Gearbox fault detection and diagnosis method, electronic equipment and storage medium

Abstract

The invention discloses a gearbox fault detection and diagnosis method, electronic equipment and a storage medium. Comprising the following steps: acquiring a target domain original signal of a full life cycle of a gearbox and source domain original signal data with a label in a source domain database; extracting time domain and frequency domain features of the target domain data, performing weighted fusion on the features to generate a target domain fusion feature vector, performing unsupervised anomaly detection, and generating an isolated forest anomaly detection model; extracting source domain data features to obtain source domain feature vectors, inputting the source domain feature vectors into a graph generation layer to construct an instance graph, and realizing fault type prediction through a classifier; and based on the target domain abnormal signal, migrating the source domain fault diagnosis model to a target domain, optimizing model adaptability through an adversarial training framework, performing fault type identification based on the optimized model, and outputting a diagnosis result. The method has the advantages of good stability, strong robustness to noise and certain generalization ability. The integrity and complementarity of feature expression are enhanced; and the fault detection precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of gearbox fault diagnosis, and in particular to a gearbox fault detection and diagnosis method, electronic equipment, and storage medium. Background Art

[0002] Gearboxes are core components in mechanical systems that enable power transmission, speed and torque conversion, and are widely used in various industrial equipment, such as wind turbines, automotive transmissions, rail transit, and industrial robots. Within these devices, gearboxes often bear high loads and operate continuously for long periods of time. Their operating status directly impacts the efficiency, stability, and safety of the entire mechanical system. Under normal operating conditions, gearboxes utilize internal gear meshing to convert speed and torque between the input and output shafts. However, long-term operation in complex environments such as high loads, variable operating conditions, and strong vibrations makes gearboxes highly susceptible to failures such as gear wear, cracks, and tooth breakage, bearing fatigue, and lubrication failure. Failure to detect and diagnose these failures can lead to decreased equipment performance, unplanned downtime, and even major safety incidents, resulting in significant economic losses and maintenance costs.

[0003] The increasingly complex and changeable operating conditions of today's rotating machinery pose a major challenge to fault diagnosis. Although data-driven fault diagnosis methods are becoming increasingly popular, they are usually heavily dependent on full-cycle data, which makes them less adaptable to different working conditions. In actual engineering applications, gearboxes frequently experience switching of operating conditions such as speed and load, which makes the same fault present differentiated feature distributions under different operating conditions. In addition, fault samples are scarce and the labeling cost is high. It is very time-consuming and energy-consuming to obtain various full-cycle gearbox fault labeling data under different working conditions. At the same time, when extracting features, most existing methods regard multidimensional features as independent vectors, lack the ability to model the intrinsic structure and correlation between features, and ignore the potential semantic relationship between features, so that the information contained in the features extracted by the deep network is incomplete, which limits the expression ability and discrimination performance of the model.

[0004] In response to the above problems, the present invention provides a gearbox fault detection and diagnosis method, an electronic device, and a storage medium. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems in the related art to a certain extent. To this end, one object of the present invention is to provide a gearbox fault detection and diagnosis method, comprising the following steps: S1: Obtain the target domain original signal of the gearbox's full life cycle vibration signal and the labeled source domain original signal data in the source domain database, and preprocess them; S2: Extract the time domain and frequency domain features of the target domain data, perform weighted fusion on the features to generate a target domain fusion feature vector, input it into the adaptive isolation forest for unsupervised anomaly detection, and generate an isolation forest anomaly detection model; S3: Extract source domain data features through convolutional neural networks to obtain source domain feature vectors, which are then input into the graph generation layer to construct an instance graph. Topological relationships are then modeled based on the graph convolutional network. Fault type prediction is achieved through a classifier to generate a source domain fault diagnosis model. S4: Based on the target domain abnormal signal output by S2, the source domain fault diagnosis model is migrated to the target domain. The model adaptability is optimized through the adversarial training framework. The fault type is identified based on the optimized model and the diagnosis result is output. The framework includes: Feature extractor: Use the pre-trained source domain feature extractor to initialize the target domain feature extractor parameters; Domain discriminator: distinguish whether the feature comes from the source domain or the target domain; Correlation alignment module: reduces the difference in feature distribution between the source domain and the target domain.

[0006] In some implementations, the preprocessing includes: performing fast Fourier transform and high-frequency noise removal on the target domain original signal and the labeled source domain original signal data.

[0007] In some implementations, the method of extracting the time domain and frequency domain features of the target domain data and performing weighted fusion on the features to generate a target domain fused feature vector in S2 is: Extract the time domain and frequency domain features of the target domain data and use them as input for multi-view learning to obtain different views; Initialize the weights of different views, dynamically adjust the view weights based on the anomaly score variance loss function, and use the stochastic gradient descent algorithm to update the weight parameters; The feature views are weighted and fused into a unified feature vector.

[0008] In some implementations, the method for performing unsupervised anomaly detection using the adaptive isolation forest in S2 is: a uniformly sample m samples from an input dataset containing N data points; b Randomly select a feature and randomly choose a split value between its maximum and minimum values; c Recursively divide the current node sample into two left and right subtrees according to the split value; d. Repeat steps bc until the preset tree height limit is reached or there is only one sample left in the node; e. Calculate the anomaly score based on the path length of the sample in the isolation tree and generate an anomaly detection model.

[0009] In some implementations, S3 specifically includes: extracting time domain and frequency domain features of the source domain data through a convolutional neural network; The convolutional neural network includes at least one convolution layer, at least one pooling layer, a fully connected layer and a Softmax module, and the at least one convolution layer corresponds to the at least one pooling layer in a one-to-one manner; wherein: Convolutional layer, extracts local feature information of the input feature matrix and outputs a multi-channel feature map; The pooling layer is set after the corresponding convolution layer to downsample the output of the convolution layer; The fully connected layer is set after the last pooling layer to output the feature vector representation; The fully connected layer is combined with the Softmax module as a classifier to predict the fault signal, and the cross entropy loss is used to estimate the difference between the true label and the predicted label.

[0010] In some implementations, the convolutional neural network further includes a graph generation layer that maps the feature matrix output by the fully connected layer into a graph structure, where: Each eigenvector is considered a node in the graph; Establish edge weights based on the similarity between different eigenvectors, generate the adjacency matrix A, and construct the instance graph; Perform convolution operations on the instance graph to embed data structure information into node features; The obtained node features are used for fault classification and domain adversarial training.

[0011] In some implementations, the domain discriminator distinguishes whether the extracted data is from the source domain or the target domain by: Binary cross entropy loss is used as the domain discrimination loss, which is expressed as: ; Where i represents the i-th sample, n is the total number of samples, represents the label of the i-th sample, the source domain is 1, the target domain is 0, and Represents the prediction result of the domain discriminator for the i-th sample.

[0012] In some implementations, the method by which the correlation alignment module reduces the difference in feature distribution between the source domain and the target domain is: The feature distribution is aligned by minimizing the norm distance between the covariance matrices of the source and target domains. The loss function is defined as: ; in, is the Frobenius norm squared, is the covariance matrix of the source domain features, is the covariance matrix of the target domain features, is the normalization coefficient; The final objective function of the optimized model for the source domain and target domain is expressed as: ; ; in, , To balance the hyperparameters, is the source domain fault classification loss, is the target domain anomaly detection loss.

[0013] On the other hand, the present invention provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; and the processor is used to implement a gearbox fault detection and diagnosis method when executing the program stored in the memory.

[0014] On the other hand, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the computer program implements the steps of a gearbox fault detection and diagnosis method.

[0015] The beneficial effects of the present application are as follows: the present invention can effectively diagnose variable operating condition faults of gearboxes, has good stability, strong robustness to noise, and certain generalization capabilities. By adaptively fusing time domain and frequency domain features, the integrity and complementarity of feature expression are enhanced; the isolation forest algorithm with adaptive parameter adjustment is used to screen abnormal samples in the target domain, thereby improving the accuracy and efficiency of fault detection; fault diagnosis adopts a combined structure of CNN and GCN, which introduces sample structure information while capturing local features, thereby improving recognition robustness under complex working conditions. In addition, the adversarial training of domain discriminators and feature extractors and the CORAL alignment mechanism are introduced to realize the unsupervised migration of source domain models to target domains, solving practical problems such as unlabeled target domain data and large differences in feature distribution. The overall solution has significant advantages in diagnostic accuracy, model generalization capability, and unsupervised adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1. It is a schematic flow chart of the detection and diagnosis method of the present invention; Figure 2 It is a schematic diagram of the graph convolutional neural network structure of the present invention; Figure 3 It is a schematic diagram of the process of the graph generation layer of the present invention; Figure 4 It is a flow chart of the present invention using the adversarial training framework to optimize the model. DETAILED DESCRIPTION

[0017] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other. In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of this application.

[0018] Examples, references Figures 1-4 This paper proposes a gearbox fault detection and diagnosis method. This method uses a two-stage approach: detection followed by diagnosis. Combining the efficiency of isolation forests in unsupervised anomaly detection with the advantages of graph convolutional networks in structured relationship modeling, the method improves the model's ability to identify abnormal conditions under different operating conditions and its accuracy in classifying unlabeled data fault types. Specifically, the method includes the following steps: S1: Obtain the target domain original signal of the gearbox's full life cycle vibration signal and the labeled source domain original signal data in the source domain database, and preprocess them; The specific performance of signal data preprocessing is: fast Fourier transform and high-frequency noise removal are performed on the original signal of the target domain and the labeled original signal data of the source domain to provide clean and unified data input for subsequent feature extraction.

[0019] The Fourier transform is specifically expressed as: Frame processing is performed to divide the original signal into segments of fixed length, and the reset ratio can be set to ensure continuity; Window function processing is performed on it, and a Hanning window is applied to each frame signal to reduce spectrum leakage, smooth signal edges, and suppress sidelobe effects.

[0020] Perform fast Fourier transform on each frame of the windowed signal to obtain a complex spectrum, and then take the modulus to obtain the amplitude spectrum; Calculate the frequency domain statistics of each frame, such as power spectral density and spectrum peak, to provide input for subsequent feature extraction.

[0021] The specific performance of high-frequency noise removal is as follows: One example is to propose a low-pass filter design: determine the cutoff frequency and set the low-pass filter cutoff frequency according to the characteristic frequency range of the gearbox (such as the rotational frequency, meshing frequency and its harmonics); design a Butterworth low-pass filter and filter directly in the frequency domain or through a time domain filter.

[0022] Another example proposes frequency domain threshold truncation: frequency domain noise identification involves analyzing the amplitude spectrum after FFT to identify the frequency range dominated by high-frequency noise. High-frequency components are then truncated and an inverse FFT is performed on the filtered spectrum to reconstruct the time domain signal. Data uniformity processing is also performed; amplitude normalization normalizes the filtered signal amplitude to the range [0, 1] or [-1, 1]. Zero-meaning is performed by subtracting the mean from each frame to eliminate DC offset.

[0023] S2: Extract the time and frequency domain features of the target domain data and perform weighted fusion on the features to generate a target domain fused feature vector to obtain more comprehensive feature information and improve the fault detection capability of the isolation forest for unlabeled target domain data. This information is then input into the adaptive isolation forest for unsupervised anomaly detection to generate an isolation forest anomaly detection model. This model performs a preliminary screening of the target domain signal. If it is a normal signal, no further processing is required. If it is judged to be an abnormal signal, the subsequent steps enter the fault diagnosis process. The specific performance of extracting the time domain and frequency domain features of the target domain data and performing weighted fusion on the features to generate the target domain fusion feature vector is as follows: Extract the time domain and frequency domain features of the target domain data and use them as input for multi-view learning to obtain different views; The time domain characteristics include at least one of a peak factor, a shape factor, an impulse factor, an edge factor and a kurtosis factor, and a combination of multiple ones; The frequency domain characteristics include at least one of gravity frequency, frequency root mean square, frequency standard deviation and frequency band energy entropy, and multiple combinations thereof; Initialize the weights of different views, dynamically adjust the view weights for each sample's anomaly score based on the sample's anomaly score variance loss function, and use the stochastic gradient descent algorithm to update the weight parameters; The variance of the anomaly score reflects the degree of separation between abnormal samples and normal samples. The larger the variance, the greater the separation between abnormal samples and normal samples. It also means that when fusing features, the model focuses on features with stronger discriminability under the current working conditions, making the abnormal sample detection effect better. By suppressing redundant feature interference and strengthening the difference in abnormal representation, the accuracy and robustness of anomaly detection are significantly improved, especially showing stronger environmental adaptability in complex working conditions or gradual fault scenarios.

[0024] The anomaly score variance loss function is: ; in, is a set of anomaly scores containing m samples, , = is the mean of the anomaly scores, is the abnormality score of the i-th sample, is the mean of the abnormal scores of m samples, is the variance of all anomaly scores in the anomaly score set; Weighted fusion of each feature view into a unified feature vector; Isolation forest is an unsupervised fault detection method based on an isolation mechanism. It identifies outliers in unlabeled data by constructing multiple binary isolation trees. The data space is recursively partitioned by randomly selecting features and split values. This random partitioning mechanism isolates outliers from normal points and those in sparse areas at shallow tree depths. To improve the adaptability of the fault detection module to different operating conditions and data distributions, this paper further proposes an adaptive isolation forest parameter adjustment mechanism, which is specifically manifested as follows: a uniformly sample m samples from an input dataset containing N data points; b Randomly select a feature and randomly choose a split value between its maximum and minimum values; c Recursively divide the current node sample into two left and right subtrees according to the split value; d. Repeat steps bc until the preset tree height limit is reached or there is only one sample left in the node; e. Calculate the anomaly score based on the path length of the sample in the isolation tree and generate an anomaly detection model.

[0025] ; in, is the path length normalization factor under n samples, is an estimate of the harmonic number, obtained by estimate, is Euler's constant 0.5772156649, Is the path length, which is the core metric of the isolation forest and is defined as the number of edges from the root node to the leaf node of the isolated sample x. is the average path length of all isolated trees in the forest for sample x; In this formula, when tends to 0, When it approaches 1, it means that the sample is very likely to be abnormal; when tend to , When it approaches 0.5, it means that the sample has no obvious abnormal characteristics; when tend to , When it approaches 0, it indicates that the sample is very likely a normal sample. By calculating the anomaly score of each test data, the anomalies are screened out and a "candidate fault sample set" is constructed to facilitate subsequent fault diagnosis.

[0026] The distribution of gearbox vibration signals varies significantly under different loads, speeds, and environmental conditions, and fixed-parameter isolation forests may perform poorly in certain operating conditions. Therefore, by maximizing the variance of the anomaly scores, the number of binary trees t in the isolation forest is adjusted inversely. The number of samples m used to train a single isolation tree can be adjusted based on the sampling frequency of the vibration signal.

[0027] The number of initial trees t is based on the amount of target domain data. Set, where N is the sample point. When the anomaly score variance When it is lower than 80% of the historical best level, the number of trees is increased by 1.2t to improve the detection sensitivity; when When the value is continuously higher than 90% of the historical best level, the number of trees is reduced by 0.9t to optimize the computational efficiency. The number of training samples m for a single tree and the signal sampling rate Automatic adaptation: When When setting ;when When setting ;when kHz, setting This design ensures that high-frequency signals retain sufficient details and low-frequency signals avoid overfitting.

[0028] S3: Extract source domain data features through a convolutional neural network to obtain source domain feature vectors. These feature vectors are then fed into a graph generation layer to construct an instance graph to obtain different views of the vibration signal. Topological relationships are then modeled based on a graph convolutional network. The adjacency matrix is ​​constructed by multiplying features by their transposes to calculate similarities. Top-k edges are retained to enhance sparsity. The graph generation layer is used to embed data structure information into the feature representation. A classifier is used to predict fault types and generate a source domain fault diagnosis model. By extracting the deep features of the signal and performing structural modeling, the model's ability to judge complex working conditions and multiple types of faults can be effectively improved. Figure 2 As shown in the figure, the deep diagnostic structure of the convolutional neural network domain and the graph convolutional network proposed in the present invention is used to learn and model the structural correlation between the local spatial features of the signal domain graph structure to achieve multi-level and global feature expression; Extract the time domain and frequency domain features of source domain data through convolutional neural network; The convolutional neural network includes at least one convolution layer, at least one pooling layer, a fully connected layer, and a Softmax module. At least one convolution layer corresponds to at least one pooling layer. After the convolution (Conv layer) and pooling (Pooling layer) operations, local feature information is extracted and a set of multi-channel feature maps (Feature maps) are obtained. The convolution layer filters the input feature matrix through the local receptive field mechanism to capture small but critical local patterns such as fluctuations and periodicity in the vibration signal. The pooling layer downsamples the convolution result to enhance the model's translation invariance and noise reduction capabilities. Finally, a set of fixed-dimensional feature vector representations is output through the fully connected layer (FC layer); specifically, it is as follows: Convolutional layer, extracts local feature information of the input feature matrix and outputs a multi-channel feature map; The pooling layer is set after the corresponding convolution layer to downsample the output of the convolution layer; The fully connected layer is set after the last pooling layer to output the feature vector representation; The fully connected layer is combined with the Softmax module as a classifier to predict the fault signal, and the cross entropy loss is used to estimate the difference between the true label and the predicted label.

[0029] Taking a double-layer convolutional layer and a double-layer pooling layer as an example: the convolutional neural network includes two sequentially connected convolution-pooling units, a fully connected layer, and a Softmax module; where: The first convolutional layer: extracts local features of the input signal and outputs a multi-channel feature map; First pooling layer: Following the first convolutional layer, the feature map is downsampled by maximum pooling; The second convolutional layer: continues to extract high-order features and outputs deeper feature maps; Second pooling layer: downsample the output of the second convolutional layer twice; Fully connected layer: flattens the two-dimensional feature map after the second pooling layer into a one-dimensional vector and outputs a fixed-length feature vector representation; The fully connected layer is combined with the Softmax module as a classifier to predict fault signals and use cross-entropy loss to estimate the difference between the true label and the predicted label. To further improve the model's ability to model the relationship between features, the convolutional neural network also includes a graph generation layer, which maps the feature matrix output by the fully connected layer into a graph structure, where: Each eigenvector is considered a node in the graph; According to the similarity between different eigenvectors, edge weights are established, adjacency matrix A is generated, and instance graph is constructed. The process is referred to Figure 3 As shown, the specific performance is: First, the extracted feature matrix X is input into the multi-layer perceptron (MLP) for nonlinear transformation. Then, the transformed feature matrix The adjacency matrix A is obtained by multiplying the similarity between different eigenvectors with their transposed matrices. Finally, according to the Top-k sorting mechanism, the first k nodes with the largest similarity are selected for each node, and the others are set to zero to obtain a sparse adjacency matrix. .

[0030] Convolution operations are performed on the instance graph to embed data structure information into node features. This operation not only retains local structural characteristics, but also strengthens the semantic connections between nodes, effectively mining topological association patterns hidden in the data.

[0031] The obtained node features are used for fault classification and domain adversarial training.

[0032] S4: Based on the target domain abnormal signal output by S2, the source domain fault diagnosis model is transferred to the target domain. The abnormal signal detected by the target domain is input and adversarial training is performed with the discriminator and feature extractor. CORAL alignment is used to further reduce the difference between domains. The model adaptability is optimized through the adversarial training framework. The fault type is identified based on the optimized model and the diagnosis result is output. Figure 4 As shown, the framework includes: Feature extractor: Initialize the target domain feature extractor parameters using a pre-trained source domain feature extractor, where the feature extractor is the convolutional neural network described in this embodiment. Domain discriminator: distinguish whether the feature comes from the source domain or the target domain; Correlation alignment module: reduces the difference in feature distribution between the source domain and the target domain.

[0033] The specific performance of the domain discriminator in distinguishing whether the extracted data comes from the source domain or the target domain is: Due to the domain shift problem, the classifier trained only with source domain data cannot process target domain data well. To solve this problem, a domain discriminator (D) is used to determine whether the extracted features are from the source domain or the target domain. Through adversarial training, the feature extractor is made to deceive the domain discriminator and generate domain-invariant features. The fully connected layer is combined with the Softmax module as a classifier for predicting fault signals. The binary cross entropy loss is used as the domain discrimination loss, which is expressed as: ; Where i represents the i-th sample, n is the total number of samples, represents the label of the i-th sample, the source domain is 1, the target domain is 0, and Represents the prediction result of the domain discriminator for the i-th sample.

[0034] The specific performance of the correlation alignment module in reducing the difference in feature distribution between the source domain and the target domain is as follows: The feature distribution is aligned by minimizing the norm distance between the covariance matrices of the source and target domains. The loss function is defined as: ; in, is the Frobenius norm squared, is the covariance matrix of the source domain features, is the covariance matrix of the target domain features, is the normalization coefficient, d is the eigenvector dimension; The final objective function of the optimized model for the source domain and target domain is expressed as: ; ; in, , To balance the hyperparameters, L S is the total loss in the source domain, is the source domain fault classification loss, is the target domain anomaly detection loss.

[0035] To achieve consistency in the feature distributions of the source and target domains, this paper introduces a domain discriminator to identify the domain to which a sample belongs. This discriminator is trained by minimizing the discriminant loss. By introducing gradient reversal during backpropagation, adversarial training is implemented against the feature extractor. During the update process, the gradient direction of the loss is reversed, so that the feature extractor aims to maximize the discriminant loss of the domain discriminator, thereby extracting domain-insensitive features. The entire network can be optimized using stochastic gradient descent (SGD), balancing model parameters between source domain task accuracy and domain alignment capabilities.

[0036] At the same time, an adaptive learning module called CORAL alignment is used in the fully connected layer. This module aligns the feature distributions between the source and target domains by minimizing the distance between the second-order statistics (i.e., the covariance matrix) of the features, further reducing inter-domain differences and improving the model's migration and generalization capabilities.

[0037] The multi-feature view fusion method for gearbox vibration signals in this paper treats time domain and frequency domain features as two independent views and realizes feature adaptive fusion under unsupervised conditions through a dynamic weight update mechanism based on the variance of abnormal scores. Under the condition of no labeled data in the target domain, the present invention adopts a two-stage method of first detection and then diagnosis. It uses multi-view fusion features to input the isolation forest, screens abnormal samples in the target domain (candidate fault samples), filters normal signals to reduce the computational load, and performs domain adaptive migration diagnosis only on abnormal samples. It also uses an adaptive parameter adjustment mechanism to dynamically adjust the number of trees in the isolation forest according to the ratio of the anomaly score variance to the historical best level, and the number of samples in a single tree is segmented according to the signal sampling rate. The present invention proposes a feature extraction and structural modeling method for the graph convolutional network module. While extracting local features, it embeds the topological relationship between samples to enhance the discriminability of fault features. It also combines the adversarial training of the domain discriminator and the feature extractor with the CORAL alignment mechanism to achieve feature distribution adaptation from the source domain to the target domain, effectively improving the diagnostic stability and accuracy under complex working conditions.

[0038] It should be understood that the method proposed in the present invention is applicable to gearboxes in various industrial equipment, not limited to those mentioned in the background art, such as wind turbines, automobile transmissions, rail transportation, and industrial robots.

[0039] Based on the same technical concept, an embodiment of the present invention further provides an electronic device, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus.

[0040] Memory for storing computer programs; The processor is used to implement the steps of the method of the present invention when executing the program stored in the memory.

[0041] The communication bus mentioned in the electronic device mentioned above can be the Peripheral Component Interconnect (PCI) bus or the Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, the figure shows only one thick line, but this does not mean that there is only one bus or only one type of bus.

[0042] The communication interface is used for communication between the above electronic device and other devices.

[0043] The memory may include Random Access Memory (RAM), or Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0044] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0045] The computer program product of the method provided in the embodiment of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the previous method embodiment. The specific implementation can be found in the method embodiment and will not be repeated here.

[0046] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device, such as a personal computer, a server, or a network device, to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0047] Unless otherwise specified, all steps of the present application may be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), indicating that the method may include steps (a) and (b) performed sequentially, or may include steps (b) and (a) performed sequentially. For example, the method may further include step (c), indicating that step (c) may be added to the method in any order, for example, the method may include steps (a), (b) and (c), or may include steps (a), (c) and (b), or may include steps (c), (a) and (b), etc.

[0048] " range " disclosed in the present application is limited in the form of lower limit and upper limit, and given range is limited by selecting a lower limit and an upper limit, and the selected lower limit and upper limit define the boundary of special range. The scope limited in this way can be to include end value or not include end value, and can be arbitrarily combined, that is, any lower limit can form a range with any upper limit combination. For example, if the scope of 60-120 and 80-110 is listed for specific parameters, it is understood that the scope of 60-110 and 80-120 is also expected. In addition, if the minimum range value 1 and 2 are listed, and if the maximum range value 3,4 and 5 are listed, then the following range can all be expected: 1-3, 1-4, 1-5, 2-3, 2-4 and 2-5. In this application, unless otherwise specified, the numerical range " ab " represents the abbreviation of any real number combination between a and b, wherein a and b are all real numbers. For example, a numerical range of "0-5" indicates that all real numbers between "0-5" are listed herein, and "0-5" is simply an abbreviation for these numerical combinations. Furthermore, when a parameter is expressed as an integer ≥ 2, this is equivalent to disclosing that the parameter is, for example, an integer of 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, etc.

[0049] If not otherwise specified, the “include” and “comprising” mentioned in this application represent open-endedness. For example, the “include” and “comprising” may mean including or comprising other components not listed.

[0050] Unless otherwise specified, all embodiments and optional embodiments of the present application can be combined with each other to form a new technical solution.

[0051] Unless otherwise specified, all technical features and optional technical features of this application can be combined with each other to form a new technical solution.

[0052] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein with equivalents. However, such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. They should all be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A gearbox fault detection and diagnosis method, characterized in that: The following steps are involved: S1: Obtain the target domain original signal of the gearbox's full life cycle vibration signal and the labeled source domain original signal data in the source domain database, and preprocess them; S2: Extract the time domain and frequency domain features of the target domain data, perform weighted fusion on the features to generate a target domain fusion feature vector, input it into the adaptive isolation forest for unsupervised anomaly detection, and generate an isolation forest anomaly detection model; S3: Extract source domain data features through convolutional neural networks to obtain source domain feature vectors, which are then input into the graph generation layer to construct an instance graph. Topological relationships are then modeled based on the graph convolutional network. Fault type prediction is achieved through a classifier to generate a source domain fault diagnosis model. S4: Based on the target domain abnormal signal output by S2, the source domain fault diagnosis model is migrated to the target domain. The model adaptability is optimized through the adversarial training framework. The fault type is identified based on the optimized model and the diagnosis result is output. The framework includes: Feature extractor: Use the pre-trained source domain feature extractor to initialize the target domain feature extractor parameters; Domain discriminator: distinguish whether the feature comes from the source domain or the target domain; Correlation alignment module: reduces the difference in feature distribution between the source domain and the target domain.

2. A gearbox fault detection and diagnosis method according to claim 1, characterized in that: Preprocessing includes: Fast Fourier transform and high-frequency noise removal are performed on the target domain original signal and the labeled source domain original signal data.

3. A gearbox fault detection and diagnosis method according to claim 1, characterized in that: The method for extracting the time domain and frequency domain features of the target domain data and performing weighted fusion on the features to generate the target domain fusion feature vector described in S2 is: Extract the time domain and frequency domain features of the target domain data and use them as input for multi-view learning to obtain different views; Initialize the weights of different views, dynamically adjust the view weights based on the anomaly score variance loss function, and use the stochastic gradient descent algorithm to update the weight parameters; The feature views are weighted and fused into a unified feature vector.

4. A gearbox fault detection and diagnosis method according to claim 1, characterized in that: The method of unsupervised anomaly detection using adaptive isolation forests described in S2 is: a uniformly sample m samples from an input dataset containing N data points; b Randomly select a feature and randomly choose a split value between its maximum and minimum values; c Recursively divide the current node sample into two left and right subtrees according to the split value; d. Repeat steps bc until the preset tree height limit is reached or there is only one sample left in the node; e. Calculate the anomaly score based on the path length of the sample in the isolation tree and generate an anomaly detection model.

5. The gearbox fault detection and diagnosis method according to claim 1, characterized in that: S3 specifically involves extracting the time domain and frequency domain features of the source domain data through a convolutional neural network; The convolutional neural network includes at least one convolution layer, at least one pooling layer, a fully connected layer and a Softmax module, and the at least one convolution layer corresponds to the at least one pooling layer in a one-to-one manner; in: Convolutional layer, extracts local feature information of the input feature matrix and outputs a multi-channel feature map; The pooling layer is set after the corresponding convolution layer to downsample the output of the convolution layer; The fully connected layer is set after the last pooling layer to output the feature vector representation; The fully connected layer is combined with the Softmax module as a classifier to predict the fault signal, and the cross entropy loss is used to estimate the difference between the true label and the predicted label.

6. A gearbox fault detection and diagnosis method according to claim 5, characterized in that: The convolutional neural network also includes a graph generation layer, which maps the feature matrix output by the fully connected layer into a graph structure, where: Each eigenvector is considered a node in the graph; Establish edge weights based on the similarity between different eigenvectors, generate the adjacency matrix A, and construct the instance graph; Perform convolution operations on the instance graph to embed data structure information into node features; The obtained node features are used for fault classification and domain adversarial training.

7. A gearbox fault detection and diagnosis method according to claim 1, characterized in that: The domain discriminator distinguishes whether the extracted data comes from the source domain or the target domain by: Binary cross entropy loss is used as the domain discrimination loss, which is expressed as: ； Where i represents the i-th sample, n is the total number of samples, represents the label of the i-th sample, the source domain is 1, the target domain is 0, and Represents the prediction result of the domain discriminator for the i-th sample.

8. The gearbox fault detection and diagnosis method according to claim 1, characterized in that: The method used by the correlation alignment module to reduce the difference in feature distribution between the source domain and the target domain is: The feature distribution is aligned by minimizing the norm distance between the covariance matrices of the source and target domains. The loss function is defined as: ； in, is the Frobenius norm squared, is the covariance matrix of the source domain features, is the covariance matrix of the target domain features, is the normalization coefficient; The final objective function of the optimized model for the source domain and target domain is expressed as: ； ； in, , To balance the hyperparameters, is the source domain fault classification loss, is the target domain anomaly detection loss.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; the memory is used to store computer programs; the processor is used to implement a gearbox fault detection and diagnosis method as described in any one of claims 1 to 8 when executing the program stored in the memory.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the gearbox fault detection and diagnosis method according to any one of claims 1 to 8 are implemented.

Images