CNN-BiLSTM fusion diagnosis method and device for bearing fault, equipment and medium

Abstract

The invention relates to the technical field of bearing fault diagnosis. By providing a bearing fault CNN-Bi LSTM fusion diagnosis method, device, equipment and medium, the method comprises the steps of abstracting bearing internal components into network nodes based on preset priori knowledge of a bearing mechanical structure, defining node connection edges according to a spatial constraint relationship between the components, and constructing a bearing internal component spatial topology network model; inputting the standardized bearing visual data set into a preset convolutional neural network, and extracting a visual feature vector set of surface damage of each component; and fusing the visual feature vector set and the time sequence features of the fault propagation topological graph, inputting the visual feature vector set and the time sequence features into a preset bidirectional long-short-term memory network, predicting a fault diffusion trend, positioning a fault source, and generating a fault diagnosis quantitative evaluation report. The technical effects of improving the bearing fault propagation modeling accuracy, enhancing the adaptation capability of the complex time sequence process and promoting the collaborative optimization level of the diagnosis process are achieved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a CNN-BiLSTM fusion diagnostic method, apparatus, equipment and medium for bearing fault diagnosis. Background Technology

[0002] With the deepening of industrial intelligence, high-precision bearings, as core components of high-end equipment, directly determine the reliability and production efficiency of the equipment through their operating status. Bearing fault diagnosis technology is a crucial link in ensuring the safe operation of industrial systems.

[0003] However, in traditional technologies, when using statistical feature extraction or a single neural network to model the bearing fault propagation process, the mechanical coupling relationship and spatial constraints between components are ignored, making it difficult to accurately capture the complex propagation logic of damage spreading from the initial point to adjacent components, often leading to misdiagnosis and missed diagnosis. Faced with the nonlinear time-series process of bearing faults from "initiation to spread to deterioration", traditional LSTM or ARIMA models can only handle the dependencies of a single time scale, making it difficult to adapt to complex time-series patterns such as sudden changes in damage propagation rate with operating conditions and parallel evolution of multiple fault sources, resulting in lagging trend prediction and untimely risk warning. In addition, traditional solutions separate "feature extraction, propagation analysis to trend prediction" into independent modules, lacking end-to-end collaborative optimization, causing information to attenuate in cross-module transmission, and failing to form a closed-loop reasoning process for diagnosis. Summary of the Invention

[0004] Therefore, it is necessary to provide a CNN-BiLSTM fusion diagnostic method, device, equipment and medium for bearing faults to address the above-mentioned technical problems, so as to improve the accuracy of bearing fault propagation modeling, enhance the adaptability to complex time-series processes, and promote the level of collaborative optimization of the diagnostic process.

[0005] Firstly, this application provides a CNN-BiLSTM fusion diagnostic method for bearing faults, which includes:

[0006] Multi-angle real-world images of the bearing were captured using an industrial-grade camera.

[0007] Image preprocessing is performed on multi-angle real-world images to obtain a standardized bearing visual dataset;

[0008] Based on prior knowledge of the bearing's mechanical structure, the internal components of the bearing are abstracted into network nodes, and the node connection edges are defined according to the spatial constraint relationship between the components to construct a spatial topology network model of the internal components of the bearing.

[0009] The standardized bearing visual dataset is input into a pre-defined convolutional neural network to extract a set of visual feature vectors of surface damage for each component.

[0010] The set of visual feature vectors is mapped to the spatial topology network model of the internal components of the bearing. A graph neural network is used to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map.

[0011] By integrating the temporal features of visual feature vector sets and fault propagation topology maps, and inputting them into a pre-defined bidirectional long short-term memory network, the system predicts fault propagation trends, locates fault sources, and generates a quantitative assessment report for fault diagnosis.

[0012] In one embodiment, a set of visual feature vectors is mapped to a spatial topology network model of the bearing's internal components. A graph neural network is used to analyze the damage propagation logic between components, identify the fault propagation path and its impact range, and generate a fault propagation topology map, including:

[0013] Perform node-feature semantic alignment processing on the visual feature vector set, match the node semantic labels of the bearing internal component spatial topology network model, and obtain the semantically aligned feature vector set.

[0014] Dynamic weight adaptive allocation processing is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain the topology network model after dynamic weight adjustment.

[0015] The semantically aligned feature vector set is subjected to hierarchical attention aggregation processing. The propagation logic of the damage between components is analyzed by multi-scale neighborhood information aggregation, and the propagation logic analysis result is obtained.

[0016] Based on the propagation logic analysis results, key fault propagation paths are identified through a preset dynamic threshold filtering algorithm for node importance. The influence range is calculated based on the connection weight diffusion of the topology network model after dynamic weight adjustment, and a fault propagation topology map is generated.

[0017] In one embodiment, a dynamic weight adaptive allocation process is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain a dynamically weighted topology network model, including:

[0018] Mechanical coupling strength features are extracted from the spatial topology network model of the internal components of the bearing. The mechanical coupling relationship between the components is matched by a pre-set bearing mechanical prior knowledge base to obtain a set of mechanical coupling strength features between the components.

[0019] The following formula is used to dynamically allocate connection weights based on the feature set of mechanical coupling strength between components. An attention mechanism is used to quantify the influence of coupling strength on the weights and adaptively update the weights, resulting in the updated set of connection weights:

[0020]

[0021] Among them, W final W represents the updated set of connection weights. init Let denot be the initial weight set, β be the adaptive learning rate, T be the number of update iterations, and ΔW be the value of the variable. t C represents the weight change in the t-th iteration. t This represents the measured coupling strength at the t-th iteration, where γ represents the scaling parameter of the sigmoid function;

[0022] The topology network model is reconstructed based on the updated set of connection weights, and the connection weights of the original topology network model are replaced to obtain a topology network model with dynamically adjusted weights.

[0023] In one embodiment, the temporal features of a set of visual feature vectors and a fault propagation topology map are fused, input into a preset bidirectional long short-term memory network, to predict the fault propagation trend and locate the fault source, generating a quantitative assessment report for fault diagnosis, including:

[0024] The temporal dimension alignment of the visual feature vector set and the fault propagation topology map is performed to unify the time step and fill in the missing temporal information, resulting in a temporally aligned multi-source feature set.

[0025] Spatiotemporal attention fusion processing is performed on the temporally aligned multi-source feature set. Visual temporal features and topological temporal features are weighted and fused through temporal attention mechanism and spatial attention mechanism to obtain spatiotemporal attention fused features.

[0026] The spatiotemporal attention fusion features are input into a pre-defined bidirectional long short-term memory network. The temporal dependencies are analyzed through forward LSTM layers and backward LSTM layers, and the fault propagation trend prediction sequence and fault source location probability distribution are output.

[0027] Based on the fault propagation trend prediction sequence and the probability distribution of fault source location, a fault risk quantitative assessment is performed. Combined with the preset bearing remaining life prediction model, the fault risk level and maintenance priority are calculated, and a fault diagnosis quantitative assessment report is generated.

[0028] In one embodiment, spatiotemporal attention fusion features are input into a pre-defined bidirectional long short-term memory network. Temporal dependencies are analyzed through forward and backward LSTM layers to output a fault propagation trend prediction sequence and a probability distribution of fault source locations, including:

[0029] The spatiotemporal attention fusion features are reshaped by adjusting the time step and feature dimension of the feature tensor to match the input specifications of the pre-defined bidirectional long short-term memory network, resulting in the reshaped spatiotemporal feature tensor.

[0030] The reshaped spatiotemporal feature tensor is input into the feedforward LSTM layer, and the feedforward temporal dependency features are extracted through the gating mechanism of forget gate, input gate and output gate to obtain the feedforward temporal feature sequence.

[0031] The reshaped spatiotemporal feature tensor is input into the backward LSTM layer, and the backward temporal dependency features are extracted through the gating mechanism of forget gate, input gate and output gate to obtain the backward temporal feature sequence.

[0032] A bidirectional feature attention fusion process is performed on the forward and backward temporal feature sequences. The temporal features are weighted, aggregated, and concatenated through a self-attention mechanism to obtain the fused temporal feature tensor.

[0033] Based on the fused temporal feature tensor, the fault propagation trend prediction sequence and the fault source location probability distribution are output through the fault propagation trend prediction head and the fault source location head, respectively.

[0034] In one embodiment, the semantically aligned feature vector set is obtained using the following formula, including:

[0035]

[0036] Among them, S align V represents the set of semantically aligned feature vectors. i N represents the i-th visual feature vector. i This represents the semantic label of the i-th node in the spatial topology network model of the internal components of the bearing. This represents the node feature semantic alignment mapping function, where M represents the total number of feature vectors.

[0037] In one embodiment, a hierarchical attention aggregation process is performed on the semantically aligned feature vector set. The propagation logic of the impairment between components is analyzed through multi-scale neighborhood information aggregation, yielding the propagation logic analysis result, including:

[0038] The semantically aligned feature vector set is subjected to intra-layer attention aggregation processing. The attention weight of the feature vector in a single-layer neighborhood is calculated and the neighborhood information is aggregated to obtain the intra-layer aggregated feature set.

[0039] Cross-layer attention aggregation is performed on the intra-layer aggregated feature set to pass attention weights between multi-layer topologies and aggregate cross-layer neighborhood information to obtain the cross-layer aggregated feature set.

[0040] Based on the cross-layer aggregated feature set, the propagation logic parsing unit is used to analyze the transfer patterns and dependencies of features between components, and the propagation logic parsing results are obtained.

[0041] Secondly, this application also provides a bearing fault CNN-BiLSTM fusion diagnostic device, which includes:

[0042] The image acquisition module is used to acquire multi-angle real-scene images of the bearing using an industrial-grade camera;

[0043] The image preprocessing module is used to preprocess multi-angle real-world images to obtain a standardized bearing visual dataset.

[0044] The topology construction module is used to abstract the internal components of the bearing into network nodes based on the prior knowledge of the bearing's mechanical structure, and to define the node connection edges according to the spatial constraint relationship between the components, thereby constructing a spatial topology network model of the internal components of the bearing.

[0045] The CNN feature extraction module is used to input a standardized bearing visual dataset into a pre-defined convolutional neural network to extract a set of visual feature vectors of surface damage of each component.

[0046] The propagation analysis module is used to map the set of visual feature vectors to the spatial topology network model of the internal components of the bearing, and uses graph neural networks to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map.

[0047] The trend prediction module is used to integrate the temporal features of the visual feature vector set and the fault propagation topology map, input a preset bidirectional long short-term memory network, predict the fault propagation trend and locate the fault source, and generate a fault diagnosis quantitative assessment report.

[0048] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods in the first aspect of this application.

[0049] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods in the first aspect of this application.

[0050] The bearing fault CNN-BiLSTM fusion diagnostic method, device, equipment, and medium provided in this application acquire multi-angle images using an industrial-grade camera and perform standardized processing. Combined with prior knowledge of the bearing's mechanical structure, a component spatial topology network is constructed. Visual features extracted by the CNN are mapped to the topology model. A graph neural network is used to analyze the damage propagation logic, identify the path and scope of influence, thereby improving the accuracy of fault propagation modeling. The temporal information of visual features and propagation topology is fused and input into BiLSTM to mine bidirectional temporal dependencies, effectively adapting to complex patterns such as nonlinear fault evolution and sudden changes in operating conditions, thus enhancing the adaptability to complex temporal processes. An end-to-end process from image acquisition to report generation is constructed, eliminating information attenuation caused by module splitting, achieving cross-module collaboration and diagnostic closed-loop, thereby promoting the collaborative optimization level of the diagnostic process. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a bearing fault CNN-Bi LSTM fusion diagnostic method according to one embodiment of the present invention;

[0053] Figure 2 The flowchart for the process of fusing visual feature vector set and fault propagation topology map temporal features in one embodiment of the present invention, inputting a preset bidirectional long short-term memory network, predicting fault propagation trend and locating fault source, and generating a fault diagnosis quantitative assessment report.

[0054] Figure 3 This is a structural diagram of a bearing fault CNN-Bi LSTM fusion diagnostic device according to one embodiment of the present invention. Detailed Implementation

[0055] To make the above-mentioned objects, features, and advantages of this application more apparent and understandable, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0056] First, the application scenarios of the embodiments of this application are described. In the embodiments of this application, a CNN-BiLSTM fusion diagnostic method, device, equipment, and medium for bearing condition monitoring in the field of high-end equipment manufacturing are provided. For example, in the daily operation and maintenance and fault early warning scenarios of core components such as wind turbine main shaft bearings, rail transit vehicle axle box bearings, and industrial robot precision bearings, this solution can achieve accurate monitoring and fault diagnosis of bearing operating status, ensuring stable equipment operation.

[0057] In illustrative purposes, the CNN-BiLSTM fusion diagnostic method, apparatus, equipment and medium for bearing faults provided in this application embodiment can also be applied to other application scenarios, such as health management of engine bearings in the aerospace field, fault diagnosis of bearings in heavy machinery transmission systems, and condition assessment of precision machine tool spindle bearings. These are only examples and are not intended to limit the specific application scenarios.

[0058] like Figure 1 As shown, this application provides a CNN-BiLSTM fusion diagnostic method for bearing faults, which includes:

[0059] S101: Captures multi-angle real-world images of bearings using an industrial-grade camera.

[0060] For example, the key bearing components and monitoring areas to be acquired are clearly defined, and multi-angle acquisition positions along the axial, radial, and circumferential directions are planned. Industrial-grade cameras are deployed and parameters such as resolution and exposure time are calibrated. Images are acquired according to a preset viewpoint sequence, and viewpoint markers are recorded simultaneously. Blurry or invalid images are filtered out to obtain multi-angle real-world images of the bearing.

[0061] S102: Perform image preprocessing on multi-angle real-world images to obtain a standardized bearing visual dataset.

[0062] For example, multi-angle real-world images are processed to unify their format, converting images from different perspectives into the same pixel format and color mode to eliminate format inconsistencies caused by differences in acquisition parameters. Then, geometric correction is performed to address potential perspective distortion, tilt deviations, and other issues in the images, correcting them based on the geometric characteristics of the bearing to ensure that the spatial proportions of the bearing components in the images match the actual dimensions.

[0063] Region cropping is performed, using the bearing outline as the boundary to remove irrelevant background areas from the image, focusing on the bearing body and key components. Next, pixel normalization is performed by adjusting parameters such as brightness and contrast to reduce the impact of lighting conditions on image quality and converting pixel values ​​to a uniform range. The processed images are then quality checked, and images with significant noise or missing information are removed, resulting in a standardized bearing visual dataset.

[0064] S103: Based on the prior knowledge of the bearing mechanical structure, the internal components of the bearing are abstracted into network nodes, and the node connection edges are defined according to the spatial constraint relationship between the components to construct a spatial topology network model of the internal components of the bearing.

[0065] For example, by reviewing the pre-existing knowledge of the bearing's mechanical structure, the types, functions, and assembly relationships of the core components inside the bearing are clarified, and key components that can be abstracted as network nodes are identified. Based on the spatial position of the components in the overall bearing structure, spatial constraints such as direct adjacency relationships and motion constraints between components are analyzed and extracted, which serve as the basis for defining the node connection edges.

[0066] Each core component is mapped to an independent network node. Connection edges are established for related nodes based on the identified spatial constraints, clarifying the specific spatial constraint types represented by each edge. Nodes and edges are integrated to form a structured network. By labeling the component attributes corresponding to nodes and the constraint characteristics corresponding to edges, a spatial topology network model of the bearing's internal components is constructed.

[0067] S104: Input the standardized bearing visual dataset into a pre-defined convolutional neural network to extract the set of visual feature vectors of surface damage for each component.

[0068] For example, the parameters of a pre-defined convolutional neural network are configured and its structure confirmed, clarifying the layers responsible for feature extraction and the extraction targets for surface damage features of each component. A standardized bearing visual dataset is input into the network according to a pre-defined format. Convolutional operations capture basic visual features such as local damage textures and edge contours in the images. Pooling is then used to retain key features and reduce data dimensionality. Through deep nonlinear transformations within the network, the basic features are integrated into more abstract higher-order features, focusing on damage features such as cracks, wear, and deformation on the surfaces of each component. The network output layer processes these features to generate damage feature representations corresponding to each component, resulting in a set of visual feature vectors representing surface damage for each component.

[0069] S105: Map the set of visual feature vectors to the spatial topology network model of the bearing internal components, use graph neural networks to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map.

[0070] For example, a mapping relationship is established between visual feature vectors and nodes in the spatial topology network model of the bearing's internal components. This ensures that the visual feature vector of each component is accurately associated with its corresponding node, guaranteeing semantic consistency between features and components. The mapped feature-bearing nodes and the connection edges of the topology network model are input into a pre-defined graph neural network. Graph convolution operations are used to aggregate the feature information of nodes and their neighborhoods to uncover potential correlations in damage propagation between components. Based on the node correlation strength output by the network, the main paths for damage propagation from the initial component to adjacent components are identified, and the set of affected components is determined according to the correlation range. By integrating path information and the scope of influence, the degree of node damage and the propagation strength of edges are graphically labeled to generate a fault propagation topology map.

[0071] S106: Integrates the temporal features of the visual feature vector set and the fault propagation topology map, inputs a preset bidirectional long short-term memory network, predicts the fault propagation trend and locates the fault source, and generates a quantitative assessment report for fault diagnosis.

[0072] For example, the temporal features of the visual feature vector set and the fault propagation topology map are dimensionally aligned and information is supplemented to unify their time steps, fill in any information gaps in the time series, and ensure the consistency of multi-source time series data. The aligned features are then fused using a spatiotemporal attention mechanism to strengthen key temporal information related to fault evolution and weaken redundant interference information, generating a fused temporal feature set. This fused temporal feature set is then input into a bidirectional long short-term memory network according to a preset format. The forward and backward layers are used to capture the forward and backward dependencies of the time series data, respectively, to analyze the evolutionary patterns of the fault over time.

[0073] Based on the network output results, combined with component damage characteristics and propagation path information, the future direction and rate of fault propagation are inferred to predict the fault propagation trend. Initial damaged components are identified through node feature tracing to locate the fault source. The fault propagation trend prediction results, fault source location information, and quantitative data on the damage degree of each component are integrated to generate a quantitative fault diagnosis assessment report containing fault level, risk probability, and maintenance recommendations, according to preset evaluation indicators.

[0074] One embodiment of this application provides a CNN-BiLSTM fusion diagnostic method for bearing faults. This method acquires multi-angle images from an industrial-grade camera and processes them using standardized methods. It then constructs a component spatial topology network based on prior knowledge of the bearing's mechanical structure. Visual features extracted by the CNN are mapped to the topology model. A graph neural network is used to analyze the damage propagation logic, identify the path and scope of influence, thereby improving the accuracy of fault propagation modeling. By fusing the temporal information of visual features and propagation topology, and inputting it into a BiLSTM to mine bidirectional temporal dependencies, the method effectively adapts to complex patterns such as nonlinear fault evolution and sudden changes in operating conditions, thus enhancing its adaptability to complex temporal processes. Furthermore, it constructs an end-to-end process from image acquisition to report generation, eliminating information attenuation caused by module splitting, achieving cross-module collaboration and a closed-loop diagnostic process, thereby promoting the collaborative optimization of the diagnostic process.

[0075] In one embodiment, a set of visual feature vectors is mapped to a spatial topology network model of the bearing's internal components. A graph neural network is used to analyze the damage propagation logic between components, identify the fault propagation path and its impact range, and generate a fault propagation topology map, including:

[0076] (1) Perform node-feature semantic alignment processing on the visual feature vector set, match the node semantic labels of the bearing internal component spatial topology network model, and obtain the semantically aligned feature vector set.

[0077] For example, when performing node-feature semantic alignment processing on the visual feature vector set, it is necessary to clarify the component attributes corresponding to the semantic labels of each node in the spatial topology network model of the bearing internal components, match and verify the representation content of each feature vector in the visual feature vector set with the semantic labels of the nodes, and ensure that the components described by the feature vectors are completely consistent with the components pointed to by the nodes, so as to obtain the semantically aligned feature vector set.

[0078] (2) Dynamic weight adaptive allocation processing is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain the topology network model after dynamic weight adjustment.

[0079] For example, when performing dynamic weight adaptive allocation processing on the spatial topology network model of the internal components of a bearing, it is necessary to extract the mechanical coupling strength characteristics between the components inside the bearing, determine the tightness of the connection relationship between different components based on the characteristics, and then adjust the initial weight of the connection edge of the corresponding node in the topology network model according to the tightness, so that the weight value can reflect the actual mechanical interaction strength between the components, and obtain the topology network model after dynamic weight adjustment.

[0080] (3) Perform hierarchical attention aggregation on the semantically aligned feature vector set, and analyze the propagation logic of the damage between components through multi-scale neighborhood information aggregation to obtain the propagation logic analysis result.

[0081] For example, when performing hierarchical attention aggregation on the semantically aligned feature vector set, it is necessary to divide the neighborhood range according to the hierarchical relationship of the components in the topology network, calculate the attention weight of each vector in the semantically aligned feature vector set in each level, aggregate the feature information in the neighborhood of the same level, and perform cross-level fusion of the aggregation results of different levels. Through comprehensive analysis of multi-scale neighborhood information, the rules and correlations of damage transmission between components are analyzed, and the propagation logic analysis results are obtained.

[0082] (4) Based on the propagation logic analysis results, key fault propagation paths are identified by a preset node importance dynamic threshold screening algorithm. The influence range is calculated based on the connection weight diffusion of the topology network model after dynamic weight adjustment, and a fault propagation topology map is generated.

[0083] For example, when generating a fault propagation topology map based on the propagation logic analysis results, a preset dynamic threshold filtering algorithm for node importance is used to select nodes and connections that play a dominant role in fault propagation from the propagation logic analysis results, thus determining key fault propagation paths. Based on the connection weights of the topology network model after dynamic weight adjustment, the range and degree of impact of the fault spreading from the initial node to surrounding nodes are calculated. The key fault propagation paths and their impact ranges are then graphically integrated and presented to generate the fault propagation topology map.

[0084] In one embodiment, a dynamic weight adaptive allocation process is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain a dynamically weighted topology network model, including:

[0085] (1) The mechanical coupling strength feature is extracted from the spatial topology network model of the internal components of the bearing. The mechanical coupling relationship between the components is matched by the preset bearing mechanical prior knowledge base to obtain the set of mechanical coupling strength features between the components.

[0086] For example, a pre-defined bearing mechanics prior knowledge base is invoked, which stores pre-defined mechanical interaction data between various internal components of the bearing. Based on the component type corresponding to the node in the spatial topology network model of the bearing's internal components, the mechanical coupling relationship between each component and its adjacent components is matched in the pre-defined bearing mechanics prior knowledge base, clarifying the mechanical interaction form and intensity level corresponding to different component combinations. Based on the matched mechanical coupling relationships, a feature data set reflecting the intensity of mechanical interaction between components is formed, resulting in a feature set of mechanical coupling intensity between components.

[0087] (2) Using the following formula, dynamic allocation of connection weights is performed based on the feature set of mechanical coupling strength between components. The influence factor of coupling strength on weights is quantified through an attention mechanism, and the weights are adaptively updated to obtain the updated set of connection weights:

[0088]

[0089] Among them, W final W represents the updated set of connection weights. init Let denot be the initial weight set, β be the adaptive learning rate, T be the number of update iterations, and ΔW be the value of the variable. t C represents the weight change in the t-th iteration. t γ represents the measured coupling strength at the t-th iteration, and γ represents the scaling parameter of the sigmoid function.

[0090] For example, an initial set of weights for the node connection edges in the spatial topology network model of the bearing internal components is determined, and the initial weight baseline value for each connection edge is defined. An attention mechanism is introduced, based on the measured coupling strength between each component in the feature set of mechanical coupling strength between components, to quantify the influence factor of the coupling strength on the weight of the corresponding connection edge, and to determine the correlation between coupling strength and weight adjustment range.

[0091] An adaptive learning rate and scaling parameters for the function are set to control the step size of weight updates and the scope of influence of the coupling strength factor. Through multiple iterations, in each iteration, the initial weight set is adjusted based on the measured coupling strength and weight changes, combined with the learning rate and scaling parameters, to gradually optimize the connection edge weights. After all iterations are completed, the updated connection weight set is obtained.

[0092] (3) Reconstruct the topology network model based on the updated set of connection weights, replace the connection weights of the original topology network model, and obtain the topology network model with dynamic weight adjustment.

[0093] For example, the original spatial topology network model of the bearing internal components is retrieved, and the current weight values ​​of all node connection edges in the model are determined. The new weight values ​​corresponding to each connection edge in the updated set of connection weights are then used to replace the old weight values ​​of the corresponding connection edges in the original spatial topology network model of the bearing internal components, ensuring that the weight of each connection edge is consistent with the updated result. The topology network model after weight replacement is then structurally validated to confirm that there are no conflicts or omissions in the node connection relationships and weight allocation, resulting in the dynamically weighted topology network model.

[0094] like Figure 2 As shown, the temporal features of the fusion of visual feature vector sets and fault propagation topology maps are input into a pre-defined bidirectional long short-term memory network to predict fault propagation trends and locate fault sources, generating a quantitative fault diagnosis assessment report, including:

[0095] S201: Perform temporal dimension alignment processing on the visual feature vector set and the temporal features of the fault propagation topology map, unify the time step and fill in the missing temporal information to obtain the temporally aligned multi-source feature set.

[0096] For example, time stamps corresponding to feature data in the visual feature vector set and time stamps corresponding to changes in each node of the fault propagation topology map are extracted respectively. Using a preset unified time interval as a benchmark, the time stamps of the two types of features are adjusted to a consistent time step to ensure the matching of the two types of features in the time dimension. For data with missing time steps in the time-series features of the visual feature vector set or the fault propagation topology map, reasonable completion is performed based on the feature change patterns of adjacent valid time steps to avoid subsequent fusion deviations due to missing information. The two types of features are integrated to obtain a time-aligned multi-source feature set.

[0097] S202: Perform spatiotemporal attention fusion processing on the multi-source feature set after temporal alignment. By weighted fusion of visual temporal features and topological temporal features through temporal attention mechanism and spatial attention mechanism, spatiotemporal attention fusion features are obtained.

[0098] For example, a temporal attention mechanism is activated to analyze the importance of visual temporal features and topological temporal features at different time steps in the temporally aligned multi-source feature set, assigning higher attention weights to features at key time steps. Simultaneously, a spatial attention mechanism is activated to focus on component feature regions related to bearing faults in the temporally aligned multi-source feature set, assigning higher attention weights to spatial features reflecting fault correlations. Combining the temporal and spatial attention weights, a weighted calculation is performed on the visual and topological temporal features, effectively integrating the key information of both types of features to obtain spatiotemporal attention fusion features.

[0099] S203: The spatiotemporal attention fusion feature is input into a pre-defined bidirectional long short-term memory network. The temporal dependency is analyzed through forward LSTM and backward LSTM layers, and the fault propagation trend prediction sequence and fault source location probability distribution are output.

[0100] For example, the input format requirements of the preset bidirectional long short-term memory network are confirmed, and the spatiotemporal attention fusion features are adjusted to feature tensors that conform to the network input specifications. The adjusted spatiotemporal attention fusion features are then input into the feedforward LSTM layer of the preset bidirectional long short-term memory network. Through the forget gate, input gate, and output gate mechanism of the feedforward LSTM layer, the temporal dependencies in the features are extracted from the forward temporal dimension to capture the pattern of fault evolution over time.

[0101] Simultaneously, the spatiotemporal attention fusion features are input into the backward LSTM layer of a pre-defined bidirectional long short-term memory network. Through the gating mechanism of the backward LSTM layer, the temporal dependencies in the features are extracted from the reverse temporal dimension to supplement the correlation information for fault reverse tracing. Combining the processing results of the forward and backward LSTM layers, the network output layer outputs a fault propagation trend prediction sequence and a probability distribution of the fault source location.

[0102] S204: Based on the fault propagation trend prediction sequence and the probability distribution of fault source location, perform fault risk quantitative assessment and processing, combine the preset bearing remaining life prediction model to calculate the fault risk level and maintenance priority, and generate a fault diagnosis quantitative assessment report.

[0103] For example, the fault propagation trend prediction sequence is analyzed to determine the future propagation direction and rate of the fault among various bearing components. The probability distribution of the fault source location is analyzed, and the component with the highest probability value is selected as a candidate fault source. A preset bearing remaining life prediction model is invoked, and the fault propagation trend data and candidate fault source information are input into the model. Combining the preset bearing life decay law and fault influence coefficient in the model, the risk level of the current fault on the overall operation of the bearing is calculated, and the fault risk level is determined.

[0104] Based on the fault risk level and the importance ranking of each bearing component, maintenance priorities are further determined. The results of fault propagation trend prediction, fault source location information, fault risk level, and maintenance priorities are integrated and compiled into a standardized document according to a pre-set report template, generating a quantitative assessment report for fault diagnosis.

[0105] In one embodiment, spatiotemporal attention fusion features are input into a pre-defined bidirectional long short-term memory network. Temporal dependencies are analyzed through forward and backward LSTM layers to output a fault propagation trend prediction sequence and a probability distribution of fault source locations, including:

[0106] (1) The spatiotemporal attention fusion features are reshaped by adjusting the time step and feature dimension of the feature tensor to match the input specifications of the preset bidirectional long short-term memory network, and the reshaped spatiotemporal feature tensor is obtained.

[0107] For example, the input specifications of a predefined bidirectional long short-term memory (LSTM) network are obtained, specifying the required feature tensor time step range and the number of feature dimensions. Based on the input specifications, the current time step sequence of the spatiotemporal attention fusion features is split or merged, while adjusting the arrangement and number of feature dimensions to ensure that the adjusted feature tensor structure fully conforms to the predefined bidirectional LSM network's acceptance criteria. After adjustment, the reshaped spatiotemporal feature tensor is obtained.

[0108] (2) Input the reshaped spatiotemporal feature tensor into the feedforward LSTM layer, and extract the feedforward temporal dependency features through the gating mechanism of forget gate, input gate and output gate to obtain the feedforward temporal feature sequence.

[0109] For example, when the reshaped spatiotemporal feature tensor is input into the feedforward LSTM layer, the feedforward LSTM layer uses a forget gate to filter historical feature information in the reshaped spatiotemporal feature tensor that is irrelevant to the forward temporal dependency analysis, discarding redundant data. New feature information from the reshaped spatiotemporal feature tensor is received through the input gate, and the new information is filtered and updated using an activation function, incorporating effective new information into the cell state. Then, the output gate generates the output features of the feedforward LSTM layer based on the current cell state, and the dependencies between features are extracted step-by-step in forward temporal order, forming a sequence reflecting the correlation of features in the forward temporal dimension, thus obtaining the forward temporal feature sequence.

[0110] (3) Input the reshaped spatiotemporal feature tensor into the backward LSTM layer, and extract the backward temporal dependency features through the gating mechanism of forget gate, input gate and output gate to obtain the backward temporal feature sequence.

[0111] For example, when the reshaped spatiotemporal feature tensor is input into the backward LSTM layer, the backward LSTM layer filters out future feature information in the reshaped spatiotemporal feature tensor that is irrelevant to backward temporal dependency analysis through a forget gate, removing invalid data. The input gate receives feature information passed from the end of time to the beginning of time in the reshaped spatiotemporal feature tensor, and the activation function filters and updates the information, integrating key information into the cell state. Then, the output gate generates the output features of the backward LSTM layer based on the current cell state, extracts the dependencies between features in reverse temporal order, forming a sequence reflecting the correlation of features in the reverse temporal dimension, thus obtaining the backward temporal feature sequence.

[0112] (4) Perform bidirectional feature attention fusion processing on the forward temporal feature sequence and the backward temporal feature sequence, and weight aggregate the temporal features through the self-attention mechanism and splice them to obtain the fused temporal feature tensor.

[0113] For example, a self-attention mechanism is initiated, calculating the attention weights of features at each time step in the forward temporal feature sequence and the attention weights of features at each time step in the backward temporal feature sequence. The weights depend on the importance of the features to the fault analysis. The forward and backward temporal feature sequences are then weighted and aggregated according to their attention weights, strengthening key feature information and weakening secondary feature information. The weighted and aggregated forward and backward temporal features are then concatenated along the time dimension to ensure consistent correspondence between the two types of features in the time dimension, resulting in a fused temporal feature tensor.

[0114] (5) Based on the fusion of temporal feature tensors, the fault propagation trend prediction sequence and the fault source location probability distribution are output through the fault propagation trend prediction head and the fault source location head, respectively.

[0115] For example, the fused temporal feature tensor is input into a preset fault propagation trend prediction head. This prediction head, through preset feature mapping and classification logic, converts the fused temporal feature tensor into sequence data reflecting the propagation state of the fault at different time points, and outputs a fault propagation trend prediction sequence. The fused temporal feature tensor is then input into a preset fault source localization head. This localization head, through preset probability calculation and distribution modeling logic, analyzes the feature information related to the fault source in the fused temporal feature tensor and generates probability values ​​for each bearing component as a fault source.

[0116] In one embodiment, the semantically aligned feature vector set is obtained using the following formula, including:

[0117]

[0118] Among them, S align V represents the set of semantically aligned feature vectors. iN represents the i-th visual feature vector. i This represents the semantic label of the i-th node in the spatial topology network model of the internal components of the bearing. This represents the node feature semantic alignment mapping function, where M represents the total number of feature vectors.

[0119] For example, the i-th visual feature vector V to be processed is defined. i Simultaneously, obtain the spatial topology network model of the internal components of the bearing and V. i The semantic label N of the corresponding i-th node i This prepares the basic data for the semantic alignment operation. The node feature semantic alignment mapping function is called. The semantic label N of the i-th node i As input, generate N i Semantic matching features represent the relationship between node semantic labels and feature space.

[0120] Calculate the i-th visual feature vector V i Alignment mapping function with node feature semantics The L2 distance between the feature representations obtained after processing reflects V i The degree of difference between the feature vectors and the feature representations generated based on node semantic labels is assessed. The above-described operation of calculating the L2 distance is performed sequentially on all M feature vectors, and these distances are summed to obtain the total distance value, which comprehensively reflects the overall difference between all visual feature vectors and the feature representations generated from the corresponding node semantic labels.

[0121] By finding the set of semantically aligned feature vectors S that minimizes the total distance value. align This achieves optimal matching between visual feature vectors and semantic labels of nodes in the spatial topology network model of bearing internal components, resulting in a semantically aligned set of feature vectors.

[0122] In one embodiment, a hierarchical attention aggregation process is performed on the semantically aligned feature vector set. The propagation logic of the impairment between components is analyzed through multi-scale neighborhood information aggregation, yielding the propagation logic analysis result, including:

[0123] (1) Perform intra-layer attention aggregation on the semantically aligned feature vector set, calculate the attention weight of the feature vector in the single-layer neighborhood and aggregate neighborhood information to obtain the intra-layer aggregated feature set.

[0124] For example, when performing intra-layer attention aggregation on the semantically aligned feature vector set, it is necessary to define the single-layer neighborhood range corresponding to each feature vector based on the direct connection relationship of the components in the spatial topology network model of the bearing internal components, ensuring that this range only includes adjacent components that have a direct physical connection with the target component. For each feature vector in the semantically aligned feature vector set, the attention weight of all adjacent feature vectors in its single-layer neighborhood is calculated by combining the component damage feature intensity reflected by the feature vector with the degree of connection between the components in the neighborhood.

[0125] Based on the calculated attention weights, the feature vectors in a single-layer neighborhood are weighted and fused to aggregate key information related to damage propagation within the neighborhood, filter out irrelevant interference information, and complete the intra-layer aggregation operation of all feature vectors to obtain an intra-layer aggregated feature set.

[0126] (2) Perform cross-layer attention aggregation processing on the intra-layer aggregated feature set, pass attention weights between multi-layer topologies and aggregate cross-layer neighborhood information to obtain the cross-layer aggregated feature set.

[0127] For example, when performing cross-layer attention aggregation processing on the intra-layer aggregated feature set, it is necessary to divide the topology into multiple layers according to the functional hierarchy of the internal components of the bearing, and clarify the functional relationships and information transmission paths between components at different levels. A cross-layer attention weight calculation mechanism is established, and the cross-layer attention weights of feature vectors at different levels in the intra-layer aggregated feature set are calculated based on the indirect interaction relationships between components at different levels.

[0128] Based on cross-layer attention weights, intra-layer aggregated features are transmitted between multi-layer topologies to achieve bidirectional interaction and aggregation of feature information between upper and lower layers. Feature information related to damage propagation within the cross-layer neighborhood is integrated to complete feature aggregation between multi-layer topologies and obtain a cross-layer aggregated feature set.

[0129] (3) Based on the cross-layer aggregated feature set, the transfer pattern and dependency relationship of features between components are analyzed through the preset propagation logic parsing unit to obtain the propagation logic parsing result.

[0130] For example, a preset propagation logic parsing unit is activated. This unit contains the basic laws and feature analysis rules for the propagation of bearing component damage. The cross-layer aggregated feature set is input into the preset propagation logic parsing unit. The pattern recognition module within the unit analyzes the order of occurrence and feature change patterns of damage features among different components in the cross-layer aggregated features, and identifies the transfer patterns of damage features.

[0131] By using the correlation analysis module within a unit to mine the mutual influence relationships of feature vectors of different components in cross-layer aggregation features, the damage dependency relationships between components are determined. By combining the transfer patterns of damage features with the dependencies between components, a complete analysis of the damage propagation path, propagation mode, and impact correlations between components is formed, yielding the propagation logic analysis results.

[0132] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0133] In one embodiment, such as Figure 3 As shown, this application also provides a bearing fault CNN-BiLSTM fusion diagnostic device 300, which includes:

[0134] Image acquisition module 301 is used to acquire multi-angle real-scene images of the bearing using an industrial-grade camera;

[0135] Image preprocessing module 302 is used to preprocess multi-angle real-scene images to obtain a standardized bearing visual dataset;

[0136] Topology building module 303 is used to abstract the internal components of the bearing into network nodes based on the prior knowledge of the bearing mechanical structure, and to define the node connection edges according to the spatial constraint relationship between the components, thereby constructing a spatial topology network model of the internal components of the bearing.

[0137] The CNN feature extraction module 304 is used to input the standardized bearing visual dataset into a preset convolutional neural network to extract a set of visual feature vectors of surface damage of each component.

[0138] The propagation analysis module 305 is used to map the set of visual feature vectors to the spatial topology network model of the internal components of the bearing, and uses a graph neural network to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map.

[0139] The trend prediction module 306 is used to fuse the temporal features of the visual feature vector set and the fault propagation topology map, input a preset bidirectional long short-term memory network, predict the fault propagation trend and locate the fault source, and generate a fault diagnosis quantitative assessment report.

[0140] Specifically, when the image acquisition module 301 acquires multi-angle real-scene images of the bearing using an industrial-grade camera, it needs to plan the shooting angles covering the key components of the bearing, control the stability of the ambient lighting, and ensure that the acquired images can clearly present the surface state of the components, thereby obtaining multi-angle real-scene images of the bearing.

[0141] When the image preprocessing module 302 preprocesses the multi-angle real-scene images, it performs denoising and enhancement operations to optimize the clarity of damage features, and then adjusts the image specifications through size unification and pixel standardization to eliminate data differences and obtain a standardized bearing visual dataset.

[0142] When constructing the spatial topology network model of the bearing internal components, the topology construction module 303 abstracts internal components such as balls and raceways into network nodes based on the preset prior knowledge of the bearing mechanical structure, defines the node connection edges according to the spatial constraint relationship between components, and generates a topology network model that can represent the spatial association of components.

[0143] The CNN feature extraction module 304 inputs the standardized bearing visual dataset into a preset convolutional neural network. The convolutional layer captures local damage features, the pooling layer filters key information, and the fully connected layer integrates features, outputting a set of visual feature vectors of surface damage for each component.

[0144] The propagation analysis module 305 semantically aligns the set of visual feature vectors with the spatial topology network model of the bearing internal components, analyzes the damage propagation logic through graph neural networks, identifies the fault propagation path and the scope of influence, and generates a fault propagation topology map.

[0145] The trend prediction module 306 aligns the visual feature vector set with the temporal features of the fault propagation topology map, inputs a bidirectional long short-term memory network to analyze the temporal dependency, predicts the fault propagation trend and locates the fault source, and generates a quantitative assessment report for fault diagnosis by combining risk assessment rules.

[0146] The propagation parsing module 305 is also used for:

[0147] Perform node-feature semantic alignment processing on the visual feature vector set, match the node semantic labels of the bearing internal component spatial topology network model, and obtain the semantically aligned feature vector set.

[0148] Dynamic weight adaptive allocation processing is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain the topology network model after dynamic weight adjustment.

[0149] The semantically aligned feature vector set is subjected to hierarchical attention aggregation processing. The propagation logic of the damage between components is analyzed by multi-scale neighborhood information aggregation, and the propagation logic analysis result is obtained.

[0150] Based on the propagation logic analysis results, key fault propagation paths are identified through a preset dynamic threshold filtering algorithm for node importance. The influence range is calculated based on the connection weight diffusion of the topology network model after dynamic weight adjustment, and a fault propagation topology map is generated.

[0151] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0152] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A CNN-BiLSTM fusion diagnostic method for bearing faults, characterized in that, The method includes: Multi-angle real-world images of the bearing were captured using an industrial-grade camera. The multi-angle real-scene images are preprocessed to obtain a standardized bearing visual dataset; Based on prior knowledge of the bearing's mechanical structure, the internal components of the bearing are abstracted into network nodes, and the node connection edges are defined according to the spatial constraint relationship between the components to construct a spatial topology network model of the internal components of the bearing. The standardized bearing visual dataset is input into a preset convolutional neural network to extract a set of visual feature vectors of surface damage of each component. The set of visual feature vectors is mapped to the spatial topology network model of the internal components of the bearing. A graph neural network is used to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map. By integrating the temporal features of the visual feature vector set and the fault propagation topology map, and inputting them into a preset bidirectional long short-term memory network, the fault propagation trend is predicted and the fault source is located, generating a quantitative assessment report for fault diagnosis.

2. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 1, characterized in that, The step of mapping the set of visual feature vectors to the spatial topology network model of the bearing's internal components, using a graph neural network to analyze the damage propagation logic between components, identifying the fault propagation path and its impact range, and generating a fault propagation topology map includes: Perform node-feature semantic alignment processing on the visual feature vector set, and match the node semantic labels of the spatial topology network model of the bearing internal component to obtain the semantically aligned feature vector set. Dynamic weight adaptive allocation processing is performed on the spatial topology network model of the internal components of the bearing. The connection weights are adjusted according to the mechanical coupling strength between the components to obtain the topology network model after dynamic weight adjustment. The semantically aligned feature vector set is subjected to hierarchical attention aggregation processing. The propagation logic of the damage between components is analyzed by multi-scale neighborhood information aggregation, and the propagation logic analysis result is obtained. Based on the propagation logic analysis results, key fault propagation paths are identified using a preset node importance dynamic threshold filtering algorithm. The influence range is calculated based on the connection weight diffusion of the dynamically weighted topology network model, and the fault propagation topology map is generated.

3. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 2, characterized in that, The step of performing dynamic weight adaptive allocation processing on the spatial topology network model of the bearing internal components, adjusting the connection weights according to the mechanical coupling strength between components, to obtain a dynamically weighted topology network model includes: The mechanical coupling strength feature extraction process is performed on the spatial topology network model of the internal components of the bearing. The mechanical coupling relationship between the components is matched by a preset bearing mechanical prior knowledge base to obtain the set of mechanical coupling strength features between the components. The following formula is used to dynamically allocate connection weights based on the set of mechanical coupling strength features between components. An attention mechanism is used to quantify the influence of coupling strength on the weights and adaptively update the weights, resulting in the updated set of connection weights: Among them, W final W represents the updated set of connection weights. init Let denot be the initial weight set, β be the adaptive learning rate, T be the number of update iterations, and ΔW be the value of the variable. t C represents the weight change in the t-th iteration. t This represents the measured coupling strength at the t-th iteration, where γ represents the scaling parameter of the sigmoid function; The updated set of connection weights is used to reconstruct the topology network model, replacing the connection weights of the original topology network model to obtain the dynamically weighted topology network model.

4. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 1, characterized in that, The process involves fusing the temporal features of the visual feature vector set and the fault propagation topology map, inputting them into a preset bidirectional long short-term memory network, predicting the fault propagation trend and locating the fault source, and generating a quantitative assessment report for fault diagnosis, including: The time-series features of the visual feature vector set and the fault propagation topology map are aligned in time dimension, the time step is unified and the missing time information is filled in, and a multi-source feature set after time alignment is obtained. The temporally aligned multi-source feature set is subjected to spatiotemporal attention fusion processing. Visual temporal features and topological temporal features are weighted and fused through temporal attention mechanism and spatial attention mechanism to obtain spatiotemporal attention fusion features. The spatiotemporal attention fusion features are input into the preset bidirectional long short-term memory network. The temporal dependencies are analyzed through forward LSTM layers and backward LSTM layers, and the fault propagation trend prediction sequence and fault source location probability distribution are output. Based on the fault propagation trend prediction sequence and the fault source location probability distribution, a fault risk quantification assessment is performed. Combined with a preset bearing remaining life prediction model, the fault risk level and maintenance priority are calculated, and the fault diagnosis quantification assessment report is generated.

5. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 4, characterized in that, The process of inputting the spatiotemporal attention fusion features into the preset bidirectional long short-term memory network, analyzing temporal dependencies through forward LSTM layers and backward LSTM layers, and outputting a fault propagation trend prediction sequence and a fault source location probability distribution includes: The spatiotemporal attention fusion features are reshaped by adjusting the time step and feature dimension of the feature tensor to match the input specifications of the preset bidirectional long short-term memory network, thus obtaining the reshaped spatiotemporal feature tensor. The reshaped spatiotemporal feature tensor is input into the feedforward LSTM layer, and the feedforward temporal dependency features are extracted through the gating mechanism of forget gate, input gate and output gate to obtain the feedforward temporal feature sequence. The reshaped spatiotemporal feature tensor is input into the backward LSTM layer, and the backward temporal dependency features are extracted through the gating mechanism of forget gate, input gate and output gate to obtain the backward temporal feature sequence. The forward temporal feature sequence and the backward temporal feature sequence are subjected to bidirectional feature attention fusion processing. The temporal features are weighted and aggregated through a self-attention mechanism and then concatenated to obtain a fused temporal feature tensor. Based on the fused temporal feature tensor, the fault propagation trend prediction sequence and the fault source location probability distribution are output through the fault propagation trend prediction head and the fault source location head, respectively.

6. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 2, characterized in that, The semantically aligned feature vector set is obtained using the following formula, including: Among them, S align V represents the set of semantically aligned feature vectors. i N represents the i-th visual feature vector. i This represents the semantic label of the i-th node in the spatial topology network model of the internal components of the bearing. This represents the node feature semantic alignment mapping function, where M represents the total number of feature vectors.

7. The bearing fault CNN-BiLSTM fusion diagnostic method according to claim 2, characterized in that, The hierarchical attention aggregation process is applied to the semantically aligned feature vector set. Through multi-scale neighborhood information aggregation, the propagation logic of the impairment between components is analyzed to obtain the propagation logic analysis result, including: The semantically aligned feature vector set is subjected to intra-layer attention aggregation processing. The attention weight of the feature vector in a single-layer neighborhood is calculated and the neighborhood information is aggregated to obtain the intra-layer aggregated feature set. Cross-layer attention aggregation processing is performed on the intra-layer aggregated feature set to transfer attention weights between multi-layer topologies and aggregate cross-layer neighborhood information to obtain a cross-layer aggregated feature set. Based on the cross-layer aggregated feature set, the transfer patterns and dependencies of features between components are analyzed through a preset propagation logic parsing unit to obtain the propagation logic parsing result.

8. A CNN-BiLSTM fusion diagnostic device for bearing faults, characterized in that, The device includes: The image acquisition module is used to acquire multi-angle real-scene images of the bearing using an industrial-grade camera; The image preprocessing module is used to preprocess the multi-angle real-scene images to obtain a standardized bearing visual dataset. The topology construction module is used to abstract the internal components of the bearing into network nodes based on the prior knowledge of the bearing's mechanical structure, and to define the node connection edges according to the spatial constraint relationship between the components, thereby constructing a spatial topology network model of the internal components of the bearing. The CNN feature extraction module is used to input the standardized bearing visual dataset into a preset convolutional neural network to extract a set of visual feature vectors of surface damage of each component. The propagation analysis module is used to map the set of visual feature vectors to the spatial topology network model of the internal components of the bearing, and to use a graph neural network to analyze the propagation logic of damage between components, identify the fault propagation path and the scope of influence, and generate a fault propagation topology map. The trend prediction module is used to fuse the temporal features of the visual feature vector set and the fault propagation topology map, input a preset bidirectional long short-term memory network, predict the fault propagation trend and locate the fault source, and generate a fault diagnosis quantitative assessment report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the bearing fault CNN-BiLSTM fusion diagnostic method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the bearing fault CNN-BiLSTM fusion diagnostic method as described in any one of claims 1 to 7.

Images