Gear pitting fault diagnosis method and system

Abstract

The invention relates to a gear pitting fault diagnosis method and system. The method comprises the following steps: acquiring a target vibration time-frequency characteristic information set when a target gear set rotates, loading target vibration time-frequency characteristic information in the target vibration time-frequency characteristic information set into a pitting corrosion fault diagnosis model, and performing diagnosis and recognition processing on each piece of target vibration time-frequency characteristic information by utilizing the pitting corrosion fault diagnosis model, and when the number of the target vibration time-frequency feature information in the target vibration time-frequency feature information group is not less than 2, performing voting processing on all the pitting fault prediction categories, so as to generate a pitting fault prediction category corresponding to each piece of target vibration time-frequency feature information in the target vibration time-frequency feature information group. And determining a corresponding pitting fault voting category after voting processing, and taking the pitting fault voting category as the pitting fault category of the target gear set. The gear pitting fault diagnosis method can efficiently diagnose the gear pitting fault under different working conditions, and improves the fault diagnosis accuracy and generalization ability of the gear pitting fault.

Description

Technical Field

[0001] The invention relates to a fault diagnosis method and system, in particular to a fault diagnosis method and system for gear pitting. Background Art

[0002] Fault prediction and health management are key management methods in the fields of industrial automation, aerospace, energy and transportation. They significantly improve the reliability and safety of industrial assets, effectively reduce equipment maintenance costs and extend the service life of assets.

[0003] Gears are important components in mechanical transmission systems. Gear pitting is a common fault, which is mainly manifested as small pits on the gear surface due to overload or material fatigue, which will significantly affect the performance of the mechanical transmission system. Therefore, early diagnosis and prevention of gear pitting can not only effectively reduce downtime and maintenance costs, but also improve production efficiency and ensure the safety of personnel and equipment.

[0004] Traditional gear pitting diagnosis mainly relies on professional technicians to disassemble and inspect the gears using professional instruments. This diagnostic method is inefficient and unsafe, and also requires the professional skills of professionals. With the advancement of sensor technology and information processing methods, gear pitting fault diagnosis has a new diagnostic perspective, but due to signal attenuation, noise interference and weak early fault signals, fault diagnosis based on vibration signals still faces challenges.

[0005] Early fault diagnosis methods combined time-frequency or statistical feature extraction (such as mean and kurtosis) with machine learning classification techniques such as support vector machines and random forests, achieving good performance. However, traditional machine learning methods rely heavily on the technical expertise of technicians in signal processing and feature extraction, and often find it difficult to capture the complex nonlinear relationships in gear vibration signals. In addition, they have poor generalization capabilities for gears in different working environments.

[0006] With the development of deep learning, gear fault diagnosis methods based on algorithms such as convolutional neural networks and recurrent neural networks can automatically extract features from raw data, and their performance is better than traditional methods that rely on manual feature extraction and domain expertise. At the same time, domain generalization technologies including adversarial learning and feature alignment improve the performance of cross-domain fault diagnosis systems and ensure the robustness of systems in different environments.

[0007] Although deep learning and domain generalization techniques avoid the manual feature extraction step and improve the performance of the model in cross-domain diagnosis, the current fault diagnosis methods mainly focus on obvious gear faults, but ignore the detailed analysis of gear pitting, especially in the early stage, weak signal features can be easily masked by noise and other interferences, highlighting the need for more sophisticated diagnostic methods.

[0008] In the existing research on gear pitting fault diagnosis, remarkable results have been achieved through physical model analysis and deep learning-based methods. However, these methods have weaknesses in feature correlation and model generalization ability, which are mainly manifested in two aspects: the existing gear pitting fault diagnosis technology based on deep learning has not fully explored the working mechanism or intrinsic correlation of gear vibration signals, and further research is needed on the representation ability and interpretability of the model; due to different working conditions, the domain generalization problem has not been fully solved, and model training still depends on target domain samples, which limits the widespread use of the model.

[0009] The application document with publication number CN118310739A records a gear fault diagnosis method based on meshing periodic spectrum. This method segments the vibration signal according to the meshing frequency, and can separate the frequency response during a single meshing. The segmented energy spectrum can reflect the changing period of the energy value within the optimal frequency band, effectively extract the fault characteristic frequency and perform fault diagnosis. However, this application mainly focuses on specific frequency band analysis and provides a method based on traditional signal processing, which is suitable for specific fault diagnosis scenarios, resulting in limited generalization ability under different working conditions.

[0010] The application document with publication number CN117909837A records a gear fault diagnosis method that collaborates with data and knowledge. First, an attribute knowledge description matrix between known gear fault labels and unknown gear fault labels is established, and then a gear fault diagnosis model driven by data and knowledge collaboration is constructed. Then, a gear fault diagnosis method based on an attribute learning classification model is designed. However, the application is highly dependent on the attribute knowledge description matrix pre-defined in the domain, which requires a clear understanding and preset of the fault type, which may be difficult for non-professional technicians to adapt to. In addition, the method mainly focuses on obvious gear faults and ignores the detailed analysis of gear pitting.

[0011] In summary, the gear pitting fault diagnosis technology based on vibration signals still has shortcomings and is difficult to meet current application needs. Summary of the invention

[0012] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a gear pitting fault diagnosis method and system, which can efficiently diagnose gear pitting faults under different working conditions, thereby improving the accuracy and generalization ability of gear pitting fault diagnosis.

[0013] According to the technical solution provided by the present invention, a gear pitting fault diagnosis method is provided, and the fault diagnosis method comprises:

[0014] A target vibration signal is obtained when the target gear set rotates, and the obtained target vibration signal is preprocessed to generate a target vibration time-frequency feature information group corresponding to the target vibration signal after the preprocessing, wherein:

[0015] The target vibration time-frequency feature information group includes a plurality of target vibration time-frequency feature information, and the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is consistent with the number of reference rotation cycles contained in the target vibration signal, so that each target vibration time-frequency feature information is used to characterize the vibration state of a reference rotation cycle of the target gear set;

[0016] The target vibration time-frequency feature information in the target vibration time-frequency feature information group is loaded into the pitting fault diagnosis model, and the pitting fault diagnosis model is used to perform diagnostic identification processing on each target vibration time-frequency feature information, so as to generate a pitting fault prediction category corresponding to each target vibration time-frequency feature information after the diagnostic identification processing, wherein:

[0017] For any target vibration time-frequency characteristic information, the diagnostic identification processing performed by the pitting fault diagnosis model includes sequential shared feature extraction processing and category classification processing, among which,

[0018] The loaded target vibration time-frequency feature information is subjected to shared feature extraction processing to generate corresponding target vibration time-frequency shared features;

[0019] The extracted target vibration time-frequency shared features are processed by category classification to generate the corresponding pitting fault prediction category;

[0020] When the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is not less than 2, all pitting fault prediction categories are subjected to voting processing to determine the corresponding pitting fault voting category after the voting processing, and the pitting fault voting category is used as the pitting fault category of the target gear set.

[0021] When preprocessing the target vibration signal, it includes:

[0022] Normalizing the target vibration signal in the time dimension to generate a target vibration normalized signal after normalization;

[0023] Based on the reference rotation period of the target gear set, the target vibration normalized signal is segmented to segment the target vibration signal into a plurality of target vibration segmented signals, wherein one target vibration segmented signal corresponds to one reference rotation period included in the target vibration signal;

[0024] Determine the control reference sampling point number based on the rotation speed state of the target gear set, and perform sampling point number control processing on each target vibration segmentation signal based on the control reference sampling point number, so as to generate a target vibration sampling point number control signal after the sampling point number control processing, wherein the sampling point number included in the target vibration sampling point number control signal is consistent with the control reference sampling point number;

[0025] Perform short-time Fourier transform on each target vibration sampling point control signal to generate a corresponding vibration time-frequency basic signal after short-time Fourier transform, where:

[0026] The basic vibration time-frequency signal includes a horizontal vibration time-frequency sub-signal, an axial vibration time-frequency sub-signal, and a vertical vibration time-frequency sub-signal;

[0027] For each vibration time-frequency basic signal, the contained horizontal vibration time-frequency sub-signal, axial vibration time-frequency sub-signal and vertical vibration time-frequency sub-signal are channel-joined to generate target vibration time-frequency characteristic information after channel-joining;

[0028] Based on all target vibration time-frequency feature information, a target vibration time-frequency feature information group is generated.

[0029] The pitting fault diagnosis model includes a shared feature extraction module and a category classifier adapted to be connected to the shared feature extraction module, wherein:

[0030] When performing diagnostic identification processing, the shared feature extraction module is used to extract shared features from the target vibration time-frequency feature information, where:

[0031] When performing shared feature extraction, the target vibration time-frequency feature information is first position-encoded to generate target vibration position encoding information after position encoding;

[0032] Extracting features from target vibration position encoding information to generate target time-frequency shared features after feature extraction;

[0033] Based on the target time-frequency shared features, the category classifier generates pitting fault prediction information corresponding to the current target vibration time-frequency feature information, and determines the corresponding pitting fault prediction category based on the pitting fault prediction information, wherein:

[0034] The pitting fault prediction information includes confidences corresponding to N pitting categories;

[0035] When determining the corresponding pitting fault prediction category based on the pitting fault prediction information, the pitting fault category with the highest confidence is used as the pitting fault prediction category.

[0036] The shared feature extraction module includes a position encoding module and a Transformer encoder for shared feature extraction, wherein:

[0037] When position encoding is performed on the target vibration time-frequency feature information, a position index is performed based on each time-frequency feature in the target vibration time-frequency feature information, and sine encoding or cosine encoding is performed based on the position of the time-frequency feature, so as to generate a corresponding time-frequency feature encoding after the position encoding, and the corresponding target time-frequency feature encoding information is generated based on the time-frequency feature encoding of all time-frequency features;

[0038] Based on the target time-frequency feature coding information and the target vibration time-frequency feature information, the target vibration position coding information is generated, wherein when the target vibration position coding information is generated, at least the time-frequency feature coding information and the target vibration time-frequency feature information are added in a position-by-position manner.

[0039] The Transformer encoder includes a plurality of encoding units connected in series, wherein:

[0040] For any encoding unit, it includes a multi-head attention unit, a multi-head attention layer normalization layer, a feedforward layer unit, and a feedforward layer normalization layer connected in sequence, where:

[0041] The multi-head attention unit includes a multi-head attention layer with residual connections;

[0042] The feed-forward layer unit includes a feed-forward layer using a residual connection;

[0043] For two serially connected coding units, along the serial connection direction of the coding units, the feedforward normalization layer of the previous coding unit is connected to the multi-head attention layer of the next coding unit.

[0044] When extracting features from target vibration position coding information, the target vibration position coding information is extracted by sequentially processing the serially connected coding units, and target time-frequency sharing features representing pitting categories and domain category sharing are output.

[0045] The category classifier includes a category classification main unit and a category classification output unit adaptively connected to the category classification main unit, wherein:

[0046] The category classification output unit adopts the LinSoftmax mapping function of approximate linear mapping. When the LinSoftmax mapping function is used for mapping output, then:

[0047]

[0048] Among them, LinSoftmax(z o ) iis the i-th confidence level of the output of the LinSoftmax mapping function, α is the slope, γ is the adjustment parameter, and z o is the logits vector output by the main unit for category classification, Output logits vector z for the category classification master unit o The i-th category classification quantity in , Output logits vector z for the category classification master unit o The j-th category classification quantity in , is the penalty weight of the i-th pitting category, is the penalty weight for the jth pitting category.

[0049] When voting, this includes:

[0050] For each pitting fault prediction category corresponding to the target vibration time-frequency characteristic information, configure the voting samples as the corresponding pitting category;

[0051] Count the number of voting samples for each pitting category, and use the pitting category with the largest number of voting samples as the pitting fault voting category of the target gear set;

[0052] When the number of voting samples for each pitting corrosion category is consistent, the pitting corrosion fault prediction category with the highest confidence is taken as the pitting corrosion fault voting category.

[0053] When building a pitting fault diagnosis model, include:

[0054] Constructing a pitting fault diagnosis basic model, wherein the pitting fault diagnosis basic model includes a shared feature extraction basic module, a category basic classifier and a domain classifier, and both the category basic classifier and the domain classifier are connected to the shared feature extraction basic module;

[0055] A pitting fault diagnosis training data set is constructed, wherein the pitting fault diagnosis training data set includes a number of training samples, wherein:

[0056] For any training sample, including the training vibration time-frequency feature information corresponding to the training gear set, the pitting category label of the current training gear set, and the label of the field speed,

[0057] The pitting fault diagnosis training data set contains at least N types of pitting category labels;

[0058] When the pitting fault diagnosis training data set is used to train the pitting fault diagnosis basic model, the model loss function used is:

[0059]

[0060] in, is the model loss function, is the category loss calculated based on the classification base classifier, is the domain loss calculated based on the domain classifier, is the association loss calculated based on the classification base classifier and the domain classifier, ω cls is the class loss weight, ω dom is the domain loss weight, ω cor is the associated loss weight;

[0061] When the training of the pitting fault diagnosis basic model reaches the target training state, the shared feature extraction basic module is configured as a shared feature extraction module, the category basic classifier is configured as a category classifier, and a pitting fault diagnosis model is formed based on the shared feature extraction module and the category classifier.

[0062] A gear pitting fault diagnosis system includes a fault diagnosis processing device, wherein the fault diagnosis processing device uses the fault diagnosis method to perform fault diagnosis on a target vibration signal obtained when a target gear set rotates, and determines the pitting fault category of the target gear set.

[0063] The advantages of the present invention are as follows: the target vibration signal is preprocessed to generate a target vibration time-frequency feature information group, and each target vibration time-frequency feature information is diagnosed and identified using a pitting fault diagnosis model to obtain a corresponding pitting fault prediction category, and a pitting fault voting category is obtained through voting processing, and the pitting fault voting category is used as the pitting fault category of the target gear set; gear pitting faults under different working conditions can be efficiently diagnosed, the fault diagnosis accuracy and generalization ability of gear pitting are improved, and low mean absolute error, high accuracy and excellent pitting fault tolerance accuracy are achieved, which significantly reduces equipment maintenance costs, improves the reliability and safety of industrial equipment, and shows great potential in actual industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The present invention is a flowchart of an embodiment of gear pitting fault diagnosis.

[0065] Figure 2 This is a structural block diagram of an embodiment of the pitting fault diagnosis model of the present invention.

[0066] Figure 3 It is a schematic diagram of an embodiment of the basic model for pitting fault diagnosis of the present invention.

[0067] Figure 4 The figure is a schematic diagram of an embodiment of the present invention for preprocessing a target vibration signal.

[0068] Figure 5 A schematic diagram of an embodiment of a coding unit of the present invention. DETAILED DESCRIPTION

[0069] The present invention will be further described below in conjunction with specific drawings and embodiments.

[0070] In order to efficiently diagnose gear pitting faults under different working conditions, the present invention provides a gear pitting fault diagnosis method. Specifically, the fault diagnosis method includes:

[0071] A target vibration signal is obtained when the target gear set rotates, and the obtained target vibration signal is preprocessed to generate a target vibration time-frequency feature information group corresponding to the target vibration signal after the preprocessing, wherein:

[0072] The target vibration time-frequency feature information group includes a plurality of target vibration time-frequency feature information, and the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is consistent with the number of reference rotation cycles contained in the target vibration signal, so that each target vibration time-frequency feature information is used to characterize the vibration state of a reference rotation cycle of the target gear set;

[0073] The target vibration time-frequency feature information in the target vibration time-frequency feature information group is loaded into the pitting fault diagnosis model, and the pitting fault diagnosis model is used to perform diagnostic identification processing on each target vibration time-frequency feature information, so as to generate a pitting fault prediction category corresponding to each target vibration time-frequency feature information after the diagnostic identification processing, wherein:

[0074] For any target vibration time-frequency characteristic information, the diagnostic identification processing performed by the pitting fault diagnosis model includes sequential shared feature extraction processing and category classification processing, among which,

[0075] The loaded target vibration time-frequency feature information is subjected to shared feature extraction processing to generate corresponding target vibration time-frequency shared features;

[0076] The extracted target vibration time-frequency shared features are processed by category classification to generate the corresponding pitting fault prediction category;

[0077] When the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is not less than 2, all pitting fault prediction categories are subjected to voting processing to determine the corresponding pitting fault voting category after the voting processing, and the pitting fault voting category is used as the pitting fault category of the target gear set.

[0078] It should be noted that the target gear set generally includes at least two gears, which mesh with each other, and the type and meshing mode of the gears in the target gear set can be consistent with the prior art. When diagnosing gear pitting faults, it is mainly aimed at the fault that only one gear in the target gear set may have pitting, that is, there will not be two or more gears in the target gear set that have pitting at the same time. Therefore, before fault diagnosis, the gears in the target gear set should be inspected and judged using technical means commonly used in the technical field.

[0079] In one embodiment of the present invention, when performing gear pitting fault diagnosis, the category of pitting fault in the target gear set is mainly determined. Specifically, the category of pitting fault refers to the level of gear pitting. For example, the gear pitting can be set to N levels. After performing gear pitting fault diagnosis, the level of gear pitting in the target gear set can be determined. For example, the gear pitting can be set to 11 levels. When the gear pitting is level 0, it indicates that all gears in the target gear set do not have pitting faults; when the gear pitting is level 1, it indicates that there is a pitting fault in one gear in the target gear set, and the level of the pitting fault is level 1. Other situations are similar and will not be explained one by one here. Of course, when the level of gear pitting is different, it indicates that the severity of gear pitting is different. For example, when the pitting level is level 1, the severity of gear pitting is less than the pitting degree of the gear when the pitting level is level 10.

[0080] Figure 1 A flow chart of an embodiment of gear pitting fault diagnosis for gears in a target gear set is shown in FIG. As can be seen from the figure, during fault diagnosis, a target vibration signal of the target gear set during rotation should be obtained. The target vibration signal can characterize the vibration state of the gears in the target gear set during rotation. The target vibration signal can be obtained in an existing commonly used manner, such as by a three-axis acceleration sensor. Specifically, the target gear set is generally assembled in a gear box, and the three-axis acceleration sensor can be installed in a position of the gear box housing near the output end bearing seat. Of course, the three-axis acceleration sensor can also be installed in other positions, so as to effectively obtain the required target vibration signal.

[0081] In order to be able to perform pitting fault diagnosis on the gears in the target gear set under different working conditions, the signal time length of the acquired target vibration signal should generally include the time length corresponding to several reference rotation cycles. At this time, the target vibration signal can be regarded as sampling and generating the vibration of the target gear set during rotation through several reference rotation cycles, that is, the target vibration signal can be regarded as the serial connection of the sampling signals under several reference rotation cycles; generally, the target vibration signal should include at least the signal time length of one reference rotation cycle. Specifically, the different working conditions of the target gear set specifically refer to the different speed conditions of the target gear set.

[0082] It should be noted that the reference rotation period generally refers to the period of one rotation of the reference gear in the target gear set. The reference gear is generally selected according to the actual situation. For example, when the target gear set includes two gears with different outer diameters, the gear with the larger outer diameter can be used as the reference gear. At this time, when the reference gear rotates one circle, the gear with the smaller outer diameter must also rotate one circle. The reference rotation period of the target gear set can be determined according to the rotation period of the gear with the larger outer diameter. It can be seen that for a determined target gear set, after determining the reference rotation period, the number of reference rotation periods contained in the target vibration signal can be determined according to the reference rotation period and the signal time length of the target vibration signal.

[0083] After obtaining the target vibration signal, the target vibration signal should be preprocessed first so that the corresponding target vibration time-frequency feature information group can be generated after the preprocessing. Specifically, the target vibration time-frequency feature information group may include one or more target vibration time-frequency feature information. The number of target vibration time-frequency feature information should generally correspond to the number of reference rotation cycles contained in the target vibration signal. If the target vibration signal can be regarded as a vibration sample generated when the target gear set rotates through one reference rotation cycle, then the target vibration time-frequency feature information group should include one target vibration time-frequency feature information. If the target vibration signal can be regarded as a vibration sample generated when the target gear set rotates through two reference rotation cycles, then the target vibration time-frequency feature information group should include two target vibration time-frequency feature information. For other corresponding situations between the target vibration time-frequency feature information and the reference rotation cycle, please refer to the description here, which will not be listed here one by one. The method and process of preprocessing the target vibration signal can refer to the corresponding description below.

[0084] Since the target vibration time-frequency characteristic information in the target vibration time-frequency characteristic information group corresponds one-to-one to the reference rotation period contained in the target vibration signal, at this time, a target vibration time-frequency characteristic information can be used to characterize the vibration state of the target gear set in a reference rotation period. The target vibration time-frequency characteristic information specifically refers to the corresponding state of vibration frequency-vibration time. The situation of the target vibration time-frequency characteristic information can refer to the following specific description.

[0085] In order to realize fault diagnosis, all target vibration time-frequency feature information in the target vibration time-frequency feature information group should be loaded into the pre-built pitting fault diagnosis model. After that, the pitting fault diagnosis model is used to perform diagnostic identification processing on each loaded target vibration time-frequency feature information, and after the diagnostic identification processing, the pitting fault prediction category corresponding to the current target vibration time-frequency feature information can be obtained, that is, for each target vibration time-frequency feature information, after the diagnostic identification processing, a corresponding pitting fault prediction category will be generated. From the above description, it can be seen that the pitting fault prediction category is the prediction level of pitting.

[0086] The pitting fault diagnosis model performs the same diagnostic identification processing on each target vibration time-frequency feature information. For example, the diagnostic identification processing can be a shared feature extraction processing and a category classification processing performed in sequence, wherein the loaded target vibration time-frequency feature information is subjected to shared feature extraction processing to generate the corresponding target vibration time-frequency shared features; the extracted target vibration time-frequency shared features are subjected to category classification processing to generate the corresponding pitting fault prediction category. The pitting fault diagnosis model and the process of performing diagnostic identification processing will be described in detail below.

[0087] It can be seen from the above description that the target vibration time-frequency feature information group may include one or more target vibration time-frequency feature information, and each target vibration time-frequency feature information will output a corresponding pitting fault prediction category after the pitting fault diagnosis model. Therefore, when the target vibration time-frequency feature information group includes multiple target vibration time-frequency feature information, there will be multiple pitting fault prediction categories. In order to obtain the pitting fault category of the target gear set, in one embodiment of the present invention, when the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is not less than 2, all pitting fault prediction categories are voted, so that the corresponding pitting fault voting category can be determined after the voting process, and the pitting fault voting category is used as the pitting fault category of the target gear set, that is, to realize the fault diagnosis of gear pitting of the target gear set. The voting processing method and process can refer to the corresponding process description below.

[0088] In one embodiment of the present invention, preprocessing the target vibration signal includes:

[0089] Normalizing the target vibration signal in the time dimension to generate a target vibration normalized signal after normalization;

[0090] Based on the reference rotation period of the target gear set, the target vibration normalized signal is segmented to segment the target vibration signal into a plurality of target vibration segmented signals, wherein one target vibration segmented signal corresponds to one reference rotation period included in the target vibration signal;

[0091] Determine the control reference sampling point number based on the rotation speed state of the target gear set, and perform sampling point number control processing on each target vibration segmentation signal based on the control reference sampling point number, so as to generate a target vibration sampling point number control signal after the sampling point number control processing, wherein the sampling point number included in the target vibration sampling point number control signal is consistent with the control reference sampling point number;

[0092] Perform short-time Fourier transform on each target vibration sampling point control signal to generate a corresponding vibration time-frequency basic signal after short-time Fourier transform, where:

[0093] The basic vibration time-frequency signal includes a horizontal vibration time-frequency sub-signal, an axial vibration time-frequency sub-signal, and a vertical vibration time-frequency sub-signal;

[0094] For each vibration time-frequency basic signal, the contained horizontal vibration time-frequency sub-signal, axial vibration time-frequency sub-signal and vertical vibration time-frequency sub-signal are channel-joined to generate target vibration time-frequency characteristic information after channel-joining;

[0095] Based on all target vibration time-frequency feature information, a target vibration time-frequency feature information group is generated.

[0096] It can be seen from the above description that the target vibration signal can be measured by a three-axis acceleration sensor. Therefore, when a three-axis acceleration sensor is used to measure and obtain the target vibration signal, the target vibration signal includes a horizontal vibration component, an axial vibration component and a vertical vibration component, wherein the horizontal vibration component, the axial vibration component and the vertical vibration component all represent the corresponding relationship between the vibration amplitude and time in the corresponding direction, and the corresponding meanings of the horizontal vibration, axial vibration and vertical vibration are consistent with the prior art and will not be repeated here. Figure 4 An embodiment of a target vibration signal is shown in FIG. Figure 4 In the target vibration signal, the X channel, Y channel, and Z channel are the horizontal vibration component, the axial vibration component, and the vertical vibration component, respectively.

[0097] Depend on Figure 4 It can be seen that when preprocessing the target vibration signal, the target vibration signal is first normalized in the time dimension. When normalizing in the time dimension, it is necessary to calculate the mean and standard deviation of the target vibration signal in all components, such as the corresponding mean and standard deviation of the horizontal vibration component, the axial vibration component, and the vertical vibration component. Thereafter, based on the calculated mean and standard deviation, a typical normalization calculation method is used for normalization to obtain a target horizontal vibration normalized sub-signal, a target axial vibration normalized sub-signal, and a target vertical vibration normalized sub-signal after normalization. It can be seen that the target vibration normalized signal should include a target horizontal vibration normalized sub-signal, a target axial vibration normalized sub-signal, and a target vertical vibration normalized sub-signal.

[0098] It can be seen from the above description that a target vibration signal can be regarded as being formed by sampling one or more reference rotation periods, and each target vibration time-frequency characteristic information corresponds to only one reference rotation period. Therefore, in order to obtain the target vibration time-frequency characteristic information, the target vibration signal should be segmented. In specific implementation, before segmentation, the segmentation reference time length is determined based on the reference rotation period of the target gear set. Generally, the segmentation reference time length is the time length corresponding to the reference rotation period of the target gear set.

[0099] After determining the reference time length for segmentation, the target vibration signal should be segmented based on the reference time length for segmentation, so as to segment the target vibration signal into a plurality of target vibration segmentation signals. At this time, a target vibration segmentation signal corresponds to the signal time of a reference rotation cycle contained in the target vibration signal. It can be understood that each target vibration segmentation signal also includes corresponding horizontal vibration components, axial vibration components and vertical vibration components. For example, the horizontal vibration component in the target vibration segmentation signal should be formed after segmentation based on the above-mentioned target horizontal vibration normalized sub-signal. The situation of the axial vibration component and the vertical vibration component in the target vibration segmentation signal can be referred to the description here and will not be repeated.

[0100] When segmenting the target vibration signal, the segmentation can be performed in the time sequence of sampling to form the target vibration signal. When the remaining segmentation portion is less than the segmentation reference time length corresponding to a reference rotation cycle, it should be discarded, such as Figure 4 As shown in the figure, an embodiment of discarding the corresponding tail of the target vibration signal is shown. In specific implementation, when the target gear set is different, the corresponding reference rotation period may be different, but because they are all divided according to the reference rotation period of the target gear set, the target vibration segmentation signal formed by the segmentation can be adapted to the current target gear set. Therefore, after the segmentation based on the reference rotation period and the corresponding segmentation reference time length, the adaptability of subsequent pitting fault diagnosis can be improved, and the reliability of the pitting fault diagnosis model for pitting fault diagnosis of different target gear sets can be improved.

[0101] It can be seen from the above description that when the target vibration signal is segmented based on the reference rotation period, each target vibration segmentation signal has the same time length, but the number of sampling points contained in each target vibration segmentation signal may not be exactly the same. In order to adapt to different target gear sets and solve the structural differences caused by speed changes, the number of control reference sampling points can be determined based on the rotation speed state of the target gear set, wherein the rotation speed state of the target gear set specifically refers to a certain interval speed of the target gear set as the reference rotation speed for aligning the sampling points, such as the reference rotation speed can be the medium speed of the target gear set, such as the 700-1500 rpm of the reference gear as the medium speed of the target gear set. At this time, the corresponding number of control reference sampling points can be determined according to the target sampling frequency and the reference rotation speed of the target gear set. The target sampling frequency is related to the rotation of the target gear set and the type of the three-axis acceleration sensor, and can be selected according to needs.

[0102] Based on the control reference sampling points, the sampling point number control processing is performed on each target vibration segmentation signal, wherein the sampling point number control processing can be a linear interpolation processing, such as Figure 4As shown; specifically, according to the target sampling frequency and the reference rotation speed of the target gear set, the position of each sampling point corresponding to the reference sampling point number can be determined; during linear interpolation, when there is no corresponding vibration component at the position of a certain sampling point, the vibration component value of the corresponding sampling point position can be generated through linear difference; in addition, when the target vibration segmentation signal has a corresponding vibration component outside the determined sampling point position, it should be discarded, so that after the sampling point number control processing, a target vibration sampling point number control signal can be generated, and the number of sampling points included in the target vibration sampling point number control signal is consistent with the control reference sampling point number, and at the same time, the sampling point positions included in all target vibration sampling point number control signals are also the same, and at this time, the target vibration sampling point number control signal is in a completely aligned state.

[0103] After the target vibration sampling point number control signal is obtained through the above-mentioned sampling point number control processing, a short-time Fourier transform is performed on each target vibration sampling point number control signal to generate a corresponding vibration time-frequency basic signal after the short-time Fourier transform, wherein the method and process of performing short-time Fourier transform on the target vibration sampling point number control signal may be consistent with the prior art, and the parameter setting of the short-time Fourier transform (STFT) may select typical parameters, such as: FFT length is 256, window length is 256, window overlap is 32, and window type is Hann.

[0104] It can be seen from the above description that the target vibration segmentation signal includes corresponding horizontal vibration components, axial vibration components and vertical vibration components at the same time. Therefore, the target vibration sampling point number control signal also includes corresponding horizontal vibration components, axial vibration components and vertical vibration components. Therefore, after short-time Fourier transform, the vibration time-frequency basic signal can be obtained, including horizontal vibration time-frequency sub-signal, axial vibration time-frequency sub-signal and vertical vibration time-frequency sub-signal. Specifically, short-time Fourier transform is performed on the horizontal vibration component of the target vibration sampling point number control signal to obtain a horizontal vibration time-frequency sub-signal. Similarly, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal can be obtained in a similar way.

[0105] For each vibration time-frequency basic signal, the horizontal vibration time-frequency sub-signal, axial vibration time-frequency sub-signal and vertical vibration time-frequency sub-signal contained therein are channel-spliced ​​to generate the target vibration time-frequency feature information after channel splicing. It should be noted that after the target vibration time-frequency feature information is generated by channel splicing, the pitting fault diagnosis model can easily learn the dependency between different frequencies and vibration directions.

[0106] In a specific implementation, when the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal are channel-joined, the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal are aligned as a whole. Figure 4 An embodiment of channel splicing to form target vibration time-frequency characteristic information is shown in the figure. In the figure, the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal are aligned in sequence, and the axial vibration time-frequency sub-signal is located between the horizontal vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal. Specifically, after the channel splicing, the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal remain unchanged.

[0107] It can be understood that the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal have the same number of sampling frequency values, and the position of each sampling frequency value is the same. Therefore, when the channels are spliced, the splicing is performed according to the position of the sampling frequency value. Generally, the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal can all be regarded as a two-dimensional frequency matrix. When the channels are spliced, a larger splicing frequency matrix can be formed. At this time, the dimension of the splicing frequency matrix in the time dimension remains unchanged, and the dimension in the frequency dimension is the sum of the corresponding dimensions of the horizontal vibration time-frequency sub-signal, the axial vibration time-frequency sub-signal and the vertical vibration time-frequency sub-signal.

[0108] Based on all the target vibration time-frequency feature information, a target vibration time-frequency feature information group can be generated, and the target vibration time-frequency feature information in the target vibration time-frequency feature information group can be regarded as independent of each other. In addition, after obtaining the target vibration time-frequency feature information, the target vibration time-frequency feature information can also be normalized in the time dimension, such as Figure 4 As shown, after the target vibration time-frequency feature information is normalized in the time dimension, the target vibration time-frequency feature information group contains the normalized target vibration time-frequency feature information. Specifically, the method of normalizing the target vibration time-frequency feature information in the time dimension can refer to the above-mentioned description of normalization in the corresponding time dimension, which will not be repeated here.

[0109] In one embodiment of the present invention, the pitting fault diagnosis model includes a shared feature extraction module and a category classifier adapted to be connected to the shared feature extraction module, wherein:

[0110] When performing diagnostic identification processing, the shared feature extraction module is used to extract shared features from the target vibration time-frequency feature information, where:

[0111] When performing shared feature extraction, the target vibration time-frequency feature information is first position-encoded to generate target vibration position encoding information after position encoding;

[0112] Extracting features from target vibration position encoding information to generate target time-frequency shared features after feature extraction;

[0113] Based on the target time-frequency shared features, the category classifier generates pitting fault prediction information corresponding to the current target vibration time-frequency feature information, and determines the corresponding pitting fault prediction category based on the pitting fault prediction information, wherein:

[0114] The pitting fault prediction information includes confidences corresponding to N pitting categories;

[0115] When determining the corresponding pitting fault prediction category based on the pitting fault prediction information, the pitting fault category with the highest confidence is used as the pitting fault prediction category.

[0116] Figure 2 An embodiment of a pitting fault diagnosis model is shown in the figure. In the figure, the pitting fault diagnosis model includes a shared feature extraction module and a category classifier. At this time, the shared feature extraction module can be used to perform shared feature extraction processing during diagnostic identification processing, and the category classifier can be used to perform category classification processing during diagnostic identification processing.

[0117] When the shared feature extraction module is used to perform shared feature extraction, the target vibration time-frequency feature information is first position-encoded, and then feature extraction is performed after position encoding. After position encoding, target vibration position encoding information is generated, and after feature extraction of the target vibration position encoding information, the target time-frequency shared feature can be obtained.

[0118] When performing the category classification process, the category classifier will generate corresponding pitting fault prediction information based on the target time-frequency shared features. It can be seen from the above description that the present invention divides the pitting fault categories of gears into N categories. Therefore, when the category classifier generates pitting fault prediction information, it should include the confidence of N pitting categories, that is, the pitting fault prediction information includes N confidences, each confidence corresponds to a pitting category, and then the pitting category with the highest confidence can be used as the pitting fault prediction category. It should be noted that the confidence of each pitting category is a value of 0 to 1. The pitting fault prediction category can be obtained by comparing the confidence values. For example, the confidence of the pitting category with level 0 is the highest, then the current pitting fault prediction category should be level 0. For other situations, please refer to the description here.

[0119] In one embodiment of the present invention, the shared feature extraction module includes a position encoding module and a Transformer encoder for shared feature extraction, wherein:

[0120] When position encoding is performed on the target vibration time-frequency feature information, a position index is performed based on each time-frequency feature in the target vibration time-frequency feature information, and sine encoding or cosine encoding is performed based on the position of the time-frequency feature, so as to generate a corresponding time-frequency feature encoding after the position encoding, and the corresponding target time-frequency feature encoding information is generated based on the time-frequency feature encoding of all time-frequency features;

[0121] Based on the target time-frequency feature coding information and the target vibration time-frequency feature information, the target vibration position coding information is generated, wherein when the target vibration position coding information is generated, at least the time-frequency feature coding information and the target vibration time-frequency feature information are added in a position-by-position manner.

[0122] Figure 3 An embodiment of a shared feature extraction module is shown in FIG. Figure 3 In the embodiment, the shared feature extraction module may include a position encoding module and a Transformer encoder, wherein the position encoding module may be used to perform the above-mentioned position encoding of the target vibration time-frequency feature information, and the Transformer encoder may be used to perform the above-mentioned feature extraction processing of the target vibration position encoding information.

[0123] The position coding module can adopt the existing commonly used position coding method, which is based on the specific method that can realize the position coding. It can be seen from the above description that the target vibration time-frequency feature information includes several time-frequency features, and the time-frequency features in the target vibration time-frequency feature information are distributed in a matrix manner. At this time, the time-frequency features in the target vibration time-frequency feature information can be indexed according to the distribution position of the time-frequency features, and sine coding or cosine coding is performed based on the position of the time-frequency features. The following examples illustrate the cases of sine coding and cosine coding.

[0124] Specifically, for sine coding and cosine coding, we have:

[0125]

[0126] Among them, pos is the position index of the time-frequency feature, PE (pos,2i) and PE (pos,2i+1) They represent the sine code and cosine code of the posth position respectively, i is the index of the dimension, d model is the encoding dimension.

[0127] In specific implementation, the dimension of the target time-frequency feature encoding information should be consistent with the dimension of the target vibration time-frequency feature information. Therefore, when the time-frequency features are indexed and position-encoded, each time-frequency feature is only sine-encoded or cosine-encoded. For example, when the time-frequency features in the target vibration time-frequency feature information exist in the form of a two-dimensional matrix, the same encoding form can be used for all time-frequency features in the same column, such as sine-encoding each time-frequency feature in the even column and cosine-encoding each time-frequency feature in the odd column. After position-encoding each time-frequency feature in this way, the target time-frequency feature encoding information can be obtained. The feature sequence can be enhanced through sine encoding and cosine encoding to enhance the expression ability.

[0128] From the above description, it can be seen that the target time-frequency feature encoding information can be a two-dimensional matrix, and each time-frequency feature is an element in the two-dimensional matrix. At this time, the position index pos of the time-frequency feature is the row position of the two-dimensional matrix, and the encoding dimension d model is the total number of columns of the two-dimensional matrix, and the index i of the dimension is the column position of the two-dimensional matrix.

[0129] After obtaining the target time-frequency feature encoding information, the time-frequency feature encoding information and the target vibration time-frequency feature information are added in parallel, so that the target vibration position encoding information can be obtained after the parallel addition. Thereafter, the Transformer encoder can perform feature extraction on the target vibration position encoding information to obtain the target time-frequency shared feature.

[0130] In one embodiment of the present invention, the Transformer encoder includes a plurality of encoding units connected in series, wherein:

[0131] For any encoding unit, it includes a multi-head attention unit, a multi-head attention layer normalization layer, a feedforward layer unit, and a feedforward layer normalization layer connected in sequence, where:

[0132] The multi-head attention unit includes a multi-head attention layer with residual connections;

[0133] The feed-forward layer unit includes a feed-forward layer using a residual connection;

[0134] For two serially connected coding units, along the serial connection direction of the coding units, the feedforward normalization layer of the previous coding unit is connected to the multi-head attention layer of the next coding unit.

[0135] When extracting features from target vibration position coding information, the target vibration position coding information is extracted by sequentially processing the serially connected coding units, and target time-frequency sharing features representing pitting categories and domain category sharing are output.

[0136] Specifically, the Transformer encoder can be composed of a number of sequentially connected encoding units. For example, the Transformer encoder can be composed of 4 sequentially connected encoding units. Figure 5 A schematic diagram of an encoding unit is shown in FIG. 1 , in which the multi-head attention layer can form a multi-head attention unit through residual connection, and the feedforward layer can form a feedforward layer unit through residual connection. Figure 5 In the figure, Add&Nordm connected to the multi-head attention layer includes a multi-head attention layer residual connection splicer and a multi-head attention layer normalization layer. The input end and the output end of the multi-head attention layer are both connected to the multi-head attention layer residual connection splicer to form a residual connection form, and the output end of the multi-head attention layer residual connection splicer is connected to the multi-head attention layer normalization layer.

[0137] In specific implementation, the multi-head attention layer can adopt the existing commonly used form. When the multi-head attention layer adopts residual connection to form a multi-head attention unit, it can prevent the gradient from disappearing and accelerate convergence. The multi-head self-attention mechanism can focus on different sequences of time-frequency features on the frequency axis, capture the long-distance dependency between different frequencies, and provide more robust features for gear pitting fault diagnosis.

[0138] The feedforward layer can adopt the existing commonly used form, such as including a feedforward first linear layer, a Dropout layer, a ReLU activation function and a feedforward second linear layer connected in sequence. The feedforward layer adopts the form of residual connection to form a feedforward layer unit, which can refer to the form of residual connection of the multi-head attention layer to form a multi-head attention unit. In this case, the feedforward linear layer serves as the input layer of the feedforward layer, and the feedforward second linear layer serves as the output end of the feedforward layer. Figure 5 The Add&Nordm connected to the feedforward layer can refer to the corresponding description above, and the specific one is based on the residual connection to form the feedforward layer unit. The formed feedforward layer unit can perform nonlinear transformation on the output of the multi-head attention unit to further extract deep features.

[0139] When the feedforward layer adopts the above structure and the dropout probability of the Dropout layer is set to drop_prob=0, for the feedforward layer, we have:

[0140] , whose input z ma The calculation formula is:

[0141] FFN(z ma )=ReLU(z ma W1+b1)W2+b2

[0142] Among them, z ma is the feature map from the input to the first linear layer of the feedforward, FFN(z ma) is the output feature map of the second feedforward linear layer, W1 and W2 are the corresponding weights of the first feedforward linear layer and the second feedforward linear layer, b1 and b2 are the corresponding bias terms, and ReLU is the ReLU activation function.

[0143] When extracting features from target vibration position coding information, the target vibration position coding information is extracted by sequentially processing the serially connected coding units, wherein a Transformer encoder formed by serially connecting multiple coding units can capture the multi-level representation of the target vibration position coding information, and finally output a global feature representation containing context awareness, thereby outputting the pitting category and the target time-frequency shared features shared by the domain category, wherein the domain category specifically refers to the speed range of the target gear set.

[0144] In one embodiment of the present invention, the category classifier includes a category classification main unit and a category classification output unit adaptively connected to the category classification main unit, wherein:

[0145] The category classification output unit adopts the LinSoftmax mapping function of approximate linear mapping. When the LinSoftmax mapping function is used for mapping output, then:

[0146]

[0147] Among them, LinSoftmax(z o ) i is the i-th confidence level of the output of the LinSoftmax mapping function, α is the slope, γ is the adjustment parameter, and z o is the logits vector output by the main unit for category classification, Output logits vector z for the category classification master unit o The i-th category classification quantity in , Output logits vector z for the category classification master unit o The j-th category classification quantity in , is the penalty weight of the i-th pitting category, is the penalty weight for the jth pitting category.

[0148] Specifically, the category classifier can adopt an existing commonly used form, such as the category classifier can take the form of a category classification main unit and a Softmax mapping function. The category classification main unit generally may include a ReLU activation function, batch normalization, a linear layer, and several fully connected layers. The form in which the category classification main unit and the Softmax mapping function constitute a classifier, and the way in which the classifier is constructed for classification can all be consistent with the existing technology and will not be repeated here. Figure 3 An embodiment of the category classification main unit is shown in Figure 3It can be seen that the category classification main unit includes three fully connected layers. The i-th category classification quantity output by the above category classification main unit is The j-th category classification quantity output by the category classification main unit They are all output by the fully connected layer connected to LinSoftmax.

[0149] It should be noted that when the existing classifier uses the Softmax mapping function, the output is converted into probability through an exponential function. This exponential growth causes the probability of large value input to increase rapidly, which easily causes the model to be overconfident in a specific category, resulting in reduced sensitivity to inter-class differences, affecting classification accuracy. In order to alleviate the problem of overconfidence in the Softmax mapping function used in the existing classifier, which leads to low accuracy in gear pitting fault diagnosis, in one embodiment of the present invention, the Softmax mapping function in the classifier is replaced with the LinSoftmax mapping function. The LinSoftmax mapping function introduces an approximately linear mapping function that has a more moderate response to the growth of the input value, thereby reducing the model's tendency to be overconfident.

[0150] From the above-mentioned LinSoftmax mapping function, it can be seen that the LinSoftmax mapping function combined with the weighting mechanism can further enhance the control of classification errors by introducing penalty weights to increase the penalty for misclassified samples. During inference, the category classifier classifies the target time-frequency shared features and uses the LinSoftmax mapping function to map the output to a probability distribution, increasing the sparsity of the pitting fault diagnosis model output and providing a more robust optimization direction.

[0151] In the specific implementation, in the above-mentioned LinSoftmax mapping function, the slope α can be set to 0.8, and the adjustment parameter γ can be set to 8. From the above description, it can be seen that the LinSoftmax mapping function can simultaneously map N confidences. Therefore, the logits vector z output by the category classification main unit o Should contain N elements, Output logits vector z for the category classification master unit o It should be noted that the logits vector z output by the main unit of category classification is o The situation is consistent with the prior art and will not be described in detail here.

[0152] In one embodiment of the present invention, the voting process includes:

[0153] For each pitting fault prediction category corresponding to the target vibration time-frequency characteristic information, configure the voting samples as the corresponding pitting category;

[0154] Count the number of voting samples for each pitting category, and use the pitting category with the largest number of voting samples as the pitting fault voting category of the target gear set;

[0155] When the number of voting samples for each pitting corrosion category is consistent, the pitting corrosion fault prediction category with the highest confidence is taken as the pitting corrosion fault voting category.

[0156] It can be seen from the above description that after each target vibration time-frequency feature information is processed by diagnostic identification, a pitting fault prediction category will be generated by the above-mentioned category classifier. At this time, the pitting fault prediction category is used as a voting sample for the corresponding pitting category. For example, if the pitting fault prediction category is level 1, the current pitting fault prediction category is used as a voting sample of level 1. The situation of other voting samples can be referred to the description here.

[0157] After preprocessing, the target vibration signal can obtain one or more target vibration time-frequency feature information. Therefore, after the same diagnostic identification processing, the corresponding pitting fault prediction category and the voting samples of the corresponding pitting category can be obtained respectively. Thereafter, the number of voting samples of each pitting category is counted, and the pitting category with the largest number of voting samples is used as the pitting fault voting category of the target gear set. For example, after preprocessing, the target vibration signal can obtain 5 target vibration time-frequency feature information. After diagnostic identification processing is performed on the 5 target vibration time-frequency feature information, 5 corresponding pitting fault prediction categories can be obtained, and the 5 pitting fault prediction categories are respectively used as 1 level 1 voting template, 3 level 2 voting samples, and 1 level 3 voting sample. At this time, the number of voting samples of level 2 is the largest, and the corresponding pitting fault is level 2, that is, the pitting fault voting category of the target gear set should be pitting level 2.

[0158] In addition, when the number of voting samples for each pitting category is consistent, since each pitting category voting sample corresponds to a confidence level, all corresponding confidence levels can be compared, and the pitting fault prediction category with the highest confidence level is used as the pitting fault voting category. It can be understood that when voting is used, the difference in pitting fault prediction categories caused by multiple target vibration time-frequency special diagnostic information can be effectively avoided, and the accuracy and reliability of pitting fault diagnosis for the target gear set can be improved.

[0159] In one embodiment of the present invention, when constructing a pitting fault diagnosis model, it includes:

[0160] Constructing a pitting fault diagnosis basic model, wherein the pitting fault diagnosis basic model includes a shared feature extraction basic module, a category basic classifier and a domain classifier, and both the category basic classifier and the domain classifier are connected to the shared feature extraction basic module;

[0161] A pitting fault diagnosis training data set is constructed, wherein the pitting fault diagnosis training data set includes a number of training samples, wherein:

[0162] For any training sample, including the training vibration time-frequency feature information corresponding to the training gear set, the pitting category label of the current training gear set, and the label of the field speed,

[0163] The pitting fault diagnosis training data set contains at least N types of pitting category labels;

[0164] When the pitting fault diagnosis training data set is used to train the pitting fault diagnosis basic model, the model loss function used is:

[0165]

[0166] in, is the model loss function, is the category loss calculated based on the classification base classifier, is the domain loss calculated based on the domain classifier, is the association loss calculated based on the classification base classifier and the domain classifier, ω cls is the class loss weight, ω dom is the domain loss weight, ω cor is the associated loss weight;

[0167] When the training of the pitting fault diagnosis basic model reaches the target training state, the shared feature extraction basic module is configured as a shared feature extraction module, the category basic classifier is configured as a category classifier, and a pitting fault diagnosis model is formed based on the shared feature extraction module and the category classifier.

[0168] It can be understood that when constructing a pitting fault diagnosis model, a pitting fault diagnosis basic model should be constructed first. Thereafter, the pitting fault diagnosis basic model is trained, and after the training reaches the target training state, a pitting fault diagnosis model can be generated based on the pitting fault diagnosis basic model. Figure 3 An embodiment of the basic model for pitting fault diagnosis is shown in FIG. Different from the pitting fault diagnosis model, the pitting fault diagnosis model also includes a domain classifier. It can be seen that the domain classifier is mainly used for training and does not participate in the reasoning process.

[0169] It should be noted that the basic module for shared feature extraction in the pitting fault diagnosis basic model can be in the same form as the shared feature extraction module of the pitting fault diagnosis model, such as the position encoding module and Transformer encoder mentioned above. The specific situation of the position encoding module and Transformer encoder can refer to the above description. The domain classifier preferably adopts the same form as the category basic classifier, and the specific situation can refer to the description of the above category classifier. Figure 3 In this paper, domain classification refers to the domain classifier, and class classification refers to the category-based classifier.

[0170] For the above-mentioned pitting fault diagnosis basic model, a pitting fault diagnosis training data set should be constructed, wherein the pitting fault diagnosis training data set should include a number of training samples, each training sample includes the training vibration time-frequency feature information corresponding to the training gear set, the pitting category label of the current training gear set, and the label of the field speed. Specifically, the training vibration time-frequency feature information can refer to the description of the above-mentioned target vibration time-frequency feature information, and the pitting category label is the pitting level of the current training gear set, such as the pitting level described above is level 0, level 1 or other situations. It can be seen from the above description that during reasoning, it can be determined that the pitting level of the target gear set is N types. Therefore, the types of pitting category labels contained in the pitting fault diagnosis training data set are at least N. The following example illustrates the method of constructing a pitting fault diagnosis training data set.

[0171] Specifically, a training gear set is provided. The training gear set can be in the same form as the target gear set. Of course, the training gear set and the target gear set can also be different, but the training gear set should also have at least two gears that mesh with each other. After the training gear set is provided, a three-axis acceleration sensor for collecting vibration signals is configured. It can be understood that the vibration signal collected at this time is a training vibration signal. The training vibration signal can refer to the description of the target vibration signal above.

[0172] In order to obtain different training samples, the gears in the training gear set should be configured with different pitting levels. As described above, the pitting level is 0 to N-1. When the pitting level is 0, it should be ensured that all gears in the training gear set have no pitting. When the pitting level is 1 to N-1, the pitting conditions of the gears in the training gear set can be specifically referred to here and the corresponding descriptions above, which will not be repeated here. After configuring the training gear sets with different pitting levels, the three-axis acceleration sensor is used to collect vibration signals to obtain training vibration signals under different pitting levels.

[0173] When collecting and generating training vibration signals, the reference gear configuration in the training gear set should also be configured in different speed states. At this time, the field speed information and field speed label corresponding to each vibration training signal can be obtained. Generally, the reference gear should be configured in different speed states, such as low speed state, medium speed state and high speed state, wherein the low speed state can be 100, 200, 300, 400, 500, 600 revolutions per minute (rpm), the medium speed state can be 700, 800, 900, 1000, 1200, 1500 revolutions per minute, and the high speed state can be 1800, 2100, 2400, 2700, 3000, 3600 revolutions per minute.

[0174] After obtaining the training vibration signal, the training vibration signal can be preprocessed using the above-mentioned preprocessing method. After preprocessing, a training vibration time-frequency feature information group corresponding to each training vibration signal can be obtained, and the training vibration time-frequency feature information group includes a number of training vibration time-frequency feature information. From the above description, it can be seen that each training vibration signal corresponds to a pitting level of the training gear set. Therefore, a pitting category label can be made for each training vibration time-frequency feature information. Based on a training vibration time-frequency feature information, the corresponding pitting category label and the field speed label, a training sample can be formed.

[0175] In specific implementation, training vibration signals under different speed states should be collected, but one training sample only includes one domain speed information and the corresponding domain speed label. The domain speed label is to classify the speed of the reference gear. For example, when there are 18 speeds mentioned above, there are 18 corresponding domain speed labels. For example, the domain speed of 100rpm can be configured with the domain speed label as speed 0, and the domain speed of 200rpm can be configured with the domain speed label as speed 1. Other situations are similar and will not be repeated here.

[0176] When using the pitting fault diagnosis training data set to train the pitting fault diagnosis basic model, it is also necessary to configure the model training parameters, such as using the RAdam optimizer, setting the batch size to 256, the total number of training epochs to 60, and the learning rate to decrease with the increase of the number of epochs. Specifically, the learning rate for the first 30 epochs is 4×10 -4 , the next 20 epochs are reduced to 2×10 -4 , the last 10 epochs are further reduced to 1×10 -4 In addition, the class loss weight ω cls , domain loss weight ω dom and the associated loss weight ω cor Both can be set to 100.

[0177] Generally, when all training samples in the pitting fault diagnosis training data set are used to train the pitting fault diagnosis basic model, it is a round of training. When training the pitting fault diagnosis basic model, when the loss function tends to be stable or the total number of training rounds is reached, the training of the pitting fault diagnosis basic model is terminated, and the target training state can be obtained. The pitting fault diagnosis basic model with the corresponding number of rounds when the loss function tends to be stable or the loss function is minimum is selected as the target model, so that the corresponding pitting fault diagnosis model can be formed.

[0178] In specific implementation, the category loss Field loss Cross entropy loss can be used. When cross entropy loss is used, a feasible calculation method can be:

[0179]

[0180] Among them, bs represents the batch size, The label encoding value of pitting category j predicted by the category-based classifier for the i-th training sample, The label encoding value of domain category j predicted by the domain classifier for the i-th training sample, is the probability that the i-th training sample belongs to the pitting category j predicted by the category-based classifier, is the probability that the i-th training sample belongs to domain category j predicted by the domain classifier, N is the number of pitting category labels contained in the pitting fault diagnosis training dataset, and M is the number of domain speed labels in the pitting fault diagnosis training dataset.

[0181] In the above formula, is the category loss during training of a batch, P c It is the pitting fault prediction information obtained by the Linsfotmax mapping function in the category base classifier. From the above description, it can be seen that P c The dimension of should be (bs, N). is the domain loss during training of a batch, P d is the domain prediction information obtained by the Linsfotmax mapping function in the domain classifier, P d The dimension of should be (bs, M). When sampling the above speed state, the number of domain speed labels M in the pitting fault diagnosis training data set should be 18.

[0182] Label encoding value Label encoding value The corresponding value can be 0 or 1. The following takes the category-based classifier as an example to illustrate the above loss function calculation process, specifically:

[0183] Taking the 11 levels of gear pitting fault classification as an example, it is assumed that a training sample corresponds to a pitting category label of 3; the output of the pitting fault diagnosis basic model for the current training sample is a logits vector of length 11, representing the logic value of each category (that is, the logits vector that has not been mapped by the LinSoftmax mapping function).

[0184] The logits vector is mapped by the LinSoftmax mapping function to obtain a probability distribution, which is also a vector of length 11. At this time, the pitting fault prediction information of the current training sample can be obtained. If the pitting fault prediction information is:

[0185] [0.143,0.007,0.032,0.387,0.086,0.019,0.003,0.052,0.235,0.012,0.004].

[0186] Since the corresponding pitting category label of the current training sample is 3, the corresponding probability is 0.387, from which the probability in the above formula can be obtained Further, the corresponding one-hot encoding can be obtained as [0,0,0,1,0,0,0,0,0,0,0]. At this time, the corresponding label encoding value is 1.

[0187] In specific implementation, when the training samples are in other situations, the specific corresponding calculation process can refer to the above description and will not be repeated here.

[0188] From the above description, it can be seen that when the above model training parameters are sampled, the batch size bs should be 256. It can be understood that during model training, the domain classifier and the category base classifier can help the pitting fault diagnosis basic model to extract different feature information, thereby improving the performance of the pitting fault diagnosis basic model on multiple tasks, and further improving the performance of the pitting fault diagnosis model on multiple tasks.

[0189] After performing domain-independent feature extraction, in order to improve the generalization ability of the model, the association loss between the domain classifier and the class base classifier is constructed To reduce the correlation between domain features and category features, we construct the association loss When , the degradation of the feature space is avoided by low-rank approximation. Specifically, for the above association loss Then we have:

[0190]

[0191] in, is the association loss of the same batch of training samples, r jis the correlation coefficient between the category feature and domain feature of the jth training sample in a batch, L1 smooth is the smoothed L1 loss.

[0192] The correlation coefficient r between the category feature and the domain feature of the jth training sample j , then:

[0193]

[0194] in, is a low-rank approximation matrix The i-th element in the j-th row, is a low-rank approximation matrix The i-th element in the j-th row, is a low-rank approximation matrix The mean of the elements in the jth row, is a low-rank approximation matrix The mean of the row corresponding to the mean of the element in the jth row, d z is a low-rank approximation matrix Low rank approximation matrix The dimensions of the corresponding row.

[0195] Depend on Figure 3 It can be seen that when calculating the associated loss When , in the category base classifier, obtain the category feature matrix Z output by the first fully connected layer c , then, for the category feature matrix Z c Perform singular value decomposition, category feature matrix Z c The singular value decomposition of can be expressed as:

[0196]

[0197] Among them, U c and V c is an orthogonal matrix, Σ c is a diagonal matrix containing singular values. In order to perform low-rank approximation, the present invention retains the first th singular values ​​and obtains a low-rank approximation matrix For low-rank approximation matrices Then we have:

[0198]

[0199] Among them, th is a hyperparameter set artificially, such as the category feature matrix Z c When the dimension of is (256, 200), th can be set to 120. c The method and process of performing singular value decomposition are consistent with those already described, and will not be described in detail here.

[0200] In the above introduction to the loss function, the corresponding loss is calculated based on a batch of training samples. After that, based on the losses corresponding to all batches of training samples, the corresponding model loss function after one round of training can be obtained. I will not go into details here.

[0201] From the above description, it can be seen that the category base classifier includes three fully connected layers. Generally, the fully connected layer connected to the LinSoftmax mapping function is the third fully connected layer, and the first fully connected layer is the farther fully connected layer.

[0202] In the specific implementation, in the domain classifier, the domain feature matrix Z output by the first fully connected layer can be obtained d , then, the domain feature matrix Z d Perform singular value decomposition and obtain a low-rank approximate matrix Generate a low-rank approximation matrix The method and process can refer to the above generation of low-rank approximate matrix The process description will not be repeated here.

[0203] It should be noted that the penalty weight of the i-th pitting category in the LinSoftmax mapping function is Generally, it should be determined through the model training process. For example, the penalty weight can be trained by the following method: Specifically:

[0204]

[0205] Among them, w represents the intensity coefficient of the penalty weight, which controls the additional penalty for misclassified samples; η represents the basic multiplication weight. Generally, the intensity coefficient w and the basic multiplication weight η are both empirically selected parameters, that is, they can be set based on experience.

[0206] Based on the above model loss function As well as the above model training instructions, the pitting fault diagnosis basic model can be trained. When the target training state is reached, the pitting fault diagnosis model can be constructed. It can be understood that when the target training state is reached, the penalty weight can be determined.

[0207] In order to prove that the pitting fault diagnosis model of the present invention performs diagnostic identification processing on the target vibration signal and determines the effectiveness of pitting fault diagnosis of the target gear set, the gear pitting fault dataset provided in the PHM North America 2023 Conference Data Challenge is verified on the pitting fault diagnosis model of the present invention. The specific verification results are shown in Table 1. Specifically:

[0208] Table 1 Indicators of the validation set at different speeds

[0209]

[0210]

[0211] In order to verify the generalization ability of the model, all samples at low speed, medium speed and high speed were retained as test sets, and the remaining samples were used for training and verification. A total of six groups of experiments were conducted, called Experiment 1 to Experiment 6, and the specific experimental results are shown in Table 1. In the six groups of experiments, the MAE of Experiment 2 to Experiment 5 was less than or equal to 0.25. In addition, when the error tolerance of the pitting level was set to 1, the average prediction error tolerance accuracy exceeded 95%. Whether in low-speed (100rpm) or high-speed (3600rpm) samples, due to the lack of nearby speed samples, the MAE and tolerance of 1 are usually better than the high-speed samples above 1800rpm, and the MAE remains below 0.2. In contrast, the MAE of high-speed samples may be as high as 0.62, but it is generally still within the acceptable range. The samples of 100rpm and 3600rpm also show significant differences, and the performance of the 100rpm sample is further reduced due to the lack of nearby speed samples. In addition, the structural differences between the 100 rpm and 200 rpm samples are more obvious than those between the 3600 rpm and 3000 rpm samples, further degrading the performance of the 100 rpm sample.

[0212] After removing the edge speed samples, the average index (the average value after removing the edge) is shown in Table 1. The three indexes reach 0.11, 92.18% and 98.06% respectively.

[0213] Table 2 reveals the differences between different feature extraction networks in the gear pothole fault diagnosis task. Compared with the pitting fault diagnosis model proposed in the present invention, traditional convolutional neural networks such as ResNet1011D, ResNet50, EfficientNet V2 M, Xception41 and ConvNeXt V2 base all perform poorly in terms of mean absolute error, accuracy and fault tolerance accuracy. Although ConvNeXt V2base performs best in CNN, it is still inferior to the pitting fault diagnosis model of the present invention in terms of fault diagnosis accuracy and robustness. This shows that the global feature dependency captured by the pitting fault diagnosis model through the self-attention mechanism has obvious advantages over the convolutional network with a local receptive field when dealing with complex gear pothole fault diagnosis problems.

[0214] In addition, the pitting fault diagnosis model also shows good efficiency in the number of parameters, achieving a balance between performance and model complexity, which not only proves the powerful ability of the pitting fault diagnosis model in feature extraction, but also highlights its potential in practical industrial applications, especially in the field of gear fault diagnosis that requires high accuracy and high fault tolerance. These results further verify the effectiveness of the proposed method, especially when dealing with gear pothole fault diagnosis problems under different working conditions, the pitting fault diagnosis model shows excellent generalization ability and robustness.

[0215] Table 2 Comparison of fault diagnosis methods under different methods

[0216]

[0217]

[0218] In the above method model, ResNet101 1D is a residual network architecture with 101 layers, consisting of residual blocks. Each block directly adds input and output through skip connections to avoid the gradient vanishing problem. Its convolution operation is one-dimensional, focusing on feature extraction of time series, and is usually used to process time-based signal data.

[0219] ResNet50 is a lightweight version of the ResNet series, containing 50 layers and composed of multiple residual blocks stacked in structure. It uses a two-dimensional convolution kernel to perform convolution operations on the input two-dimensional feature map, maintains feature flow through jump connections, and strikes a balance between network depth and computational efficiency.

[0220] EfficientNet V2M is a medium-sized model in the EfficientNet family. It optimizes depth, width, and resolution based on a compound scaling method, uses grouped convolution and mobile convolution modules, combined with Swish activation functions and deep feature hierarchy structures to process time-frequency inputs in an efficient manner.

[0221] Xception41 is a network based on depthwise separable convolutions. It consists of 41 layers and replaces the traditional convolution operation. By dividing the convolution into two independent operations in the depth direction and the spatial direction, it greatly improves the computational efficiency while maintaining strong feature modeling capabilities.

[0222] ConvNeXt V2Base is a modern convolutional network that is optimized based on the standard convolutional architecture. It includes grouped convolutions, normalization layers, and deep feature fusion modules. It draws on the design ideas of Transformer and can efficiently model the local and global information of input features.

[0223] In summary, a gear pitting fault diagnosis system can be obtained. In one embodiment of the present invention, it includes a fault diagnosis processing device, wherein, to obtain a target vibration signal when the target gear set rotates, the fault diagnosis processing device adopts the above-mentioned fault diagnosis method to perform fault diagnosis and determine the pitting fault category of the target gear set.

[0224] Specifically, the fault diagnosis and processing device can adopt existing commonly used computer equipment. The way and process of the fault diagnosis and processing device to diagnose the target vibration signal can refer to the above corresponding instructions and will not be repeated here.

Claims

1. A gear pitting fault diagnosis method, characterized in that: The fault diagnosis method comprises: A target vibration signal is obtained when the target gear set rotates, and the obtained target vibration signal is preprocessed to generate a target vibration time-frequency feature information group corresponding to the target vibration signal after the preprocessing, wherein: The target vibration time-frequency feature information group includes a plurality of target vibration time-frequency feature information, and the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is consistent with the number of reference rotation cycles contained in the target vibration signal, so that each target vibration time-frequency feature information is used to characterize the vibration state of a reference rotation cycle of the target gear set; The target vibration time-frequency feature information in the target vibration time-frequency feature information group is loaded into the pitting fault diagnosis model, and the pitting fault diagnosis model is used to perform diagnostic identification processing on each target vibration time-frequency feature information, so as to generate a pitting fault prediction category corresponding to each target vibration time-frequency feature information after the diagnostic identification processing, wherein: For any target vibration time-frequency characteristic information, the diagnostic identification processing performed by the pitting fault diagnosis model includes sequential shared feature extraction processing and category classification processing, among which, The loaded target vibration time-frequency feature information is subjected to shared feature extraction processing to generate corresponding target vibration time-frequency shared features; The extracted target vibration time-frequency shared features are processed by category classification to generate the corresponding pitting fault prediction category; When the number of target vibration time-frequency feature information in the target vibration time-frequency feature information group is not less than 2, all pitting fault prediction categories are subjected to voting processing to determine the corresponding pitting fault voting category after the voting processing, and the pitting fault voting category is used as the pitting fault category of the target gear set.

2. The gear pitting fault diagnosis method according to claim 1 is characterized in that: When preprocessing the target vibration signal, it includes: Normalizing the target vibration signal in the time dimension to generate a target vibration normalized signal after normalization; Based on the reference rotation period of the target gear set, the target vibration normalized signal is segmented to segment the target vibration signal into a plurality of target vibration segmented signals, wherein one target vibration segmented signal corresponds to one reference rotation period included in the target vibration signal; Determine the control reference sampling point number based on the rotation speed state of the target gear set, and perform sampling point number control processing on each target vibration segmentation signal based on the control reference sampling point number, so as to generate a target vibration sampling point number control signal after the sampling point number control processing, wherein the sampling point number included in the target vibration sampling point number control signal is consistent with the control reference sampling point number; Perform short-time Fourier transform on each target vibration sampling point control signal to generate a corresponding vibration time-frequency basic signal after short-time Fourier transform, where: The basic vibration time-frequency signal includes a horizontal vibration time-frequency sub-signal, an axial vibration time-frequency sub-signal, and a vertical vibration time-frequency sub-signal; For each vibration time-frequency basic signal, the contained horizontal vibration time-frequency sub-signal, axial vibration time-frequency sub-signal and vertical vibration time-frequency sub-signal are channel-joined to generate target vibration time-frequency characteristic information after channel-joining; Based on all target vibration time-frequency feature information, a target vibration time-frequency feature information group is generated.

3. The gear pitting fault diagnosis method according to claim 1 is characterized in that: The pitting fault diagnosis model includes a shared feature extraction module and a category classifier adapted to be connected to the shared feature extraction module, wherein: When performing diagnostic identification processing, the shared feature extraction module is used to extract shared features from the target vibration time-frequency feature information, where: When performing shared feature extraction, the target vibration time-frequency feature information is first position-encoded to generate target vibration position encoding information after position encoding; Extracting features from target vibration position encoding information to generate target time-frequency shared features after feature extraction; Based on the target time-frequency shared features, the category classifier generates pitting fault prediction information corresponding to the current target vibration time-frequency feature information, and determines the corresponding pitting fault prediction category based on the pitting fault prediction information, wherein: The pitting fault prediction information includes confidences corresponding to N pitting categories; When determining the corresponding pitting fault prediction category based on the pitting fault prediction information, the pitting fault category with the highest confidence is used as the pitting fault prediction category.

4. The gear pitting fault diagnosis method according to claim 3 is characterized in that: The shared feature extraction module includes a position encoding module and a Transformer encoder for shared feature extraction, wherein: When position encoding is performed on the target vibration time-frequency feature information, a position index is performed based on each time-frequency feature in the target vibration time-frequency feature information, and sine encoding or cosine encoding is performed based on the position of the time-frequency feature, so as to generate a corresponding time-frequency feature encoding after the position encoding, and the corresponding target time-frequency feature encoding information is generated based on the time-frequency feature encoding of all time-frequency features; Based on the target time-frequency feature coding information and the target vibration time-frequency feature information, the target vibration position coding information is generated, wherein when the target vibration position coding information is generated, at least the time-frequency feature coding information and the target vibration time-frequency feature information are added in a position-by-position manner.

5. The gear pitting fault diagnosis method according to claim 4 is characterized in that: The Transformer encoder includes a plurality of encoding units connected in series, wherein: For any encoding unit, it includes a multi-head attention unit, a multi-head attention layer normalization layer, a feedforward layer unit, and a feedforward layer normalization layer connected in sequence, where: The multi-head attention unit includes a multi-head attention layer with residual connections; The feed-forward layer unit includes a feed-forward layer using a residual connection; For two serially connected coding units, along the serial connection direction of the coding units, the feedforward normalization layer of the previous coding unit is connected to the multi-head attention layer of the next coding unit. When extracting features from target vibration position coding information, the target vibration position coding information is extracted by sequentially processing the serially connected coding units, and target time-frequency sharing features representing pitting categories and domain category sharing are output.

6. The gear pitting fault diagnosis method according to any one of claims 3 to 5, characterized in that: The category classifier includes a category classification main unit and a category classification output unit adaptively connected to the category classification main unit, wherein: The category classification output unit adopts the LinSoftmax mapping function of approximate linear mapping. When the LinSoftmax mapping function is used for mapping output, then: Among them, LinSoftmax(z o ) i is the i-th confidence level of the output of the LinSoftmax mapping function, α is the slope, γ is the adjustment parameter, and z o is the logits vector output by the main unit for category classification, Output logits vector z for the category classification master unit o The i-th category classification quantity in , Output logits vector z for the category classification master unit o The j-th category classification quantity in , is the penalty weight of the i-th pitting category, is the penalty weight for the jth pitting category.

7. The gear pitting fault diagnosis method according to any one of claims 3 to 5, characterized in that: When voting, this includes: For each pitting fault prediction category corresponding to the target vibration time-frequency characteristic information, configure the voting samples as the corresponding pitting category; Count the number of voting samples for each pitting category, and use the pitting category with the largest number of voting samples as the pitting fault voting category of the target gear set; When the number of voting samples for each pitting corrosion category is consistent, the pitting corrosion fault prediction category with the highest confidence is taken as the pitting corrosion fault voting category.

8. The gear pitting fault diagnosis method according to any one of claims 3 to 5, characterized in that: When building a pitting fault diagnosis model, include: Constructing a pitting fault diagnosis basic model, wherein the pitting fault diagnosis basic model includes a shared feature extraction basic module, a category basic classifier and a domain classifier, and both the category basic classifier and the domain classifier are connected to the shared feature extraction basic module; A pitting fault diagnosis training data set is constructed, wherein the pitting fault diagnosis training data set includes a number of training samples, wherein: For any training sample, including the training vibration time-frequency feature information corresponding to the training gear set, the pitting category label of the current training gear set, and the label of the field speed, The pitting fault diagnosis training data set contains at least N types of pitting category labels; When the pitting fault diagnosis training data set is used to train the pitting fault diagnosis basic model, the model loss function used is: L=ω cls ·L cls +oh dom ·L dom +oh cor ·L cor Among them, L is the model loss function, L cls is the category loss calculated based on the classification base classifier, L dom is the domain loss calculated based on the domain classifier, L cor is the association loss calculated based on the classification base classifier and the domain classifier, ω cls is the class loss weight, ω dom is the domain loss weight, ω cor is the associated loss weight; When the training of the pitting fault diagnosis basic model reaches the target training state, the shared feature extraction basic module is configured as a shared feature extraction module, the category basic classifier is configured as a category classifier, and a pitting fault diagnosis model is formed based on the shared feature extraction module and the category classifier.

9. A gear pitting fault diagnosis system, characterized in that: It comprises a fault diagnosis processing device, wherein, to obtain a target vibration signal when the target gear set rotates, the fault diagnosis processing device adopts the fault diagnosis method described in any one of claims 1 to 8 to perform fault diagnosis and determine the pitting fault category of the target gear set.

Images