Variable speed planetary gearbox fault diagnosis method, system and device and medium

Abstract

The invention discloses a variable-speed planetary gearbox fault diagnosis method, and relates to the technical field of mechanical fault diagnosis, and the method comprises the steps: processing a vibration signal of a variable-speed planetary gearbox through optimal generalized S transformation to obtain an optimal time-frequency diagram, and representing time-frequency characteristics; extracting enhanced features by combining a time-frequency enhanced attention mechanism with a plurality of receptive fields, aggregating the enhanced features through a space attention mechanism to form joint weighted features, and inputting the joint weighted features into a deep residual network to obtain final feature representation; and finally outputting a fault prediction value by constructing a domain adaptive network and adopting cross entropy loss and multi-kernel maximum mean difference domain adaptive loss combined training. The space attention weight is adjusted through a time-frequency attention mechanism, fault key feature representation is strengthened, and redundant information interference is reduced; deep features are extracted in combination with a deep residual network, and multi-core maximum mean difference is adopted to align source domain and target domain feature distribution, so that distribution difference is reduced, and domain adaptive diagnosis of the variable-speed planetary gearbox is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mechanical transmission component maintenance, and particularly relates to a variable-speed planetary gearbox fault diagnosis method, system, device and medium. BACKGROUND

[0002] As a core key component in a mechanical transmission system, a planetary gearbox is widely used in large key equipment such as a drilling top drive system and a fracturing pump in the oil and gas industry due to its high transmission efficiency, strong bearing capacity and compact structure, and its running state is directly related to the safe and stable operation and production efficiency of the entire equipment system. However, the internal structure of the planetary gearbox is complex, and contains multiple gear meshing, planetary carrier revolution and rotation and other compound motion forms, which makes it prone to gear wear, tooth surface crack, bearing failure and other faults. If the fault cannot be diagnosed in time and accurately, it may not only cause equipment downtime and production stoppage, resulting in significant economic losses, but also induce safety accidents and produce disastrous consequences.

[0003] In the process of planetary gearbox fault diagnosis, vibration signal analysis is the most commonly used technical means, but in actual application, technical problems are faced: since the planetary gearbox usually works in harsh conditions such as oil and gas drilling and fracturing, the collected vibration signals not only contain a large amount of environmental noise and interference signals, but also have extremely complex signal components, covering gear rotation frequency, meshing frequency and side frequency band generated by modulation effect and other frequency components, so that the effective features related to the fault in time-frequency analysis are covered by a large amount of redundant information, resulting in great difficulty in fault feature extraction. At the same time, the planetary gearbox is often in variable speed working condition in actual operation, and the data characteristics distribution collected at different speeds has significant difference, that is, there is a cross-domain distribution deviation problem between known working condition data and unknown speed working condition data. Traditional fault diagnosis methods are generally based on the assumption that data obeys independent and identically distributed, and when facing cross-domain data under variable speed, the generalization ability of the diagnosis model is greatly reduced, and it is difficult to realize accurate and reliable fault diagnosis. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the defect that the distribution difference between the source domain and the target domain in the existing planetary gearbox fault diagnosis method is large, resulting in poor fault diagnosis reliability, so as to provide a variable-speed planetary gearbox fault diagnosis method, system, device and medium.

[0005] A variable-speed planetary gearbox fault diagnosis method, characterized in that it comprises the following steps: obtaining a vibration signal of a variable-speed planetary gearbox; processing the vibration signal by optimal generalized S transform to obtain an optimal time-frequency diagram of the vibration signal; based on a time-frequency enhanced attention mechanism, extracting enhanced features of the optimal time-frequency diagram through multiple receptive fields; The enhanced features obtained by different receptive fields in the spatial dimension are aggregated by a spatial attention mechanism to form a joint weighted feature; The joint weighted feature is input into a deep residual network to obtain a final feature representation; A domain adaptive network is constructed, and the final feature representations of the source domain and the target domain are input into the domain adaptive network; The domain adaptive network is trained by a loss function, and the loss function includes a cross-entropy loss and a domain adaptive loss; The domain adaptive loss is calculated based on the final feature representations of the source domain and the target domain by a multi-kernel maximum mean difference method; The final feature representation is input into the domain adaptive network to output a fault prediction value of the corresponding planetary gear box.

[0006] Further, the method for obtaining the optimal time-frequency graph comprises: processing the vibration signal by a generalized S transform to obtain a time-frequency graph; calculating a time-frequency aggregation measure of the time-frequency graph to obtain an adjustment factor of the generalized S transform with the minimum time-frequency aggregation measure as an optimal adjustment factor; and processing the vibration signal by the generalized S transform based on the optimal adjustment factor to obtain the optimal time-frequency graph of the vibration signal.

[0007] Further, the method for forming the joint weighted feature comprises: A serial convolution sequence is constructed by connecting a convolution kernel and a dilated convolution kernel in sequence to extract a fault feature of the optimal time-frequency graph; the fault feature includes an initial feature and a dilated convolution feature, the initial feature is formed by the convolution kernel, and the dilated convolution feature is formed by the dilated convolution kernel; The initial feature and the dilated convolution feature are reduced in channel, the initial feature and the dilated convolution feature after channel reduction are aggregated by average pooling and maximum pooling and are spliced in channel, spatial attention weights are obtained by a convolution layer and an activation function, the initial feature and the dilated convolution feature are weighted by the spatial attention weights, and a preliminary attention feature is formed by convolution. The optimal time-frequency graph is processed by a convolution layer and an activation function to obtain time-frequency attention weights, the optimal time-frequency graph is multiplied element by element by the time-frequency attention weights to obtain time-frequency dimension weighting, and the joint weighted feature is obtained by multiplying the preliminary attention feature element by element.

[0008] Further, the deep residual network is a ResNet18 network; the ResNet18 network includes an initial convolution layer, a maximum pooling layer, a residual block, an average pooling layer, and a fully connected layer connected; the ResNet18 network includes four groups of the residual block, each group of the residual block includes two basic blocks for realizing deep convolution.

[0009] Further, the method further comprises the following steps: constructing a multi-kernel feature kernel by a plurality of base kernel weighted combination; calculating distribution distance of final feature representation of source domain and target domain in reproducing kernel Hilbert space based on the multi-kernel feature kernel by multi-kernel maximum mean discrepancy method, forming domain adaptive loss; constructing domain adaptive network, inputting the final feature representation of source domain and target domain into the domain adaptive network; optimizing the domain adaptive network by loss function, the loss function including cross-entropy loss and domain adaptive loss; inputting the final feature representation into the domain adaptive network, outputting fault prediction value of variable speed planetary gearbox.

[0010] Further, the domain adaptive loss is a square form of reconstruction Hilbert space distance of embedding mean value in Hilbert space, the Hilbert space including the multi-kernel feature kernel; The domain adaptive loss is represented as: ; Wherein, represents the Hilbert space with multi-kernel feature kernel , and respectively represent the probability distribution of source domain and target domain, represents feature mapping, and represent features respectively sampled from source domain and target domain.

[0011] Further, the loss function includes cross-entropy loss and domain adaptive loss, represented as: ; Wherein, represents cross-entropy loss, represents domain adaptive loss, represents trade-off coefficient;The cross-entropy loss is represented as: ; Wherein, represents batch size of source domain, represents predicted probability of source domain sample by classifier, represents real label of source domain sample.

[0012] A variable speed planetary gearbox fault diagnosis system is used to realize variable speed planetary gearbox fault diagnosis according to the variable speed planetary gearbox fault diagnosis method described above, comprising: The data preprocessing module is used for processing the vibration signal through an optimal generalized S transform to obtain an optimal time-frequency graph of the vibration signal. The time-frequency enhanced attention network is used for extracting a joint weighted feature of the optimal time-frequency graph based on a time-frequency enhanced attention mechanism and a spatial attention mechanism. The deep residual network is used for processing the joint weighted feature to form a final feature representation. The domain adaptive network is used for realizing network training through a loss function comprising a cross-entropy loss and a domain adaptive loss, inputting the final feature representation, and outputting a fault prediction value of the corresponding variable planetary gearbox.

[0013] A computer device comprises a processor and a memory connected with the processor, and the memory stores one or more programs to be executed by the processor to implement the steps in the variable planetary gearbox fault diagnosis method.

[0014] A computer readable storage medium stores one or more programs to be executed by a processor to implement the steps in the variable planetary gearbox fault diagnosis method.

[0015] Beneficial effects: The variable planetary gearbox fault diagnosis method provided by the application solves the problems of redundant information interference, low diagnosis precision and poor generalization ability caused by variable speed cross-domain distribution difference in the prior art, and achieves the following beneficial effects through the design of the enhanced attention mechanism and the domain adaptive diagnosis method: The application adopts the optimal generalized S transform to process the vibration signal, realizes dynamic adaptation of time-frequency resolution through parameter optimization, can accurately capture key components such as gear rotation frequency, meshing frequency and modulation side frequency band in the vibration signal, effectively suppresses environmental noise and interference signals under harsh working conditions, and generates an optimal time-frequency graph that completely retains time-domain dynamic changes and frequency-domain distribution characteristics, solves the problems of fuzzy representation and effective information loss in traditional time-frequency transformation, and provides high signal-to-noise ratio and high recognition degree data support for subsequent feature extraction.

[0016] The application innovatively designs a collaborative architecture of the time-frequency enhanced attention mechanism and the spatial attention aggregation, comprehensively covers fault features of different scales and different frequency ranges through multiple receptive fields, then adaptively allocates weights through the spatial attention mechanism, strengthens the expression of key fault features such as gear wear and tooth surface crack, accurately suppresses redundant information and noise interference, completely solves the technical pain point that fault features are covered by redundant information, and significantly improves the recognition degree and extraction efficiency of the fault features.

[0017] The application combines a deep residual network to extract a final feature representation, effectively avoids the performance degradation problem of the deep network by using a residual block structure, can mine the hidden fault essence features in the time-frequency graph layer by layer, and compared with the traditional convolution network, the feature expression is more hierarchical and robust, and provides more reliable feature support for subsequent fault classification and cross-domain adaptation.

[0018] The application constructs a domain adaptive network based on multi-core maximum mean difference, realizes accurate alignment of source domain and target domain feature distribution through joint training of cross-entropy loss and domain adaptive loss, effectively overcomes the cross-domain distribution deviation problem under variable speed conditions, greatly improves the generalization ability of the diagnostic model under different speed conditions, and realizes the domain adaptive diagnosis of the variable speed planetary gear box. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Fig. 1 The main method flowchart of the present application is shown in the figure. Fig. 2 The system structure schematic diagram of the present application is shown in the figure. Fig. 3 The network structure schematic diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below. In the following description, a lot of specific details are set forth in order to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the connotation of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0022] In the description of the present application, it should be understood that the terms "first", "second" are only used for description purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first", "second" can be explicitly or implicitly included at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0023] In this application, unless otherwise clearly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements or the interaction relationship of two elements, unless otherwise clearly limited. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0024] Embodiment one: Referring to Figs. 1-3 The embodiment provides a fault diagnosis method for a variable planetary gearbox, characterized by comprising the following steps: Step S1: obtaining a vibration signal of a variable planetary gearbox; processing the vibration signal by optimal generalized S transform (OGST) to obtain an optimal time-frequency graph of the vibration signal; Step S2: based on a time-frequency enhanced attention mechanism, extracting enhanced features of the optimal time-frequency graph through multiple receptive fields; Through a spatial attention mechanism, the enhanced features obtained by different receptive fields are aggregated in the spatial dimension to form a joint weighted feature; Step S3: inputting the joint weighted feature into a deep residual network to obtain a final feature representation; Step S4: constructing a domain adaptive network, and inputting the final feature representation of the source domain and the target domain into the domain adaptive network; The domain adaptive network is trained by a loss function, and the loss function includes a cross-entropy loss and a domain adaptive loss; The domain adaptive loss is based on the final feature representation of the source domain and the target domain, and is calculated by a multi-kernel maximum mean difference method; The final feature representation is input into the domain adaptive network, and a fault prediction value corresponding to the variable planetary gearbox is output.

[0025] Specifically, step S1 comprises: Processing the vibration signal by generalized S transform (GST) to obtain a time-frequency graph; In this embodiment, the generalized S transform is represented as: ; Wherein, is a time series signal, i.e. the vibration signal of the variable planetary gearbox collected and obtained; and are time shift factor and frequency respectively; is an adjustment factor, and when is the standard S transform.

[0026] To optimize the performance of the time-frequency representation, the optimal value is needed to realize the OGST; in this embodiment, a time-frequency concentration measure of the time-frequency graph is calculated, and the adjustment factor of the generalized S transform at the time when the time-frequency concentration measure is the smallest is taken as the optimal adjustment factor; according to the concentration degree of the time-frequency distribution, the time-frequency concentration measure is taken as the optimization criterion, which is expressed as: ; wherein, denotes the normalized energy of the GST at the point, is a constant greater than 1, the smaller the value is, the better the time-frequency concentration is.

[0027] Based on the optimal adjustment factor, the vibration signal is processed by the generalized S transform to obtain the optimal time-frequency graph of the vibration signal.

[0028] In step S2, after obtaining the optimal time-frequency graph of the vibration signal, a time-frequency enhanced attention mechanism (TFEAM) is designed to realize feature extraction for enhancing the representation of the fault feature, including the following steps: A serial convolution sequence is constructed by connecting the convolution kernel and the dilated convolution kernel in sequence to extract the fault feature of the optimal time-frequency graph; the fault feature includes an initial feature and a dilated convolution feature, the initial feature is formed by the convolution kernel, and the dilated convolution feature is formed by the dilated convolution kernel; in this embodiment, the serial convolution sequence includes a 5x5 convolution kernel and a 7x7 dilated convolution kernel with a dilated rate of 3, which realizes step-by-step extraction of the fault feature, and the dilated convolution can effectively expand the receptive field of the network without significantly increasing the parameter amount, so as to more accurately capture the large-range feature information in the time-frequency graph.

[0029] The serial convolution sequence is expressed as: ; wherein, denotes convolution; denotes dilated convolution; denotes batch normalization; denotes activation function; denotes the extracted initial feature; denotes the extracted dilated convolution feature.

[0030] After extracting the initial feature and the dilated convolution feature, a spatial attention mechanism is used to aggregate the features obtained by different receptive fields in the spatial dimension. The initial feature and the dilated convolution feature are reduced in channel dimension, the reduced initial feature and the reduced dilated convolution feature are aggregated by average pooling (AvgPool) and maximum pooling (MaxPool) and are spliced in channel, spatial attention weights are obtained by a convolution layer and an activation function, the initial feature and the dilated convolution feature are weighted by the spatial attention weights, and a preliminary attention feature is formed by convolution. The time-frequency attention weights are obtained by a convolution layer and an activation function on the optimal time-frequency graph, are multiplied element by element with the optimal time-frequency graph to obtain time-frequency dimension weighting, and are multiplied element by element with the preliminary attention feature to obtain a joint weighted feature.

[0031] The process is represented as: ; wherein, and represent the initial feature and the dilated convolution feature after channel dimension reduction respectively; Concat represents channel splicing; represents a Sigmoid function, i.e. an activation function; represents a spatial attention weight; represents a preliminary attention feature; represents the result output by TFEAM, i.e. a joint weighted feature.

[0032] Specifically, in step S3, the deep residual network is a ResNet18 network; the ResNet18 network includes a connected initial convolution layer, a maximum pooling layer, a residual block, an average pooling layer and a full connection layer; the ResNet18 network includes four groups of the residual block, each group of the residual block includes two basic blocks for realizing deep convolution.

[0033] In this embodiment, the input of the ResNet18 network is the joint weighted feature; the initial convolutional layer is Conv-BN-ReLU, including a 7*7 convolutional layer, which is accelerated by batch normalization (BN) to stabilize the feature distribution, and introduces non-linear activation by ReLU to filter negative response features. The max pooling layer is represented as MaxPool, including a 3*3 pooling kernel, which is used for down-sampling the feature map, retaining key features while reducing computational complexity. Four residual blocks are connected in turn, and each residual block is represented as two basic blocks (Basic Block), each of which includes a 3*3 convolution kernel (Conv), which makes the features pass through the main branch of convolution transformation and the direct transmission residual branch and then add them, avoiding deep gradient vanishing. The average pooling layer (AvgPool) is used to realize the global average pooling of the features. The fully connected layer (FC) is used for mapping of the feature vector to generate the final feature representation, thereby realizing extraction of the key fault features.

[0034] Specifically, in step S4, the following steps are included: In the final feature representation, the source domain feature and the target domain feature are input into the domain adaptation network, where and represent the batch size of the source domain and the target domain, respectively, represents the feature dimension.

[0035] In order to improve the adaptability to complex distribution, a multi-kernel feature kernel is constructed by multiple basic kernel weighting combination, represented as: ; wherein is a constraint enhancement weight, which can ensure the generation of a typical multi-feature kernel, is the number of feature kernels.

[0036] Based on the multi-kernel feature kernel, the distribution distance of the final feature representations of the source domain and the target domain in the reproducing kernel Hilbert space is calculated by a multi-kernel maximum mean difference method (MK-MMD), forming a domain adaptation loss; The domain adaptation network is constructed, and the final feature representations of the source domain and the target domain are input into the domain adaptation network; the domain adaptation network is optimized by a loss function, and the loss function includes a cross-entropy loss and a domain adaptation loss; The loss function includes a cross-entropy loss and a domain adaptation loss, and the total loss is optimized, which can improve the classification accuracy of the source domain and the domain adaptability of the target domain, represented as: ; wherein represents the cross-entropy loss, represents a domain adaptation loss, represents a trade-off coefficient for balancing the contributions of cross-entropy loss and distribution alignment; cross-entropy loss is calculated by supervised classification of source domain samples, and is represented as: ; wherein, represents the batch size of the source domain, represents the predicted probability of the classifier for the source domain sample , represents the true label of the source domain sample.

[0037] The domain adaptation loss is in the form of the square of the reconstruction Hilbert space distance of the embedding mean value in the Hilbert space, and the Hilbert space includes the multi-kernel feature kernel; The domain adaptation loss is represented as: ; wherein, represents a Hilbert space with a multi-kernel feature kernel , and respectively represent the probability distribution of the source domain and the target domain, represents a feature mapping, and represent the features sampled from the source domain and the target domain, respectively.

[0038] By using the multi-kernel maximum mean difference method (MK-MMD), the difference in data distribution under variable speed can be solved, and domain adaptation from the source domain to the target domain can be realized.

[0039] The domain adaptation network maps the extracted features using a fully connected layer, and combines a Softmax function to realize prediction of the fault class.

[0040] The final feature representation is input into the domain adaptation network, and a fault prediction value of the variable planetary gearbox is obtained.

[0041] The embodiment also provides a variable planetary gearbox fault diagnosis system, which is used to realize variable planetary gearbox fault diagnosis according to the variable planetary gearbox fault diagnosis method. The data preprocessing module is configured to process the vibration signal by using the optimal generalized S transform to obtain an optimal time-frequency graph of the vibration signal. The time-frequency enhanced attention network is configured to extract a joint weighted feature of the optimal time-frequency graph based on a time-frequency enhanced attention mechanism and a spatial attention mechanism. The deep residual network is configured to process the joint weighted feature to form a final feature representation. Domain adaptive network for network training by a loss function comprising a cross-entropy loss and a domain adaptive loss, inputting a final feature representation input and outputting an obtained failure prediction value for a corresponding planetary gearbox.

[0042] The embodiment also provides a computer device, comprising a processor and a memory connected to the processor, the memory storing one or more programs to be executed by the processor to implement the steps in the planetary gearbox fault diagnosis method as described above.

[0043] The embodiment also provides a computer-readable storage medium storing one or more programs to be executed by a processor to implement the steps in the planetary gearbox fault diagnosis method as described above.

[0044] Embodiment two: The embodiment provides an example of planetary gearbox fault diagnosis by a planetary gearbox fault diagnosis method.

[0045] The vibration signals monitored by the acceleration sensor are collected by the planetary gearbox fault simulation test bench, and the sampling frequency is 12 kHz. To comprehensively evaluate the diagnosis performance, multiple typical faults of the planetary wheel and the sun wheel bearing are simulated, including planetary wheel peeling, cage cracking, single-sided wear, and five common fault types of the sun wheel bearing rolling body pitting and inner ring cracking. Data under normal operating conditions are also collected. Three data sets are constructed at rotational speeds of 1200 rpm, 1500 rpm and 2400 rpm, as shown in Table 1. A total of six diagnostic tests are designed in this paper, for example, in D1→D2, the source domain rotational speed is 1200 rpm and the target domain rotational speed is 1500 rpm.

[0046] Table 1: Test data sets

[0047] For the vibration signals collected in the test, the sampling length is 1024 and the sliding window moving step is 500, and the signal is converted to a time-frequency graph by OGST. In this embodiment, 200 pictures are generated for each type of sample, and are divided into a training set, a validation set and a test set according to a ratio of 6:2:2.

[0048] In this embodiment, the hyperparameters involved in the method model include the number of iterations, the learning rate and the batch size, which are directly related to the accuracy and stability of the diagnosis result. In the training process, the number of iterations is set to 50 times, and the model that performs best in the validation set is applied to the test set to evaluate the effect. The learning rate and batch size The parameters need to be determined in advance before training, so it is necessary to conduct sensitivity analysis. The grid search method is used to determine the optimal parameter combination, and the search space is defined as: , The training environment of the embodiment is based on python3.9 and pytorch2.5. Considering the diagnostic accuracy and the computational resource occupancy of the model of the embodiment, 0.0001 and 64 are preferably set as and respectively. and .

[0049] To analyze the influence of each module on the performance of the method of the application, three network structures are designed: ① baseline network ResNet18 without using MK-MMD and TFEAM; ② RT18 network without using MK-MMD; and ③ DAR18 network without using TFEAM.

[0050] The experimental results are shown in Table 2. The results show that the average accuracy of DART18 is 97.85%, which is significantly better than the network without using MK-MMD or TFEAM. Specifically, if TFEAM and MK-MMD are not used, the average accuracy of the Resnet18 network decreases to 86.39%, and the performance of the model is significantly affected; if MK-MMD is not used, the average accuracy of the RT18 network decreases to 91.81%, which verifies the key role of MK-MMD in aligning the feature distribution of the source domain and the target domain; if TFEAM is not used, the average accuracy of the DAR18 network decreases to 95.07%, and the decrease is most obvious in D3→D2, which shows the important role of TFEAM in enhancing the representation of fault key features. The experimental results fully verify the important contribution of the MK-MMD and TFEAM modules in the diagnosis process.

[0051] Table 2: Ablation test results

[0052] To verify the performance and advantages of DART18, six commonly used domain adaptation methods are selected for comparison test. To ensure fairness of comparison, all methods use the same feature extraction and classification structure. The results of each group of diagnosis tests are shown in Table 3. The results show that compared with DANN, Coral, DSAN, BNM, JAN and MCD methods, the average accuracy of DART18 is increased by 11.61, 7.57, 7.78, 7.16, 8.43 and 11.12 percentage points respectively, and it shows the best performance in each test, fully embodying the robustness and superiority in fault diagnosis.

[0053] It can be seen that the improvement of the diagnosis ability of the method is due to the TFEAM, which can focus on the key features of the fault and reduce the overlap of feature distribution; further, the consistency of the feature distribution of the source domain and the target domain is realized through the MK-MMD, and then the fault diagnosis performance is improved.

[0054] Table 3: Comparative test results

[0055] The technical features of the above-described embodiments can be combined arbitrarily, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the description.

[0056] The above-described embodiments only express several implementation manners of the application, the description is more specific and detailed, but it should not be understood as a limitation on the patent application scope. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made, which belong to the protection scope of the application. Therefore, the patent protection scope of the application should be subject to the appended claims.

Claims

1. A fault diagnosis method for a variable speed planetary gearbox, characterized in that, Includes the following steps: The vibration signal of the variable speed planetary gearbox is acquired; the vibration signal is processed by the optimal generalized S-transform to obtain the optimal time-frequency diagram of the vibration signal; Based on the time-frequency enhanced attention mechanism, the enhanced features of the optimal time-frequency map are extracted through multiple receptive fields; The enhanced features acquired from different receptive fields are aggregated in the spatial dimension through a spatial attention mechanism to form a joint weighted feature; The joint weighted features are input into a deep residual network to obtain the final feature representation; Construct a domain adaptive network by inputting the final feature representations of the source domain and the target domain into the domain adaptive network; The domain adaptive network is trained using a loss function, which includes cross-entropy loss and domain adaptive loss. The domain adaptive loss is calculated based on the final feature representations of the source and target domains using the multi-kernel maximum mean difference method. The final feature representation is input into the domain adaptive network, and the output is the fault prediction value of the corresponding speed planetary gearbox.

2. The method for fault diagnosis of a variable speed planetary gearbox according to claim 1, characterized in that, The method for obtaining the optimal time-frequency diagram includes: processing the vibration signal through a generalized S-transform to obtain a time-frequency diagram; calculating the time-frequency clustering metric of the time-frequency diagram, and obtaining the adjustment factor of the generalized S-transform when the time-frequency clustering metric is minimized as the optimal adjustment factor; and processing the vibration signal through a generalized S-transform based on the optimal adjustment factor to obtain the optimal time-frequency diagram of the vibration signal.

3. The method for diagnosing faults in a variable speed planetary gearbox according to claim 1, characterized in that, The method for forming the joint weighted features includes: A serial convolution sequence is constructed by sequentially connecting convolution kernels and dilated convolution kernels to extract fault features from the optimal time-frequency map; the fault features include initial features and dilated convolution features, wherein the initial features are formed by the convolution kernels and the dilated convolution features are formed by the dilated convolution kernels; Channel dimensionality reduction is performed on the initial features and the dilated convolutional features. The dimensionality-reduced initial features and the dilated convolutional features are aggregated by average pooling and max pooling and concatenated by channels. Spatial attention weights are obtained through convolutional layers and activation functions. The initial features and the dilated convolutional features are weighted by spatial attention weights and formed by convolution. The optimal time-frequency map is processed through convolutional layers and activation functions to obtain time-frequency attention weights. These weights are then multiplied element-wise with the optimal time-frequency map to obtain time-frequency dimension weights, and finally multiplied element-wise with the initial attention features to obtain joint weighted features.

4. The fault diagnosis method for a variable speed planetary gearbox according to claim 1, characterized in that, The deep residual network is a ResNet18 network; the ResNet18 network includes a connected initial convolutional layer, a max pooling layer, residual blocks, an average pooling layer, and a fully connected layer; the ResNet18 network includes four groups of residual blocks, each group of residual blocks including two basic blocks, used to implement deep convolution.

5. The method for fault diagnosis of a variable speed planetary gearbox according to claim 1, characterized in that, It also includes the following steps: A multi-kernel feature kernel is constructed by weighted combination of multiple basic kernels; Based on the multi-kernel feature kernel, the distribution distance of the final feature representations of the source domain and the target domain in the regenerated kernel Hilbert space is calculated using the multi-kernel maximum mean difference method, forming a domain adaptive loss; Construct a domain adaptive network by inputting the final feature representations of the source and target domains into the domain adaptive network; optimize the domain adaptive network using a loss function, which includes cross-entropy loss and domain adaptive loss; The final feature representation is input into the domain adaptive network, and the output is the fault prediction value of the variable speed planetary gearbox.

6. The method for fault diagnosis of a variable speed planetary gearbox according to claim 5, characterized in that, The domain adaptive loss is the squared form of the reconstructed Hilbert space distance from the embedding mean in the Hilbert space, where the Hilbert space includes the multi-kernel feature kernel; The domain adaptive loss is expressed as: ； in, Indicates a kernel with multi-core characteristics Hilbert space, and Let represent the probability distributions of the source domain and the target domain, respectively. Represents feature mapping, and This represents features sampled from the source domain and the target domain, respectively.

7. The method for fault diagnosis of a variable speed planetary gearbox according to claim 5, characterized in that, The loss function includes cross-entropy loss and domain adaptive loss, expressed as: ； in, Represents cross-entropy loss, Representation domain adaptive loss, Indicates the trade-off coefficient; The cross-entropy loss is expressed as: ； in, Indicates the batch size of the source domain. This indicates that the classifier evaluates the source domain samples. The predicted probability, This represents the true label of the source domain sample.

8. A fault diagnosis system for a variable speed planetary gearbox, characterized in that, The method for diagnosing faults in a variable speed planetary gearbox according to any one of claims 1-7 includes: The data preprocessing module is used to process the vibration signal through the optimal generalized S-transform to obtain the optimal time-frequency diagram of the vibration signal; A time-frequency enhanced attention network is used to extract the joint weighted features of the optimal time-frequency map based on the time-frequency enhanced attention mechanism and the spatial attention mechanism; A deep residual network is used to process the joint weighted features to form the final feature representation; A domain-adaptive network is used to train the network using a loss function that includes cross-entropy loss and domain-adaptive loss. The input is the final feature representation, and the output is the fault prediction value of the corresponding speed planetary gearbox.

9. A computer device, characterized in that, The computer device includes a processor and a memory connected to the processor. The memory stores one or more programs, which are executed by the processor to perform the steps in the fault diagnosis method for a variable speed planetary gearbox as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which are executed by a processor to implement the steps in the fault diagnosis method for a variable speed planetary gearbox as described in any one of claims 1-7.

Images