A gearbox fault detection and diagnosis method, an electronic device and a storage medium

Abstract

The application discloses a kind of gearbox fault detection diagnosis method, electronic equipment and storage medium.It includes: obtaining the target domain original signal of the whole life cycle of gearbox and the source domain original signal data with label in source domain database;Extract the time domain, frequency domain features of target domain data, generate target domain fusion feature vector by weighting fusion, carry out unsupervised anomaly detection, generate isolated forest anomaly detection model;Extract source domain data features to obtain source domain feature vector, input graph generation layer to construct instance graph, and realize fault type prediction through classifier;Based on target domain abnormal signal, the source domain fault diagnosis model is migrated to target domain, and the adaptability of the model is optimized through the adversarial training framework, and the fault type is identified based on the optimized model, and the diagnosis result is output.The application has good stability, strong robustness to noise, and certain generalization ability.Enhance the integrity and complementarity of feature expression;Improve the accuracy and efficiency of fault detection.

Description

Technical Field

[0001] This invention relates to the field of gearbox fault diagnosis. In particular, it relates to a gearbox fault detection and diagnosis method, electronic device, and storage medium. Background Technology

[0002] Gearboxes are core components in mechanical systems, enabling power transmission and speed / torque conversion. They are widely used in various industrial equipment, such as wind turbines, automotive transmissions, rail transportation, and industrial robots. In these devices, gearboxes typically bear the burden of high loads and long-term continuous operation, and their operating status directly affects the efficiency, stability, and safety of the entire mechanical system. Under normal operating conditions, gearboxes achieve speed and torque conversion between the input and output shafts through internal gear meshing. However, gearboxes operate under complex environments with high loads, variable operating conditions, and strong vibrations, making them highly susceptible to failures such as gear wear, cracks, tooth breakage, bearing fatigue, and lubrication failure. Failure to detect and diagnose these faults in a timely manner can lead to equipment performance degradation, unplanned downtime, and even major safety accidents, resulting in significant economic losses and maintenance costs.

[0003] The increasingly complex and variable operating conditions of modern rotating machinery pose significant challenges to fault diagnosis. While data-driven fault diagnosis methods are gaining popularity, they typically rely heavily on full-cycle data, resulting in limited adaptability to diverse operating conditions. In practical engineering applications, gearboxes frequently undergo changes in speed, load, and other operating conditions, causing the same fault to exhibit differentiated characteristic distributions under different conditions. Furthermore, fault samples are scarce, and annotation costs are high, making the acquisition of various full-cycle gearbox fault labeling data under different operating conditions extremely time-consuming and energy-intensive. Simultaneously, in feature extraction, most existing methods treat multidimensional features as independent vectors, lacking the ability to model the intrinsic structure and correlations between features and ignoring potential semantic relationships. Consequently, the information contained in the features extracted by deep networks is incomplete, limiting the model's expressive power and discriminative performance.

[0004] To address the above problems, this invention proposes a gearbox fault detection and diagnosis method, electronic device, and storage medium. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, one object of this invention is to provide a gearbox fault detection and diagnosis method, comprising the following steps:

[0006] S1: Obtain the target domain raw signal and the source domain raw signal data with labels from the source domain database of the vibration signal throughout the entire life cycle of the gearbox, and preprocess them;

[0007] S2: Extract the time domain and frequency domain features of the target domain data, perform weighted fusion of the features to generate the target domain fusion feature vector, input it into the adaptive isolated forest for unsupervised anomaly detection, and generate the isolated forest anomaly detection model;

[0008] S3: Extract source domain data features through a convolutional neural network to obtain source domain feature vectors, input them into a graph generation layer to construct an instance graph, model topological relationships based on graph convolutional networks, and use a classifier to predict fault types to generate a source domain fault diagnosis model;

[0009] S4: Based on the target domain abnormal signal output by S2, the source domain fault diagnosis model is transferred to the target domain. The model adaptability is optimized through an adversarial training framework. Based on the optimized model, fault type identification is performed, and the diagnosis result is output. The framework includes:

[0010] Feature extractor: Initialize the target domain feature extractor parameters using a pre-trained source domain feature extractor;

[0011] Domain discriminator: distinguishes whether features originate from the source domain or the target domain;

[0012] Related alignment module: Reduces the difference in feature distribution between the source and target domains.

[0013] In some implementations, preprocessing includes performing Fast Fourier Transform and high-frequency noise removal on the target domain raw signal and the labeled source domain raw signal data.

[0014] In some implementations, the method described in S2 for extracting the time-domain and frequency-domain features of the target domain data and then weighting and fusing these features to generate a target domain fused feature vector is as follows:

[0015] 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;

[0016] Initialize the weights of different views, dynamically adjust the view weights based on the outlier score variance loss function, and update the weight parameters using the stochastic gradient descent algorithm;

[0017] The various feature views are weighted and merged into a unified feature vector.

[0018] In some implementations, the method for unsupervised anomaly detection in the adaptive isolated forest described in S2 is as follows:

[0019] a. Uniformly sample m samples from an input dataset containing N data points;

[0020] b. Randomly select features and randomly choose a segmentation value between their maximum and minimum values;

[0021] c recursively divides the current node sample into two subtrees, left and right, based on the segmentation value;

[0022] d. Repeat steps b and c until the preset tree height limit is reached or only one sample remains in the node;

[0023] e calculates the anomaly score based on the path length of the sample in the isolation tree and generates an anomaly detection model.

[0024] In some implementations, S3 specifically refers to: extracting the time-domain and frequency-domain features of the source domain data through a convolutional neural network;

[0025] The convolutional neural network includes at least one convolutional layer, at least one pooling layer, a fully connected layer, and a Softmax module, wherein the at least one convolutional layer corresponds one-to-one with the at least one pooling layer; wherein:

[0026] Convolutional layers extract local feature information from the input feature matrix and output multi-channel feature maps;

[0027] Pooling layers are set after the corresponding convolutional layers to downsample the output of the convolutional layers.

[0028] A fully connected layer is set after the last pooling layer to output a feature vector representation.

[0029] The fully connected layer is combined with the Softmax module as a classifier to predict fault signals, and the cross-entropy loss is used to estimate the difference between the true and predicted labels.

[0030] In some implementations, convolutional neural networks also include graph generation layers, which map the feature matrices output by the fully connected layers into a graph structure, where:

[0031] Each feature vector is considered a node in the graph;

[0032] Edge weights are established based on the similarity between different feature vectors, an adjacency matrix A is generated, and an instance graph is constructed.

[0033] Convolution operations are performed on the instance graph to embed data structure information into node features;

[0034] The obtained node features are used for fault classification and domain adversarial training.

[0035] In some implementations, the domain discriminator distinguishes whether the extracted data comes from the source domain or the target domain using the following methods:

[0036] Using binary cross-entropy loss as the domain discrimination loss, it is expressed as:

[0037] ;

[0038] Where i represents the i-th sample, and n is the total number of samples. Let represent the label of the i-th sample, with a source domain of 1 and a target domain of 0, and The domain discriminant predicts the i-th sample.

[0039] In some implementations, the relevant alignment module reduces the difference in feature distribution between the source and target domains by:

[0040] The loss function is defined as follows: By minimizing the norm distance between the covariance matrices of the source and target domains to align the feature distribution, the loss function is:

[0041] ;

[0042] in, It is the square of the Frobenius norm. It is the covariance matrix of the source domain features. It is the covariance matrix of the target domain features. It is the normalization coefficient;

[0043] The optimized model's final objective function for both the source and target domains is expressed as:

[0044] ;

[0045] ;

[0046] in, , To balance hyperparameters, For source domain fault classification loss, The loss is for anomaly detection in the target domain.

[0047] 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.

[0048] On the other hand, the present invention proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a gearbox fault detection and diagnosis method.

[0049] The beneficial effects of this application are as follows: This invention can effectively diagnose gearbox faults under varying operating conditions, exhibits good stability, strong robustness to noise, and a certain degree of generalization ability. By adaptively fusing time-domain and frequency-domain features, the integrity and complementarity of feature representation are enhanced; the use of an isolated forest algorithm with adaptive parameter adjustment for screening abnormal samples in the target domain improves the accuracy and efficiency of fault detection; the fault diagnosis employs a combination of CNN and GCN structures, capturing local features while introducing sample structure information, thus improving the robustness of recognition under complex operating conditions. Furthermore, the introduction of adversarial training between the domain discriminator and feature extractor, along with a CORAL alignment mechanism, enables unsupervised transfer of the source domain model to the target domain, 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 ability, and unsupervised adaptability. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the detection and diagnosis method of the present invention;

[0051] Figure 2 This is a schematic diagram of the convolutional neural network structure of the present invention;

[0052] Figure 3 This is a schematic diagram of the graph generation layer of the present invention;

[0053] Figure 4 This is a schematic diagram of the process of optimizing the model using the adversarial training framework of this invention. Detailed Implementation

[0054] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. To enable those skilled in the art to better understand the solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0055] Examples, References Figures 1-4 This invention proposes a gearbox fault detection and diagnosis method. The method employs a two-stage approach: detection followed by diagnosis. It combines the efficiency of isolated forests in unsupervised anomaly detection with the advantages of graph convolutional networks in structured relationship modeling, thereby improving the model's ability to identify abnormal states under different operating conditions and its classification accuracy for unlabeled data fault types. Specifically, it includes the following steps:

[0056] S1: Obtain the target domain raw signal and the source domain raw signal data with labels from the source domain database of the vibration signal throughout the entire life cycle of the gearbox, and preprocess them;

[0057] The specific manifestation of signal data preprocessing is as follows: performing Fast Fourier Transform and high-frequency noise removal on the original signal data of the target domain and the original signal data of the labeled source domain, so as to provide clean and uniform data input for subsequent feature extraction.

[0058] The Fourier transform is specifically manifested as follows:

[0059] The signal is segmented into frames, with the original signal divided into segments of fixed length. A reset ratio can be set to ensure continuity.

[0060] The signal is processed using a window function, and a Hanning window is applied to each frame of the signal to reduce spectral leakage, smooth signal edges, and suppress sidelobe effects.

[0061] Perform a Fast Fourier Transform on each frame of the windowed signal to obtain the complex spectrum, and take the modulus of the spectrum to obtain the amplitude spectrum.

[0062] Calculate frequency domain statistics for each frame, such as power spectral density and spectral peak value, to provide input for subsequent feature extraction.

[0063] High-frequency noise removal specifically manifests as follows:

[0064] One example is to propose a low-pass filter design: determine the cutoff frequency, set the low-pass filter cutoff frequency according to the gearbox's characteristic frequency range (such as rotational frequency, meshing frequency and its harmonics); design a Butterworth low-pass filter to filter directly in the frequency domain, or implement it through a time-domain filter.

[0065] In another example, frequency domain threshold truncation is proposed: frequency domain noise identification involves analyzing the amplitude spectrum after FFT to identify the frequency range dominated by high-frequency noise; then, high-frequency components are truncated, and an inverse FFT is performed on the filtered spectrum to reconstruct the time-domain signal. Data uniformity processing is also included; amplitude normalization: the amplitude of the filtered signal is unified to the range of [0,1] or [-1,1]. Zero-mean normalization: the mean is subtracted from each frame of the signal to eliminate DC offset.

[0066] S2: Extract the time-domain and frequency-domain features of the target domain data, and perform weighted fusion of the features to generate a target domain fusion feature vector to obtain more comprehensive feature information and improve the fault detection capability of the isolated forest for unlabeled target domain data; then input it into the adaptive isolated forest for unsupervised anomaly detection to generate an isolated forest anomaly detection model. This model performs preliminary screening of the target domain signals. If the signals are normal, no further processing is required; if they are determined to be anomalies, the subsequent steps are taken to enter the fault diagnosis process.

[0067] The specific process of extracting time-domain and frequency-domain features from the target domain data and then weighting and fusing these features to generate a target domain fused feature vector is as follows:

[0068] 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;

[0069] The time-domain features include at least one and multiple combinations of peak factor, waveform factor, impulse factor, edge factor and kurtosis factor;

[0070] Frequency domain characteristics include at least one and multiple combinations of gravity frequency, root mean square frequency, frequency standard deviation, and frequency band energy entropy;

[0071] Initialize the weights of different views, dynamically adjust the view weights for each sample based on the outlier score variance loss function, and update the weight parameters using the stochastic gradient descent algorithm.

[0072] The variance of the anomaly score reflects the degree of separation between abnormal samples and normal samples. The larger the variance, the more separated the abnormal samples are from the normal samples. It also means that when the feature is fused, the model focuses on the features that are more discriminative under the current working conditions, resulting in better detection of abnormal samples. By suppressing the interference of redundant features and strengthening the difference in anomaly representation, the accuracy and robustness of anomaly detection are significantly improved, especially in complex working conditions or gradual failure scenarios, where it shows stronger environmental adaptability.

[0073] The outlier score variance loss function is:

[0074] ;

[0075] in, Let m be the set of outlier scores containing m samples. , = The mean of the outlier scores. Let i be the anomaly score for the i-th sample. Let m be the mean of the anomaly scores of the m samples. The variance of all outlier scores in the outlier score set;

[0076] The various feature views are weighted and merged into a unified feature vector;

[0077] 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, placing them in sparse regions at a relatively shallow tree depth. To improve the adaptability of the fault detection module to different operating conditions and data distributions, this invention further proposes an adaptive isolation forest parameter adjustment mechanism, specifically:

[0078] a. Uniformly sample m samples from an input dataset containing N data points;

[0079] b. Randomly select features and randomly choose a segmentation value between their maximum and minimum values;

[0080] c recursively divides the current node sample into two subtrees, left and right, based on the segmentation value;

[0081] d. Repeat steps b and c until the preset tree height limit is reached or only one sample remains in the node;

[0082] e calculates the anomaly score based on the path length of the sample in the isolation tree and generates an anomaly detection model.

[0083] ;

[0084] in, It is the path length normalization factor for n samples. It is an estimate of the harmonic number, obtained through... estimate, It is Euler's constant, 0.5772156649. Path length is the core metric for isolated forests, defined as the number of edges traversed from the root node to the leaf node of an isolated sample x. It is the average path length of sample x across all isolated trees in the forest;

[0085] In this formula, when Approaching 0, When the value approaches 1, it indicates that the sample is highly likely to be an anomaly; when... tending to , When the value approaches 0.5, it indicates that the sample has no obvious abnormal characteristics; when... tending to , When the score approaches 0, it indicates that the sample is highly likely to be a normal sample. By calculating the anomaly score for each test data point, outliers are filtered out to construct a "candidate fault sample set" for subsequent fault diagnosis.

[0086] The vibration signal distribution of a gearbox varies significantly under different loads, speeds, and environmental conditions, and an isolated forest with fixed parameters may perform poorly under certain operating conditions. Therefore, by maximizing the variance of outlier scores, the number of binary trees t in the isolated forest is adjusted inversely, and the number of samples m when training a single isolated tree can be adjusted by the sampling frequency of the vibration signal.

[0087] The initial number of trees t is based on the amount of data in the target domain. Let N be the number of sample points. When the variance of outlier scores... When the level is below 80% of the historical best, increase the number of trees by 1.2t to improve detection sensitivity; when When the performance consistently exceeds 90% of the historical best, the number of trees is reduced by 0.9t to optimize computational efficiency. The number of training samples per tree, m, is related to the signal sampling rate. Automatic adaptation: When When, set ;when When, set ;when kHz, settings This design ensures that high-frequency signals retain sufficient detail while avoiding overfitting of low-frequency signals.

[0088] S3: Extract source domain data features through a convolutional neural network to obtain source domain feature vectors, and input them into a graph generation layer to construct instance graphs to obtain different views of vibration signals. Model topological relationships based on graph convolutional networks, construct an adjacency matrix by multiplying features by their transposes to calculate similarity between features, retain the top-k edges to enhance sparsity, embed data structure information into feature representations using the graph generation layer, and use a classifier to predict fault types to generate a source domain fault diagnosis model.

[0089] By extracting deep signal features and using structured modeling, the model's ability to judge complex operating conditions and multiple types of faults can be effectively improved. Figure 2 As shown, this invention proposes a deep diagnostic structure that combines convolutional neural networks and domain graph convolutional networks. This structure models the structural relationships between the local spatial features of the signal through domain graph learning, thereby achieving multi-level and global feature representation.

[0090] Extracting time-domain and frequency-domain features of source domain data using convolutional neural networks;

[0091] The convolutional neural network includes at least one convolutional layer, at least one pooling layer, a fully connected layer, and a softmax module, with each convolutional layer corresponding to at least one pooling layer. Through convolution and pooling operations, local feature information is extracted, resulting in a set of multi-channel feature maps. The convolutional layer filters the input feature matrix using a local receptive field mechanism, capturing subtle but crucial local patterns such as fluctuations and periodicity in the vibration signal. The pooling layer downsamples the convolutional results, enhancing the model's translation invariance and noise reduction capabilities. Finally, a set of fixed-dimensional feature vectors is output through the fully connected layer (FC layer); specifically:

[0092] Convolutional layers extract local feature information from the input feature matrix and output multi-channel feature maps;

[0093] Pooling layers are set after the corresponding convolutional layers to downsample the output of the convolutional layers.

[0094] A fully connected layer is set after the last pooling layer to output a feature vector representation.

[0095] The fully connected layer is combined with the Softmax module as a classifier to predict fault signals, and the cross-entropy loss is used to estimate the difference between the true and predicted labels.

[0096] Taking a two-layer convolutional layer and a two-layer pooling layer as an example: the convolutional neural network includes two sequentially connected convolutional-pooling units, a fully connected layer, and a Softmax module; wherein:

[0097] First convolutional layer: Extracts local features from the input signal and outputs a multi-channel feature map;

[0098] First pooling layer: Following the first convolutional layer, the feature map is downsampled using max pooling;

[0099] The second convolutional layer continues to extract higher-order features and outputs a deeper feature map.

[0100] Second pooling layer: performs secondary downsampling on the output of the second convolutional layer;

[0101] Fully connected layer: flattens the two-dimensional feature map after pooling in the second pooling layer into a one-dimensional vector, and outputs a feature vector representation of fixed length;

[0102] The fully connected layers are combined with a Softmax module as a classifier to predict fault signals, and cross-entropy loss is used to estimate the difference between the true and predicted labels. To further enhance the model's ability to model relationships between features, the convolutional neural network also includes a graph generation layer, which maps the feature matrix output by the fully connected layers into a graph structure, where:

[0103] Each feature vector is considered a node in the graph;

[0104] Edge weights are established based on the similarity between different eigenvectors, an adjacency matrix A is generated, and an instance graph is constructed. The process is described in the following reference. Figure 3 As shown, the specific manifestation is as follows:

[0105] First, the extracted feature matrix X is input into a multilayer perceptron (MLP) for nonlinear transformation. Then, the transformed feature matrix... The adjacency matrix A is obtained by multiplying it by its transpose to calculate the similarity between different eigenvectors. Finally, according to the Top-k sorting mechanism, the top k most similar neighboring nodes are selected for each node, and the others are set to zero, thus obtaining the sparse adjacency matrix. .

[0106] By performing convolution operations on instance graphs, data structure information is embedded into node features. This operation not only preserves local structural characteristics but also strengthens the semantic connections between nodes, effectively uncovering the topological association patterns hidden in the data.

[0107] The obtained node features are used for fault classification and domain adversarial training.

[0108] S4: Based on the target domain abnormal signal output from S2, the source domain fault diagnosis model is transferred to the target domain. The abnormal signal detected in the target domain is input, and adversarial training is performed using the discriminator and feature extractor. CORAL alignment further reduces inter-domain differences. The model's adaptability is optimized through the adversarial training framework. Based on the optimized model, fault type identification is performed, and the diagnostic results are output. Figure 4 As shown, the framework includes:

[0109] Feature extractor: Initialize the target domain feature extractor parameters using a pre-trained source domain feature extractor, wherein the feature extractor is the convolutional neural network described in this embodiment;

[0110] Domain discriminator: distinguishes whether features originate from the source domain or the target domain;

[0111] Related alignment module: Reduces the difference in feature distribution between the source and target domains.

[0112] The specific behavior of the domain discriminator in distinguishing whether the extracted data comes from the source domain or the target domain is as follows:

[0113] Due to the domain offset problem, classifiers trained solely on source domain data cannot effectively handle target domain data. To address this issue, a domain discriminator (D) is used to determine whether extracted features originate from the source or target domain. Furthermore, adversarial training is employed to trick the feature extractor into generating domain-invariant features. A fully connected layer combined with a Softmax module is used as the classifier for predicting fault signals, employing binary cross-entropy loss as the domain discriminant loss, expressed as:

[0114] ;

[0115] Where i represents the i-th sample, and n is the total number of samples. Let represent the label of the i-th sample, with a source domain of 1 and a target domain of 0, and The domain discriminant predicts the i-th sample.

[0116] The specific ways in which the correlation alignment module reduces the difference in feature distribution between the source and target domains are as follows:

[0117] The loss function is defined as follows: By minimizing the norm distance between the covariance matrices of the source and target domains to align the feature distribution, the loss function is:

[0118] ;

[0119] in, It is the square of the Frobenius norm. It is the covariance matrix of the source domain features. It is the covariance matrix of the target domain features. is the normalization coefficient, and d is the dimension of the feature vector;

[0120] The optimized model's final objective function for both the source and target domains is expressed as:

[0121] ;

[0122] ;

[0123] in, , To balance the hyperparameters, L S For the total loss of the source domain, For source domain fault classification loss, The loss is for anomaly detection in the target domain.

[0124] To achieve consistency in feature distribution between the source and target domains, this invention introduces a domain discriminator to determine the domain to which a sample belongs. The discriminator is trained by minimizing the discriminative loss. Adversarial training of the feature extractor is achieved by introducing gradient reversal during backpropagation. Reversing the gradient direction of the loss during the update process allows the feature extractor to maximize the discriminative loss of the domain discriminator, thereby extracting domain-insensitive features. The overall network can be optimized using stochastic gradient descent (SGD) to balance the model parameters between source domain task accuracy and domain alignment capability.

[0125] Meanwhile, a CORAL-aligned adaptive learning module is used in the fully connected layer. This module achieves feature distribution alignment between the two domains by minimizing the distance between the second-order statistics (i.e., covariance matrices) of the features in the source and target domains, thereby further reducing inter-domain differences and improving the model's transfer and generalization capabilities.

[0126] This invention provides a multi-feature view fusion method for gearbox vibration signals, which treats time-domain and frequency-domain features as two independent views and achieves adaptive feature fusion under unsupervised conditions through a dynamic weight update mechanism based on anomaly score variance.

[0127] This invention employs a two-stage approach of detection followed by diagnosis in the absence of labeled data in the target domain. It uses multi-view fusion feature input to an isolated forest to screen abnormal samples (candidate fault samples) in the target domain and filter normal signals to reduce computational load. It performs domain adaptive migration diagnosis only on abnormal samples and uses an adaptive parameter adjustment mechanism to dynamically adjust the number of trees in the isolated forest based on the ratio of the variance of the abnormal score to the historical best level. Furthermore, the number of samples per tree is configured in segments according to the signal sampling rate.

[0128] This invention proposes a feature extraction and structural modeling method for graph convolutional network modules. While extracting local features, it embeds the topological relationships between samples to enhance the discriminative power of fault features. Furthermore, by combining adversarial training between the domain discriminator and the feature extractor with the CORAL alignment mechanism, it achieves feature distribution adaptation from the source domain to the target domain, effectively improving the diagnostic stability and accuracy under complex working conditions.

[0129] It should be understood that the method proposed in this invention is applicable to gearboxes in various industrial equipment, and is not limited to those mentioned in the background art, such as wind turbine generators, automobile transmissions, rail transportation, and industrial robots.

[0130] Based on the same technical concept, embodiments of the present invention also provide 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.

[0131] Memory, used to store computer programs;

[0132] A processor, when executing a program stored in memory, implements the steps of the method of the present invention.

[0133] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0134] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0135] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0136] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0137] The computer program product of the method provided in the embodiments 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 preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0138] If the aforementioned functions are implemented as 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 this invention, essentially, or the part that contributes to the prior art, or a portion 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 several instructions to cause a computer device, such as a personal computer, server, or network device, to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0139] Unless otherwise specified, all steps in this 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 it may include steps (b) and (a) performed sequentially. For example, the mention that the method may also include step (c) indicates that step (c) may be added to the method in any order. For example, the method may include steps (a), (b), and (c), or it may include steps (a), (c), and (b), or it may include steps (c), (a), and (b), etc.

[0140] The "range" disclosed in this application is defined by a lower limit and an upper limit. A given range is defined by selecting a lower limit and an upper limit, which define the boundaries of a particular range. Ranges defined in this way can include or exclude endpoints and can be arbitrarily combined; that is, any lower limit can be combined with any upper limit to form a range. For example, if ranges of 60-120 and 80-110 are listed for a specific parameter, it is expected that ranges of 60-110 and 80-120 are also included. Furthermore, if minimum range values ​​of 1 and 2 are listed, and if maximum range values ​​of 3, 4, and 5 are listed, then the following ranges are all expected: 1-3, 1-4, 1-5, 2-3, 2-4, and 2-5. In this application, unless otherwise stated, the numerical range "ab" represents a shortened representation of any combination of real numbers between a and b, where a and b are real numbers. For example, the numerical range "0-5" indicates that all real numbers between "0-5" have been listed in this article; "0-5" is simply a shortened representation of these numerical combinations. Furthermore, when a parameter is stated as an integer ≥2, it is equivalent to disclosing that the parameter is, for example, an integer such as 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, etc.

[0141] Unless otherwise specified, the terms "comprising" and "including" as used in this application are open-ended. For example, "comprising" and "including" may mean including or containing other components not listed.

[0142] Unless otherwise specified, all embodiments and optional embodiments of this application can be combined to form new technical solutions.

[0143] Unless otherwise specified, all technical features and optional technical features of this application may be combined to form new technical solutions.

[0144] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. 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 foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting and diagnosing gearbox faults, characterized in that, Includes the following steps: S1: Obtain the target domain raw signal and the source domain raw signal data with labels from the source domain database of the vibration signal throughout the entire life cycle of the gearbox, and preprocess them; S2: Extract the time domain and frequency domain features of the target domain data, perform weighted fusion of the features to generate the target domain fusion feature vector, input it into the adaptive isolated forest for unsupervised anomaly detection, and generate the isolated forest anomaly detection model; S3: Extract source domain data features through a convolutional neural network to obtain source domain feature vectors, input them into a graph generation layer to construct an instance graph, model topological relationships based on graph convolutional networks, and use a classifier to predict fault types 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 transferred to the target domain. The model adaptability is optimized through an adversarial training framework. Based on the optimized model, fault type identification is performed, and the diagnosis result is output. The framework includes: Feature extractor: Initialize the target domain feature extractor parameters using a pre-trained source domain feature extractor; Domain discriminator: distinguishes whether features originate from the source domain or the target domain; Related alignment module: Reduces the difference in feature distribution between the source and target domains.

2. The 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 original signal data in the target domain and the original signal data in the labeled source domain.

3. The gearbox fault detection and diagnosis method according to claim 1, characterized in that, The method described in S2 for extracting time-domain and frequency-domain features of the target domain data and then weighting and fusing these features to generate a target domain fused 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; Initialize the weights of different views, dynamically adjust the view weights based on the outlier score variance loss function, and update the weight parameters using the stochastic gradient descent algorithm; The various feature views are weighted and merged into a unified feature vector.

4. The gearbox fault detection and diagnosis method according to claim 1, characterized in that, The method for unsupervised anomaly detection using adaptive isolated forests described in S2 is as follows: a. Uniformly sample m samples from an input dataset containing N data points; b. Randomly select features and randomly choose a segmentation value between their maximum and minimum values; c recursively divides the current node sample into two subtrees, left and right, based on the segmentation value; d. Repeat steps b and c until the preset tree height limit is reached or only one sample remains in the node; e calculates the anomaly score based on the path length of the sample in the isolation tree and generates 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 using a convolutional neural network. The convolutional neural network includes at least one convolutional layer, at least one pooling layer, a fully connected layer, and a Softmax module, with at least one convolutional layer corresponding to at least one pooling layer; in: Convolutional layers extract local feature information from the input feature matrix and output multi-channel feature maps; Pooling layers are set after the corresponding convolutional layers to downsample the output of the convolutional layers. A fully connected layer is set after the last pooling layer to output a feature vector representation. The fully connected layer is combined with the Softmax module as a classifier to predict fault signals, and the cross-entropy loss is used to estimate the difference between the true and predicted labels.

6. The gearbox fault detection and diagnosis method according to claim 5, characterized in that, Convolutional neural networks also include graph generation layers, which map the feature matrices output by the fully connected layers into a graph structure, where: Each feature vector is considered a node in the graph; Edge weights are established based on the similarity between different feature vectors, an adjacency matrix A is generated, and an instance graph is constructed. Convolution operations are performed 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. The 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 using the following method: Using binary cross-entropy loss as the domain discrimination loss, it is expressed as: ； in, Indicates the first One sample, It is the total number of samples. Indicates the first The labels of each sample, with the source domain being... The target domain is 0, and The domain discriminator represents the first domain. The prediction results for each sample.

8. A gearbox fault detection and diagnosis method according to claim 7, characterized in that, The method by which the correlation alignment module reduces the difference in feature distribution between the source and target domains is: The loss function is defined as follows: By minimizing the norm distance between the covariance matrices of the source and target domains to align the feature distribution, the loss function is: ； in, yes Norm square, It is the covariance matrix of the source domain features. It is the covariance matrix of the target domain features. It is the normalization coefficient; The optimized model's final objective function for both the source and target domains is expressed as: ； in, To balance hyperparameters, For source domain fault classification loss, The loss is for anomaly detection in the target domain.

9. An electronic device, characterized in that, The device 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; and the processor, when executing the program stored in the memory, implements the gearbox fault detection and diagnosis method according to any one of claims 1-8.

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

Images