] where is a hyperparameter, which is set to 0.1 in the experiment. is the predicted output of the j-th sample after SoftMax normalization. When this index is higher than the threshold, the system considers that the sample belongs to an old fault, and then uses the old model to diagnose the fault. Otherwise, it is considered that the sample belongs to a new fault, and then 50 typical examples are randomly selected Construct a training set with the new fault samples for the new stage model training. The threshold is set to the confidence level at the 90% position after sorting the confidence levels in the training set from large to small. Experiments show that such a threshold setting can judge most of the old class samples in the test as old faults and can accurately detect most of the new fault samples. Table 3 lists the accuracy of detecting new faults using the generalized entropy index in each stage. Since the last stage contains all category data, that is, there are no new faults, the accuracy is not listed. It should be emphasized that there is a great contingency in a single judgment of whether a fault belongs to a new fault. In the experiment, the system will generally alarm only when it is judged as a new fault three times in a row. Therefore, the experimental progress listed in Table 3 can basically meet the needs of new fault detection.

[0097] Table 3. Experimental Results of the New Fault Detection Algorithm

[0098]

[0099] In the fourth step, the old samples in the training set constructed in the third step are input into the old model for inference, and the network output results of all old classes are stored. For the old class samples in the training set, the new model aims to reproduce the predicted output stored in the old model; while for the new class samples in the training set, the new model encourages the network to output the correct class indicators of the new class samples. Therefore, the designed loss is:

[0100]

[0101] where the former term is the cross-entropy loss for new class data classification, and the latter term is the distillation loss for old class prediction. is the predicted output of the previous step for the old class. represents the designed diagnostic model. and represent the samples and corresponding labels in the sample set. To further alleviate the catastrophic forgetting phenomenon in incremental learning, the designed attention distillation loss requires that the attention responses of the new model and the old model to the old class data be consistent, which can be achieved by minimizing the cosine similarity of the attention weights of the new model and the old model:

[0102]

[0103] where is the cosine distance between the attention maps of the j-th layer of the new model and the old model. represents the cosine similarity. represents the second norm. is the feature of the j-th layer. represents the attention map of the feature of the j-th layer. The symbols t and t-1 represent the new model and the old model respectively. Reducing the cosine distance between the attention maps of the two models increases their similarity. The attention distillation loss can be expressed as:

[0104]

[0105] Among them, the superscripts c and s represent spatial attention and channel attention respectively, and N represents the number of layers for distillation. The loss function for incremental training can be expressed as:

[0106]

[0107] By jointly optimizing the parameters with the above two losses, the diagnostic model can not only quickly identify new fault categories but also maintain the prediction accuracy for old fault categories.

[0108] Figure 2 is a schematic diagram of a traditional system provided by an embodiment of the present disclosure. The main components of the mechanical system shown on the left include: a drive motor, a bevel gearbox, a reduction gearbox, a gear transmission system, and a vibration sensor; power is provided by the drive motor, and the power is transmitted through the bevel gearbox and the reduction gearbox to rotate the measured gear below, driving the entire mechanical part above to rotate, and the movement is controlled by the motor. The vibration sensor collects vibration signals during the movement process for fault diagnosis.

[0109] Figure 3 shows pinions in 6 fault states, including: tooth surface wear, half tooth breakage, uniform wear, crack, tooth missing, and foreign object bulge.

[0110] Figure 4 shows examples of vibration signals collected from 6 fault gears.

[0111] Figure 5 shows the specific composition of the network model.

[0112] Figure 6 shows a schematic diagram of the incremental training process after detecting a new fault. The old data is input into the old model for inference, and the network prediction outputs and attention responses of all old classes are stored. Then, the new class data and the old data are uniformly input into the new model for training. The designed task distillation loss and attention distillation loss are used to guide the network to maintain the memory ability for old fault classes while learning the new fault class features.

[0113] Figure 7 shows the new fault detection results of the stage 1 model. The horizontal axis represents the sample index. The first 424 samples are old faults with labels from 0 to 2; while the subsequent samples are new faults with labels from 3 to 6. The orange legend represents the corresponding generalized entropy index for each sample. It can be seen that the learned threshold can accurately distinguish new faults from old faults, providing guidance for the transition in the incremental stage.

[0114] Figure 8It shows the comparison of the accuracy rates of the baseline method and the method proposed in the present invention at different training stages. The baseline method only uses the cross-entropy loss for classification without using the distillation loss, so the accuracy rate will decrease rapidly as the number of fault categories increases. The method proposed in the present invention jointly optimizes the model using the task distillation loss and the attention distillation loss, and the accuracy rate in each stage after the initial stage is higher than that of the baseline method.

[0115] The above are only the preferred embodiments of the present disclosure, and do not limit the implementation manners and protection scope of the present disclosure. For those skilled in the art, it should be realized that all the solutions obtained by equivalent substitution and obvious changes made by using the content of the present disclosure specification should be included in the protection scope of the present disclosure.

Claims

1. A transmission system gear fault diagnosis method based on incremental learning, characterized in that: The method comprises the following steps: Step 1: Collect the monitoring data of the transmission system gears, build a data set, and divide it into fault diagnosis tasks at different stages. Each fault diagnosis task at different stages includes fault category data at different stages. Step 2: constructing an initial model for transmission system gear fault diagnosis, and training the initial model using initial fault category data in the fault diagnosis task to obtain a trained model; Step 3: Input the test data into the trained model for fault diagnosis to determine whether it is a new fault; when a new fault occurs, the test data at this time is regarded as new fault category data, and the fault category data that has occurred is regarded as old fault category data. A small number of samples are randomly selected from the old fault category data and fused with the new fault category data to update the training set, so that the initial model performs incremental learning to obtain a new model. Otherwise, the trained model continues to be applied to diagnose the fault.

2. The method according to claim 1, characterized in that Preferably, the monitoring data in step 1 is collected by a vibration sensor.

3. The method according to claim 1, characterized in that The monitoring data in step 1 includes: a vibration signal of a gear in a normal state and a vibration signal of a gear in a faulty state.

4. The method according to claim 3, characterized in that The faulty gears include: tooth surface wear, half broken teeth, uniform wear, cracks, missing teeth and foreign body protrusions.

5. The method according to claim 1, characterized in that The initial model adds a channel attention module and a spatial attention module after each residual block to enhance the model's attention to key fault features.

6. The method according to claim 1, characterized in that The step 3 also includes: defining a generalized entropy index to determine whether the fault is a new fault; When a new sample is inferred in the old model, the generalized entropy index defines the confidence that the sample belongs to the old fault category. When the index is higher than the threshold, the system considers that the sample belongs to the old fault category, otherwise it considers that the sample belongs to a new fault.

7. The method according to claim 6, characterized in that The threshold is set to the confidence level at the 90% position after the confidence levels in the training set are sorted from large to small.

8. Transmission system gear fault diagnosis system based on incremental learning, including: A collection module is used to collect monitoring data of transmission system gears; Building module, used to build the initial model; The training module is used to train the initial model and update the training set after new faults are generated; The judgment module is used to judge whether the test data belongs to a new fault.

9. A computer storage medium, wherein: The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 7.

10. An electronic device, wherein: The electronic device comprises: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 7 when executing the program.

Images