Gearbox fault detection method, device, equipment and storage medium

Abstract

The embodiment of the application provides a kind of gearbox fault detection method, device, equipment and storage medium, the method comprises: obtaining the target runtime sequence data of gearbox, the target runtime sequence data is divided into multiple target data segments;Respectively each target data segment is input into multi-scale feature extraction model and is mapped, obtains the multi-scale initial feature corresponding to each target data segment, respectively each multi-scale initial feature is carried out multi-scale weighted fusion, obtains the target operating characteristic corresponding to each target data segment;Each target operating characteristic is input into time sequence model and is mapped, obtains the target time domain feature corresponding to each target data segment;After each target time domain feature is processed and input into classification model, classification processing is carried out, obtains the target fault detection result of the gearbox.The embodiment of the application can improve the accuracy of target fault detection result.

Description

Technical Field

[0001] This application relates to, but is not limited to, the field of fault detection technology, and in particular to a method, apparatus, equipment, and storage medium for detecting gearbox faults. Background Technology

[0002] Gearboxes are one of the core components of rotating machinery. Gearboxes often operate under harsh conditions of high speed and high load. Under the long-term influence of alternating strain forces, key components such as gears and bearings in gearboxes are prone to structural damage such as wear, fracture, pitting, and cracks, posing significant safety hazards. Therefore, it is necessary to inspect gearboxes for faults.

[0003] Currently, traditional gearbox fault detection methods mainly rely on convolutional neural networks (CNNs) to extract spatial features from gearbox operating data for classification and prediction to determine the fault category. However, the spatial features that CNNs focus on cannot effectively capture the complex information in gearbox operating data, resulting in low accuracy of fault detection results. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a gearbox fault detection method, apparatus, device, and storage medium, which can improve the accuracy of target fault detection results.

[0006] To achieve the above objectives, a first aspect of this application proposes a gearbox fault detection method, characterized by comprising: acquiring target runtime sequence data of a gearbox; dividing the target runtime sequence data into multiple target data segments; inputting each target data segment into a multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment; performing multi-scale weighted fusion on each multi-scale initial feature to obtain target runtime features corresponding to each target data segment; inputting each target runtime feature into a time-series model for mapping to obtain target time-domain features corresponding to each target data segment; and performing attention processing on each target time-domain feature before inputting it into a classification model for classification processing to obtain the target fault detection result of the gearbox.

[0007] In some embodiments, the multi-scale feature extraction model includes a first feature extraction network and a second feature extraction network. The first feature extraction network includes N sequentially connected first convolutional layers, and the second feature extraction network includes M sequentially connected second convolutional layers, where N is less than M, and both N and M are positive integers. The multi-scale initial features include first-scale initial features and second-scale initial features. The step of inputting each target data segment into the multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment, and performing multi-scale weighted fusion on each of the multi-scale initial features to obtain target running features corresponding to each target data segment, includes: for each target data segment, inputting the target data segment into the first feature extraction network for mapping to obtain the first-scale initial feature corresponding to the target data segment; inputting the target data segment into the second feature extraction network for mapping to obtain the second-scale initial feature corresponding to the target data segment; and performing multi-scale weighted fusion on the first-scale initial feature and the second-scale initial feature corresponding to each target data segment to obtain target running features corresponding to the target data segment.

[0008] In some embodiments, the first feature extraction network includes a first convolutional layer, a second convolutional layer, and a first pooling layer connected in sequence; the second feature extraction network includes a third convolutional layer, a fourth convolutional layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, and a third pooling layer connected in sequence; the kernel size of the first convolutional layer is greater than or equal to the kernel size of the second convolutional layer; the kernel size of the first convolutional layer is greater than the kernel size of the third convolutional layer; the kernel size of the third convolutional layer is greater than or equal to the kernel size of the fourth convolutional layer; the kernel size of the third convolutional layer is greater than the kernel size of the fifth convolutional layer; and the kernel size of the fifth convolutional layer is greater than or equal to the kernel size of the sixth convolutional layer. The target data segment is input into the first feature extraction network for mapping to obtain the first-scale initial features corresponding to the target data segment. The target data segment is input into the second feature extraction network for mapping to obtain the second-scale initial feature corresponding to the target data segment. This includes: calling the first convolutional layer and the second convolutional layer to perform convolution processing on the target data segment to obtain a first intermediate feature; calling the first pooling layer to perform a pooling operation on the first intermediate feature to obtain the first-scale initial feature corresponding to the target data segment; calling the third convolutional layer and the fourth convolutional layer to perform convolution processing on the target data segment to obtain a second intermediate feature; calling the second pooling layer to perform a pooling operation on the second intermediate feature to obtain a third intermediate feature; calling the fifth convolutional layer and the sixth convolutional layer to perform convolution processing on the third intermediate feature to obtain a fourth intermediate feature; and calling the third pooling layer to perform a pooling operation on the fourth intermediate feature to obtain the second-scale initial feature corresponding to the target data segment.

[0009] In some embodiments, the step of inputting each of the target running features into a time series model for mapping to obtain the target time-domain features corresponding to each of the target data segments includes: inputting each of the target running features into a time series model, determining the forward hidden layer state and the backward hidden layer state corresponding to each of the target running features based on the time series model; and for each target running feature, performing a weighted summation based on the forward hidden layer state and the backward hidden layer state corresponding to the target running feature to determine the target time-domain features corresponding to the target data segment.

[0010] In some embodiments, the step of performing attention processing on each of the target temporal features and then inputting them into a classification model for classification processing to obtain the target fault detection result of the gearbox includes: performing attention processing on each of the target temporal features to obtain attention features; obtaining the working type identifier of the gearbox; concatenating the working type identifier and the attention features and then inputting them into a classification model for classification processing to obtain the target fault detection result of the gearbox.

[0011] In some embodiments, the step of concatenating the work type identifier and the attention feature and inputting them into a classification model for classification processing to obtain the target fault detection result of the gearbox includes: obtaining the historical fault probability of the gearbox; inputting the attention feature and the historical fault probability into a probability prediction model for regression processing to obtain the current fault probability of the gearbox; and concatenating the work type identifier, the current fault probability, and the attention feature and inputting them into a classification model for classification processing to obtain the target fault detection result of the gearbox.

[0012] In some embodiments, before performing multi-scale feature extraction on each of the target data segments, the gearbox fault detection method further includes: acquiring sample runtime sequence data of the gearbox and sample detection labels corresponding to the sample runtime sequence data; dividing the sample runtime sequence data into multiple sample data segments, wherein the sample detection labels are used to indicate whether the sample runtime sequence data belongs to a fault category or a normal type; inputting each of the sample data segments into a multi-scale feature extraction model for mapping to obtain multi-scale sample features corresponding to each of the sample data segments; performing multi-scale weighted fusion on each of the multi-scale sample features to obtain sample runtime features corresponding to each of the target data segments; inputting each of the sample runtime features into a time-series model for mapping to obtain sample time-domain features corresponding to each of the sample data segments; performing attention processing on each of the sample time-domain features and then inputting them into a classification model for classification processing to obtain the sample fault detection result of the gearbox; determining the model loss based on the difference between the sample fault detection result and the sample detection label; and jointly training the multi-scale feature extraction model, the time-series model, and the classification model based on the model loss.

[0013] To achieve the above objectives, a second aspect of this application proposes a gearbox fault detection device, comprising: an acquisition module for acquiring runtime sequence data of a gearbox; a multi-scale feature extraction module for dividing the runtime sequence data into multiple data segments, performing multi-scale feature extraction on each data segment to obtain multi-scale initial features corresponding to each data segment, and performing multi-scale weighted fusion on each multi-scale initial feature to obtain target runtime features corresponding to each data segment; a temporal feature extraction module for inputting each target runtime feature into a temporal model for mapping to obtain temporal features corresponding to each data segment; and a classification module for performing attention processing on each temporal feature and then inputting it into a classification model for classification processing to obtain the fault detection result of the gearbox.

[0014] To achieve the above objectives, a third aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the gearbox fault detection method described in the first aspect.

[0015] To achieve the above objectives, a fourth aspect of the present application provides a storage medium, which is a computer-readable storage medium storing a computer program that, when executed by a processor, implements the gearbox fault detection method described in the first aspect.

[0016] The embodiments of this application include at least the following beneficial effects: By dividing the target runtime sequence data of the gearbox into multiple target data segments, and then inputting each of the multiple target data segments into a multi-scale feature extraction model for mapping, multi-scale initial features corresponding to the multiple target data segments are obtained. Since each target data segment belongs to a subsequence of the target runtime sequence data, and each target data segment contains multi-scale feature information of the corresponding subsequence, the local spatial features of the target runtime sequence data are fully captured. Then, multi-scale weighted fusion is performed on each multi-scale initial feature to obtain the target runtime features corresponding to each target data segment. This can retain fault-related features in the multi-scale initial features and remove invalid features. Finally, each target runtime feature is input into the time series model for mapping. The algorithm captures the target temporal features corresponding to each target data segment, and then applies attention processing to each target temporal feature. This is equivalent to using the feature-capturing ability of attention processing to focus on different parts of the input sequence composed of each target temporal feature, thereby enhancing the feature extraction capability in the temporal domain. This achieves full capture of the global temporal features of the target runtime sequence data. Finally, the data is input into a classification model for classification processing to obtain the target fault detection result of the gearbox. Therefore, when processing the target runtime sequence data of the gearbox, it can fully capture both the spatial features of the data and the dynamic changes and temporal dependencies of the data, effectively capturing the complex information in the target runtime sequence data of the gearbox. This improves the reliability of the target runtime sequence data analysis and thus improves the accuracy of the target fault detection result.

[0017] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0018] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0019] Figure 1 This is a schematic flowchart of an optional gearbox fault detection method provided in an embodiment of this application;

[0020] Figure 2 This is a schematic diagram of an optional process for feature extraction and weighted fusion provided in an embodiment of this application;

[0021] Figure 3 A schematic diagram of an optional multi-scale feature extraction process provided in an embodiment of this application;

[0022] Figure 4 A schematic diagram of an optional specific process for temporal feature extraction provided in an embodiment of this application;

[0023] Figure 5 This is a schematic diagram of an optional flow for attention processing and classification processing provided in an embodiment of this application;

[0024] Figure 6 This is a schematic diagram of an optional specific process for the classification process provided in an embodiment of this application;

[0025] Figure 7 A schematic diagram of an optional process for the joint training model provided in an embodiment of this application;

[0026] Figure 8 This is a schematic diagram of an optional structure of the multi-scale feature extraction model and the weighted fusion model provided in the embodiments of this application;

[0027] Figure 9 An optional flowchart illustrating the reasoning process provided in an embodiment of this application;

[0028] Figure 10 A schematic diagram of an optional training process provided for an embodiment of this application;

[0029] Figure 11 A schematic diagram of an optional structure of the gearbox fault detection device provided in an embodiment of this application;

[0030] Figure 12 This is a schematic diagram of an optional hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0032] It should be noted that in various specific embodiments of this application, when processing data related to the characteristics of the target object, such as target object attribute information or attribute information sets, is required, the permission or consent of the target object will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. The target object can be a user. In addition, when embodiments of this application need to obtain target object attribute information, separate permission or consent from the target object will be obtained through pop-ups or redirection to a confirmation page. Only after obtaining the target object's separate permission or consent will the necessary target object-related data for the normal operation of the embodiments of this application be obtained.

[0033] In the description of this application, "several" means one or more, "multiple" means two or more, "greater than", "less than", "exceeding" etc. are understood to exclude the number itself, and "above", "below", "within" etc. are understood to include the number itself.

[0034] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, or the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0035] Currently, traditional gearbox fault detection methods mainly rely on convolutional neural networks (CNNs) to extract spatial features from gearbox operating data for classification and prediction to determine the fault category. However, the spatial features that CNNs focus on cannot effectively capture the complex information in gearbox operating data, resulting in low accuracy of fault detection results.

[0036] To address the issue of low accuracy in fault detection results, this application provides a gearbox fault detection method, apparatus, device, and storage medium. The method includes: acquiring target runtime sequence data of the gearbox; dividing the target runtime sequence data into multiple target data segments; inputting each target data segment into a multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment; performing multi-scale weighted fusion on each multi-scale initial feature to obtain target runtime features corresponding to each target data segment; inputting each target runtime feature into a time-series model for mapping to obtain target time-domain features corresponding to each target data segment; and performing attention processing on each target time-domain feature before inputting it into a classification model for classification processing to obtain the target fault detection result of the gearbox.

[0037] According to the solution provided in this application, the target runtime sequence data of the gearbox is divided into multiple target data segments. These segments are then input into a multi-scale feature extraction model for mapping, yielding multi-scale initial features corresponding to each segment. Since each target data segment belongs to a sub-sequence of the target runtime sequence data, and each segment contains multi-scale feature information of its corresponding sub-sequence, the local spatial features of the target runtime sequence data are fully captured. Then, multi-scale weighted fusion is performed on each initial feature to obtain the target runtime features corresponding to each segment. This process retains fault-related features and removes invalid features from the initial features. Finally, each target runtime feature is input into a time-series model for mapping. The method obtains the target temporal features corresponding to each target data segment, and then applies attention processing to each target temporal feature. This is equivalent to using the feature-capturing ability of attention processing to focus on different parts of the input sequence composed of each target temporal feature, thereby enhancing the feature extraction capability in the temporal domain. This achieves full capture of the global temporal features of the target runtime sequence data. Finally, the data is input into a classification model for classification processing to obtain the target fault detection result of the gearbox. Therefore, when processing the target runtime sequence data of the gearbox, it can fully capture both the spatial features of the data and the dynamic changes and temporal dependencies of the data. This effectively captures the complex information in the target runtime sequence data of the gearbox, improves the reliability of the target runtime sequence data analysis, and thus improves the accuracy of the target fault detection result.

[0038] The gearbox fault detection method, apparatus, device, and storage medium provided in this application are specifically described through the following embodiments. First, the gearbox fault detection method in this application embodiment is described.

[0039] The gearbox fault detection method provided in this application relates to the field of computer technology. This gearbox fault detection method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the gearbox fault detection method, but is not limited to the above forms.

[0040] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0041] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0042] like Figure 1 As shown, Figure 1 This is a schematic flowchart of an optional gearbox fault detection method provided in an embodiment of this application. The gearbox fault detection method can be executed by a server, a terminal, or a server in conjunction with a terminal. The gearbox fault detection method includes, but is not limited to, the following steps S110 to S140:

[0043] A method for detecting gearbox faults, characterized in that it includes:

[0044] Step S110: Obtain the target runtime sequence data of the gearbox and divide the target runtime sequence data into multiple target data segments.

[0045] Step S120: Input each target data segment into the multi-scale feature extraction model for mapping to obtain the multi-scale initial features corresponding to each target data segment. Perform multi-scale weighted fusion on each multi-scale initial feature to obtain the target running features corresponding to each target data segment.

[0046] Step S130: Input the running features of each target into the time series model for mapping to obtain the target time domain features corresponding to each target data segment.

[0047] Step S140: After attention processing of the temporal features of each target, input them into the classification model for classification processing to obtain the target fault detection result of the gearbox.

[0048] Vibration signals can be collected through a gearbox experimental platform to obtain target runtime sequence data. The vibration signals or target runtime sequence data reflect the change of the gearbox's vibration acceleration over time.

[0049] Specifically, the time period corresponding to the target runtime sequence data is divided into multiple time steps, and the target runtime sequence data corresponding to each time step is further divided into multiple target data segments.

[0050] Among them, the target fault detection result is used to indicate whether the target runtime sequence data belongs to the fault category or the normal type. For example, the fault category may include gear tooth breakage, gear damage, tooth surface wear, tooth root crack, bearing ball failure, inner and outer ring combined failure, outer ring failure, and inner ring failure, etc.

[0051] Based on this, the target runtime sequence data of the gearbox is divided into multiple target data segments. These segments are then input into a multi-scale feature extraction model for mapping, yielding multi-scale initial features corresponding to each segment. Since each segment belongs to a subsequence of the target runtime sequence data and contains multi-scale feature information of its corresponding subsequence, the local spatial features of the target runtime sequence data are fully captured. Multi-scale weighted fusion is then performed on each initial feature to obtain the target runtime features corresponding to each segment. This process retains fault-related features and removes invalid features. Finally, these target runtime features are input into a time-series model for mapping, yielding the corresponding target runtime features. The target time-domain features corresponding to the target data segments are then processed by attention processing. This is equivalent to focusing on different parts of the input sequence composed of various target time-domain features through the feature-capturing ability of attention processing, thereby enhancing the feature extraction capability in the time domain. This achieves full capture of the global time-domain features of the target runtime sequence data. Finally, the data is input into a classification model for classification processing to obtain the target fault detection result of the gearbox. Therefore, when processing the target runtime sequence data of the gearbox, it can fully capture both the spatial features of the data and the dynamic changes and temporal dependencies of the data. This effectively captures the complex information in the target runtime sequence data of the gearbox, improves the reliability of the target runtime sequence data analysis, and thus improves the accuracy of the target fault detection result.

[0052] Additionally, refer to Figure 2In one embodiment, the multi-scale feature extraction model includes a first feature extraction network and a second feature extraction network. The first feature extraction network includes N sequentially connected first convolutional layers, and the second feature extraction network includes M sequentially connected second convolutional layers, where N is less than M and both N and M are positive integers. The multi-scale initial features include first-scale initial features and second-scale initial features. Each target data segment is input into the multi-scale feature extraction model for mapping to obtain the multi-scale initial features corresponding to each target data segment. The multi-scale initial features are then multi-scale weighted and fused to obtain the target running features corresponding to each target data segment, including but not limited to the following steps:

[0053] Step S210: For each target data segment, input the target data segment into the first feature extraction network for mapping to obtain the first scale initial feature corresponding to the target data segment, and input the target data segment into the second feature extraction network for mapping to obtain the second scale initial feature corresponding to the target data segment.

[0054] Step S220: For each target data segment, perform multi-scale weighted fusion of the first-scale initial features and the second-scale initial features corresponding to the target data segment to obtain the target running features corresponding to the target data segment.

[0055] Among them, reference Figure 8 The first feature extraction network and the second feature extraction network work in parallel.

[0056] In one possible implementation, a weighted fusion model is used to perform multi-scale weighted fusion of features from various multi-scale samples.

[0057] It should be noted that the weighting process for the initial features at the first scale is as follows: One-dimensional global average pooling is performed on the initial features at the first scale, which is equivalent to averaging all elements of the initial features at the first scale. All elements are compressed into an average value to determine the first mean. Then, the first mean is input into two fully connected layers for linear transformation to determine the first weight vector. Finally, the first weight vector is multiplied by the initial features at the first scale to determine the first weighted feature vector.

[0058] In addition, the weighting process for the initial features at the second scale is as follows: one-dimensional global average pooling is performed on the initial features at the second scale, which is equivalent to averaging all elements of the initial features at the second scale, compressing all elements into an average value to determine the second mean, and then the second mean is input into two fully connected layers for linear transformation to determine the second weight vector. Finally, the second weight vector is multiplied by the initial features at the second scale to determine the second weighted feature vector.

[0059] The formula for one-dimensional global average pooling is as follows:

[0060]

[0061] Before obtaining the initial features at the first scale, max pooling is required. j Here, z is the pooling kernel of the max pooling layer, j is the size of the pooling kernel, and i refers to the feature extraction network at the i-th scale in the multi-scale feature extraction model. For example, (i) = 1 represents the first feature extraction network, (i) = 2 represents the second feature extraction network, and when (i) = 1, z (i) Let L be the first mean, and L be the vector length of the initial feature at the first scale. For the nth element in the initial features of the first scale, when (i) = 2, z (i) Let L be the second mean, and L be the vector length of the initial features at the second scale. It is the nth element in the initial features of the second scale.

[0062] The formula for determining each element in the first weight vector or the second weight vector is as follows:

[0063]

[0064] Where i refers to the feature extraction network at the i-th scale in the multi-scale feature extraction model. For example, (i) = 1 represents the first feature extraction network, (i) = 2 represents the second feature extraction network, and when (i) = 1, s (i) Let s be the first weight vector. When (i) = 2, s (i) This is the second weight vector; as well as These are the weight matrices corresponding to the two fully connected layers. as well as These are the biases corresponding to the two fully connected layers, σ is the sigmoid activation function, and ReLU is the ReLU activation function.

[0065] The formula for determining each element in the first weighted feature vector or the second weighted feature vector is as follows:

[0066]

[0067] Before obtaining the initial features at the first scale, max pooling is required. j Here, j represents the pooling kernel of the max pooling layer, j is the size of the pooling kernel, and i refers to the feature extraction network at the i-th scale in the multi-scale feature extraction model. For example, (i) = 1 represents the first feature extraction network, (i) = 2 represents the second feature extraction network, and when (i) = 1, s (i) This is the first weight vector. This is the nth element of the initial feature at the first scale. It is the nth element of the first weighted eigenvector. When (i) = 2, s (i) This is the second weight vector. This is the nth element of the initial feature at the second scale. It is the nth element of the second weighted eigenvector.

[0068] It should be noted that the matrix sizes of the first-scale initial features and the second-scale initial features are the same. The first-scale initial features and the second-scale initial features are fused. Specifically, the first weighted feature vector and the second weighted feature vector are multiplied by matrix or element-wise to determine the target running features corresponding to the target data segment.

[0069] The formula for determining the target runtime characteristics corresponding to the target data segment is as follows:

[0070]

[0071] Where x(n) is the nth element of the target running feature, The nth element of the first weighted eigenvector. It is the nth element of the second weighted eigenvector.

[0072] Based on this, for each target data segment, the first-scale initial features are determined by a first feature extraction network with fewer convolutional layers, which reduces computational load and network complexity. The second-scale initial features are determined by a second feature extraction network with more convolutional layers, enabling the second-scale initial features to represent more complex feature combinations. Then, the first and second feature extraction networks understand the spatial features of the target data segment at the two scales respectively. Finally, the first-scale initial features and the second-scale initial features are weighted and fused to obtain the target running features corresponding to the target data segment. This allows the target running features to more comprehensively describe the spatial features corresponding to the target data segment, thereby avoiding the limitations of a single feature extraction network in understanding spatial features.

[0073] Additionally, refer to Figure 3In one embodiment, the first feature extraction network includes a first convolutional layer, a second convolutional layer, and a first pooling layer connected in sequence. The second feature extraction network includes a third convolutional layer, a fourth convolutional layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, and a third pooling layer connected in sequence. The kernel size of the first convolutional layer is greater than or equal to the kernel size of the second convolutional layer, the kernel size of the first convolutional layer is greater than the kernel size of the third convolutional layer, the kernel size of the third convolutional layer is greater than or equal to the kernel size of the fourth convolutional layer, the kernel size of the third convolutional layer is greater than the kernel size of the fifth convolutional layer, and the kernel size of the fifth convolutional layer is greater than or equal to the kernel size of the sixth convolutional layer. The target data segment is input into the first feature extraction network for mapping to obtain the first-scale initial features corresponding to the target data segment. The target data segment is then input into the second feature extraction network for mapping to obtain the second-scale initial features corresponding to the target data segment. This includes, but is not limited to, the following steps:

[0074] Step S310: The first convolutional layer and the second convolutional layer are called to perform convolution processing on the target data segment to obtain the first intermediate feature. The first pooling layer is called to perform pooling operation on the first intermediate feature to obtain the first scale initial feature corresponding to the target data segment.

[0075] Step S320: The third and fourth convolutional layers are called to perform convolution processing on the target data segment to obtain the second intermediate feature. The second pooling layer is then called to perform pooling operation on the second intermediate feature to obtain the third intermediate feature.

[0076] Step S330: The fifth and sixth convolutional layers are called to perform convolution processing on the third intermediate feature to obtain the fourth intermediate feature. The third pooling layer is then called to perform pooling operation on the fourth intermediate feature to obtain the second-scale initial feature corresponding to the target data segment.

[0077] Among them, reference Figure 8 The convolutional kernels in the first, second, third, fourth, fifth, and sixth convolutional layers can all be one-dimensional. The convolutional kernels in the first feature extraction network are wider, while those in the second feature extraction network are narrower. For example, the kernel size of the first convolutional layer can be 1×32, the kernel size of the second convolutional layer can be 1×16, the kernel size of the third convolutional layer can be 1×16, the kernel size of the fourth convolutional layer can be 1×16, the kernel size of the fifth convolutional layer can be 1×5, and the kernel size of the sixth convolutional layer can be 1×5.

[0078] For the first convolutional layer in the feature extraction network of each size, the corresponding formula for feature extraction is as follows:

[0079]

[0080] Where i refers to the feature extraction network at the i-th scale in the multi-scale feature extraction model, for example, (i) = 1 represents the first feature extraction network, and (i) = 2 represents the second feature extraction network; h n =X (i) (n), h n The nth element of the target runtime sequence data is equivalent to X. (i) (n) represents the nth target data segment of the target runtime sequence data, k j Here, j represents the convolution kernel corresponding to the convolutional layer, and j is the size of the convolution kernel. denoted as the feature extraction result of the nth target data segment, where f is the convolution operation, δ is the activation function, and τ is the time step.

[0081] It should be noted that the first pooling layer, the second pooling layer, and the third pooling layer are all max pooling layers. Therefore, the dimensionality of the first scale initial feature and the first intermediate feature decreases sequentially, and the amount of data decreases sequentially. The dimensionality reduction of the second scale initial feature, the third intermediate feature, and the second intermediate feature decreases sequentially, and the amount of data decreases sequentially.

[0082] The formula for the max pooling layer is as follows:

[0083]

[0084] Where, m j Here, j represents the pooling kernel of the max-pooling layer, j is the size of the pooling kernel, and i refers to the feature extraction network at the i-th scale in the multi-scale feature extraction model. For example, (i) = 1 represents the first feature extraction network, (i) = 2 represents the second feature extraction network, and when (i) = 1... For the nth element of the initial feature at the first scale, when (i) = 2, This is the nth element of the initial feature at the second scale; For the feature extraction result of the nth target data segment, the max pooling layer also has a pooling window. t is the index variable of the pooling window of the max pooling layer. t is used to traverse each position of the pooling window. p is the size of the pooling window. max means taking the maximum value.

[0085] Based on this, the first feature extraction network extracts features from the target data segment through two relatively wide convolutional layers, enabling it to extract relatively overall features of the target data segment and obtain the first intermediate features. Then, pooling is performed on the first intermediate features to reduce their dimensionality and the computational cost of the first feature extraction network, resulting in the first-scale initial features corresponding to the target data segment. Meanwhile, the second feature extraction network first extracts features from the target data segment through two relatively narrow convolutional layers combined with pooling operations, enabling it to initially extract more detailed features of the target data segment and obtain the second intermediate features. Then, two even narrower convolutional layers combined with max pooling are used to extract features from the second intermediate features, accumulating and combining the previously more detailed features into complex and hierarchical features, resulting in the second-scale initial features. Therefore, the first and second feature extraction networks can increase feature diversity, improve overall generalization ability, and improve feature fusion results.

[0086] Additionally, refer to Figure 4 In one embodiment, the runtime features of each target are input into a time-series model for mapping to obtain the target time-domain features corresponding to each target data segment, including but not limited to the following steps:

[0087] Step S410: Input the running features of each target into the time series model, and determine the forward hidden layer state and backward hidden layer state corresponding to each running feature of the target based on the time series model.

[0088] Step S420: For each target running feature, perform a weighted summation based on the forward hidden state and the reverse hidden state corresponding to the target running feature to determine the target time domain feature corresponding to the target data segment.

[0089] It should be noted that the temporal model includes a forward hidden layer and a backward hidden layer. The time steps of the forward hidden layer are progressively increased from front to back. For each time step of the forward hidden layer, there is a forward sub-network, which includes a first update module, a first reset module, a first candidate module, and a first state update module. Each target running feature corresponds to the input of the forward hidden layer from the first time step to the last time step. For each time step, the forward hidden layer state of each time step is determined based on the target running feature of the current time step and the forward hidden layer state of the previous time step.

[0090] Specifically, the forward hidden state of the previous time step is pre-initialized as a zero vector, and the corresponding forward hidden states are determined sequentially from the first time step to the last time step. For each time step, the target running feature of the current time step and the forward hidden state of the previous time step are input together into the first update module and the first reset module corresponding to the current time step. The first update module determines the first activation value corresponding to the current time step, and the first reset module determines the second activation value corresponding to the current time step. The second activation value, the target running feature of the current time step, and the forward hidden state of the previous time step are input together into the first candidate module to obtain the first candidate output. Finally, the first activation value, the first candidate output, and the forward hidden state of the previous time step are input together into the first state update module to determine the forward hidden state of the current time step.

[0091] The first activation value determines the extent to which the forward hidden state of the previous time step is passed to the forward hidden state of the current time step, and the second activation value determines how much information from the forward hidden state of the previous time step is used to compute the first candidate output of the current time step.

[0092] The formula for determining the forward hidden layer state is as follows:

[0093]

[0094] Where t is the time step number incremented from front to back. Let e ​​be the forward hidden state at time step t. t The target running features are input at the t-th time step. The state of the forward hidden layer at time step t-1.

[0095] In addition, the time steps of the backward hidden layer are progressively increased from back to front. For each time step of the backward hidden layer, there is a backward sub-network, which includes a second update module, a second reset module, a second candidate module, and a second state update module. Each target running feature corresponds to the input of the backward hidden layer from the last time step to the first time step. For each time step, the backward hidden layer state of each time step is determined based on the target running features of the current time step and the backward hidden layer state of the previous time step.

[0096] Specifically, the backward hidden state of the previous time step is pre-initialized as a zero vector, and the corresponding backward hidden state is determined sequentially from the first time step to the last time step. For each time step, the target running feature of the current time step and the backward hidden state of the previous time step are input together into the second update module and the second reset module corresponding to the current time step. The second update module determines the second activation value corresponding to the current time step, and the second reset module determines the second activation value corresponding to the current time step. The second activation value, the target running feature of the current time step, and the backward hidden state of the previous time step are input together into the second candidate module to obtain the second candidate output. Finally, the second activation value, the second candidate output, and the backward hidden state of the previous time step are input together into the second state update module to determine the backward hidden state of the current time step.

[0097] The second activation value is used to determine the extent to which the backward hidden state of the previous time step is passed to the backward hidden state of the current time step, and the second activation value is used to determine how much information from the backward hidden state of the previous time step is used to compute the second candidate output of the current time step.

[0098] The formula for determining the backward hidden state is as follows:

[0099]

[0100] Where t is the time step number incremented from back to front. Let e ​​be the backward hidden state at time step t. t The target running features are input at the t-th time step. The state of the forward hidden layer at time step t-1.

[0101] It should be noted that for each time step, a first weight matrix corresponding to the forward hidden state, a second weight matrix corresponding to the backward hidden state, and a hidden layer bias corresponding to the current time step are pre-set. The target time domain features are determined by summing the results of multiplying the first weight matrix with the forward hidden state, multiplying the second weight matrix with the backward hidden state, and the hidden layer bias.

[0102] The formula for determining the target time-domain features is as follows:

[0103]

[0104] Where t is the time step number incremented from front to back, h t Let be the target time-domain feature at time step t. Let be the forward hidden state at time step t. Let w be the backward hidden state at time step t. tLet v be the first weight matrix corresponding to the t-th time step. t Let b be the second weight matrix corresponding to the t-th time step. t This is the hidden layer bias corresponding to the t-th time step.

[0105] Among them, reference Figure 9 , from e1 to e t That is, the target running features from the first time step to the t-th time step are input into the time series model for mapping to obtain h1 to ht. t That is, the target time-domain features from the first time step to the t-th time step.

[0106] Based on this, by inputting the operational features of each target into the time series model, the forward hidden state and backward hidden state corresponding to each operational feature of the target are determined based on the time series model. Then, the forward hidden state and backward hidden state corresponding to the operational features of the target are weighted and summed to determine the target time domain features corresponding to the target data segment. This can capture the data dependencies from both the past to the future and the future to the past, making the data analysis more comprehensive. Furthermore, the introduction of weighted summation can automatically adjust the weights during training, enhancing the end-to-end learning capability.

[0107] Additionally, refer to Figure 5 In one embodiment, the temporal features of each target are processed by attention and then input into a classification model for classification to obtain the target fault detection result of the gearbox, including but not limited to the following steps:

[0108] Step S510: Perform attention processing on the temporal features of each target to obtain attention features.

[0109] Step S520: Obtain the gearbox's working type identifier, concatenate the working type identifier and attention features, and input them into the classification model for classification processing to obtain the gearbox's target fault detection result.

[0110] Among them, reference Figure 9 The various target temporal features are concatenated into a target temporal feature matrix, which is then input into a multi-head attention module. Each attention head in the multi-head attention module has its own corresponding learnable query parameter matrix W. Q Key-value parameter matrix W K Numerical parameter matrix W V Then query the parameter matrix W Q Multiplying the query matrix Q by the input matrix yields the query matrix Q, and the key-value parameter matrix W is then used. K Multiplying the input matrix yields the key value matrix K, and the value parameter matrix W is then used. V Multiplying the query matrix Q by the input matrix yields the numerical matrix V. Finally, the query matrix Q is transformed into the transpose of the key matrix K.T Multiplying the matrix yields an intermediate reference matrix. After reducing the dimensionality of the intermediate reference matrix and normalizing it, it is multiplied by the numerical matrix V to obtain the corresponding attention sub-features. After concatenating multiple attention sub-features and performing a linear transformation, the attention features are obtained.

[0111] The calculation formulas for the input matrix, query matrix, key matrix, and numerical matrix are as follows:

[0112] X = BiGRU(x) T )

[0113] Q = XW Q

[0114] K = XW K

[0115] V = XW V

[0116] Where, x T For the time series model, X is the input matrix, Q is the query matrix, and W is the input matrix. Q For querying the parameter matrix, K is the key-value matrix, and W is the value matrix. K V is the key-value parameter matrix, and W is the numerical matrix. V This is a matrix of numerical parameters.

[0117] In addition, the formula for determining each attention sub-feature is as follows:

[0118]

[0119] Where Q is the query matrix, K is the key matrix, and K T K is the transpose of the key matrix K. T d k Let K be the dimension of the key-value matrix, V be the numerical matrix, and Attention(Q,K,V) be the attention sub-features of each attention head.

[0120] as well as:

[0121] head i =Attention(QW i Q ,KW i K VW i V )

[0122] Among them, head i Let QW be the i-th attention head, where i is the attention head number. i Q Let KW be the query matrix for the i-th attention head. iK Let VW be the query matrix for the i-th attention head. i V ) is the query matrix for the i-th attention head.

[0123] Furthermore, after concatenating multiple attention sub-features and performing a linear transformation, the formula for obtaining the attention features is as follows:

[0124] MHA(Q,K,V)=Concat(head1,…,head h W o

[0125] Where MHA(Q,K,V) represents the attention features, head1,…,head h These are the attention sub-features corresponding to the first attention head from the last attention head. For example, head1 is the attention sub-feature corresponding to the first attention head, and head... h W is the attention sub-feature corresponding to the last attention head. o It is a linear transformation matrix.

[0126] as well as:

[0127] Y = MHA(XW) Q XW K XW V )

[0128] Where Y is the attention feature, X is the input matrix, Q is the query matrix, and W is the input matrix. Q For querying the parameter matrix, K is the key-value matrix, and W is the value matrix. K V is the key-value parameter matrix, and W is the numerical matrix. V This is the value parameter matrix.

[0129] It should be noted that the gearbox's working type identifier can be a string of numbers used to indicate the type or working condition of the gearbox. For example, the working type identifier of a fixed shaft gearbox can be [0,1,1], the working type identifier of a planetary gearbox can be [1,1,0], and the working type identifier of a differential gearbox can be [0,1,0]. The embodiments of this application are not limited here.

[0130] Understandably, the identification of working type can reflect the ease with which different types or working conditions of gearboxes are prone to failure, thus aiding in fault detection.

[0131] One approach is to directly concatenate the job type identifier and the attention feature, or to expand the job type identifier to the same size as the attention feature before concatenating them.

[0132] Based on this, by performing attention processing on the temporal features of each target, attention features are obtained. This allows us to capture the dependencies between different parts of the data and the unique features corresponding to each part, enhancing feature extraction capabilities and preventing information loss due to excessively long data. This improves the accuracy of fault detection results. Then, the gearbox's working type identifier is obtained. The working type identifier and attention features are concatenated and input into the classification model for classification processing to obtain the target fault detection result of the gearbox. The concatenation of the working type identifier and attention features enhances the richness of feature representation, helping the classification model to capture more complex data patterns and relationships. This avoids the classification model making a relatively simple judgment based solely on attention features, thus improving the accuracy of fault detection results.

[0133] Additionally, refer to Figure 6 In one embodiment, the work type identifier and attention feature are concatenated and input into a classification model for classification processing to obtain the target fault detection result of the gearbox, including but not limited to the following steps:

[0134] Step S610: Obtain the historical failure probability of the gearbox.

[0135] Step S620: Input the attention features and historical failure probabilities into the probability prediction model for regression processing to obtain the current failure probability of the gearbox.

[0136] Step S630: The working type identifier, current fault probability and attention feature are concatenated and input into the classification model for classification processing to obtain the target fault detection result of the gearbox.

[0137] The historical failure probability can be the average failure rate of the gearbox from the past to the present, the current failure probability predicted by the probability prediction model, or the failure probability calculated by other statistical or calculation methods. The embodiments of this application are not limited here.

[0138] It is understandable that both historical failure probability and current failure probability are used to indicate the ease with which a gearbox may fail, and they play an auxiliary role in fault detection. For example, when the current failure probability of a gearbox is high, that is, when a gearbox is not likely to fail at present, the corresponding fault detection result will tend to be judged as normal or other fault categories with a higher probability of occurrence.

[0139] Based on this, by obtaining the historical failure probability of the gearbox, we essentially obtain the variable factors affecting the current failure probability. The attention feature and historical failure probability are input into the probability prediction model for regression processing to obtain the current failure probability of the gearbox. Since the historical failure probability represents the ease with which the gearbox will fail in the past, it can describe the current failure probability of the gearbox from one dimension. The attention feature represents the dependency between different parts of the gearbox's target runtime sequence data, and can describe the current failure probability from another dimension. Therefore, both the attention feature and the historical failure probability guide the determination of the current failure probability, increasing its reliability. Then, the working type identifier, the current failure probability, and the attention feature are concatenated and input into a classification model for classification processing to obtain the target failure detection result of the gearbox. The working type identifier represents the type or operating condition of the gearbox, affecting the failure detection result from one dimension. Furthermore, the current failure probability and the attention feature affect the failure detection result from two other dimensions. Therefore, the working type identifier, the current failure probability, and the attention feature guide the determination of the failure detection result, avoiding the classification model's reliance solely on the attention feature for judgment, and improving the accuracy of the failure detection result.

[0140] Additionally, refer to Figure 7 In one embodiment, before performing multi-scale feature extraction on each target data segment, the gearbox fault detection method further includes, but is not limited to, the following steps:

[0141] Step S710: Obtain the sample runtime sequence data of the gearbox and the sample detection label corresponding to the sample runtime sequence data, and divide the sample runtime sequence data into multiple sample data segments.

[0142] Step S720: Input each sample data segment into the multi-scale feature extraction model for mapping to obtain the multi-scale sample features corresponding to each sample data segment. Perform multi-scale weighted fusion on each multi-scale sample feature to obtain the sample running features corresponding to each target data segment.

[0143] Step S730: Input the running features of each sample into the time series model for mapping to obtain the sample time domain features corresponding to each sample data segment.

[0144] Step S740: After attention processing of the temporal features of each sample, input them into the classification model for classification processing to obtain the sample fault detection results of the gearbox.

[0145] Step S750: Determine the model loss based on the difference between the sample fault detection result and the sample detection label, and jointly train the multi-scale feature extraction model, the temporal model and the classification model based on the model loss.

[0146] Among them, the sample detection label is used to indicate whether the sample runtime sequence data belongs to the fault category or the normal type. For example, the fault category may include gear tooth breakage, gear damage, tooth surface wear, tooth root crack, bearing ball failure, inner and outer ring combined failure, outer ring failure, and inner ring failure, etc.

[0147] The model loss is backpropagated sequentially to the classification model, the time series model, and the multi-scale feature extraction model to calculate the parameter gradients in each model. The parameters of each model are then adjusted based on the parameter gradients to minimize the model loss.

[0148] Based on this, firstly, by acquiring the sample runtime sequence data of the gearbox and the corresponding sample detection labels, the sample runtime sequence data is divided into multiple sample data segments. Then, each sample data segment is input into a multi-scale feature extraction model for mapping to obtain the multi-scale sample features corresponding to each sample data segment. Multi-scale weighted fusion is then performed on each multi-scale sample feature to obtain the sample runtime features corresponding to each target data segment. These sample runtime features are then input into a time-series model for mapping to obtain the sample time-domain features corresponding to each sample data segment. Finally, after attention processing, the sample time-domain features are input into a classification model for further processing. The system performs classification processing to obtain sample fault detection results for the gearbox. Finally, it determines the model loss based on the difference between the sample fault detection results and the sample detection labels. Based on the model loss, it jointly trains the multi-scale feature extraction model, the temporal model, and the classification model. This enables the multi-scale feature extraction model to learn and understand the local spatial information of the target runtime sequence data, the temporal model to learn and understand the global temporal information of the target runtime sequence data, and the classification model to learn and understand the intrinsic relationship between the local spatial information and the global temporal information of the target runtime sequence data. Thus, in the inference stage, accurate fault detection results are obtained by analyzing the target runtime sequence data of the gearbox.

[0149] Specifically, refer to Figure 10 , Figure 10 This is an optional flowchart illustrating the training process provided in an embodiment of this application.

[0150] First, acquire data from the planetary gearbox experimental platform. Then, preprocess the data from the planetary gearbox experimental platform. Next, divide the preprocessed data into a training dataset and a test dataset. The training dataset includes sample runtime sequence data and the sample detection labels corresponding to the sample runtime sequence data. For example, the ratio of the training dataset to the test dataset is 8:2.

[0151] Then, all models and networks are initialized with parameters. Then, sample data fragments of sample runtime sequence data are input into the multi-scale feature extraction model for mapping. Multi-scale weighted fusion is performed on each multi-scale sample feature to obtain the sample runtime features corresponding to each target data fragment.

[0152] Then, the running features of each sample are input into the time series model for the first mapping, and the mapping results are input into the time series model for the second mapping to obtain the sample time domain features corresponding to each sample data segment;

[0153] Then, the temporal features of the samples are regularized, and the regularization result is processed by a multi-head attention model and then input into a fully connected layer for mapping and normalization to obtain the sample fault detection results of the gearbox. Here, the fully connected layer is equivalent to a classification model.

[0154] Then, the model loss is determined based on the difference between the sample fault detection results and the sample detection labels, and the multi-scale feature extraction model, time series model and classification model are jointly trained based on the model loss.

[0155] In one possible implementation, a weighted fusion model is used to perform multi-scale weighted fusion of the features of each multi-scale sample, and a multi-head attention module is used to perform attention processing on the temporal features of each sample.

[0156] The model loss can also be backpropagated to the multi-head attention processing model and the weighted fusion model to calculate the parameter gradient of each model. Then, the parameters of each model are adjusted according to the parameter gradient to minimize the model loss.

[0157] Additionally, refer to Figure 11 This application also provides a gearbox fault detection device 1100, comprising:

[0158] The acquisition module 1110 is used to acquire the runtime sequence data of the gearbox;

[0159] The multi-scale feature extraction module 1120 is used to divide the runtime sequence data into multiple data segments, extract multi-scale features from each data segment to obtain the multi-scale initial features corresponding to each data segment, and perform multi-scale weighted fusion on each multi-scale initial feature to obtain the target runtime features corresponding to each data segment.

[0160] The temporal feature extraction module 1130 is used to input the running features of each target into the time series model for mapping, so as to obtain the temporal features corresponding to each data segment.

[0161] The classification module 1140 performs attention processing on each time-domain feature and then inputs it into the classification model for classification to obtain the gearbox fault detection result. It is understood that the specific implementation of this gearbox fault detection device 1100 is basically the same as the specific embodiment of the gearbox fault detection method described above, and will not be repeated here.

[0162] Additionally, refer to Figure 12 , Figure 12 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0163] The processor 1201 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0164] The memory 1202 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1202 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1202 and is called and executed by the processor 1201 to execute the gearbox fault detection method of the embodiments of this application.

[0165] The input / output interface 1203 is used to implement information input and output;

[0166] The communication interface 1204 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0167] Bus 1205 transmits information between various components of the device (e.g., processor 1201, memory 1202, input / output interface 1203, and communication interface 1204);

[0168] The processor 1201, memory 1202, input / output interface 1203 and communication interface 1204 are connected to each other within the device via bus 1205.

[0169] This application also provides a storage medium, which is a computer-readable storage medium for computer-readable storage. The storage medium stores one or more programs, which can be executed by one or more processors to implement the gearbox fault detection method described above.

[0170] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0171] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0172] It will be understood by those skilled in the art that Figures 1 to 12 The technical solutions shown do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0173] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0174] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0175] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0176] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0177] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0178] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0180] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for detecting gearbox faults, characterized in that, include: Obtain the target runtime sequence data of the gearbox, and divide the target runtime sequence data into multiple target data segments; Each target data segment is input into a multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment. Multi-scale weighted fusion is then performed on each multi-scale initial feature to obtain target running features corresponding to each target data segment. Each of the target operational features is input into a time series model for mapping to obtain the target time domain features corresponding to each of the target data segments; After attention processing of the temporal features of each target, the data are input into a classification model for classification to obtain the target fault detection result of the gearbox. The multi-scale feature extraction model includes a first feature extraction network and a second feature extraction network. The first feature extraction network includes N sequentially connected first convolutional layers, and the second feature extraction network includes M sequentially connected second convolutional layers. N is less than M, and both N and M are positive integers. The multi-scale initial features include first-scale initial features and second-scale initial features. The step of inputting each of the target data segments into a multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment, and then performing multi-scale weighted fusion on each of the multi-scale initial features to obtain target running features corresponding to each target data segment includes: For each target data segment, the target data segment is input into the first feature extraction network for mapping to obtain the first scale initial feature corresponding to the target data segment, and the target data segment is input into the second feature extraction network for mapping to obtain the second scale initial feature corresponding to the target data segment. For each target data segment, the first-scale initial feature and the second-scale initial feature corresponding to the target data segment are fused using multi-scale weighting to obtain the target running feature corresponding to the target data segment.

2. The gearbox fault detection method according to claim 1, characterized in that, The first feature extraction network includes a first convolutional layer, a second convolutional layer, and a first pooling layer connected in sequence. The second feature extraction network includes a third convolutional layer, a fourth convolutional layer, a second pooling layer, a fifth convolutional layer, a sixth convolutional layer, and a third pooling layer connected in sequence. The kernel size of the first convolutional layer is greater than or equal to the kernel size of the second convolutional layer. The kernel size of the first convolutional layer is greater than the kernel size of the third convolutional layer. The kernel size of the third convolutional layer is greater than or equal to the kernel size of the fourth convolutional layer. The kernel size of the third convolutional layer is greater than the kernel size of the fifth convolutional layer. The kernel size of the fifth convolutional layer is greater than or equal to the kernel size of the sixth convolutional layer. The step of inputting the target data segment into the first feature extraction network for mapping to obtain the first-scale initial feature corresponding to the target data segment, and inputting the target data segment into the second feature extraction network for mapping to obtain the second-scale initial feature corresponding to the target data segment, includes: The first convolutional layer and the second convolutional layer are called to perform convolution processing on the target data segment to obtain the first intermediate feature. The first pooling layer is called to perform pooling operation on the first intermediate feature to obtain the first scale initial feature corresponding to the target data segment. The third and fourth convolutional layers are called to perform convolution processing on the target data segment to obtain the second intermediate feature. The second pooling layer is then called to perform pooling operation on the second intermediate feature to obtain the third intermediate feature. The fifth and sixth convolutional layers are invoked to perform convolution processing on the third intermediate feature to obtain the fourth intermediate feature. The third pooling layer is then invoked to perform pooling operation on the fourth intermediate feature to obtain the second scale initial feature corresponding to the target data segment.

3. The gearbox fault detection method according to claim 1, characterized in that, The step of inputting each of the target operational features into a time-series model for mapping to obtain the target time-domain features corresponding to each of the target data segments includes: Each of the target operation features is input into the time series model, and the forward hidden layer state and backward hidden layer state corresponding to each of the target operation features are determined based on the time series model. For each of the target operating features, a weighted sum is performed based on the forward hidden layer state and the reverse hidden layer state corresponding to the target operating feature to determine the target temporal feature corresponding to the target data segment.

4. The gearbox fault detection method according to claim 1, characterized in that, The step of performing attention processing on the temporal features of each target and then inputting them into a classification model for classification to obtain the target fault detection result of the gearbox includes: Each of the target temporal features is subjected to attention processing to obtain attention features; The working type identifier of the gearbox is obtained, and the working type identifier and the attention feature are concatenated and input into the classification model for classification processing to obtain the target fault detection result of the gearbox.

5. The gearbox fault detection method according to claim 4, characterized in that, The step of concatenating the work type identifier and the attention feature and inputting them into a classification model for classification processing to obtain the target fault detection result of the gearbox includes: Obtain the historical failure probability of the gearbox; The attention features and the historical failure probabilities are input into the probability prediction model for regression processing to obtain the current failure probability of the gearbox. The working type identifier, the current fault probability, and the attention feature are concatenated and input into the classification model for classification processing to obtain the target fault detection result of the gearbox.

6. The gearbox fault detection method according to claim 1, characterized in that, Before performing multi-scale feature extraction on each of the target data segments, the gearbox fault detection method further includes: The sample runtime sequence data of the gearbox and the sample detection tag corresponding to the sample runtime sequence data are obtained. The sample runtime sequence data is divided into multiple sample data segments, wherein the sample detection tag is used to indicate whether the sample runtime sequence data belongs to the fault category or the normal type. Each of the sample data segments is input into a multi-scale feature extraction model for mapping to obtain multi-scale sample features corresponding to each of the sample data segments. Multi-scale weighted fusion is then performed on each of the multi-scale sample features to obtain sample running features corresponding to each of the target data segments. The running features of each sample are input into a time series model for mapping to obtain the temporal features of each sample data segment; After attention processing of the temporal features of each sample, the data are input into a classification model for classification to obtain the sample fault detection results of the gearbox. The model loss is determined based on the difference between the sample fault detection result and the sample detection label, and the multi-scale feature extraction model, the temporal model and the classification model are jointly trained based on the model loss.

7. A gearbox fault detection device, characterized in that, include: The acquisition module is used to acquire the runtime sequence data of the gearbox; The multi-scale feature extraction module is used to divide the runtime sequence data into multiple target data segments, input each target data segment into the multi-scale feature extraction model for mapping, obtain the multi-scale initial features corresponding to each target data segment, and perform multi-scale weighted fusion on each multi-scale initial feature to obtain the target runtime features corresponding to each target data segment. The temporal feature extraction module is used to input the running features of each target into the time series model for mapping, so as to obtain the temporal features corresponding to each data segment; The classification module performs attention processing on each of the time-domain features and then inputs them into the classification model for classification processing to obtain the fault detection results of the gearbox. The multi-scale feature extraction model includes a first feature extraction network and a second feature extraction network. The first feature extraction network includes N sequentially connected first convolutional layers, and the second feature extraction network includes M sequentially connected second convolutional layers. N is less than M, and both N and M are positive integers. The multi-scale initial features include first-scale initial features and second-scale initial features. The step of inputting each of the target data segments into a multi-scale feature extraction model for mapping to obtain multi-scale initial features corresponding to each target data segment, and then performing multi-scale weighted fusion on each of the multi-scale initial features to obtain target running features corresponding to each target data segment includes: For each target data segment, the target data segment is input into the first feature extraction network for mapping to obtain the first scale initial feature corresponding to the target data segment, and the target data segment is input into the second feature extraction network for mapping to obtain the second scale initial feature corresponding to the target data segment. For each target data segment, the first-scale initial feature and the second-scale initial feature corresponding to the target data segment are fused using multi-scale weighting to obtain the target running feature corresponding to the target data segment.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the gearbox fault detection method according to any one of claims 1 to 6.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the gearbox fault detection method according to any one of claims 1 to 6.

Images