A bearing fault diagnosis method based on multi-scale feature fusion

Abstract

This invention relates to the field of data processing and pattern recognition technology, specifically to a bearing fault diagnosis method based on multi-scale feature fusion. The method includes: fusing multi-source data such as vibration, acoustic emission, and rotational speed; resampling the rotational speed data in the angle domain to generate a two-dimensional order spectrum; stacking these spectrograms to construct a three-dimensional operating condition information tensor; employing a master-slave modulated heterogeneous neural network, where a main branch three-dimensional convolutional network extracts high-dimensional spatiotemporal features, and a secondary branch one-dimensional convolutional network extracts temporal details from the original sequence, generating affine transformation parameters to dynamically modulate the high-dimensional features; mapping the output state vector to a fault evolution knowledge graph; using a graph attention network and incorporating a Monte Carlo dropout mechanism for probability prediction; calculating the probability mean as the fault classification result; and quantifying the diagnostic confidence using the probability variance. This invention addresses the problem of insufficient feature discrimination power caused by scale mismatch under varying operating conditions through multi-scale feature fusion and dynamic modulation.

Description

Technical Field

[0001] This invention relates to the field of data processing and pattern recognition technology, specifically to a bearing fault diagnosis method based on multi-scale feature fusion. Background Technology

[0002] In the field of industrial big data analytics and pattern recognition, time-series data analysis of the operating status of rotating machinery is a core task. Existing data processing technologies typically treat the acquired signals as data pattern objects to be identified, using signal processing algorithms or machine learning models to construct a mapping relationship from input data to state categories. The general process of these methods involves transforming a one-dimensional time series into a feature vector through specific mathematical transformations, and then dividing the category boundaries in the feature space to achieve automatic classification and recognition of data patterns.

[0003] However, existing fault diagnosis methods mostly rely on single-scale feature extraction techniques. When faced with complex operating conditions such as varying speeds and loads in industrial settings, fault signals often exhibit non-stationary and nonlinear dynamic characteristics. Since the statistical characteristics of equipment operating data often dynamically drift with changing operating conditions, the duration of effective feature components in the time domain and the distribution range in the frequency domain are uncertain. Existing data analysis methods often use single-scale observation windows or fixed-resolution feature extraction units for calculation. This static scale setting cannot adapt to the dynamic changes in signal characteristics, leading to scale mismatch in the algorithm's processing. Specifically, when the time span of data features does not match the algorithm's observation scale, it can lead to the blurring and loss of local transient details or the breakage of long-distance global dependencies, resulting in insufficient discriminative power in the constructed feature space and difficulty in achieving high-precision pattern differentiation in variable data distributions.

[0004] Therefore, this invention proposes a bearing fault diagnosis method based on multi-scale feature fusion. Summary of the Invention

[0005] The purpose of this invention is to provide a bearing fault diagnosis method based on multi-scale feature fusion. By constructing a multi-scale time-frequency feature matrix set based on a logarithmic distribution window function and a three-dimensional operating condition information tensor with fused rotational speed weights, a master-slave modulation heterogeneous neural network is used to achieve affine transformation modulation of one-dimensional temporal details on high-dimensional spatiotemporal features. Furthermore, Bayesian probabilistic inference is performed in conjunction with a fault evolution knowledge graph. This solves the problems of existing single-scale analysis methods mentioned in the background art, which suffer from loss of local transient details, breakage of global dependencies, and insufficient discriminative power in feature space due to the mismatch between feature scale and observation window under varying operating conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A bearing fault diagnosis method based on multi-scale feature fusion includes:

[0008] The radial vibration sequence, axial vibration sequence, acoustic emission sequence, and rotational speed data of the bearing are obtained; the vibration and acoustic emission sequences are transformed using short-time Fourier transform to generate a two-dimensional order spectrum; a weight matrix is ​​generated based on the rotational speed data, and the two-dimensional order spectrum is stitched together to construct a three-dimensional operating condition information tensor.

[0009] A master-slave modulation heterogeneous neural network is constructed; a high-dimensional spatiotemporal feature map is extracted from the three-dimensional working condition information tensor through the three-dimensional convolutional layer of the main branch; a time-series vector is extracted from the vibration and acoustic emission sequences through the one-dimensional convolutional layer of the auxiliary branch, and the time-series vector is mapped to an affine transformation parameter set; based on the affine transformation parameter set, a feature affine transformation operation is performed on the high-dimensional spatiotemporal feature map in the main branch, and the state embedding vector is output.

[0010] A fault evolution knowledge graph containing pattern state nodes is constructed, and the state embedding vector is mapped to the initial node features. A graph attention network is used to perform node feature aggregation calculation on the knowledge graph, and a Monte Carlo dropout mechanism is introduced to perform probabilistic sampling prediction. Based on the posterior probability distribution of the nodes output by the probabilistic sampling prediction, the probability mean is calculated as the fault classification result, and the probability variance is calculated as the diagnostic confidence index.

[0011] Preferably, the two-dimensional order spectrum specifically includes: converting rotational speed data into an instantaneous angular velocity sequence; performing time integration on the instantaneous angular velocity sequence to resolve the mapping relationship between sampling time and cumulative rotation phase of the shaft in the sequence; predefining a fixed angle increment and generating an arithmetic sequence; applying the mapping relationship to the radial vibration sequence, axial vibration sequence, and acoustic emission sequence respectively; applying a cubic spline interpolation algorithm to the radial vibration sequence, axial vibration sequence, and acoustic emission sequence to reconstruct the signal amplitude at each angle point of the predefined arithmetic sequence, generating corresponding angle domain sequences with equal angular interval sampling for the vibration and acoustic emission sequences; applying a short-time Fourier transform to the angle domain sequence to generate a two-dimensional complex order spectrum with the center position of the transform window as one dimension and the order as the other dimension; the order is defined as the signal frequency divided by the shaft rotation frequency; performing a modulo operation on the two-dimensional complex order spectrum to generate a two-dimensional order spectrum.

[0012] Preferably, the construction of the three-dimensional operating condition information tensor specifically includes: inputting rotational speed data into a multilayer perceptron network to extract rotational speed feature vectors corresponding to the number of channels in the two-dimensional order spectrum; and using tensor broadcasting technology to copy and expand the rotational speed feature vectors in the time and frequency dimensions to construct a spatial dimension and a two-dimensional order spectrum. Figure 1The rotational speed feature matrix is ​​obtained; element-wise multiplication is performed between the rotational speed feature matrix and the matrix in the two-dimensional order spectrum to generate a weighted time-frequency feature matrix; the weighted time-frequency feature matrix is ​​stacked according to the channel dimension to construct a three-dimensional operating condition information tensor containing the channel dimension, frequency dimension and time dimension.

[0013] Preferably, the construction of the master-slave modulation heterogeneous neural network specifically includes: an input layer configured to receive the three-dimensional working condition information tensor and the original vibration and acoustic emission sequences in parallel; a processing layer connected to the input layer, including: a main branch network architecture composed of cascaded three-dimensional convolutional layers, configured to process the three-dimensional working condition information tensor and extract a high-dimensional spatiotemporal feature map; an auxiliary branch network architecture composed of cascaded one-dimensional convolutional layers and parallel linear mapping layers at the ends, configured to process the original sequence and generate an affine transformation parameter set; the processing layer is further configured to perform a feature affine transformation operation on the high-dimensional spatiotemporal feature map extracted by the main branch network architecture based on the affine transformation parameter set generated by the auxiliary branch network architecture; and an output layer connected to the processing layer, configured to perform global flattening processing on the feature map after the feature affine transformation operation, generating and outputting a one-dimensional state embedding vector.

[0014] Preferably, the extraction of high-dimensional spatiotemporal feature maps from the three-dimensional working condition information tensor through the three-dimensional convolutional layer of the main branch specifically includes: inputting the three-dimensional working condition information tensor into the first three-dimensional convolutional layer of the main branch network, performing three-dimensional convolution operations to extract shallow spatial texture features; sequentially performing batch normalization, nonlinear activation, and three-dimensional max pooling operations on the shallow spatial texture features to reduce the spatial resolution of the feature map and increase the number of feature channels layer by layer; obtaining the tensor data output by the last convolutional layer of the main branch network and defining it as the high-dimensional spatiotemporal feature map to be modulated.

[0015] Preferably, the step of extracting temporal vectors from the vibration and acoustic emission sequences through the one-dimensional convolutional layer of the auxiliary branch and mapping the temporal vectors to an affine transformation parameter set specifically includes: concatenating the radial vibration sequence, axial vibration sequence, and acoustic emission sequence into channels, inputting them into the one-dimensional convolutional layer of the auxiliary branch network, and extracting local temporal features; performing global average pooling on the local temporal features to compress the convolutionally processed temporal features into a fixed-length global context feature vector; splitting the global context feature vector into two independent linear mapping layers; calculating and outputting a scaling factor vector through the first linear mapping layer, and calculating and outputting a translation factor vector through the second linear mapping layer; and combining the scaling factor vector and the translation factor vector to form the affine transformation parameters.

[0016] Preferably, the output state embedding vector specifically includes: obtaining the number of channels in the high-dimensional spatiotemporal feature map; verifying that the dimensions of the scaling factor vector and the translation factor vector are consistent with the number of channels in the high-dimensional spatiotemporal feature map; performing element-wise multiplication on each channel feature map of the high-dimensional spatiotemporal feature map based on the scaling factor vector; performing element-wise addition on each channel feature map after the multiplication operation based on the translation factor vector; performing global flattening on the feature map after the addition operation, and outputting a one-dimensional state embedding vector.

[0017] Preferably, the construction of a fault evolution knowledge graph containing pattern state nodes, and mapping the state embedding vector to initial node features, specifically includes: predefining a node set of the fault evolution knowledge graph, the node set including the observation node to be diagnosed, normal state nodes, inner ring fault nodes, outer ring fault nodes, and rolling element fault nodes; predefining an edge set of the graph, the edge set establishing a fully connected topological relationship between the observation node to be diagnosed and the normal state nodes, inner ring fault nodes, outer ring fault nodes, and rolling element fault nodes; and assigning the state embedding vector to the observation node to be diagnosed as the input node feature of the graph attention network.

[0018] Preferably, the step of using a graph attention network to perform node feature aggregation calculation on the knowledge graph and introducing a Monte Carlo dropout mechanism to perform probabilistic sampling prediction specifically includes: calculating the feature correlation coefficients of connected nodes in the graph; performing a linear rectified activation operation with leakage on the feature correlation coefficients and normalizing them using the Softmax function to generate attention weight coefficients; performing a weighted summation on the feature vectors of adjacent nodes based on the attention weight coefficients to update the feature representation of the target node; enabling a dropout layer before the fully connected classification layer of the graph attention network and setting a predetermined dropout probability; keeping the network model parameters fixed, repeatedly performing a predetermined number of forward propagation processes on the same sample, and filtering out some neurons according to the dropout probability in each forward propagation process to generate multiple sets of different node classification probability distribution vectors.

[0019] Preferably, the step of calculating the probability mean as the fault classification result and calculating the probability variance as the diagnostic confidence index based on the node posterior probability distribution output by probability sampling prediction specifically includes: obtaining multiple sets of node classification probability distribution vectors generated by probability sampling prediction, wherein each set of vectors contains the probability value of nodes corresponding to each mode state; performing an arithmetic mean operation on the probability values ​​of nodes corresponding to the same mode state in the multiple sets of vectors, and determining the mode state corresponding to the maximum value in the obtained set of average probability values ​​as the final fault classification result; performing variance calculation on the probability values ​​of nodes corresponding to the same mode state in the multiple sets of vectors, and using the mean variance of all nodes as a confidence index to measure the uncertainty of the current diagnostic result; when the confidence index exceeds a preset threshold, the current diagnostic result is determined to be unreliable.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention fundamentally eliminates the influence of speed variations on fault characteristic frequencies by performing angle-domain resampling on the original time-domain signal and transforming it to generate an order-time spectrum, thus constructing a feature representation with operating condition information. This method transforms the non-stationary variable speed diagnosis problem into the identification problem of stable characteristic patterns, effectively overcoming the difficulties of characteristic frequency drift and energy dissipation caused by speed fluctuations in traditional time-frequency analysis methods. Furthermore, by introducing speed-related operating condition weights to enhance the spectrum, it further captures the modulation effect of implicit operating condition changes such as load on the characteristic amplitude, significantly improving the model's diagnostic robustness and generalization ability under wide speed range operating conditions.

[0022] 2. This invention utilizes the main branch to extract high-dimensional order-spatiotemporal features from the three-dimensional working condition information tensor, while simultaneously using the auxiliary branch to extract an affine transformation parameter set containing scaling and translation factors from the sequence, and performs channel-by-channel dynamic affine transformation modulation on the high-dimensional features. This mechanism of fine-tuning high-dimensional abstract features with original temporal details solves the problem of transient impact information attenuation caused by layers of convolution and pooling in deep networks. This allows the model to retain a high degree of sensitivity to the microstructure of the signal while learning working condition invariant patterns, thereby constructing a state embedding vector that is more sensitive to early and subtle faults and has higher discriminative power.

[0023] 3. This invention maps data-driven state embedding vectors to a knowledge graph with physical topological relationships, aggregates node features using a graph attention network, and introduces a probabilistic sampling mechanism to calculate the mean and variance of the posterior probability distribution. This method not only uses the graph structure to constrain the reasoning process to conform to the fault evolution logic, but also quantifies the confidence of the diagnostic results through variance calculation. It can effectively identify unreliable predictions caused by no faults or abnormal noise, greatly improving the safety and interpretability of the diagnostic method in complex industrial environments. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the overall method of a bearing fault diagnosis method based on multi-scale feature fusion proposed in this invention.

[0025] Figure 2 This is a flowchart illustrating the data processing and three-dimensional operating condition information tensor construction method for a bearing fault diagnosis method based on multi-scale feature fusion proposed in this invention.

[0026] Figure 3 This is a flowchart of a master-slave modulated heterogeneous neural network method for bearing fault diagnosis based on multi-scale feature fusion proposed in this invention.

[0027] Figure 4 This is a flowchart of a knowledge graph-based probabilistic diagnosis method for bearing fault diagnosis proposed in this invention, based on multi-scale feature fusion. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] A bearing fault diagnosis method based on multi-scale feature fusion, referenced Figure 1 The specific implementation steps of the method proposed in this invention include:

[0030] S1. Obtain the radial vibration sequence, axial vibration sequence, acoustic emission sequence, and rotational speed data of the bearing; use the rotational speed data to perform angle domain resampling on the vibration and acoustic emission sequences, transform the resampled angle domain sequence into a two-dimensional order spectrum through time-frequency analysis, and stack the spectrum corresponding to each signal along the channel dimension to construct a three-dimensional working condition information tensor;

[0031] S2. Construct a master-slave modulation heterogeneous neural network; extract high-dimensional spatiotemporal feature maps from the three-dimensional working condition information tensor through the three-dimensional convolutional layer of the main branch; extract time-series vectors from the vibration and acoustic emission sequences through the one-dimensional convolutional layer of the auxiliary branch, and map the time-series vectors to an affine transformation parameter set; perform feature affine transformation operation on the high-dimensional spatiotemporal feature maps in the main branch based on the affine transformation parameter set, and output the state embedding vector;

[0032] S3. Construct a fault evolution knowledge graph containing pattern state nodes, and map the state embedding vectors to initial node features; use a graph attention network to perform node feature aggregation calculation on the knowledge graph, and introduce a Monte Carlo dropout mechanism to perform probabilistic sampling prediction; based on the node posterior probability distribution output by the probabilistic sampling prediction, calculate the probability mean as the fault classification result, and calculate the probability variance as the diagnostic confidence index.

[0033] Example 1:

[0034] This embodiment provides a specific application of a bearing fault diagnosis method based on multi-scale feature fusion.

[0035] Further, the radial vibration sequence, axial vibration sequence, acoustic emission sequence, and rotational speed data of the bearing are acquired; the vibration and acoustic emission sequences are resampled in the angle domain using the rotational speed data; the resampled angle domain sequences are transformed into two-dimensional order spectra through time-frequency analysis; and the spectra corresponding to each signal are stacked along the channel dimension to construct a three-dimensional operating condition information tensor, corresponding to step S1 above, refer to... Figure 2 The specific process is as follows:

[0036] Before performing data processing, a multi-source signal acquisition environment is first established. High-sensitivity piezoelectric accelerometers are installed at the radial and axial positions of the bearing housing to be diagnosed, respectively, to acquire radial and axial vibration sequences. Simultaneously, a broadband acoustic emission sensor is installed near the bearing housing's load-bearing area to acquire high-frequency acoustic emission sequences reflecting early microcrack propagation. Furthermore, a photoelectric encoder or tachometer mounted on the shaft end is used to synchronously acquire rotational speed pulse signals in real time, which are then converted to obtain rotational speed data. All sensors are connected to the industrial control computer via a high-speed data acquisition card to ensure time synchronization of the multi-channel signals.

[0037] Before performing the transformation using the short-time Fourier transform kernel, the acquired radial vibration sequence, axial vibration sequence, and acoustic emission sequence are subjected to Z-score normalization. Specifically, the mean and standard deviation of each sequence segment are calculated, and the mean is subtracted from each sampling point in the sequence and divided by the standard deviation to eliminate the dimensional differences between signals acquired by different sensors and ensure the numerical stability of subsequent feature extraction.

[0038] Generating a two-dimensional order spectrum involves three core steps: instantaneous phase calculation, angle domain resampling, and order spectrum analysis. It should be noted that since the input signal has been converted into an angle domain sequence, the horizontal axis of the generated two-dimensional spectrum physically represents the center position of the transform window in the angle domain, and the vertical axis represents the order. In this specification, to maintain correspondence with conventional time-frequency analysis terminology, it is collectively referred to as a two-dimensional order spectrum. Specifically, when performing instantaneous phase calculation, the synchronously acquired rotational speed data in revolutions per minute (RPM) is first converted to a unit by multiplying by a coefficient of π / 30 to obtain an instantaneous angular velocity sequence in radians per second (rad / s). Subsequently, discrete integration is performed on the instantaneous angular velocity sequence. In this embodiment, the summation method is used to calculate the total cumulative rotational phase of the axis of rotation relative to the starting point at each original sampling time point, thereby establishing a precise nonlinear mapping relationship between the time axis and the angle axis. Before performing angle domain resampling, a high-resolution sampling reference, i.e., a fixed angle increment, needs to be pre-set; in this embodiment, it is set to 0.1 degrees. Based on this angle increment, an arithmetic sequence of angles covering the entire analysis period is generated, forming a new sampling coordinate axis. Next, the established mapping relationship is applied to the radial vibration sequence, axial vibration sequence, and acoustic emission sequence, respectively. This embodiment preferably employs a cubic spline interpolation algorithm, which ensures the smoothness and high fidelity of the signal after resampling. For each angle point in the arithmetic sequence, the cubic spline interpolation algorithm accurately calculates the signal amplitude that should be present at that angle point based on its nearest preceding and following original sampling points and their phase-amplitude relationship. Through this process, three original equal-time interval sequences are successfully converted into three angle domain sequences sampled at equal angle intervals. Finally, when performing order spectrum analysis, a short-time Fourier transform is applied to each generated angle domain sequence. Since the independent variable of the input sequence has changed from time to angle, the physical meaning of the resulting two-dimensional spectrum also changes accordingly: its horizontal axis represents the center angle position of the transform window, reflecting the phase information of the fault impact within the rotation period; its vertical axis represents the order, i.e., the multiple relationship between the fault characteristic frequency and the current shaft rotation frequency. For example, the first order represents the component synchronized with the rotational frequency, while the theoretical order of the outer race fault is a fixed non-integer value. After performing a modulo operation on the transformed complex numerical matrix, a final two-dimensional order spectrum is generated that can maintain the stability of the characteristic mode under variable speed conditions.

[0039] By performing angle-domain resampling, the original non-stationary time-domain signal is converted into a quasi-stationary angle-domain sequence, and the resulting order-time spectrum can keep the order of fault characteristics constant under variable speed operating conditions. This technique, which decouples fault modes from operating conditions at the data level, solves the problems of characteristic spectral line ambiguity and drift caused by speed fluctuations in traditional time-frequency analysis, greatly improving the accuracy and robustness of diagnosis.

[0040] The rotational speed data is converted into a mapping relationship between the instantaneous phase of the rotating shaft and the sampling time. Using the mapping relationship, the original vibration and acoustic emission time-domain sequences sampled at equal time intervals are resampled into angle-domain sequences sampled at equal angle intervals. A short-time Fourier transform is applied to the angle-domain sequences, and the transformation result is subjected to modulo operation to finally generate a two-dimensional order spectrum.

[0041] In its specific implementation, this method first converts the synchronously acquired rotational speed data into an instantaneous angular velocity sequence. Then, through time integration, it accurately analyzes the nonlinear mapping relationship between the original time-domain sampling time and the cumulative rotational phase of the shaft. Subsequently, this mapping relationship is applied to the radial vibration, axial vibration, and acoustic emission sequences, respectively. A cubic spline interpolation algorithm is used to reconstruct the signal amplitude at each sampling point of a predefined arithmetic angle sequence, thereby converting each original time-domain signal into an angle-domain sequence sampled at equal angular intervals. Finally, a short-time Fourier transform is applied to each angle-domain sequence to generate a two-dimensional complex order spectrum with the center angular position of the transform window as one dimension and the order (i.e., the ratio of the signal frequency to the shaft rotation frequency) as the other dimension. A modulo operation is then performed on this spectrum to finally obtain the two-dimensional order-time spectrum.

[0042] By transforming non-stationary time-domain signals into quasi-stationary angle-domain signals, the modulation effect of speed changes on fault characteristic frequencies is eliminated. This solves the problems of characteristic spectral line drift, ambiguity, and energy dispersion that are common in traditional time-frequency analysis methods under varying operating conditions, greatly improving the accuracy, robustness, and generalization ability of the diagnostic model under wide speed-domain operating conditions.

[0043] Rotational speed data is input into a multilayer perceptron network (MLPF) to extract rotational speed feature vectors corresponding to the number of channels in the two-dimensional order spectrum. The MPF consists of an input layer, two hidden layers, and an output layer. The input layer receives a normalized rotational speed scalar; the first hidden layer contains 32 neurons using the ReLU activation function; the second hidden layer contains 64 neurons using the ReLU activation function; the output layer contains neurons equal to the number of channels in the two-dimensional order spectrum (the number of input signals, 3 in this embodiment), using the Sigmoid activation function to generate normalized rotational speed feature weights. Using tensor broadcasting, the rotational speed feature vectors are copied and expanded in the time and frequency dimensions to construct a spatial dimension and a two-dimensional order spectrum. Figure 1The rotational speed feature matrix is ​​obtained; element-wise multiplication is performed between the rotational speed feature matrix and the matrix in the two-dimensional order spectrum to generate a weighted time-frequency feature matrix; the weighted time-frequency feature matrix is ​​stacked according to the channel dimension to construct a three-dimensional operating condition information tensor containing the channel dimension, frequency dimension, and time dimension. It should be noted that although angle domain resampling and order analysis eliminate the drift effect of rotational speed fluctuations on the fault characteristic frequency (horizontal axis position), there are still significant nonlinear differences in the amplitude of mechanical vibration energy (vertical axis intensity) at different rotational speeds (for example, the vibration energy under high speed and heavy load is much greater than that under low speed and light load). Therefore, this invention further introduces element-wise multiplication between the rotational speed feature matrix and the order spectrum. Its physical meaning is not to reintroduce frequency interference, but to construct an amplitude attention mechanism based on the operating condition. This mechanism allows the network to adaptively adjust the attention weight of different regions in the spectrum according to the current rotational speed level, thereby compensating for the nonlinear amplitude modulation effect caused by rotational speed changes. Specifically, to construct a three-dimensional operating condition information tensor, the synchronously acquired one-dimensional scalar rotational speed data is first input into a pre-configured multilayer perceptron network. The scalar rotational speed is then mapped to a one-dimensional rotational speed feature vector, the dimension of which is the same as the number of channels in the two-dimensional order spectrum. Subsequently, using a tensor broadcasting mechanism, the rotational speed feature vector is copied and expanded in both the time and frequency dimensions to construct a tensor that is spatially similar to the two-dimensional order spectrum. Figure 1 The output layer of the multilayer perceptron network is designed to generate a two-dimensional rotational speed feature matrix, with the number of rows and columns matching the dimensions of the two-dimensional order spectrum generated in the previous step. After generating the rotational speed feature matrix, an element-wise Hadamard product is performed between the matrix and the three two-dimensional order spectra generated in the previous step for radial vibration, axial vibration, and acoustic emission signals, respectively, to obtain three feature matrices weighted by the operating condition information. Finally, these three weighted two-dimensional spectra are stacked along the new channel dimension to construct a three-dimensional operating condition information tensor. The final dimensional arrangement of the operating condition information tensor is (number of channels, frequency / resolution dimension, time dimension).

[0044] The proposed solution deeply embeds operating condition information into the initial stage of feature extraction. Through nonlinear mapping and feature modulation, the calculation process can adaptively adjust the feature response according to the operating conditions, thereby effectively decoupling the coupling relationship between fault features and operating conditions and enhancing the robustness of diagnosis.

[0045] Furthermore, a master-slave modulation heterogeneous neural network is constructed; a high-dimensional spatiotemporal feature map is extracted from the three-dimensional working condition information tensor through the three-dimensional convolutional layer of the main branch; a time-series vector is extracted from the vibration and acoustic emission sequences through the one-dimensional convolutional layer of the auxiliary branch, and the time-series vector is mapped to an affine transformation parameter set; based on the affine transformation parameter set, a feature affine transformation operation is performed on the high-dimensional spatiotemporal feature map in the main branch, and the state embedding vector is output, corresponding to step S2 above. The specific process is as follows:

[0046] The construction of the master-slave modulation heterogeneous neural network specifically includes: an input layer configured to receive the three-dimensional working condition information tensor and the original vibration and acoustic emission sequences in parallel; a processing layer connected to the input layer, including: a main branch network architecture composed of cascaded three-dimensional convolutional layers, configured to process the three-dimensional working condition information tensor and extract high-dimensional spatiotemporal feature maps; an auxiliary branch network architecture composed of cascaded one-dimensional convolutional layers and parallel linear mapping layers at the ends, configured to process the original sequences and generate an affine transformation parameter set; the processing layer is also configured to perform feature affine transformation operations on the high-dimensional spatiotemporal feature maps extracted by the main branch network architecture based on the affine transformation parameter set generated by the auxiliary branch network architecture; and an output layer connected to the processing layer, configured to perform global flattening processing on the feature maps after feature affine transformation operations, generating and outputting a one-dimensional state embedding vector. Specifically, before data processing, the dual-stream heterogeneous neural network topology needs to be configured as follows: Configure the main processing branch, which consists of multiple three-dimensional convolutional layers connected in sequence. Simultaneously, an auxiliary processing branch is configured, consisting of multiple cascaded one-dimensional convolutional layers, with two independent linear mapping layers running in parallel at the network's ends. A parameter transfer path is established between the two branches: the output data of the two linear mapping layers at the end of the auxiliary branch are defined as a set of computational parameters, and this set of parameters is directed to the output feature map of a specific three-dimensional convolutional layer in the main processing branch. During implementation, the parameter transfer path must ensure that the dimension of the parameter set generated by the auxiliary branch is strictly equal to the number of channels in the feature map of the main branch it is intended to interact with.

[0047] The dual-stream parallel computational flow design provides optimized processing paths for both high-order spatiotemporal features and raw temporal details, while the defined parameter passing method establishes a direct computational link for the deep interaction between these two data modalities.

[0048] The three-dimensional working condition information tensor is input into the first three-dimensional convolutional layer of the main branch network, and three-dimensional convolution operation is performed to extract shallow spatial texture features. Batch normalization operation, nonlinear activation operation and three-dimensional max pooling operation are performed on the shallow spatial texture features in sequence to reduce the spatial resolution of the feature map and increase the number of feature channels layer by layer. The tensor data output by the last convolutional layer of the main branch network is obtained and defined as the high-dimensional spatiotemporal feature map to be modulated.

[0049] During the forward propagation of data in the main branch, the 3D working condition information tensor is first fed into the first 3D convolutional layer as input, and shallow features are extracted through convolution operations. Subsequently, the feature map output from the 3D convolutional layer is sequentially fed into a batch normalization layer, a nonlinear activation layer, and a 3D max pooling layer for processing. The computational unit of "convolution-normalization-activation-pooling" is repeatedly stacked multiple times in the main branch. During the layer-by-layer propagation, the spatial dimension (frequency and temporal dimension) of the feature map decreases due to the pooling operation, while the channel dimension increases due to the setting of the number of convolutional kernels. After processing by all convolutional blocks, the 3D tensor data output by the last convolutional layer of the main branch is defined as a high-dimensional spatiotemporal feature map awaiting modulation in the next step.

[0050] By processing information tensors through three-dimensional convolution, it is possible to automatically extract and abstract spatiotemporal dynamic features that transcend static two-dimensional textures and reflect the fault evolution process, thereby generating higher-order feature representations with greater discriminative power.

[0051] The radial vibration sequence, axial vibration sequence, and acoustic emission sequence are concatenated and input into a one-dimensional convolutional layer of the auxiliary branch network to extract local temporal features. Global average pooling is performed on the local temporal features to compress the convolutional temporal features into a fixed-length global context feature vector. The global context feature vector is then split and input into two independent linear mapping layers. The scaling factor vector is calculated and output through the first linear mapping layer, and the translation factor vector is calculated and output through the second linear mapping layer. The scaling factor vector and the translation factor vector are combined to form the affine transformation parameters.

[0052] During the forward propagation of data in the auxiliary branch, the original three sets of one-dimensional time-series sequences—radial vibration, axial vibration, and acoustic emission—are first concatenated along the channel dimension to form three-channel one-dimensional input data. This data is then input into the one-dimensional convolutional layer sequence of the auxiliary branch for feature extraction. Global average pooling is performed on the variable-length time-series features output from the last one-dimensional convolutional layer, averaging the time-series features and compressing them into a fixed-length global context feature vector. This global context feature vector is then used as a bifurcation point and simultaneously input into two parallel, structurally independent linear mapping layers. The first linear mapping layer calculates and outputs a scaling factor vector (…). The second linear mapping layer calculates and outputs a translation factor vector. These two vectors ( and Together, they constitute the final affine transformation parameter set.

[0053] By directly applying it to the original one-dimensional data, high-fidelity temporal information is preserved, and complex temporal dynamics can be efficiently condensed into a set of low-dimensional affine transformation parameters, providing information-rich and structurally simple input for subsequent feature modulation.

[0054] Obtain the number of channels in the high-dimensional spatiotemporal feature map; verify that the dimensions of the scaling factor vector and the translation factor vector are consistent with the number of channels in the high-dimensional spatiotemporal feature map; perform element-wise multiplication on each channel feature map of the high-dimensional spatiotemporal feature map based on the scaling factor vector; perform element-wise addition on each channel feature map after multiplication based on the translation factor vector; perform global flattening on the feature map after addition and output a one-dimensional state embedding vector.

[0055] To complete feature modulation, the number of channels C of the high-dimensional spatiotemporal feature map output by the main branch is first obtained, and the scaling factor vector generated by the auxiliary branch is verified. Translation factor vector Are all dimensions C? First, the scaling factor vector and translation factor vector are broadcast and expanded in the spatial dimension to make their size consistent with the spatial resolution (frequency dimension × time dimension) of the high-dimensional spatiotemporal feature map, thus constructing scaling and translation tensors that perfectly match the dimensions of the feature map. After verification, a channel-by-channel affine transformation is performed on the high-dimensional spatiotemporal feature map: for the i-th channel (i from 1 to C) in the map, all elements are compared with the i-th element of the scaling factor vector. Multiply, and then multiply the result by the i-th element of the translation factor vector. Add. After performing this "multiply-add" operation on all C channels, the resulting 3D feature map is globally flattened, and all elements are rearranged into a one-dimensional long vector, which is the final output state embedding vector.

[0056] By adaptively calibrating the abstract features extracted by the deep network channel by channel-by-channel using the detailed information of the original signal, a dynamic feature modulation calculation is achieved, which effectively overcomes the problem of attenuation of microscopic transient information during deep feature extraction.

[0057] A master-slave modulation heterogeneous neural network is constructed. An auxiliary branch extracts and generates an affine transformation parameter set containing scaling and translation factors from the original one-dimensional time-series signal. Using this affine transformation parameter set, a channel-by-channel feature affine transformation operation is performed on the high-dimensional spatiotemporal feature map extracted from the three-dimensional working condition information tensor by the master branch network. This dynamically injects temporal details into deep semantic features, outputting a state embedding vector.

[0058] In the interaction phase of the master-slave heterogeneous network, global average pooling is first performed on the local temporal features extracted by one-dimensional convolution, compressing the variable-length temporal signal into a fixed-length global context vector. Subsequently, the global context vector is split and input to two parallel linear mapping layers, which respectively calculate a positive scaling factor vector and a real-valued translation factor vector. The dimensions of these two sets of vectors are strictly designed to be equal to the number of channels N of the high-dimensional feature map output by the main branch. During feature fusion, the scaling factor vector is expanded into a mask with the same spatial size as the feature map, and pixel-level multiplication is performed on each of the N channels of the main branch to adjust the activation intensity of each channel. Next, the translation factor vector is similarly expanded and superimposed onto the feature map to adjust the baseline offset of the features. The entire process utilizes the complete temporal context information preserved in the original signal to dynamically gate the highly abstract spatial feature channels in the deep network.

[0059] By constructing cross-modal channel attention and self-calibration logic, the system can adaptively suppress background noise-dominated channels in the main branch and enhance fault-related channels based on the physical temporal structure of the current signal. This strategy of modulating three-dimensional spatiotemporal space with one-dimensional details solves the problem of micro-detail loss caused by layers of convolution and pooling in deep networks, and significantly improves the model's sensitivity to capturing weak fault features in strong noise backgrounds.

[0060] Furthermore, a fault evolution knowledge graph containing pattern state nodes is constructed, and the state embedding vectors are mapped to initial node features. A graph attention network is used to perform node feature aggregation calculation on the knowledge graph, and a Monte Carlo dropout mechanism is introduced to perform probabilistic sampling prediction. Based on the posterior probability distribution of the nodes output by the probabilistic sampling prediction, the probability mean is calculated as the fault classification result, and the probability variance is calculated as the diagnostic confidence index. Corresponding to step S3 above, the specific process is as follows:

[0061] A predefined set of nodes for a fault evolution knowledge graph is provided, comprising nodes to be diagnosed, nodes in normal state, nodes with inner race faults, nodes with outer race faults, and nodes with rolling element faults. A predefined set of edges is also provided, establishing a fully connected topology between the nodes to be diagnosed and the nodes in normal state, nodes with inner race faults, nodes with outer race faults, and nodes with rolling element faults. Although the basic topology of the graph is a predefined star topology, this invention models fault evolution as a dynamic drift process of the attention distribution of the observed nodes among nodes in different modes of state. Through continuous reasoning on the time-series data stream using a graph attention network, the distribution of attention weights between nodes changes over time. This trajectory change of probability distribution characterizes the evolution path of the bearing from normal to faulty, and between different fault degrees. This structure not only allows the calculation of the distance (edge ​​weight) between the node to be diagnosed and the faulty node, but also reserves an interface to introduce associated edges between faulty nodes in the future (edges indicating inner race wear evolving into inner race spalling), thereby supporting more complex reasoning tasks. Therefore, this scheme utilizes a graph attention network to process this topology, essentially performing a structured reasoning based on metric learning. The state embedding vector is assigned to the observation node to be diagnosed, serving as the input node feature of the graph attention network. In this embodiment, the graph attention network is configured as a two-layer stacked structure. The first layer adopts a multi-head attention mechanism, setting up 8 independent attention heads, each with an output feature dimension of 8, which are concatenated to output 64-dimensional features; the second layer is a single-head attention layer, with the output dimension corresponding to the number of fault categories, and a Dropout operation with a probability of 0.6 is introduced between layers to prevent overfitting.

[0062] Before performing the diagnosis, a graph structure needs to be predefined. The node set of the graph structure consists of two parts: the observation node to be diagnosed as input, and N pattern state nodes representing preset states ("normal", "inner ring fault", "outer ring fault", "rolling body fault"). The edge set of the graph is defined by establishing connections between the observation node to be diagnosed and all N pattern state nodes, forming a star topology. During computation, the one-dimensional state embedding vector output from the previous neural network step is directly used as the initial node feature of the observation node to be diagnosed. For each pattern state node in the graph, a learnable embedding vector of fixed dimensions is initialized. This learnable embedding vector is assigned values ​​using a random normal distribution during network initialization and is updated synchronously with other network parameters through backpropagation during subsequent model training, thereby learning the typical feature representation representing the corresponding fault mode.

[0063] By mapping diagnostic problems to reasoning tasks on a knowledge graph, the decision-making process can be constrained by pre-defined, physically-compliant topological relationships between states, thereby elevating diagnosis from independent classification to structured reasoning.

[0064] Calculate the feature correlation coefficients of connected nodes in the graph, perform a linear rectified activation operation with leakage on the feature correlation coefficients, and normalize them using the Softmax function to generate attention weight coefficients; based on the attention weight coefficients, perform a weighted summation on the feature vectors of adjacent nodes to update the feature representation of the target node; enable a dropout layer before the fully connected classification layer of the graph attention network and set a predetermined dropout probability; keep the network model parameters fixed, and repeat the forward propagation process a predetermined number of times for the same sample. In each forward propagation process, some neurons are masked according to the dropout probability to generate multiple sets of different node classification probability distribution vectors.

[0065] The computation is performed using a graph attention network, introducing randomness. First, the cross-correlation of features between nodes is calculated, and then normalized using a linear rectified function with leakage and Softmax to obtain the attention weights on the connecting edges. Based on these weights, the features of neighboring nodes are weighted and summed to update the feature representation of the central node. This process is iteratively performed within the graph. Crucially, a random dropout layer is set up and enabled before the final output layer of the graph attention network (the fully connected layer used for classification), with a dropout probability of 0.5. During prediction, all weight parameters of the network are fixed, and the complete forward propagation computation is repeated N times (50 times) for the same input sample. Since the dropout layer randomly masks different neurons in each propagation, these N computations generate N slightly different classification probability distribution vectors for each mode state node.

[0066] The application of graph attention networks enables the computation process to have associative reasoning capabilities, while the Monte Carlo dropout operation introduced in the reasoning stage provides the core computational means to transform deterministic forward propagation into Bayesian approximate inference.

[0067] Obtain multiple sets of node classification probability distribution vectors generated by probability sampling prediction, where each set of vectors contains the probability value of nodes corresponding to each mode state; perform an arithmetic mean operation on the probability values ​​of nodes corresponding to the same mode state in the multiple sets of vectors, and determine the mode state corresponding to the maximum value in the obtained set of average probability values ​​as the final fault classification result; perform variance calculation on the probability values ​​of nodes corresponding to the same mode state in the multiple sets of vectors, and use the mean variance of all nodes as a confidence index to measure the uncertainty of the current diagnosis result; when the confidence index exceeds a preset threshold, the current diagnosis result is determined to be unreliable.

[0068] After obtaining the N sets of node classification probability distribution vectors generated in the previous step, the following process is followed: To determine the final fault category, for each mode state node, extract the N probability values ​​corresponding to the N sets of vectors and calculate the arithmetic mean of these N values. After obtaining the average probability of all mode state nodes, select the mode state represented by the node with the largest average probability value as the final fault classification result. To calculate the confidence level of the diagnosis, also for each mode state node, calculate the statistical variance of the N probability values. Then, perform an arithmetic mean on the variance values ​​calculated for all nodes to obtain a single variance mean representing the overall uncertainty; this variance mean is the confidence level index. Finally, compare the confidence level index with a preset threshold (0.02). If the index value is greater than the threshold, the current diagnosis result is marked as unreliable.

[0069] The Monte Carlo dropout mechanism from deep learning is introduced into the diagnostic inference stage. This involves maintaining the fully connected classification layer before dropout during prediction and repeatedly performing forward propagation on the same sample. The probabilistic sampling process generates a set of posterior probability distributions. The variance mean of this distribution is calculated to quantify the model's uncertainty as a diagnostic confidence index. Unreliable diagnostic results are proactively identified and judged based on whether they exceed a preset threshold.

[0070] In the inference and prediction phase of the graph attention network, instead of using the conventional deterministic inference model, the random dropout layer in the model is kept active with a dropout probability set to 0.5. For each input sample to be diagnosed, 50 independent forward propagation calculations are performed. Since the neurons that are blocked during each propagation change randomly, these 50 calculations will generate 50 slightly different node classification probability distribution vectors. These node classification probability distribution vectors are collected, and the arithmetic mean of the 50 predictions is calculated for each fault mode to serve as the final classification probability after eliminating random errors. More importantly, the statistical variance of these 50 predictions for each class is calculated, and the mean variance of all classes is extracted and quantified as the current diagnostic confidence index. Once the real-time calculated confidence index exceeds the preset variance threshold (0.02), the circuit breaker logic is triggered, marking the current diagnosis as unreliable.

[0071] By quantifying the model's hesitation when faced with unknown data using statistical methods, the risk of statistical characteristic deviation in data distribution across different operating conditions is effectively addressed. It can automatically identify and intercept abnormal samples under unknown faults or extreme operating conditions, avoiding the erroneous judgments made by traditional deep learning models when faced with data outside their cognitive scope, thus greatly improving the safety and reliability of industrial diagnostic methods.

[0072] By statistically averaging multiple random predictions, a more robust diagnostic conclusion is obtained. Furthermore, by calculating the statistical variance, a quantitative confidence index is provided for the diagnostic results, enabling the entire method to self-assess the reliability of the output.

[0073] This invention integrates multimodal and multiscale data and processes it using a unique detail-fidelity network, ultimately performing graph-based probabilistic inference. This not only improves diagnostic accuracy under varying operating conditions but also provides quantifiable confidence levels for each diagnosis, significantly enhancing the safety and practicality of the diagnosis.

[0074] Example 2:

[0075] This embodiment aims to explain in detail the specific architecture and working mechanism of the master-slave modulation heterogeneous neural network. For details, please refer to [reference needed]. Figure 3 The core idea of ​​the master-slave modulation heterogeneous neural network is to use a lightweight auxiliary branch (auxiliary branch) to extract dynamic modulation parameters that reflect transient changes in the signal from the original one-dimensional time-series data, and use these parameters to finely adjust the deep and abstract features extracted by the main processing branch (main branch) from the complex two-dimensional time-frequency image, thereby achieving deep fusion and information complementarity of different modal features.

[0076] In the specific network construction, the main branch is designed as a deep 3D convolutional network, specifically for processing the 3D condition information tensor constructed in the preceding steps. The main branch can be composed of three 3D convolutional blocks connected in series. The first convolutional block uses 16 convolutional kernels of size (3, 3, 3) to process the input tensor, sets the convolutional stride to (1, 1, 1) and adopts a boundary zero-padding strategy to maintain the spatial resolution of the feature map, and extracts shallow texture and contour information; the second and third convolutional blocks use 32 and 64 convolutional kernels of size (3, 3, 3) respectively, to deepen the level of feature abstraction layer by layer. Within each convolutional block, after the 3D convolutional layer, a batch normalization layer is sequentially connected to accelerate convergence, a ReLU nonlinear activation function to enhance the model's expressive power, and a 3D max-pooling layer of size (2, 2, 2). The pooling stride is explicitly set to be consistent with the pooling kernel size (stride of 2) to perform downsampling, halving the size of the feature map in the length, width, and height dimensions, thereby reducing spatial resolution and expanding the receptive field. After processing by these three convolutional blocks, the main branch finally outputs a high-dimensional spatiotemporal feature map with 64 channels. This high-dimensional spatiotemporal feature map contains rich macroscopic spatiotemporal structural information related to fault modes, but may lose some fine transient details of the original signal during layer-by-layer abstraction.

[0077] Meanwhile, the auxiliary branch is a lightweight one-dimensional convolutional network focused on processing the original one-dimensional vibration and acoustic emission sequences. First, the radial, axial vibration, and acoustic emission sequences are concatenated along the channel dimension to form an input of (3, L), where L is the sequence length. The input then passes through two one-dimensional convolutional layers, for example, with kernel sizes of 7 and 5, and the number of channels of 16 and 32, respectively, to capture local temporal patterns at different time scales. After convolutional processing, a global average pooling layer is applied to compress the variable-length temporal features into a fixed-dimensional vector of length 32, which represents the global context information of the entire time series. Finally, this global context vector is fed in parallel to two independent fully connected (linear mapping) layers. The first fully connected layer outputs a 64-dimensional scaling factor vector. The second fully connected layer outputs a 64-dimensional translation factor vector. The dimensions of these two vectors are precisely designed to match the number of channels in the high-dimensional feature map output by the main branch.

[0078] The final feature modulation (i.e., feature affine transformation) occurs at the main branch output. For each channel in the 64-channel high-dimensional spatiotemporal feature map output by the main branch, an independent affine transformation operation is performed: all eigenvalues ​​of the channel are first multiplied element-wise by the scaling factor vector generated by the auxiliary branch. The i-th element Then add the translation factor vector element by element. The i-th element This operation is equivalent to using the global context information of the original time-series signal to perform independent, adaptive calibration and fine-tuning on each deep abstract feature channel, enhancing fault-sensitive feature channels and suppressing irrelevant or noise-dominated channels. The modulated feature map is then globally flattened to form the final one-dimensional state embedding vector. This one-dimensional state embedding vector retains the macroscopic spatiotemporal structure while re-injecting microscopic temporal details, resulting in extremely high discriminative power.

[0079] Example 3:

[0080] This embodiment focuses on how to utilize fault evolution knowledge graphs and graph attention networks, and introduces a Monte Carlo dropout mechanism to achieve the transformation from deterministic state vectors to probabilistic diagnostic conclusions and confidence assessments. For details, please refer to [link / reference]. Figure 4 This process elevates the traditional black-box classification problem to a Bayesian approximate inference process based on graph reasoning and possessing uncertainty quantification capabilities, greatly enhancing the reliability and interpretability of the diagnostic method.

[0081] Before executing the fault diagnosis inference step, joint supervised training of the master-slave modulation heterogeneous neural network and the graph attention network is required. The specific training steps are as follows: 1. Construct the training dataset: Collect historical bearing operation data containing known fault labels and divide it into training and validation sets. 2. Define the loss function: Use the cross-entropy loss function to measure the difference between the node classification probability distribution output by the network and the true fault labels. 3. Optimization strategy: Use the Adam optimizer, set the initial learning rate (0.001), and calculate the gradient of the loss function relative to the network weights based on the backpropagation algorithm. 4. Iterative update: Input the training data into the network in batches, and use the calculated gradient to synchronously update the weight parameters in the main branch 3D convolution kernel, the auxiliary branch 1D convolution kernel, the affine transformation generation layer, and the graph attention mechanism until the loss function converges. The training process adopts an end-to-end joint optimization strategy. The cross-entropy loss calculated by the output layer of the graph attention network is used as the global loss function. Using the chain rule, the gradient is propagated backward through the graph attention layer and the node feature mapping interface, and then back to the main and auxiliary branches of the master-slave modulation heterogeneous neural network to achieve synchronous parameter updates between the feature extractor and the graph inference model.

[0082] First, during the model building phase, a fault evolution knowledge graph with clear physical meaning is predefined. This graph contains the observation node to be diagnosed and four mode state nodes, representing "normal state," "inner loop fault," "outer loop fault," and "rolling body fault," respectively. The graph's topology is set as a star topology, meaning there are connecting edges between the observation node to be diagnosed and all four mode state nodes. At the start of diagnosis, state embedding vectors generated by a master-slave modulated heterogeneous neural network are used as the initial features of the observation node to be diagnosed. The initial features of the four mode state nodes can be set as randomly initialized embedding vectors that can be trained along with the network; they will eventually learn ideal feature representations representing their respective fault modes.

[0083] Next, a graph attention network is used to aggregate information from the knowledge graph. The core mechanism of the graph attention network lies in its ability to dynamically calculate the correlation between nodes. For the observation node to be diagnosed, the network calculates the attention coefficient between its features and the features of each connected pattern state node. Specifically, the feature vector of the observation node to be diagnosed is concatenated with the feature vector of the pattern state node, and the concatenated vector is input into a single-layer feedforward neural network with shared weights, outputting a scalar correlation score. Subsequently, a linear rectified function with leakage is applied to the scalar correlation score for activation, and the scores of all neighboring nodes are normalized using the Softmax function to obtain the final attention weight coefficient. This coefficient reflects the similarity or correlation strength between the current observation state and a specific fault mode. Based on these attention coefficients, the network performs a weighted summation of the features of all pattern state nodes, thereby updating the feature representation of the observation node to be diagnosed. After one or more layers of graph attention computation, the final feature vector of the observation node to be diagnosed will be rich in information related to the most relevant fault mode. Then, the final feature vector is fed into a classifier consisting of a drop-out layer, a fully connected layer, and a Softmax layer, which outputs the classification probabilities for the four fault modes and the normal state.

[0084] The key innovation of this embodiment lies in the handling of uncertainty during the inference phase. A Monte Carlo dropout layer is enabled before the fully connected classification layer of the graph attention network. In conventional model testing, the dropout layer is usually turned off to make deterministic predictions. However, here, the dropout layer is kept active during prediction (dropout probability p=0.5). For the same input sample (i.e., the same state embedding vector), the entire forward propagation process is repeated 50 times. Since the neurons that are randomly dropped in each propagation are different, these 50 calculations will yield 50 slightly different classification probability distribution vectors.

[0085] Finally, statistical analysis was performed based on these 50 probability distribution vectors:

[0086] Fault Classification: For each possible fault state (inner loop fault), the average probability across 50 predictions is calculated. Finally, the state with the highest average probability is selected as the final diagnosis. This ensemble averaging approach makes the diagnosis more robust than a single prediction.

[0087] Confidence Assessment: For each fault state, the variance of the predicted probability is calculated 50 times. A larger variance indicates that the model's prediction is less stable under minor network disturbances, meaning the model is less certain about the prediction. The variances of all states are averaged to obtain a single, quantified confidence index. This confidence index is compared to a preset empirical threshold (e.g., 0.02). If the calculated mean variance exceeds this threshold, the current diagnostic result is deemed unreliable, and an alarm may be triggered or manual review may be recommended. This mechanism effectively identifies novel faults never seen by the model before or samples affected by strong noise, preventing the model from making erroneous judgments based on blind confidence beyond its cognitive range.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on multi-scale feature fusion, characterized in that, The method comprises the following steps: obtaining a radial vibration sequence, an axial vibration sequence, an acoustic emission sequence and a rotating speed data of a bearing; performing angular domain resampling on the vibration and acoustic emission sequences by using the rotating speed data, transforming the resampled angular domain sequences into two-dimensional order spectrum graphs by time-frequency analysis, and stacking the spectrum graphs corresponding to each signal along the channel dimension to construct a three-dimensional working condition information tensor; constructing a master-slave modulation heterogeneous neural network; extracting a high-dimensional spatiotemporal feature spectrum from the three-dimensional working condition information tensor through a three-dimensional convolution layer of the master branch; extracting a time series vector from the vibration and acoustic emission sequences through a one-dimensional convolution layer of the auxiliary branch, and mapping the time series vector into an affine transformation parameter set; performing a feature affine transformation operation on the high-dimensional spatiotemporal feature spectrum in the master branch based on the affine transformation parameter set, and outputting a state embedding vector; constructing a fault evolution knowledge graph containing mode state nodes, and mapping the state embedding vector into an initial node feature; performing node feature aggregation calculation on the knowledge graph by using a graph attention network, and introducing a Monte Carlo dropout mechanism to perform probability sampling prediction; calculating a probability mean value as a fault classification result based on a node posterior probability distribution output by the probability sampling prediction, and calculating a probability variance as a diagnosis confidence index.

2. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The two-dimensional order spectrum graph specifically comprises: converting the rotating speed data into an instantaneous angular velocity sequence; time-integrating the instantaneous angular velocity sequence to analyze the mapping relationship between the sampling time and the cumulative rotation phase of the rotating shaft in the sequence; predefining a fixed angle increment and generating an equidifferent angle sequence; applying the mapping relationship to the radial vibration sequence, the axial vibration sequence and the acoustic emission sequence; for the radial vibration sequence, the axial vibration sequence and the acoustic emission sequence, applying a cubic spline interpolation algorithm to reconstruct the signal amplitude at each angle point of the pre-defined equidifferent angle sequence, and generating corresponding angle domain sequences of the vibration and acoustic emission sequences sampled at equal angles; applying a short-time Fourier transform to the angle domain sequence to generate a two-dimensional complex order spectrum with the center position of the transform window as one dimension and the order as the other dimension; the order is defined as the value of the signal frequency divided by the rotating frequency of the rotating shaft; performing a modulo operation on the two-dimensional complex order spectrum to generate a two-dimensional order spectrum graph.

3. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The three-dimensional working condition information tensor is specifically constructed by: inputting the rotating speed data into a multi-layer perception network to extract a rotating speed feature vector corresponding to the channel number of the two-dimensional order spectrum graph; using tensor broadcasting technology to copy and expand the rotating speed feature vector in the time dimension and the frequency dimension to construct a rotating speed feature matrix with the same spatial size as the two-dimensional order spectrum graph; performing element-wise multiplication operation on the rotating speed feature matrix and the matrix in the two-dimensional order spectrum graph to generate a weighted time-frequency feature matrix; stacking the weighted time-frequency feature matrix along the channel dimension to construct a three-dimensional working condition information tensor containing the channel dimension, the frequency dimension and the time dimension.

4. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The constructed master-slave modulation heterogeneous neural network specifically comprises: an input layer configured to receive the three-dimensional working condition information tensor and the original vibration and acoustic emission sequence in parallel; a processing layer connected with the input layer, comprising: a master branch network architecture composed of three-dimensional convolution layers connected in series, configured to process the three-dimensional working condition information tensor and extract high-dimensional spatiotemporal feature maps; an auxiliary branch network architecture composed of one-dimensional convolution layers connected in series and linear mapping layers connected in parallel at the end, configured to process the original sequence and generate an affine transformation parameter set; the processing layer is further configured to perform a feature affine transformation operation on the high-dimensional spatiotemporal feature maps extracted by the master branch network architecture based on the affine transformation parameter set generated by the auxiliary branch network architecture; and an output layer connected with the processing layer, configured to perform global flattening processing on the feature maps processed by the feature affine transformation operation and generate and output a one-dimensional state embedding vector.

5. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The three-dimensional convolution layer of the master branch extracts high-dimensional spatiotemporal feature maps from the three-dimensional working condition information tensor, specifically comprising: inputting the three-dimensional working condition information tensor into the first three-dimensional convolution layer of the master branch network, performing three-dimensional convolution operation to extract shallow layer spatial texture features; sequentially performing batch normalization operation, nonlinear activation operation and three-dimensional maximum pooling operation on the shallow layer spatial texture features, reducing the spatial resolution of the feature map and increasing the number of feature channels layer by layer; and obtaining the tensor data output by the last convolution layer of the master branch network, which is defined as the high-dimensional spatiotemporal feature map to be modulated.

6. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The one-dimensional convolution layer of the auxiliary branch extracts a time sequence vector from the vibration and acoustic emission sequence, and maps the time sequence vector into an affine transformation parameter set, specifically comprising: channel splicing the radial vibration sequence, the axial vibration sequence and the acoustic emission sequence, inputting them into the one-dimensional convolution layer of the auxiliary branch network, and extracting local time sequence features; performing global average pooling operation on the local time sequence features to compress the convolution-processed time sequence features into a fixed-length global context feature vector; inputting the global context feature vector into two independent linear mapping layers; calculating a scaling factor vector through the first linear mapping layer and a translation factor vector through the second linear mapping layer, and combining the scaling factor vector and the translation factor vector to form an affine transformation parameter.

7. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The output state embedding vector specifically comprises: obtaining the number of channels of the high-dimensional spatiotemporal feature map; verifying that the dimensions of the scaling factor vector and the translation factor vector are consistent with the number of channels of the high-dimensional spatiotemporal feature map; performing element-wise multiplication operation on each channel feature map of the high-dimensional spatiotemporal feature map based on the scaling factor vector; performing element-wise addition operation on each channel feature map after the multiplication operation based on the translation factor vector; and performing global flattening processing on the feature map after the addition operation to output a one-dimensional state embedding vector.

8. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The constructing the fault evolution knowledge graph of the mode state node comprises: predefining a node set of the fault evolution knowledge graph, the node set comprising an observation node to be diagnosed, a normal state node, an inner ring fault node, an outer ring fault node and a rolling body fault node; predefining an edge set of the graph, the edge set establishing a full connection topological relationship between the observation node to be diagnosed and the normal state node, the inner ring fault node, the outer ring fault node and the rolling body fault node; and assigning a state embedding vector to the observation node to be diagnosed as an input node feature of the graph attention network.

9. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The node feature aggregation calculation of the knowledge graph by the graph attention network and the introduction of the Monte Carlo dropout mechanism for probability sampling prediction specifically comprise: calculating a feature correlation coefficient of connected nodes in the graph, performing a leaky linear rectified activation operation on the feature correlation coefficient, and performing normalization processing by using a Softmax function to generate an attention weight coefficient; based on the attention weight coefficient, performing weighted summation on the feature vectors of adjacent nodes to update the feature representation of the target node; enabling a dropout layer before a fully connected classification layer of the graph attention network and setting a predetermined dropout probability; keeping the network model parameters fixed, repeatedly performing a forward propagation process a predetermined number of times for the same sample, and shielding part of the neurons according to the dropout probability in each forward propagation process to generate multiple groups of different node classification probability distribution vectors.

10. The bearing fault diagnosis method based on multi-scale feature fusion according to claim 1, characterized in that, The node posterior probability distribution output based on the probability sampling prediction, the calculation of the probability mean as the fault classification result, and the calculation of the probability variance as the diagnosis confidence index specifically comprise: obtaining multiple groups of node classification probability distribution vectors generated by the probability sampling prediction, wherein each group of vectors contains probability values corresponding to each mode state node; performing an arithmetic average operation on the probability values corresponding to the same mode state node in the multiple groups of vectors, determining the mode state corresponding to the maximum value in the obtained group of average probability values as the final fault classification result; performing variance calculation on the probability values corresponding to the same mode state node in the multiple groups of vectors, and taking the mean value of the variances of all nodes as a confidence index for measuring the uncertainty of the current diagnosis result; when the confidence index exceeds a preset threshold, determining that the current diagnosis result is not reliable.

Images