Dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern image

Abstract

The application relates to the technical field of fault diagnosis, and discloses a double-flow confrontation cross-device fault diagnosis method based on a symmetric point mode image, which comprises the following steps: converting a vibration signal of a device to be diagnosed into a two-dimensional image sample based on a symmetric point mode; constructing a double-flow feature extraction network, wherein the double-flow feature extraction network comprises a first feature extraction unit for extracting shallow texture features and a second feature extraction unit for extracting deep global features; fusing the shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit to obtain fusion feature information; generating domain-invariant features with consistent distribution between a source domain device and a target device through domain confrontation training based on the fusion feature information; and inputting the domain-invariant features into a fault classification network to obtain a fault diagnosis result. The application effectively improves the accuracy of cross-device fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and more specifically, to a dual-stream countermeasures cross-device fault diagnosis method based on symmetrical point pattern images. Background Technology

[0002] Existing fault diagnosis technologies in mechanical equipment health management generally face a problem: in practical applications, they need to process data from multiple devices simultaneously, which often vary greatly in model, operating conditions, and field environment. The vibration signals collected by different devices may have sampling rates ranging from several kilohertz to tens of kilohertz, and the amplitude range also varies depending on sensor sensitivity and installation location. The signal forms themselves are also diverse, including periodic signals from stable operation, signals during variable speed transitions, and fault signals containing sudden impacts. Furthermore, it is difficult to completely standardize the data acquisition card model, sensor mounting method, and duration of each sampling in the field.

[0003] Currently, most mainstream fault diagnosis methods rely on traditional machine learning models or deep learning networks based on a single data source. These typically require manually designed features or model training and validation using labeled data specific to the device. However, this approach is prone to problems such as feature distribution shifts, insufficient model generalization ability, and decreased fault identification accuracy in cross-device diagnostic scenarios. When there are significant differences in operating conditions or data distribution between the target and source devices, model performance often degrades significantly, affecting the detection and identification of key fault types such as bearing failures, gear wear, and rotor imbalance. Furthermore, existing systems lack automatic adaptation mechanisms to differences in data distribution across different domains during model training, often requiring manual re-labeling of target domain data or adjustment of the model structure to maintain diagnostic accuracy.

[0004] Therefore, it is necessary to provide a dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images, which aims to solve the problems of feature distribution shift, insufficient domain adaptability, poor model generalization performance, and lack of adaptive transfer capability to data distribution differences under different operating conditions in current fault diagnosis in cross-device scenarios. Summary of the Invention

[0005] In view of this, the present invention proposes a dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images, aiming to solve the problems of existing fault diagnosis technologies in cross-device scenarios, such as feature distribution shift, insufficient domain adaptability, poor model generalization performance, and lack of adaptive transfer capability to data distribution differences under different operating conditions.

[0006] This invention proposes a dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images, comprising the following steps: S100: Based on the symmetric point pattern, the vibration signal of the device to be diagnosed is converted into a two-dimensional image sample. S200. Construct a dual-stream feature extraction network, the dual-stream feature extraction network including a first feature extraction unit and a second feature extraction unit; wherein, the first feature extraction unit is used to extract the shallow texture features of the two-dimensional image sample, and the second feature extraction unit is used to extract the deep global features of the two-dimensional image sample; S300: The shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit are fused to obtain fused feature information; S400. Based on the fused feature information, domain-invariant features with consistent distribution between the source and target devices are generated through domain adversarial training. S500. Input the domain-invariant features into the fault classification network to obtain the fault diagnosis results.

[0007] Furthermore, the specific steps of converting the vibration signal of the device to be re-diagnosed into a two-dimensional image sample based on the symmetric point pattern are as follows: The vibration signal of the device to be diagnosed is normalized to obtain a standard vibration signal sequence; Map each data point in the standard vibration signal sequence to a two-dimensional Cartesian coordinate system, with the time series index set as the horizontal axis and the signal amplitude set as the vertical axis, to generate an initial vibration waveform. Using the horizontal axis of the two-dimensional Cartesian coordinate system as the axis of symmetry, the initial vibration waveform is mirrored to generate a symmetric point pattern diagram containing a symmetric structure. The symmetrical point pattern diagram is converted into an image format to obtain the two-dimensional image sample.

[0008] Furthermore, the first feature extraction unit includes: The first convolutional layer is used to extract edge texture information from the two-dimensional image sample; The first pooling layer is used to reduce the feature dimension.

[0009] Furthermore, the second feature extraction unit includes: The second convolutional layer is used to perform preliminary feature extraction on the two-dimensional image samples to obtain intermediate feature information; The third convolutional layer is used to extract deep features from the intermediate feature information to obtain a deep feature map; The second pooling layer is used to reduce the dimensionality of the deep feature map and extract key feature information; The second convolutional layer, the third convolutional layer, and the second pooling layer are connected in sequence.

[0010] Furthermore, the fusion of the shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit is specifically as follows: The shallow texture features output by the first feature extraction unit are converted into a one-dimensional vector form to obtain the first feature vector; The deep global features output by the second feature extraction unit are converted into a one-dimensional vector form to obtain the second feature vector; The first feature vector and the second feature vector are concatenated to generate a fused feature vector as the fused feature information.

[0011] Furthermore, the specific steps of generating domain-invariant features through domain adversarial training are as follows: Construct a domain discriminator to determine whether the input feature information belongs to the source domain device or the target device; A gradient inversion layer is set between the fused feature information and the neighborhood discriminator; During model training, the gradient reversal layer adjusts the update direction of network parameters, so that when the fused feature information is fed into the domain discriminator, it can confuse the judgment result of the domain discriminator, thereby generating domain-invariant features that are consistently distributed between the source and target devices.

[0012] Furthermore, the domain-invariant features are input into the fault classification network to obtain the fault diagnosis results as follows: Construct a fault classifier, which includes a fully connected layer and an output layer; The domain-invariant features are input into the fully connected layer for mapping processing to obtain the vector to be classified; The output layer calculates the probability value of the vector to be classified belonging to each preset fault category, and selects the fault category with the highest probability value as the final fault diagnosis result.

[0013] Furthermore, it also includes: when the fault diagnosis result is a fault state, transmitting the fault diagnosis result to the display terminal.

[0014] Compared with existing technologies, the advantages of this invention are as follows: By converting the vibration signal of the device to be diagnosed into a two-dimensional image sample through a symmetrical point pattern, the temporal characteristics of the signal can be preserved and the visualization of fault modes can be enhanced. The resulting benefit is that mapping a one-dimensional vibration signal into a two-dimensional image with a symmetrical structure makes fault modes easier for deep networks to identify in the image domain, while improving the accuracy of fault characterization. Secondly, a dual-stream feature extraction network is constructed to extract shallow texture features and deep global features from the two-dimensional image sample and fuse them to obtain fused feature information that combines local details and global semantics. The resulting benefit is that shallow texture features can capture local patterns such as edges, corners, and textures formed by fault impacts in the image, while deep global features can perceive the overall distribution of the fault in the time-frequency domain. The fusion of these two features provides a more comprehensive fault feature representation, thereby improving the identification accuracy of fault types such as bearing faults and gear wear, and reducing the risk of information loss under complex operating conditions with a single feature extraction method. Furthermore, based on fused feature information, domain-invariant features with consistent distribution across source and target devices are generated through domain adversarial training. The beneficial effects are that the collaborative mechanism between the domain discriminator and the gradient inversion layer effectively obfuscates feature distribution differences between device domains, enabling the model to automatically suppress device-type-related domain offset information during training. This allows the model to learn fault feature representations that are universal across different devices. This mechanism improves the model's generalization ability and transfer performance across device scenarios without requiring a large amount of labeled target domain data, overcoming the diagnostic performance degradation caused by differences in data distribution across different models and operating conditions. Finally, the domain-invariant features are input into the fault classification network to output fault diagnosis results, achieving end-to-end automatic fault diagnosis across devices. The beneficial effects are that the entire diagnostic process requires no manual intervention, can adaptively handle vibration signals from different sources, and maintains high diagnostic accuracy and stability even when device operating conditions change.

[0015] On the other hand, the present invention also provides a dual-stream adversarial cross-device fault diagnosis system based on symmetrical point pattern images, comprising: The data preprocessing module is used to acquire the vibration signal of the device to be diagnosed and convert the vibration signal into a symmetrical point pattern image; The dual-stream feature extraction module is used to input the symmetric point pattern image into the dual-stream feature extraction network to extract shallow texture features and deep global features respectively; The feature fusion module is used to fuse shallow texture features and deep global features to obtain fused feature information; The domain adversarial training module is used to perform domain adversarial training based on fused feature information and generate domain-invariant features; The fault classification module is used to input domain-invariant features into the fault classification network and output fault diagnosis results. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a dual-stream, cross-device fault diagnosis method based on symmetric point pattern images provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images provided in an embodiment of the present invention. Figure 3 A schematic diagram of the original vibration signal for the dual-stream anti-cross-device fault diagnosis method based on symmetrical point pattern images provided in an embodiment of the present invention; Figure 4 A schematic diagram of a normalized signal for a dual-stream anti-cross-device fault diagnosis method based on symmetric point pattern images provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the initial waveform of the dual-stream anti-cross-device fault diagnosis method based on symmetric point pattern images provided in an embodiment of the present invention. Figure 6 A schematic diagram of a symmetric point pattern image for a dual-stream anti-cross-device fault diagnosis method based on symmetric point pattern images provided in an embodiment of the present invention; Figure 7 This is a functional block diagram of a dual-stream anti-cross-device fault diagnosis system based on symmetric point pattern images, provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey its scope to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] like Figures 1-6 As shown in some embodiments of this application, this embodiment provides a dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images, including the following steps: Step S100: Convert the vibration signal of the device to be diagnosed into a two-dimensional image sample based on the symmetric point pattern.

[0019] Specifically, the process of converting the vibration signal of the device to be diagnosed into a two-dimensional image sample based on the symmetric point pattern involves: normalizing the vibration signal of the device to be diagnosed to obtain a standard vibration signal sequence. The normalization process uses a minimum-maximum scaling method, and its calculation formula is as follows: in, The original vibration signal sequence, For the first The amplitude of each sampling point The normalized standard amplitude is obtained by linearly mapping all amplitudes to the interval [0,1].

[0020] Map each data point in the standard vibration signal sequence to a two-dimensional Cartesian coordinate system, with the time series index set as the horizontal axis and the signal amplitude set as the vertical axis, to generate an initial vibration waveform. This mapping relationship can be represented as a sequence of coordinate points. , where the x-coordinate =i(1,2,...N), where N is the signal length, and the y-axis is... = This forms an initial discrete waveform point set; using the horizontal axis of the two-dimensional Cartesian coordinate system as the axis of symmetry, the initial vibration waveform is mirrored to generate a symmetrical point pattern diagram containing a symmetrical structure; the symmetrical point pattern diagram is then converted into an image format to obtain the two-dimensional image sample.

[0021] Understandably, converting one-dimensional vibration signals into two-dimensional image samples using symmetric point patterns can preserve the temporal characteristics of the original vibration signals and enhance the visualization of fault modes. In practical cross-device fault diagnosis scenarios, vibration signals collected by different devices often differ significantly in sampling frequency, amplitude range, and noise level. If the original signal is directly input into the diagnostic model, it is easily affected by factors such as inconsistent signal lengths, different amplitude scales, and noise interference, reducing the model's diagnostic accuracy and generalization ability. Specifically, normalization processing can unify the vibration signals from different devices to the same amplitude scale range, eliminating the interference of dimensional differences on subsequent feature extraction. The normalized signal sequence is mapped to a two-dimensional Cartesian coordinate system, with the time series index as the x-axis and the signal amplitude as the y-axis to generate an initial vibration waveform. This preserves the time-domain waveform characteristics of the signal while providing a foundation for subsequent symmetric transformations. A mirror mapping process is then performed with the x-axis as the axis of symmetry. The resulting symmetric point pattern retains the positive half-cycle information of the original waveform and supplements the mirror information of the negative half-cycle, making the periodic impact features and harmonic components form a more significant and regular texture pattern in the image. Finally, the symmetric point pattern is converted into a standard two-dimensional image format, such as a grayscale image or a red-green-blue (RGB) image, facilitating feature extraction by subsequent deep convolutional networks. Through this method, the original one-dimensional vibration signal is transformed into a two-dimensional image sample with rich texture information, enhancing and highlighting the subtle fault impact features that are difficult to detect in the time domain in the image domain, providing high-quality input for subsequent dual-stream feature extraction.

[0022] In a specific embodiment of this application, the above steps are implemented as follows: First, vibration signals are collected from an accelerometer mounted on a rotating machinery bearing housing at a sampling frequency of 12kHz, with each collection lasting 1 second, resulting in a one-dimensional vibration signal sequence containing 12,000 data points. This signal may include vibration data under various operating conditions, such as normal operation, inner race fault, outer race fault, and rolling element fault. Second, the collected raw vibration signals are normalized, mapping the amplitude range to the (0,1) interval to obtain a standard vibration signal sequence, thereby eliminating the influence of amplitude scale differences under different operating conditions. Then, each data point in the standard vibration signal sequence is sequentially mapped to a two-dimensional Cartesian coordinate system. The initial vibration waveform is plotted using the time index of the data points (i.e., the 1st point, 2nd point, ..., the 12000th point) as the x-axis and the normalized amplitude as the y-axis. Then, the initial vibration waveform is mirrored along the x-axis to generate a symmetrical point pattern diagram containing positive and negative symmetry. In the symmetrical point pattern diagram, the original waveform is above the x-axis, and its mirrored waveform is below the x-axis, together forming a complete pattern symmetrical about the x-axis. Finally, the generated symmetrical point pattern diagram is converted into a 224×224 pixel grayscale image as input samples for the subsequent dual-stream feature extraction network. Through this conversion, the periodic impact caused by bearing failure is represented in the symmetrical point pattern diagram as regularly textured spots with symmetrical top and bottom edges. Compared to the weak impact pulses in the original one-dimensional signal, this provides stronger visualization and recognizability in the image domain.

[0023] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0024] Step S200: Construct a dual-stream feature extraction network, which includes a first feature extraction unit and a second feature extraction unit; wherein, the first feature extraction unit is used to extract the shallow texture features of the two-dimensional image sample, and the second feature extraction unit is used to extract the deep global features of the two-dimensional image sample.

[0025] Specifically, the first feature extraction unit includes: a first convolutional layer for extracting edge texture information from the two-dimensional image sample; and a first pooling layer for reducing the feature dimension. The second feature extraction unit includes: a second convolutional layer for performing preliminary feature extraction on the two-dimensional image sample to obtain intermediate feature information; a third convolutional layer for performing deep feature extraction on the intermediate feature information to obtain a deep feature map; and a second pooling layer for performing dimensionality reduction and compression on the deep feature map to extract key feature information; wherein the second convolutional layer, the third convolutional layer, and the second pooling layer are connected sequentially.

[0026] Understandably, by constructing a dual-stream feature extraction network, shallow texture features and deep global features of two-dimensional image samples are extracted separately, thereby achieving multi-scale and multi-level representation of fault image information. In practical applications of fault diagnosis, different types of faults exhibit different feature morphologies in images: local pitting faults in bearings may manifest as local edge and corner texture changes in images, while faults such as gear wear or rotor imbalance may show differences in the overall distribution and global morphology of the image. If only a single feature extraction path is used, it is difficult to effectively capture both local details and global semantic information simultaneously. Specifically, in the first feature extraction unit, the first convolutional layer uses a small convolutional kernel, which can sensitively capture shallow local patterns such as edges, corners, and textures formed by fault impacts in the image. These patterns are of great significance for distinguishing the microscopic differences of different fault types in the image domain; the first pooling layer, such as max pooling or average pooling, downsamples the feature map, which can reduce the feature dimension, reduce the amount of computation, and enhance the translation invariance and robustness to local perturbations of the features. In the second feature extraction unit, the second convolutional layer uses a larger kernel or fewer channels to perform preliminary feature extraction on the input image, obtaining intermediate features containing basic contour information. The third convolutional layer performs depthwise convolution on the intermediate features to further expand the receptive field and extract higher-level abstract semantic information, resulting in a deep feature map containing global distribution information. The second pooling layer performs global pooling or adaptive pooling on the deep feature map, compressing the spatial dimension of the feature map and extracting the most discriminative global statistical features. Through this dual-stream parallel feature extraction architecture, complementary information of local details and global semantics can be obtained simultaneously at different depths of the network, providing richer discriminative basis for subsequent feature fusion.

[0027] In a specific embodiment of this application, the above steps are implemented as follows: After obtaining the symmetric point pattern image, it is input into the constructed dual-stream feature extraction network. The first feature extraction unit adopts a lightweight convolutional structure, including a first convolutional layer with a kernel size of 3×3, a stride of 1, and 32 output channels, used to extract the edge texture information of the image; then a max pooling layer with a kernel size of 2×2 and a stride of 2 is connected as the first pooling layer, which halves the spatial size of the feature map to obtain a shallow texture feature map. Meanwhile, the second feature extraction unit employs a deeper convolutional network, including a second convolutional layer with a kernel size of 5×5, a stride of 1, and 64 output channels. This second layer performs preliminary feature extraction on the input image and retains more spatial structural information. Subsequently, a third convolutional layer with a kernel size of 3×3, a stride of 1, and 128 output channels is connected to further extract deeper abstract features based on the preliminary features, expanding the receptive field to the global range. Finally, a global average pooling layer is connected as the second pooling layer, compressing the spatial dimension of the deep feature map to 1×1, resulting in a deep global feature vector. Through this structure, the shallow texture features output by the first feature extraction unit retain the local edges and texture details formed by the fault impact in the image, such as the regularly spaced texture spots presented in the symmetric point pattern map of a bearing outer ring fault. The deep global features output by the second feature extraction unit encode the distribution information of the entire image at the macroscopic level, such as the overall texture density and symmetry changes in the symmetric point pattern map under different fault states. The two types of features complement each other, providing multi-dimensional information support for subsequent fusion and diagnosis.

[0028] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0029] Step S300: The shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit are fused to obtain fused feature information.

[0030] Specifically, fusing the shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit involves: converting the shallow texture features output by the first feature extraction unit into a one-dimensional vector to obtain a first feature vector; converting the deep global features output by the second feature extraction unit into a one-dimensional vector to obtain a second feature vector; and concatenating the first feature vector and the second feature vector to generate a fused feature vector as the fused feature information. The concatenation operation is expressed as follows: in, Let be the first feature vector, with dimension d1; The second feature vector has a dimension of d2; the concatenated and merged feature vector The dimension is d1+d2, which realizes the merging of local details and global information.

[0031] It is understandable that fusing shallow texture features and deep global features can organically combine local detail information with global semantic information, forming a more discriminative and robust comprehensive feature representation. In practical fault diagnosis scenarios, relying solely on local texture features is easily affected by noise interference and local anomalies, leading to unstable diagnostic results; while relying solely on global features may lose local pattern information sensitive to subtle faults, resulting in missed diagnoses of early, weak faults. Therefore, effectively fusing the two types of features can balance local sensitivity and global stability, improving the accuracy and robustness of fault diagnosis. Specifically, since the shallow texture features output by the first feature extraction unit are usually in the form of multi-dimensional feature maps, while the deep global features output by the second feature extraction unit may be in the form of vectors or feature maps, their data dimensions and structural forms are not entirely the same. Therefore, they need to be uniformly converted into a one-dimensional vector form first. The conversion method can use flattening operations, global pooling, or adaptive pooling, etc., to compress the spatial dimension while retaining the channel dimension information. After obtaining the first and second feature vectors in a unified format, they are concatenated along the feature dimension to generate a fused feature vector containing both types of feature information. Compared with weighted summation or other nonlinear fusion methods, the concatenation operation has the advantages of being simple to implement, lossless in information, and capable of end-to-end training, and can retain the original feature information to the maximum extent for subsequent domain adversarial and classification tasks.

[0032] In a specific embodiment of this application, the above steps are implemented as follows: After extracting shallow texture features through the first feature extraction unit, assuming the feature map is a three-dimensional tensor with shape (height h=16, width d=16, number of channels c=32), it is flattened into a one-dimensional vector to obtain a first feature vector with dimension 16×16×32=8192. Simultaneously, the deep global features output by the second feature extraction unit are assumed to be a feature map with shape (height h=1, width d=1, number of channels c=128), and a second feature vector with dimension 128 is obtained after global average pooling. Subsequently, the first and second feature vectors are concatenated along the feature dimensions to generate a fused feature vector with dimension 8192+128=8320, which serves as the input to the subsequent domain adversarial training module. Through the above concatenation operation, the fused feature vector contains both rich detailed information about the local fault impact patterns in the shallow texture features and comprehensive semantic information about the overall image distribution state in the deep global features. For example, for a symmetrical point pattern image of a bearing inner ring fault, shallow texture features may capture the density of local texture spots formed by periodic impacts, while deep global features reflect the overall offset of the entire image in terms of symmetry. The fused feature formed by stitching the two together can comprehensively describe the fault mode from different granularities, thus providing a more complete basis for subsequent domain-invariant feature learning and fault classification.

[0033] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0034] Step S400: Based on the fused feature information, generate domain-invariant features with consistent distribution between the source and target devices through domain adversarial training.

[0035] Specifically, the generation of domain-invariant features through domain adversarial training involves: constructing a domain discriminator to determine whether the input feature information belongs to the source domain device or the target device; setting a gradient reversal layer between the fused feature information and the domain discriminator; and performing an identity mapping during forward propagation in the gradient reversal layer. During backpropagation, the gradient of the loss with respect to the input is multiplied by a negative constant. ,Right now: in, It is the inversion constant. The partial differential symbol is used to represent the partial derivative of a function. This is the loss of the domain discriminator. This mechanism causes the update direction of the feature extraction network to be opposite to that of the domain discriminator, thus forcing the network to learn domain-invariant features.

[0036] During model training, the gradient reversal layer adjusts the update direction of network parameters, so that when the fused feature information is fed into the domain discriminator, it can confuse the judgment result of the domain discriminator, thereby generating domain-invariant features that are consistently distributed between the source and target devices.

[0037] Understandably, generating domain-invariant features through domain adversarial training can effectively overcome the feature distribution shift problem caused by differences in operating conditions across devices, enabling the model to have good generalization ability across different devices. In actual cross-device fault diagnosis scenarios, there are often significant differences in data distribution between source domain devices, such as fault simulation benches in a laboratory environment, and target devices, such as actual industrial equipment operating in the field. These differences include different sampling frequencies, sensor models, load conditions, and background noise. If the model trained directly on the source domain data is applied to the target domain, the diagnostic accuracy often drops significantly. The core idea of ​​domain adversarial training is to introduce a domain discriminator and have it engage in adversarial play against a feature extraction network. The feature extraction network attempts to extract features that can deceive the domain discriminator, making it unable to distinguish whether the feature comes from the source domain or the target domain, thereby forcing the feature extraction network to learn feature representations that are universal and invariant across different device domains. Specifically, in the implementation process, a domain discriminator network is first constructed, whose input is a fused feature vector, and whose output is the probability that the feature belongs to a source domain device or a target domain device. The neighborhood discriminator can employ a fully connected neural network structure, containing several hidden layers and a final binary classification output layer. Secondly, a gradient inversion layer is set before the fused feature information is input into the neighborhood discriminator. The gradient inversion layer behaves as an identity mapping during forward propagation, meaning it does not modify the features, but during backpropagation, it automatically multiplies the gradient by a negative constant, typically 1 / 2. This reverses the gradient direction. Thus, during model training, the parameter updates of the feature extraction network are guided towards increasing the loss of the neighborhood discriminator; that is, the features learned by the feature extraction network make it difficult for the neighborhood discriminator to correctly identify the domain label. Meanwhile, the neighborhood discriminator's own parameter updates are directed towards reducing its own loss, i.e., striving to improve the accuracy of its domain label identification. Through this adversarial game, the feature extraction network ultimately generates domain-invariant features that tend to be uniformly distributed between the source and target domain devices.

[0038] In a specific embodiment of this application, the above steps are implemented as follows: After constructing the dual-stream feature extraction network and obtaining the fused feature vector, a neighborhood discriminator is constructed to distinguish the feature source domain. This neighborhood discriminator employs a three-layer fully connected network structure with 512, 128, and 1 hidden layer neurons, respectively. The last layer uses a sigmoid activation function to output a probability value between 0 and 1, representing the probability that the input feature belongs to the target domain. Before inputting the fused feature vector into the neighborhood discriminator, a gradient inversion layer (GRL) is set, with its inversion coefficient... The value is set to 0.1. During model training, samples from source domain devices (such as laboratory bearing fault simulation benches) and target domain devices (such as actual wind turbine bearings in operation) are input into the network. The feature extraction network first extracts the fused features of each sample, and then inputs them into the neighborhood discriminator via a gradient inversion layer. During training, the neighborhood discriminator continuously updates its parameters to distinguish features between the two domains, while the feature extraction network, after receiving the backpropagation gradient through the gradient inversion layer, updates its parameters in the direction of increasing the loss of the neighborhood discriminator, making it increasingly difficult for the discriminator to distinguish the source of its extracted features. After multiple rounds of adversarial training, the distribution difference of the fused features extracted by the feature extraction network between the source and target domains is effectively suppressed. For example, the fault features extracted from the source domain device and the same fault features extracted from the target domain device are far apart in the feature space. After adversarial training, the feature distribution centers of the two tend to align, thus generating a set of domain-invariant features with consistent distribution between the source and target domains.

[0039] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0040] Step S500: Input the domain-invariant features into the fault classification network to obtain the fault diagnosis results.

[0041] Specifically, inputting the domain-invariant features into the fault classification network to obtain the fault diagnosis result involves: constructing a fault classifier, which includes a fully connected layer and an output layer; inputting the domain-invariant features into the fully connected layer for mapping processing to obtain a vector to be classified; calculating the probability value of the vector to be classified belonging to each preset fault category through the output layer, and selecting the fault category with the highest probability value as the final fault diagnosis result.

[0042] Understandably, inputting domain-invariant features into a fault classification network enables end-to-end automatic fault diagnosis across devices. The task of the fault classification network is to establish a mapping from the feature space to fault categories based on the learned domain-invariant features. First, a fault classifier is constructed, typically consisting of one or more fully connected layers and an output layer. The fully connected layers perform a non-linear mapping on the domain-invariant features, projecting them from the feature space to a high-dimensional space suitable for classification, and extracting the correlation weights between each dimension of the features and the fault category. The output layer uses a normalized exponential function (Softmax) activation function to transform the vector to be classified output by the fully connected layer into a probability distribution for each preset fault category. In practical applications, the preset fault categories can be set according to specific diagnostic needs; for example, for bearing fault diagnosis, they can be set to categories such as normal, inner race fault, outer race fault, rolling element fault, and cage fault. Finally, the fault category with the highest probability value is selected as the final fault diagnosis result output.

[0043] In a specific embodiment of this application, the above steps are implemented as follows: After obtaining the domain-invariant features, they are input into a fault classification network. This fault classification network contains a fully connected layer and a Softmax output layer. The fully connected layer has 64 neurons and uses the Modified Line Unit (ReLU) activation function to further map the domain-invariant features into vectors to be classified. The number of neurons in the Softmax output layer is the same as the number of fault categories. For example, for bearing fault diagnosis, it is set to 5 neurons, corresponding to five categories: normal, inner race fault, outer race fault, rolling element fault, and cage fault. During the model training phase, the fault classifier is supervised and trained using the cross-entropy loss function with labeled sample data from the source domain device to correctly identify each type of fault. During the testing phase, i.e., the cross-device diagnosis phase, the vibration signals collected from the target domain device are processed sequentially through steps S100 to S400 to obtain the corresponding domain-invariant features, which are then input into the trained fault classification network. The Softmax output layer calculates the probability value of the feature belonging to each fault category. For example, the output results are as follows: normal: 0.1, inner ring fault: 0.8, outer ring fault: 0.05, rolling element fault: 0.02, cage fault: 0.03. Among them, the probability value of inner ring fault is the largest at 0.8. Therefore, the final fault diagnosis result output by the network is inner ring fault.

[0044] Furthermore, in some preferred embodiments of this application, the method further includes: when the fault diagnosis result is a fault state, transmitting the fault diagnosis result to a display terminal.

[0045] Understandably, when diagnostic results indicate that equipment is in a faulty state, timely transmission of these results to the display terminal helps on-site maintenance personnel obtain the equipment's health status immediately, enabling them to take timely repair or shutdown measures and prevent more serious equipment damage or production accidents caused by the malfunction worsening. In practice, the display terminal can be an on-site industrial control computer screen, a mobile terminal such as a mobile phone or tablet, or a large display screen in a remote monitoring center. Transmission methods can utilize wired networks such as industrial Ethernet or wireless networks such as WiFi to ensure real-time push and visual presentation of diagnostic results.

[0046] The above scenarios are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0047] In the above embodiments, the vibration signal of the device to be diagnosed is converted into a two-dimensional image sample by using a symmetrical point pattern. This preserves the temporal characteristics of the signal and enhances the visualization of fault modes. The beneficial effect is that mapping a one-dimensional vibration signal into a two-dimensional image with a symmetrical structure makes the fault mode easier for deep networks to identify in the image domain, while improving the accuracy of fault characterization. Secondly, a dual-stream feature extraction network is constructed to extract shallow texture features and deep global features from the two-dimensional image sample and fuse them to obtain fused feature information that combines local details and global semantics. The beneficial effect is that shallow texture features can capture local patterns such as edges, corners, and textures formed by fault impacts in the image, while deep global features can perceive the overall distribution of the fault in the time-frequency domain. The fusion of the two provides a more comprehensive fault feature representation, thereby improving the identification accuracy of fault types such as bearing faults and gear wear, and reducing the risk of information loss under complex working conditions by a single feature extraction method. Furthermore, based on fused feature information, domain-invariant features with consistent distribution across source and target devices are generated through domain adversarial training. The beneficial effects are that the collaborative mechanism between the domain discriminator and the gradient inversion layer effectively obfuscates feature distribution differences between device domains, enabling the model to automatically suppress device-type-related domain offset information during training. This allows the model to learn fault feature representations that are universal across different devices. This mechanism improves the model's generalization ability and transfer performance across device scenarios without requiring a large amount of labeled target domain data, overcoming the diagnostic performance degradation caused by differences in data distribution across different models and operating conditions. Finally, the domain-invariant features are input into the fault classification network to output fault diagnosis results, achieving end-to-end automatic fault diagnosis across devices. The beneficial effects are that the entire diagnostic process requires no manual intervention, can adaptively handle vibration signals from different sources, and maintains high diagnostic accuracy and stability even when device operating conditions change.

[0048] like Figure 7 As shown, this embodiment provides a dual-stream adversarial cross-device fault diagnosis system based on symmetric point pattern images, including: a data preprocessing module, a dual-stream feature extraction module, a feature fusion module, a domain adversarial training module, and a fault classification module.

[0049] Specifically, the data preprocessing module acquires the vibration signal of the device to be diagnosed and converts it into a symmetrical point pattern image; the dual-stream feature extraction module inputs the symmetrical point pattern image into a dual-stream feature extraction network to extract shallow texture features and deep global features respectively; the feature fusion module fuses the shallow texture features and the deep global features to obtain fused feature information; the domain adversarial training module performs domain adversarial training based on the fused feature information to generate domain-invariant features; and the fault classification module inputs the domain-invariant features into a fault classification network and outputs the fault diagnosis result. All modules are connected sequentially.

[0050] It is understood that the dual-stream anti-cross-device fault diagnosis method and system based on symmetric point pattern images in the above embodiments have the same beneficial effects, and will not be described in detail here.

[0051] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0052] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dual-stream, cross-device fault diagnosis method based on symmetrical point pattern images, characterized in that, Includes the following steps: Based on the symmetric point pattern, the vibration signal of the device to be diagnosed is converted into a two-dimensional image sample; A dual-stream feature extraction network is constructed, comprising a first feature extraction unit and a second feature extraction unit; wherein the first feature extraction unit is used to extract shallow texture features of the two-dimensional image sample, and the second feature extraction unit is used to extract deep global features of the two-dimensional image sample. The shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit are fused to obtain fused feature information. Based on the fused feature information, domain-invariant features with consistent distribution between the source and target devices are generated through domain adversarial training. The domain-invariant features are input into the fault classification network to obtain the fault diagnosis results.

2. The dual-stream cross-device fault diagnosis method based on symmetrical point pattern images as described in claim 1, characterized in that, The specific steps for converting the vibration signal of the device to be diagnosed into a two-dimensional image sample based on the symmetric point pattern are as follows: The vibration signal of the device to be diagnosed is normalized to obtain a standard vibration signal sequence; Map each data point in the standard vibration signal sequence to a two-dimensional Cartesian coordinate system, set the time series index as the horizontal axis and the signal amplitude as the vertical axis to generate an initial vibration waveform. Using the horizontal axis of the two-dimensional Cartesian coordinate system as the axis of symmetry, the initial vibration waveform is mirrored to generate a symmetric point pattern diagram containing a symmetric structure. The symmetrical point pattern diagram is converted into an image format to obtain the two-dimensional image sample.

3. The dual-stream anti-inter-device fault diagnosis method based on symmetrical point pattern images as described in claim 1, characterized in that, The first feature extraction unit includes: The first convolutional layer is used to extract edge texture information from the two-dimensional image sample; The first pooling layer is used to reduce the feature dimension.

4. The dual-stream cross-device fault diagnosis method based on symmetrical point pattern images as described in claim 1, characterized in that, The second feature extraction unit includes: The second convolutional layer is used to perform preliminary feature extraction on the two-dimensional image samples to obtain intermediate feature information; The third convolutional layer is used to extract deep features from the intermediate feature information to obtain a deep feature map; The second pooling layer is used to reduce the dimensionality of the deep feature map and extract key feature information. Furthermore, the second convolutional layer, the third convolutional layer, and the second pooling layer are connected in sequence.

5. The dual-stream anti-inter-device fault diagnosis method based on symmetric point pattern images as described in claim 1, characterized in that, The fusion of the shallow texture features extracted by the first feature extraction unit and the deep global features extracted by the second feature extraction unit is specifically as follows: The shallow texture features output by the first feature extraction unit are converted into a one-dimensional vector form to obtain the first feature vector; The deep global features output by the second feature extraction unit are converted into a one-dimensional vector form to obtain the second feature vector; The first feature vector and the second feature vector are concatenated to generate a fused feature vector as the fused feature information.

6. The dual-stream adversarial cross-device fault diagnosis method based on symmetric point pattern images as described in claim 1, characterized in that, The specific method for generating domain-invariant features through domain adversarial training is as follows: Construct a domain discriminator to determine whether the input feature information belongs to the source domain device or the target device; A gradient inversion layer is set between the fused feature information and the neighborhood discriminator; During model training, the gradient reversal layer adjusts the update direction of the network parameters, allowing the fused feature information to be fed into the domain discriminator, thus confusing the judgment results of the domain discriminator and generating domain-invariant features that are consistently distributed between the source and target devices.

7. The dual-stream cross-device fault diagnosis method based on symmetrical point pattern images as described in claim 1, characterized in that, The domain-invariant features are input into the fault classification network to obtain the fault diagnosis results as follows: Construct a fault classifier, which includes a fully connected layer and an output layer; The domain-invariant features are input into the fully connected layer for mapping processing to obtain the vector to be classified; The output layer calculates the probability value of the vector to be classified belonging to each preset fault category, and selects the fault category with the highest probability value as the final fault diagnosis result.

8. The dual-stream cross-device fault diagnosis method based on symmetric point pattern images as described in claim 1, characterized in that, It also includes transmitting the fault diagnosis result to a display terminal when the fault diagnosis result is a fault state; wherein the display terminal includes at least one of an on-site industrial control computer screen and a mobile terminal.

9. A dual-stream countermeasure cross-device fault diagnosis system based on symmetrical point pattern images, applicable to the dual-stream countermeasure cross-device fault diagnosis method based on symmetrical point pattern images as described in any one of claims 1-8, characterized in that, include: The data preprocessing module is used to acquire the vibration signal of the device to be diagnosed and convert the vibration signal into a symmetrical point pattern image; The dual-stream feature extraction module is used to input the symmetric point pattern image into the dual-stream feature extraction network to extract shallow texture features and deep global features respectively; The feature fusion module is used to fuse shallow texture features and deep global features to obtain fused feature information; The domain adversarial training module is used to perform domain adversarial training based on fused feature information and generate domain-invariant features.

10. The dual-stream adversarial cross-device fault diagnosis system based on symmetric point pattern images as described in claim 9, characterized in that, It also includes a fault classification module connected to the domain adversarial training module, which is used to input domain-invariant features into the fault classification network and output fault diagnosis results.

Images