Planetary gearbox crack fault classification method, system, equipment and medium

Abstract

The invention discloses a method, a system, equipment and a medium for classifying crack faults of a planetary gearbox, belongs to the classification of the crack faults of the planetary gearbox, and aims to solve the problems that a large number of complete source domain samples are difficult to obtain and the model classification accuracy is low. The method comprises the following steps: S1, acquiring a planetary gearbox crack fault acoustic signal sample data set and hard tag data; s2, constructing a crack fault classification model; s3, pre-training the crack fault classification model by using the sample data set and the hard tag data; s4, inputting sample data in the sample data set into the pre-trained crack fault classification model and obtaining a category probability; aggregating the category probabilities of all sample data belonging to the same crack fault type, and weighting to generate soft label data; using the sample data set and the soft label data to update and retrain the crack fault classification model; and S5, the updated crack fault classification model is used for real-time classification of crack faults.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and relates to automatic identification and classification of fault types, in particular to a planetary gearbox crack fault classification method, system, device and medium. BACKGROUND

[0002] The planetary gearbox is also called a planetary reduction gearbox or a gear box (Planetary Gearbox). It is a transmission device based on a planetary gear structure, which realizes transmission speed ratio reduction and torque amplification through the meshing of a sun gear, a planetary gear and an inner ring gear. Due to its excellent transmission efficiency and load capacity, it is widely used in wind power generation, industrial equipment, aerospace and other fields.

[0003] Under long-term harsh working environment and complex working conditions, the key components of the planetary gearbox are prone to vibration wear, fatigue wear and other wear and fatigue, thereby causing the gearbox crack fault. If these risks are not discovered and diagnosed in time, it may cause the equipment to stop, cause serious economic losses, and even cause safety hazards.

[0004] Since the shape of the gearbox crack has certain uncertainty, the diagnosis, monitoring / detection of the gearbox crack is mainly realized by manual naked eye or handheld detector, which is affected by the responsibility, professional quality and other factors of the staff. There are certain "missed detection and wrong detection" risks in the monitoring / detection process, and the manual monitoring efficiency is low. Therefore, it is necessary to detect the crack of the planetary gearbox by a more automated method.

[0005] The invention patent application with the application number 202311349855.0 discloses a CRH5 motor train unit gearbox crack edge detection method, and the detection steps include: S1, collecting the CRH5 motor train unit gearbox crack image; S2, preprocessing the gearbox crack image to obtain a crack gray image; S3, performing filter processing on the crack gray image, and the filter processing adopts a median filter algorithm to perform filter processing on the crack gray image to obtain a crack image after filter processing; S4, using an Otsu adaptive threshold algorithm to perform image segmentation on the crack image after filter processing to separate the background from the image to obtain a target image; and S5, using a multi-scale morphological operation combined with a Canny operator to perform edge detection on the target image to obtain a clear contour.

[0006] With the rapid development of artificial intelligence technology in recent years, artificial intelligence technology has been applied to various industries, including gearbox crack diagnosis and monitoring / detection.

[0007] The application patent application with the application number 202510328335.4 also discloses an intelligent fault diagnosis method for a planetary gearbox based on ITD and LSTM-CNN, which comprises the following steps: first, collecting the vibration signals of the planetary gearbox through a sensor, performing ITD decomposition on each sampling signal to obtain a plurality of PR components, selecting the PR component with larger instantaneous frequency fluctuation by using Hilbert transform, converting the reconstructed signal into a frequency spectrum function by using Fourier transform, extracting the amplitude of the characteristic frequency, constructing a feature vector, inputting the feature vector into the LSTM-CNN for intelligent diagnosis, combining the classification cross-entropy loss function to ensure the stability of the calculation and the efficiency of the training, and providing a comprehensive and efficient solution for the fault type and degree diagnosis of the sun gear of the gear transmission system.

[0008] The fault classification task based on artificial intelligence (neural network model) often needs to train the network model. However, when the model is trained by using hard labels in the conventional method, the model is difficult to capture the complex correlation and internal combination structure between classes, and the generalization performance of the model is affected. To solve this problem, the conventional method is to obtain multi-source domain data and use it for model training. However, in some industrial applications, due to the limitation of equipment operating conditions and data acquisition cost, it is difficult to obtain a large number of complete source domain sample data, which ultimately affects the accuracy of the planetary gearbox crack fault diagnosis and classification in industrial applications. SUMMARY

[0009] The purpose of the present application is to solve the technical problem that the model generalization performance is low due to the difficulty in obtaining a large number of complete source domain sample data in existing industrial applications due to the limitation of equipment operating conditions, data acquisition cost, etc., and to provide a planetary gearbox crack fault classification method, system, device and medium.

[0010] The present application adopts the following technical solutions to achieve the above-mentioned purposes: A planetary gearbox crack fault classification method comprises the following steps: Step S1, obtaining sample data; Obtaining planetary gearbox crack fault acoustic signal sample data to obtain a sample data set; marking the crack fault type of the acoustic signal in the sample data set to obtain hard label data; Step S2, constructing a crack fault classification model; Constructing a crack fault classification model for classifying the acoustic signals of the planetary gearbox; Step S3, pre-training the crack fault classification model; Using the sample data set and the hard label data to pre-train the crack fault classification model; Step S4, label refinement and model iteration update; inputting the sample data in the sample data set into the pre-trained crack fault classification model, and outputting corresponding category probability; aggregating and weighting the category probability of all sample data belonging to the same crack fault type, and generating soft label data; updating and retraining the crack fault classification model using the sample data set and the soft label data as new supervision signals; Step S5, real-time fault classification; obtaining the planetary gearbox crack fault acoustic signal to be classified, and inputting the crack fault classification model updated iteratively in step S4, and the crack fault classification model outputs the crack fault category.

[0011] Further, the crack fault classification model comprises a convolution layer, a normalization layer, a pooling layer, and a full connection layer. The acoustic signal is inputted into the normalization layer after convolution processing by the convolution layer, and is inputted into the pooling layer after normalization and activation function, and the output of the pooling layer is inputted into the full connection layer after stretching processing, and the output of the full connection layer is calculated by Softmax as the output of the crack fault classification model.

[0012] Further, in step S3, when the pre-trained crack fault classification model, the loss function is: ; wherein, represents the true probability of the acoustic signal being the crack fault type, represents the predicted probability of the acoustic signal being the crack fault type, represents the total number of categories.

[0013] Further, in step S4, the specific steps of label refinement and model iterative updating are: Step S4-1, inputting the sample data in the sample data set into the pre-trained crack fault classification model, and the crack fault classification model outputs corresponding category prediction probability; Step S4-2, weighting the category prediction probability to obtain a weighting coefficient : ; using the weighting coefficient aggregating the category prediction probability of all samples belonging to the same fault category, and obtaining a soft label : ; ; Step S4-3, retraining and updating the crack fault classification model using the sample data set and the soft label data as new supervision signals;​ Step S4-4: Input the sample data from the sample dataset into the crack fault classification model updated in step S4-3, and output the corresponding category prediction probability; weight and aggregate the category prediction probabilities to obtain the updated soft label. Step S4-5: Determine whether the updated soft label is sufficiently refined. If it is sufficiently refined, output the current crack fault classification model and use it for real-time fault classification. If it is not sufficiently refined, repeat steps S4-3 to S4-5 for iterative updates until the soft label is sufficiently refined. in, This represents the predicted probability vector of the b-th sample of the j-th fault category at the k-th iteration. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category in the k-th iteration. This represents the label vector for the j-th fault category at the (k+1)-th iteration. This represents the label indicator value indicating that the j-th fault category belongs to the i-th fault category at the k-th iteration. This represents the total number of category i. This represents the total number of samples in the j-th fault category.

[0014] Furthermore, in steps S4-5, the following criteria are used to determine whether the label refinement is sufficient: For the j-th fault category, calculate the category prediction probability for all samples of that fault category. With the current tag The KL divergence between samples is calculated, and the average KL divergence of all samples in that class is taken to obtain the average KL divergence. : ; Calculate the average of the maximum and minimum KL divergence for different categories. ratio : ; ratio Whether it tends to stabilize is used as the basis for iterative convergence, that is... If the process is not stable, proceed to the next iteration of label refinement and update; if the process is stable, terminate the iteration and use the final soft label as a new supervision signal to retrain and update the crack fault classification model. in, Indicates the total number of categories. This represents the total number of samples in the j-th fault category. This indicates the label value indicating that the j-th fault category belongs to the i-th fault category. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category. This represents the stability threshold.

[0015] A planetary gearbox crack fault classification system includes the following steps: The sample data acquisition module is used to acquire sample data of acoustic signals of planetary gearbox crack faults to obtain a sample dataset; and to label the crack fault types of the acoustic signals in the sample dataset to obtain hard-labeled data. The crack fault classification model construction module is used to build a crack fault classification model for classifying the acoustic signals of a planetary gearbox. The crack fault classification model pre-training module is used to pre-train the crack fault classification model using sample datasets and hard-labeled data. The label refinement and model iteration update module is used to input sample data from the sample dataset into the pre-trained crack fault classification model and output the corresponding class probability. The category probabilities of all sample data belonging to the same crack failure type are aggregated, weighted, and soft-label data is generated. The crack fault classification model is updated and retrained using sample datasets and soft-label data as new supervision signals. The real-time fault classification module is used to acquire the acoustic signal of the planetary gearbox crack fault to be classified, and input the crack fault classification model updated by the label refinement and model iteration update module. The crack fault classification model outputs the crack fault category.

[0016] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method described above.

[0017] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method.

[0018] The beneficial effects of this invention are as follows: 1. In this invention, a small amount of acoustic signal sample data and labels (hard labels) are first used to pre-train the crack fault classification model. Then, the model's predicted probabilities for the sample data are adaptively refined to generate corresponding soft labels, which are then used for model retraining and updating. By using soft labels obtained through adaptive label refinement with a small amount of sample data to retrain and update the model, a large amount of complete source domain sample data can still be obtained even under limitations such as equipment operating conditions and data acquisition costs. This greatly reduces the difficulty of obtaining a large number of complete source domain samples, ensuring sufficient samples for model training. This enhances the model's generalization performance and improves the accuracy of the model in diagnosing and classifying cracks in planetary gearbox housings. Experiments show that, when trained using only one or two source domains, the model based on adaptive label refinement achieves a classification accuracy that is approximately 22% higher than traditional convolutional neural networks under unseen operating conditions.

[0019] 2. In this invention, during adaptive label refinement, the new soft label for each category is generated by weighted aggregation of the model output probability distribution of all samples in that category; the weight of each sample is determined by its training loss, ensuring that samples that are more challenging to the current model can play a more important role in the optimized label definition, thereby breaking through the limitations of static one-hot encoding, capturing rich inter-class semantic relationships, enhancing the model's ability to discriminate difficult samples, and thus effectively improving the model's domain generalization performance; in addition, the entire process is autonomously guided by a stability coefficient based on KL divergence, which serves as a convergence criterion to ensure optimization quality. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a flowchart illustrating the present invention; Figure 3 This is a schematic diagram of the crack fault classification model in this invention; Figure 4 This is a schematic diagram illustrating the change in classification accuracy with the number of refinement iterations in this invention; Figure a represents a single-source domain scenario, and Figure b represents a dual-source domain scenario. Figure 5 This is a schematic diagram illustrating the changing trend of the label refinement stability coefficient in this invention; In the figure, Figure a shows the trend of the maximum and minimum KL divergence values ​​as a function of the number of refinements, and Figure b shows the trend of the ratio as a function of the number of refinements. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0022] Therefore, all other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] Example 1 This embodiment provides a method for classifying crack faults in planetary gearboxes, which analyzes and processes the acoustic and optical signals of the planetary gearbox to obtain the corresponding crack fault types. For example... Figure 1 , Figure 2 As shown, the specific steps include: Step S1: Obtain sample data; Acquire acoustic signal sample data of planetary gearbox crack faults to obtain sample dataset; label the crack fault types of the acoustic signals in the sample dataset to obtain hard-labeled data.

[0024] When collecting acoustic signals of planetary gearbox faults, a planetary gearbox fault acoustic signal acquisition system is used. This system includes a motor, a planetary gear reducer, and a magnetic powder brake arranged sequentially. An acoustic testing system is mounted on the outer casing of the planetary gearbox. By adjusting the motor and the magnetic powder brake, the planetary gear reducer can be operated under different speeds and torque loads. The acoustic testing system includes a detachable acoustic sensor (existing sensors can be used, and those skilled in the art can choose accordingly), positioned near the planetary gearbox. The acoustic sensor has a sampling frequency of 10kHz and is equipped with a data acquisition device (existing data acquisition devices can be used, and those skilled in the art can choose accordingly). The construction of this planetary gearbox fault acoustic signal acquisition system can utilize existing technology, requiring no creative effort.

[0025] When acquiring sample data, the samples were labeled with eight categories (seven types of crack failures and one normal type). These labels (crack failure types) included three types of single-component failures and four types of compound failures. Single-component failures included bearing rolling element damage (R), bearing outer ring damage (O), and planetary gear damage (P). The four compound failures included bearing outer ring damage + bearing rolling element damage (OR), bearing rolling element damage + planetary gear damage (RP), bearing outer ring damage + planetary gear damage (OP), and bearing outer ring damage + bearing rolling element damage + planetary gear damage (ORP).

[0026] When acquiring sample data, a total of six different operating conditions were collected. Each sample contained 4096 data points. The specific parameters of each operating condition are shown in Table 1 (each operating condition includes 7 fault categories and one normal category, so the total number of samples is 8 times the number of class samples).

[0027] Table 1 Data Acquisition Operating Parameters

[0028] After obtaining the above sample data, data preprocessing can be performed. Conventional methods in the art can be used for data preprocessing, requiring no inventive effort from those skilled in the art.

[0029] Step S2: Construct a crack fault classification model; A crack fault classification model is constructed to classify the acoustic signals of planetary gearboxes and determine whether the planetary gearboxes corresponding to the acoustic signals have crack faults and what kind of crack faults they are.

[0030] like Figure 3 As shown, the crack fault classification model includes a convolutional layer, a normalization layer, a pooling layer, and a fully connected layer. The acoustic signal is processed by convolution in the convolutional layer and then input into the normalization layer. After normalization and activation by the normalization layer, it is input into the pooling layer. The output of the pooling layer is stretched and then input into the fully connected layer. The output of the fully connected layer is calculated by Softmax and used as the output of the crack fault classification model.

[0031] Step S3: Pre-train the crack fault classification model; The crack fault classification model was pre-trained using sample datasets and hard-labeled data.

[0032] The specific method for pre-training the crack fault classification model can be directly applied to existing technologies, and those skilled in the art do not need to make any creative effort.

[0033] However, the loss function for pre-training is specifically: ; in, This represents the true probability that the acoustic signal represents a crack fault type. This indicates that the acoustic signal represents a crack fault type. Predicting probabilities This indicates the total number of categories.

[0034] Step S4: Label refinement and model iterative update; Input the sample data from the sample dataset into the pre-trained crack fault classification model and output the corresponding class probability; The category probabilities of all sample data belonging to the same crack failure type are aggregated, weighted, and soft-label data is generated. The crack fault classification model was updated and retrained using sample datasets and soft-labeled data as new supervision signals.

[0035] During adaptive label refinement, the model is iterated and updated based on the soft labels obtained from the refinement. Because iterative updates and assessments of adequacy are required during adaptive label refinement, the entire process involves generating soft labels at least twice. The specific process is as follows: Figure 2 As shown, the process includes initial adaptive label refinement and second adaptive label refinement. If the soft label obtained after the second adaptive label refinement meets the conditions, no iterative update is required. If the soft label obtained after the second adaptive label refinement does not meet the conditions, iterative update is performed until the iteration termination condition is met.

[0036] The specific steps for the initial adaptive label refinement (i.e., k=1) are as follows: Step S4-1: Input the sample data from the sample dataset into the pre-trained / iteratively updated crack fault classification model. The crack fault classification model outputs the predicted probability of the corresponding sample's category.

[0037] Step S4-2: Weight the predicted probabilities of the categories to obtain the weighting coefficients. : ; Then use weighting coefficients The predicted class probabilities of all samples belonging to the same fault category are aggregated to obtain soft labels. Specifically: First, based on the weighting coefficients and category prediction probability Obtain the label corresponding to the fault category. : ; Then, for each category of labels probability value Collect the data to obtain soft tags. : ; Step S4-3: Use the sample dataset and soft-label data as new supervision signals to retrain and update the crack fault classification model.

[0038] Step S4-4: Perform the second adaptive label refinement. Input the sample data in the sample dataset into the crack fault classification model updated in step S4-3. The crack fault classification model outputs the corresponding category prediction probability. Then, the category prediction probabilities are weighted and aggregated (i.e., the calculation in step S4-2 is performed) to obtain the updated soft label.

[0039] Step S4-5: Determine whether the updated soft label is sufficiently refined, and use this as the basis for terminating the soft label iteration update. If the updated soft label is sufficiently refined, output the current crack fault classification model and use it for real-time fault classification in step S5; if the updated soft label is not sufficiently refined, repeat steps S4-3 to S4-5 for iterative updates until the latest generated soft label is sufficiently refined, then stop the iteration and output the current crack fault classification model.

[0040] in, This represents the predicted probability vector of the b-th sample of the j-th fault category at the k-th iteration. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category in the k-th iteration. This represents the label vector for the j-th fault category at the (k+1)-th iteration. This represents the label indicator value indicating that the j-th fault category belongs to the i-th fault category at the k-th iteration. This represents the total number of category i. This represents the total number of samples in the j-th fault category.

[0041] In steps S4-5, when determining whether the soft label is sufficiently detailed, the following criteria will be used to judge whether the label detailing is sufficient: For the j-th fault category, calculate the category prediction probability for all samples of that fault category. With the current tag The KL divergence between samples is calculated, and the average KL divergence of all samples in that class is taken to obtain the average KL divergence. : ; Calculate the average of the maximum and minimum KL divergence for different categories. ratio : ; ratio Whether it tends to stabilize is used as the basis for iterative convergence, that is... If it is not stable (i.e. If the condition is not met, proceed to the next tag refinement iteration; if the condition is stable (i.e., ... If the condition is met, the iteration terminates, and the final soft label is used as a new supervision signal to retrain and update the crack fault classification model. in, Indicates the total number of categories. This represents the total number of samples in the j-th fault category. This indicates the label value indicating that the j-th fault category belongs to the i-th fault category. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category. This represents the stability threshold.

[0042] Step S5, real-time fault classification; Acquire the acoustic signal of the planetary gearbox crack fault to be classified, and input it into the crack fault classification model after iterative update in step S4. The crack fault classification model outputs the crack fault category.

[0043] Example 2 This embodiment provides a planetary gearbox crack fault classification system, used to analyze and process the acoustic and optical signals of the planetary gearbox to obtain the corresponding crack fault type. Specifically, it includes: The sample data acquisition module is used to acquire sample data of acoustic signals of planetary gearbox crack faults to obtain a sample dataset; and to label the crack fault types of the acoustic signals in the sample dataset to obtain hard-labeled data.

[0044] When collecting acoustic signals of planetary gearbox faults, a planetary gearbox fault acoustic signal acquisition system is used. This system includes a motor, a planetary gear reducer, and a magnetic powder brake arranged sequentially. An acoustic testing system is mounted on the outer casing of the planetary gearbox. By adjusting the motor and the magnetic powder brake, the planetary gear reducer can be operated under different speeds and torque loads. The acoustic testing system includes a detachable acoustic sensor (existing sensors can be used, and those skilled in the art can choose accordingly), positioned near the planetary gearbox. The acoustic sensor has a sampling frequency of 10kHz and is equipped with a data acquisition device (existing data acquisition devices can be used, and those skilled in the art can choose accordingly). The construction of this planetary gearbox fault acoustic signal acquisition system can utilize existing technology, requiring no creative effort.

[0045] When acquiring sample data, the samples were labeled with eight categories (seven types of crack failures and one normal type). These labels (crack failure types) included three types of single-component failures and four types of compound failures. Single-component failures included bearing rolling element damage, bearing outer ring damage, and planetary gear damage. The four compound failures included a combination of bearing outer ring damage and bearing rolling element damage, a combination of bearing rolling element damage and planetary gear damage, a combination of bearing outer ring damage and planetary gear damage, and a combination of bearing outer ring damage, bearing rolling element damage, and planetary gear damage.

[0046] When acquiring sample data, a total of six different operating conditions were collected. Each sample contained 4096 data points. The specific parameters of each operating condition are shown in Table 1 (each operating condition includes 7 fault categories and one normal category, so the total number of samples is 8 times the number of class samples).

[0047] Table 1 Data Acquisition Operating Parameters

[0048] After obtaining the above sample data, data preprocessing can be performed. Conventional methods in the art can be used for data preprocessing, requiring no inventive effort from those skilled in the art.

[0049] The crack fault classification model construction module is used to build a crack fault classification model, which is used to classify the acoustic signals of the planetary gearbox and determine whether the planetary gearbox corresponding to the acoustic signals has a crack fault and what kind of crack fault it is.

[0050] like Figure 3 As shown, the crack fault classification model includes a convolutional layer, a normalization layer, a pooling layer, and a fully connected layer. The acoustic signal is processed by convolution in the convolutional layer and then input into the normalization layer. After normalization and activation by the normalization layer, it is input into the pooling layer. The output of the pooling layer is stretched and then input into the fully connected layer. The output of the fully connected layer is calculated by Softmax and used as the output of the crack fault classification model.

[0051] The crack fault classification model pre-training module is used to pre-train the crack fault classification model using sample datasets and hard-labeled data.

[0052] The specific method for pre-training the crack fault classification model can be directly applied to existing technologies, and those skilled in the art do not need to make any creative effort.

[0053] However, the loss function for pre-training is specifically: ; in, This represents the true probability that the acoustic signal represents a crack fault type. This indicates that the acoustic signal represents a crack fault type. Predicting probabilities This indicates the total number of categories.

[0054] The label refinement and model iteration update module is used to input sample data from the sample dataset into the pre-trained crack fault classification model and output the corresponding class probability. The category probabilities of all sample data belonging to the same crack failure type are aggregated, weighted, and soft-label data is generated. The crack fault classification model was updated and retrained using sample datasets and soft-labeled data as new supervision signals.

[0055] During adaptive label refinement, the model is iterated and updated based on the soft labels obtained from the refinement. Because iterative updates and assessments of adequacy are required during adaptive label refinement, the entire process involves generating soft labels at least twice. The specific process is as follows: Figure 2 As shown, the process includes initial adaptive label refinement and second adaptive label refinement. If the soft label obtained after the second adaptive label refinement meets the conditions, no iterative update is required. If the soft label obtained after the second adaptive label refinement does not meet the conditions, iterative update is performed until the iteration termination condition is met.

[0056] The specific steps for the initial adaptive label refinement (i.e., k=1) are as follows: Step S4-1: Input the sample data from the sample dataset into the pre-trained / iteratively updated crack fault classification model. The crack fault classification model outputs the predicted probability of the corresponding sample's category.

[0057] Step S4-2: Weight the predicted probabilities of the categories to obtain the weighting coefficients. : ; Then use weighting coefficients The predicted class probabilities of all samples belonging to the same fault category are aggregated to obtain soft labels. Specifically: First, based on the weighting coefficients and category prediction probability Obtain the label corresponding to the fault category. : ; Then, for each category of labels probability value Collect the data to obtain soft tags. : ; Step S4-3: Use the sample dataset and soft-label data as new supervision signals to retrain and update the crack fault classification model.

[0058] Step S4-4: Perform the second adaptive label refinement. Input the sample data in the sample dataset into the crack fault classification model updated in step S4-3. The crack fault classification model outputs the corresponding category prediction probability. Then, the category prediction probabilities are weighted and aggregated (i.e., the calculation in step S4-2 is performed) to obtain the updated soft label.

[0059] Step S4-5: Determine whether the updated soft label is sufficiently refined, and use this as the basis for terminating the soft label iteration update. If the updated soft label is sufficiently refined, output the current crack fault classification model and use it for real-time fault classification in step S5; if the updated soft label is not sufficiently refined, repeat steps S4-3 to S4-5 for iterative updates until the latest generated soft label is sufficiently refined, then stop the iteration and output the current crack fault classification model.

[0060] in, This represents the predicted probability vector of the b-th sample of the j-th fault category at the k-th iteration. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category in the k-th iteration. This represents the label vector for the j-th fault category at the (k+1)-th iteration. This represents the label indicator value indicating that the j-th fault category belongs to the i-th fault category at the k-th iteration. This represents the total number of category i. This represents the total number of samples in the j-th fault category.

[0061] In steps S4-5, when determining whether the soft label is sufficiently detailed, the following criteria will be used to judge whether the label detailing is sufficient: For the j-th fault category, calculate the category prediction probability for all samples of that fault category. With the current tag The KL divergence between samples is calculated, and the average KL divergence of all samples in that class is taken to obtain the average KL divergence. : ; Calculate the average of the maximum and minimum KL divergence for different categories. ratio : ; ratio Whether it tends to stabilize is used as the basis for iterative convergence, that is... If it is not stable (i.e. If the condition is not met, proceed to the next tag refinement iteration; if the condition is stable (i.e., ... If the condition is met, the iteration terminates, and the final soft label is used as a new supervision signal to retrain and update the crack fault classification model. in, Indicates the total number of categories. This represents the total number of samples in the j-th fault category. This indicates the label value indicating that the j-th fault category belongs to the i-th fault category. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category. This represents the stability threshold.

[0062] The real-time fault classification module is used to acquire the acoustic signal of the planetary gearbox crack fault to be classified, and input the crack fault classification model updated by the label refinement and model iteration update module. The crack fault classification model outputs the crack fault category.

[0063] Example 3 A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of a planetary gearbox crack fault classification method.

[0064] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0065] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the memory may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the planetary gearbox crack fault classification method. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0066] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run the program code for the planetary gearbox crack fault classification method.

[0067] Example 4 A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of a planetary gearbox crack fault classification method.

[0068] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the planetary gearbox crack fault classification method described above.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the planetary gearbox crack fault classification method described in the embodiments of this application.

[0070] Test case The acoustic signal acquisition system for planetary gearbox faults described in Example 1 was used to acquire acoustic signals of the reducer (i.e., the reducer composed of planetary gears) under different speeds and torque loads in a semi-anechoic experimental environment. The experimental framework included two typical scenarios: (1) single-source domain generalization; and (2) dual-source domain generalization. The robustness of the model was examined in the presence of significant distribution differences between the source and target domains (especially under conditions of drastic speed changes). The specific experimental configurations for the two scenarios are detailed in Table 2.

[0071] Table 2 Test Configuration Instructions

[0072] Experiments showed that the classification accuracy of this crack fault classification model varies with the number of refinement iterations as follows: Figure 4 As shown, in both experimental scenarios, the classification performance of the crack fault classification model reached its optimal level in the 18-22 iteration range. After 20 label refinement iterations, the average accuracy of the single-source domain scenario significantly improved from 68.75% of the baseline model (the existing convolutional neural network) to 91.73%, and the average accuracy of the dual-source domain scenario improved from 61.01% to 87.50%. In the single-source domain scenario G4, a slight decrease in initial accuracy was observed, but it gradually improved with further refinement iterations and eventually stabilized.

[0073] Figure 5 The label refinement stability coefficients for tasks G1 and G2 are shown. The changing trend. From Figure 5 It can be seen that in the initial stage of the conversion from hard labels to soft labels (i.e., the 0th to 1st iterations). The value changes drastically due to instability in the initial transformation process, but this perturbation does not affect the overall convergence; after the first iteration, The value monotonically decreases from 239.70, finally converging to 2.54 on the 20th iteration, at which point the classification accuracy reaches its peak; when the convergence threshold... When the value is set to 0.1, the absolute difference in the 20th iteration has decreased to 0.08. Therefore, the experiment shows that the proposed stability coefficient is effective. It can effectively quantify the reliability of label allocation, and its convergence characteristics can serve as a reliable criterion for terminating the refinement process.

Claims

1. A method for classifying crack faults in a planetary gearbox, characterized in that, Includes the following steps: Step S1: Obtain sample data; Acquire acoustic signal sample data of planetary gearbox crack faults to obtain sample dataset; label the crack fault types of the acoustic signals in the sample dataset to obtain hard-labeled data; Step S2: Construct a crack fault classification model; A crack fault classification model was constructed to classify the acoustic signals of a planetary gearbox. Step S3: Pre-train the crack fault classification model; The crack fault classification model was pre-trained using sample datasets and hard-labeled data; Step S4: Label refinement and model iterative update; Input the sample data from the sample dataset into the pre-trained crack fault classification model and output the corresponding class probability; The category probabilities of all sample data belonging to the same crack failure type are aggregated, weighted, and soft-label data is generated. The crack fault classification model is updated and retrained using sample datasets and soft-label data as new supervision signals. Step S5, real-time fault classification; Acquire the acoustic signal of the planetary gearbox crack fault to be classified, and input it into the crack fault classification model after iterative update in step S4. The crack fault classification model outputs the crack fault category.

2. The method for classifying crack faults in a planetary gearbox as described in claim 1, characterized in that, The crack fault classification model includes convolutional layers, normalization layers, pooling layers, and fully connected layers; The acoustic signal is processed by convolution in the convolutional layer and then input into the normalization layer. After normalization and activation, it is input into the pooling layer. The output of the pooling layer is stretched and then input into the fully connected layer. The output of the fully connected layer is calculated by Softmax and used as the output of the crack fault classification model.

3. The method for classifying crack faults in a planetary gearbox as described in claim 1, characterized in that, In step S3, when pre-training the crack fault classification model, the loss function is: ； in, This represents the true probability that the acoustic signal represents a crack fault type. This indicates that the acoustic signal represents a crack fault type. Predicting probabilities This indicates the total number of categories.

4. The method for classifying crack faults in a planetary gearbox as described in claim 1, characterized in that, In step S4, the specific steps for label refinement and model iterative update are as follows: Step S4-1: Input the sample data from the sample dataset into the pre-trained crack fault classification model, and the crack fault classification model outputs the corresponding category prediction probability. Step S4-2: Weight the predicted probabilities of the categories to obtain the weighting coefficients. : ； Using weighting coefficients The predicted class probabilities of all samples belonging to the same fault category are aggregated to obtain soft labels. : ； ； Step S4-3: Use the sample dataset and soft-label data as new supervision signals to retrain and update the crack fault classification model; Step S4-4: Input the sample data from the sample dataset into the crack fault classification model updated in step S4-3, and output the corresponding category prediction probability; weight and aggregate the category prediction probabilities to obtain the updated soft label. Step S4-5: Determine whether the updated soft label is sufficiently detailed; If the refinement is sufficient, the current crack fault classification model is output and used for real-time fault classification; if the refinement is insufficient, steps S4-3 to S4-5 are repeated for iterative updates until the soft labels are sufficiently refined. in, This represents the predicted probability vector of the b-th sample of the j-th fault category at the k-th iteration. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category in the k-th iteration. This represents the label vector for the j-th fault category at the (k+1)-th iteration. This represents the label indicator value indicating that the j-th fault category belongs to the i-th fault category at the k-th iteration. This represents the total number of category i. This represents the total number of samples in the j-th fault category.

5. The method for classifying crack faults in a planetary gearbox as described in claim 4, characterized in that, In steps S4-5, determine whether the label refinement is sufficient according to the following criteria: For the j-th fault category, calculate the category prediction probability for all samples of that fault category. With the current tag The KL divergence between samples is calculated, and the average KL divergence of all samples in that class is taken to obtain the average KL divergence. : ； Calculate the average of the maximum and minimum KL divergence for different categories. ratio : ； ratio Whether it tends to stabilize is used as the basis for iterative convergence, that is... ; If it is not stable, then proceed with the next tag refinement iteration update; If the model has stabilized, the iteration is terminated, and the final soft label is used as a new supervision signal to retrain and update the crack fault classification model. in, Indicates the total number of categories. This represents the total number of samples in the j-th fault category. This indicates the label value indicating that the j-th fault category belongs to the i-th fault category. This represents the probability that the b-th sample of the j-th fault category is predicted to be of the i-th fault category. This represents the stability threshold.

6. A planetary gearbox crack fault classification system, characterized in that, Includes the following steps: The sample data acquisition module is used to acquire sample data of acoustic signals of planetary gearbox crack faults to obtain a sample dataset; and to label the crack fault types of the acoustic signals in the sample dataset to obtain hard-labeled data. The crack fault classification model construction module is used to build a crack fault classification model for classifying the acoustic signals of a planetary gearbox. The crack fault classification model pre-training module is used to pre-train the crack fault classification model using sample datasets and hard-labeled data. The label refinement and model iteration update module is used to input sample data from the sample dataset into the pre-trained crack fault classification model and output the corresponding class probability. The category probabilities of all sample data belonging to the same crack failure type are aggregated, weighted, and soft-label data is generated. The crack fault classification model is updated and retrained using sample datasets and soft-label data as new supervision signals. The real-time fault classification module is used to acquire the acoustic signal of the planetary gearbox crack fault to be classified, and input the crack fault classification model updated by the label refinement and model iteration update module. The crack fault classification model outputs the crack fault category.

7. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 5.

Images