Remaining service life prediction method based on graph neural network and time-frequency analysis

Abstract

The invention relates to a residual service life prediction method based on a graph neural network and time-frequency analysis, which is applied to the technical field of equipment health management and predictive maintenance, and comprises the following steps: obtaining original sensor signal data of a target sensor; performing feature selection and recombination on the original sensor signal data to generate sensor time sequence features; performing graph structure conversion on the sensor time sequence characteristics to obtain a space-time diagram; an attention mechanism and multiplication interaction are introduced into the space-time diagram, and a processing space-time diagram is generated; performing spatio-temporal feature processing on the processing spatio-temporal diagram based on a preset processing rule to obtain global spatio-temporal representation; and predicting the service life of the target sensor based on the global spatio-temporal representation. The method has the effect of effectively capturing local and global space-time dependence so as to improve the accuracy of residual life prediction.

Description

Technical Field

[0001] The present application relates to the technical field of equipment health management and predictive maintenance, and in particular to a method for predicting remaining useful life based on graph neural networks and time-frequency analysis. Background Art

[0002] With the increasing deployment of sensors in industrial equipment and the growing demand for operational monitoring, predicting the remaining useful life of sensors has become a crucial component of predictive maintenance. Existing deep learning-based methods, such as CNNs and LSTMs, can effectively extract temporal dependencies but struggle to explicitly model the spatial correlations between sensors. While graph neural network-based methods introduce spatial structure, they often rely on predefined static graphs and are unable to adapt to the dynamically changing sensor relationships that occur during equipment degradation.

[0003] While graph attention networks dynamically model node relationships through an attention mechanism, they still have significant shortcomings. For example, attention scoring functions often employ linear transformations or shallow MLPs, which have limited nonlinear expression capabilities and are difficult to characterize high-order nonlinear coupling between sensors in complex industrial scenarios. Furthermore, most methods employ a "time-space separation" processing framework, making it difficult to effectively and uniformly capture the evolving dependencies of sensors across timestamps. Furthermore, at the feature extraction level, traditional methods often lack the collaborative utilization of multi-scale time-frequency features, making it impossible to simultaneously capture long-term degradation trends and short-term transient changes, limiting the completeness of degradation representation.

[0004] Therefore, existing methods have significant bottlenecks in dynamic dependency modeling, nonlinear relationship expression and multi-scale feature fusion. There is an urgent need for a technology that can effectively capture local and global spatiotemporal dependencies to improve the accuracy of remaining life prediction. Summary of the Invention

[0005] In order to effectively capture local and global spatiotemporal dependencies to improve the accuracy of remaining useful life prediction, this application provides a remaining useful life prediction method based on graph neural network and time-frequency analysis.

[0006] In the first aspect, the present application provides a method for predicting remaining useful life based on graph neural networks and time-frequency analysis, which adopts the following technical solutions: A method for predicting remaining useful life based on graph neural network and time-frequency analysis, including: Obtaining raw sensor signal data of the target sensor; Perform feature selection and reorganization on the original sensor signal data to generate sensor time series features; Perform graph structure conversion on the sensor time series features to obtain a spatiotemporal graph; Introducing attention mechanism and multiplication interaction on the spatiotemporal graph to generate and process the spatiotemporal graph; performing spatiotemporal feature processing on the spatiotemporal graph based on a preset processing rule to obtain a global spatiotemporal representation; predicting the service life of the target sensor based on the global spatiotemporal representation.

[0007] By adopting the above technical solutions, the time sequence features are converted into a unified graph structure, the limitations of traditional pre-defined static graphs are abandoned, the dynamic changes in spatial correlation and complex dependence across timestamps between sensors in different running states and degradation stages can be adaptively captured, the representation ability of the model for the evolving degradation mode in a real industrial scene is greatly improved, the multiplication interaction and attention mechanism are introduced to replace the traditional shallow MLP scoring function, the expression bottleneck of linear or simple nonlinear transformation is broken, the original signals are subjected to multi-scale decomposition and adaptive reconstruction, the one-sidedness of simple time domain or frequency domain analysis is overcome, through spatiotemporal feature processing, efficient compression of local salient features and effective fusion of global spatiotemporal dependence are realized, the generalization ability of the model is enhanced, and thus the accuracy and reliability of the remaining life prediction are improved.

[0008] Optionally, the feature selection and reorganization of the original sensor signal data to generate sensor time sequence features comprises: normalizing the original sensor signal data to generate processed sensor data; performing fixed-length segmentation processing on the processed sensor data based on a preset segmentation length to generate a plurality of time windows; performing discrete wavelet transformation processing on each of the time windows and introducing a frequency modulation gating mechanism to obtain feature components; performing feature reconstruction on the feature components based on a preset inverse wavelet transformation to generate reconstructed feature components; performing original time sequence preservation processing on the reconstructed feature components by using residual connection to obtain sensor time sequence features.

[0009] Optionally, the graph structure conversion of the sensor time sequence features to obtain a spatiotemporal graph comprises: dividing the sensor time sequence features into a plurality of overlapping windows according to a sliding window; determining spatial edges, temporal edges and cross-sensor time edges based on the overlapping windows; generating a spatiotemporal graph based on the spatial edges, the temporal edges and the cross-sensor time edges.

[0010] Optionally, the introduction of an attention mechanism and a multiplication interaction on the spatiotemporal graph to generate a processed spatiotemporal graph comprises: inputting node features in the spatiotemporal graph into the attention mechanism for transformation processing to output sub-node features; The child node features are converted through the multiplication interaction to obtain converted child node features; The converted child node features are introduced into three types of regularization terms for complex regularization term processing to generate a processing result; The processing result is subjected to attention weight calculation and normalization processing to obtain a processing spatio-temporal graph.

[0011] Optionally, the inputting of the node features in the spatio-temporal graph into the attention mechanism for transformation processing and outputting of child node features comprises: The node features are input into a MultKAA module in the attention mechanism, and B-spline basis functions are used as scoring functions in a KAN layer for transformation to obtain output child node features.

[0012] Optionally, the spatio-temporal feature processing of the processing spatio-temporal graph based on a preset processing rule to obtain a global spatio-temporal representation comprises: A sliding window is used to traverse the processing spatio-temporal graph to generate a plurality of local spatio-temporal subgraphs; A MultKAA module is used to perform graph convolution on the local spatio-temporal subgraphs to update node features and generate an update result; The update result is used to generate a global spatio-temporal representation.

[0013] Optionally, the generation of the global spatio-temporal representation based on the update result comprises: An SAGPOOL graph pooling operation is applied to the update result to generate a compressed graph structure; The compressed graph structure is subjected to merging processing to generate a global graph; Final graph convolution is performed on the global graph to extract global spatio-temporal dependency relationships to generate a global spatio-temporal representation.

[0014] In a second aspect, the present application provides a residual useful life prediction device based on a graph neural network and time-frequency analysis, which adopts the following technical solution: A residual useful life prediction device based on a graph neural network and time-frequency analysis comprises: A signal data acquisition module is configured to acquire original sensor signal data of a target sensor; A time sequence feature generation module is configured to perform feature selection and reorganization on the original sensor signal data to generate sensor time sequence features; A time sequence feature conversion module is configured to perform graph structure conversion on the sensor time sequence features to obtain a spatio-temporal graph; A spatio-temporal data processing module is configured to introduce an attention mechanism and a multiplication interaction into the spatio-temporal graph to generate a processing spatio-temporal graph; A global representation generation module is configured to perform spatio-temporal feature processing on the processing spatio-temporal graph based on a preset processing rule to obtain a global spatio-temporal representation. A service life prediction module is used to predict the service life of the target sensor based on the global spatio-temporal representation.

[0015] By adopting the technical scheme, the time sequence features are converted into a unified graph structure, the limitation of a traditional pre-defined static graph is abandoned, the dynamic changes in space correlation and complex dependence across timestamps between sensors in different running states and degradation stages are adaptively captured, the representation ability of the model for an evolving degradation mode in a real industrial scene is greatly improved, the multiplication interaction and attention mechanism are introduced to replace a traditional shallow MLP scoring function, the expression bottleneck of linear or simple nonlinear transformation is broken, the original signals are subjected to multi-scale decomposition and adaptive reconstruction, the one-sidedness of simple time domain or frequency domain analysis is overcome, through spatio-temporal feature processing, efficient compression of local salient features and effective fusion of global spatio-temporal dependence are realized, the generalization ability of the model is enhanced, and the accuracy and reliability of the remaining life prediction are improved.

[0016] In a third aspect, the present application provides an electronic device, which adopts the following technical scheme: An electronic device includes a processor coupled with a memory; The processor is configured to execute a computer program stored in the memory, so that the electronic device executes the computer program of the remaining service life prediction method based on the graph neural network and the time-frequency analysis of any one of the first aspect.

[0017] In a fourth aspect, the present application provides a computer readable storage medium, which adopts the following technical scheme: A computer readable storage medium stores a computer program capable of being loaded and executed by a processor to execute the remaining service life prediction method based on the graph neural network and the time-frequency analysis of any one of the first aspect. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of a remaining service life prediction method based on a graph neural network and time-frequency analysis provided by an embodiment of the present application.

[0019] Figure 2 is a structural diagram of a graph neural network provided by an embodiment of the present application.

[0020] Figure 3 is a flowchart of a MultKAA attention mechanism provided by an embodiment of the present application.

[0021] Figure 4 is a structural block diagram of a remaining service life prediction device based on a graph neural network and time-frequency analysis provided by an embodiment of the present application.

[0022] Figure 5This is a structural block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] The present application is further described in detail below with reference to the accompanying drawings.

[0024] The present application provides a method for predicting remaining useful life based on a graph neural network and time-frequency analysis. This method can be executed by an electronic device, which can be a server or a terminal device. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smartphone, a tablet computer, a desktop computer, etc., but is not limited thereto.

[0025] Figure 1 A flowchart of a method for predicting remaining useful life based on graph neural networks and time-frequency analysis is provided in an embodiment of the present application.

[0026] like Figure 1 As shown, the main process of the method is described as follows (steps S101 to S106): Step S101: acquiring raw sensor signal data of a target sensor.

[0027] In this embodiment, the original sensor signal data of the target sensor for which the remaining service life prediction is required is collected, and the original sensor signal data and the Figure 2 The graph neural network shown in FIG1 performs data processing to predict the remaining service life, wherein the original sensor signal data is the data generated by the sensors installed in the aircraft engine, including temperature signal data, pressure signal data, speed signal data, etc.

[0028] Step S102 : performing feature selection and reorganization on the original sensor signal data to generate sensor time series features.

[0029] For step S102, the original sensor signal data is normalized to generate processed sensor data; the processed sensor data is segmented into fixed lengths based on a preset segmentation length to generate multiple time windows; each time window is subjected to discrete wavelet transform processing and a frequency modulation gating mechanism is introduced to obtain characteristic components; the characteristic components are reconstructed based on a preset inverse wavelet transform to generate reconstructed characteristic components; the reconstructed characteristic components are subjected to original time series preservation processing using residual connections to obtain sensor time series features.

[0030] In this example, the C-MAPSS aircraft engine degradation dataset is used as input. This dataset contains multiple sensor signals reflecting the engine's operating status under different operating conditions and degradation modes. The raw sensor signal data is first normalized to generate processed sensor data. This processed sensor data is then segmented into fixed-length segments according to a preset segment length, resulting in multiple time windows. Each time window is converted into a multidimensional representation that simultaneously reflects both fine-grained transient states and coarse-grained degradation trends over time.

[0031] Apply a trainable discrete wavelet transform to each time window: ; in, is the input signal segment; For press Scale and press The wavelet basis function of the translation amount; is the convolution filter, is the convolution operator; is element-wise multiplication; is the Hadamard product; is a nonlinear activation function; is the trainable projection matrix.

[0032] After discrete wavelet transform, a frequency modulation gating mechanism is introduced to adaptively weight the decomposed components, emphasizing the low-frequency components that represent long-term degradation trends and selectively retaining the high-frequency components that represent short-term performance changes to obtain the characteristic components.

[0033] The characteristic components are reconstructed using a trainable inverse discrete wavelet transform to obtain the reconstructed characteristic components. The trainable inverse discrete wavelet transform is: ; in, is the detail coefficient of the first layer decomposition, is the approximate coefficient of the first-level decomposition, is the inverse wavelet filter corresponding to the detail coefficient, is the inverse wavelet filter corresponding to the approximate coefficients.

[0034] Add residual connections to the reconstructed feature components, which retain the original structure and input the module Directly add it to the reconstructed output to obtain the sensor timing characteristics: ; in, is the module input, Reconstructed output.

[0035] In step S103, the sensor time sequence features are converted into a graph structure to obtain a space-time graph.

[0036] In step S103, the sensor time sequence features are divided into multiple overlapping windows according to a sliding window; the spatial edges, the time edges and the cross-sensor time edges are determined based on the overlapping windows; and the space-time graph is generated based on the spatial edges, the time edges and the cross-sensor time edges.

[0037] In this embodiment, each time window is mapped into a space-time graph, and the nodes in the graph represent the sensor segment features. The sensor time sequence features output by the frequency modulation gating wavelet module are divided into multiple overlapping windows, and the space-time graph is generated based on the sensor time sequence features in each window. The window division is performed, where N is the number of sensors, T is the number of time steps, the window length f and the step size are , and N overlapping windows are generated . Each window has N sensor patches ; the m-th sensor segment in each time window is subjected to nonlinear mapping to obtain a node feature vector ; and a node set is formed . , where the node is defined as the m-th sensor in the window , and the feature is ; in each time window , a spatial edge set is constructed according to the feature correlation of the sensors ; a time edge set is constructed between the same sensor across windows ; a cross-time cross-sensor edge set is constructed between different sensors across windows ; that is, the spatial correlation edges represent the spatial dependence of the sensors in the same window ; the time correlation edges represent the time dependence of the same sensor between windows ; and the cross-sensor time dependence edges represent the cross-time stamp dependence relationship between different sensors between different windows. All the spatial, time and cross-time cross-sensor edge sets are combined into a complete space-time graph: . . Step S104 introduces a notice mechanism and a multiplication interaction on the space-time graph to generate a processed space-time graph. .

[0038] Step S104 introduces a notice mechanism and a multiplication interaction on the space-time graph to generate a processed space-time graph.

[0039] ​​​For step S104, the node features in the spatio-temporal graph are input into the attention mechanism for transformation processing, and sub-node features are output; the sub-node features are converted through multiplication interaction to obtain converted sub-node features; the converted sub-node features are introduced into three types of regularization terms for complex regularization term processing to generate a processing result; the processing result is subjected to attention weight calculation and normalization processing to obtain a processed spatio-temporal graph.

[0040] Further, the node features in the spatio-temporal graph are input into the attention mechanism for transformation processing, and sub-node features are output, including: inputting the node features into the MultKAA module in the attention mechanism, transforming through the KAN layer using the B-spline basis function as the scoring function to obtain output sub-node features.

[0041] In this embodiment, the node features in the spatio-temporal graph are subjected to feature alignment , to obtain paired input features, which are input into the MultKAA module as shown in Figure 3 , the first layer input is the node is transformed through the KAN (Kolmogorov-Arnold Network) layer to obtain the output , that is, in the MultKAA, the input node features are processed through the KAN layer to obtain output sub-node features , the KAN layer uses a nonlinear mapping based on the B-spline basis function to replace the traditional attention scoring function, and then performs conversion through the multiplication interaction layer to capture high-order nonlinear coupling, the MultKAA structure is multi-layer cascade, each layer is composed of a standard KAN layer and a multiplication layer

[0042] satisfies : ; wherein: ; After conversion, the multi-layer MultKAA is combined, and the multi-layer MultKAA is represented as: ; To suppress overfitting and enhance structural sparsity, three types of regularization terms are introduced on the converted sub-node features for complex regularization terms: constraint B-spline coefficient sparsity; balance feature channel utilization rate; The constrained multiplicative interaction weights only preserve physically important sensor couplings, resulting in a processed result, and finally attention weights are computed: ; where, is the feature alignment, the unnormalized attention scores are computed by a linear transformation and a LeakyReLU activation, and then normalized using a softmax over neighboring nodes, and the normalized attention weights are computed by the following equation: .

[0043] Step S105, based on the preset processing rule, the spatio-temporal graph is processed to obtain a global spatio-temporal representation.

[0044] For step S105, a sliding window traversal is used to process the spatio-temporal graph to generate a plurality of local spatio-temporal subgraphs; the MultKAA module is used to update the node features of the local spatio-temporal subgraphs to generate an update result; and the update result is used to generate a global spatio-temporal representation.

[0045] Further, generating a global spatio-temporal representation based on the update result includes: applying a SAGPOOL graph pooling operation to the update result to generate a compressed graph structure; merging the compressed graph structure to generate a global graph; and performing a final graph convolution on the global graph to extract global spatio-temporal dependency relationships to generate a global spatio-temporal representation.

[0046] In this embodiment, a time sliding mechanism with a window size of 2 and a step size of 1 is used to divide and process the processed spatio-temporal graph to generate a plurality of local spatio-temporal subgraphs. On each local subgraph, a graph convolution is performed based on the MultKAA attention weight to update the node features to generate an update result. The graph convolution based on the MultKAA attention weight is: ; where W is a trainable weight matrix.

[0047] The update result is applied to a SAGPOOL graph pooling operation to reduce the complexity of the graph by selecting top-k nodes based on the saliency score, thereby generating a compressed high-level representation to obtain a compressed graph structure. All the pooled local features are merged into a global graph, i.e., all the compressed graph structures are merged to generate a global graph. A final graph convolution is performed on the global graph to capture global dependency patterns across time and sensors, extract global spatio-temporal dependency relationships, and generate a final representation Step S106, based on the global spatio-temporal representation, the service life of the target sensor is predicted.

[0048] In this embodiment, the final graph convolution is performed on the global space-time, or the readout function is directly used, the features of the entire graph are aggregated into a single vector, and the final target sensor residual service life value is input into the full connection layer and regressed.

[0049] Figure 4 A structure block diagram of a residual service life prediction device 200 based on a graph neural network and time-frequency analysis is provided for the application embodiment.

[0050] As shown in Figure 4 The residual service life prediction device 200 based on the graph neural network and the time-frequency analysis mainly includes: The signal data acquisition module 201 is configured to acquire original sensor signal data of a target sensor. The time sequence feature generation module 202 is configured to perform feature selection and reorganization on the original sensor signal data to generate sensor time sequence features. The time sequence feature conversion module 203 is configured to perform graph structure conversion on the sensor time sequence features to obtain a space-time graph. The space-time data processing module 204 is configured to introduce a attention mechanism and a multiplication interaction on the space-time graph to generate a processed space-time graph. The global representation generation module 205 is configured to perform space-time feature processing on the processed space-time graph based on a preset processing rule to obtain a global space-time representation. The service life prediction module 206 is configured to predict the service life of the target sensor based on the global space-time representation.

[0051] As an optional implementation manner of the present embodiment, the time sequence feature generation module 202 is specifically configured to perform normalization processing on the original sensor signal data to generate processed sensor data, perform fixed-length segmentation processing on the processed sensor data based on a preset segmentation length to generate a plurality of time windows, introduce a frequency modulation gating mechanism after performing discrete wavelet transformation processing on each time window to obtain feature components, perform feature reconstruction on the feature components based on a preset inverse wavelet transformation to generate reconstructed feature components, and perform original time sequence preservation processing on the reconstructed feature components by using residual connection to obtain sensor time sequence features.

[0052] As an optional implementation manner of the present embodiment, the time sequence feature conversion module 203 is specifically configured to divide the sensor time sequence features into a plurality of overlapping windows according to a sliding window, determine a spatial edge, a time edge and a cross-sensor time edge based on the overlapping windows, and generate a space-time graph based on the spatial edge, the time edge and the cross-sensor time edge.

[0053] As an optional implementation manner of the present embodiment, the time sequence feature conversion module 203 includes: The node feature output module is configured to input the node feature in the spatio-temporal graph into the attention mechanism for transformation processing, and output a sub-node feature. The node feature conversion module is configured to convert the sub-node feature through multiplication interaction to obtain a converted sub-node feature. The processing result generation module is configured to introduce the converted sub-node feature into three types of regularization terms for complex regularization term processing to generate a processing result. The result calculation processing module is configured to perform attention weight calculation and normalization processing on the processing result to obtain a processed spatio-temporal graph.

[0054] In this optional embodiment, the node feature conversion module is specifically configured to input the node feature into a MultKAA module in the attention mechanism, and transform the node feature through a KAN layer using a B-spline basis function as a scoring function to obtain an output sub-node feature.

[0055] As an optional embodiment of this embodiment, the global representation generation module 205 includes: The local sub-graph generation module is configured to perform sliding window traversal processing on the spatio-temporal graph to generate a plurality of local spatio-temporal sub-graphs. The update result generation module is configured to perform graph convolution on the local spatio-temporal sub-graphs based on the MultKAA module to update the node features and generate an update result. The spatio-temporal representation generation module is configured to generate a global spatio-temporal representation based on the update result.

[0056] In this optional embodiment, the spatio-temporal representation generation module is specifically configured to apply a SAGPOOL graph pooling operation to the update result to generate a compressed graph structure, perform merging processing on the compressed graph structure to generate a global graph, perform final graph convolution on the global graph to extract global spatio-temporal dependency relationships, and generate a global spatio-temporal representation.

[0057] In one example, the modules in any of the above apparatuses can be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASICs), or one or more digital signal processors (DSPs), or one or more field programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0058] For example, when the modules in the apparatus can be implemented in the form of a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can invoke a program. For another example, these modules can be integrated together to be implemented in the form of a system-on-a-chip (SOC).

[0059] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the apparatus and modules described above can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0060] Figure 5 The structural block diagram of the electronic device 300 provided in the embodiments of the present application is shown.

[0061] As shown in Figure 5 The electronic device 300 includes a processor 301 and a memory 302, and can further include one or more of an information input / output (I / O) interface 303, a communication component 304, and a communication bus 305.

[0062] The processor 301 is configured to control the overall operation of the electronic device 300 to complete all or part of the steps of the remaining useful life prediction method based on the graph neural network and time-frequency analysis described above; the memory 302 is configured to store various types of data to support the operation of the electronic device 300, which can include, for example, instructions for operating any application or method on the electronic device 300, and application-related data. The memory 302 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as one or more of a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0063] The I / O interface 303 provides an interface between the processor 301 and other interface modules, which can be a keyboard, a mouse, a button, etc. These buttons can be virtual buttons or physical buttons. The communication component 304 is configured to perform wired or wireless communication between the electronic device 300 and other devices. The wireless communication, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them, so the corresponding communication component 104 can include a Wi-Fi component, a Bluetooth component, an NFC component.

[0064] The electronic device 300 can be implemented by one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic elements for performing the remaining useful life prediction method based on graph neural network and time-frequency analysis given in the above embodiments.

[0065] The communication bus 305 can include a path for transmitting information between the above components. The communication bus 305 can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 305 can be divided into an address bus, a data bus, a control bus, etc.

[0066] The electronic device 300 can include, but is not limited to, a mobile terminal of a mobile phone, a notebook computer, a digital broadcast receiver, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a PMP (Portable Multimedia Player), a car terminal (e.g., a car navigation terminal), etc., and a fixed terminal such as a digital TV, a desktop computer, etc., and can be a server, etc.

[0067] The present application also provides a computer readable storage medium, and the computer readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of the remaining useful life prediction method based on graph neural network and time-frequency analysis are implemented.

[0068] The computer readable storage medium can include a U disk, a mobile hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and the like various media that can store program codes.

[0069] The terms "comprising", "containing" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0070] The above description is merely preferred embodiments of the present application and a description of the principles of the technology used. Those skilled in the art should understand that the scope of the application involved in the present application is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combinations of the above technical features or their equivalent features without departing from the above application concept. For example, the above features are replaced with the technical features applied in the present application (but not limited to) having similar functions to form technical solutions.

Claims

1. A method for predicting remaining useful life based on graph neural network and time-frequency analysis, characterized in that: include: Acquiring raw sensor signal data of the target sensor; Performing feature selection and reorganization on the original sensor signal data to generate sensor time series features; Performing graph structure conversion on the sensor time series features to obtain a spatiotemporal graph; Introducing an attention mechanism and multiplication interaction on the spatiotemporal graph to generate a processing spatiotemporal graph; Performing spatiotemporal feature processing on the processed spatiotemporal graph based on preset processing rules to obtain a global spatiotemporal representation; The service life of the target sensor is predicted based on the global spatiotemporal representation.

2. The method according to claim 1, characterized in that The performing feature selection and reorganization on the original sensor signal data to generate sensor time series features includes: performing normalization processing on the raw sensor signal data to generate processed sensor data; The processed sensor data is segmented into fixed lengths based on a preset segmentation length to generate a plurality of time windows; After performing discrete wavelet transform processing on each of the time windows, a frequency modulation gating mechanism is introduced to obtain characteristic components; Performing feature reconstruction on the feature component based on a preset inverse wavelet transform to generate a reconstructed feature component; Residual connection is used to perform original time series preservation processing on the reconstructed feature components to obtain sensor time series features.

3. The method according to claim 2, characterized in that The converting the sensor time series features into a graph structure to obtain a spatiotemporal graph includes: Dividing the sensor time series features into multiple overlapping windows according to the sliding window; determining a spatial edge, a temporal edge, and a cross-sensor temporal edge based on the overlapping windows; A spatiotemporal graph is generated based on the spatial edges, the temporal edges, and the cross-sensor temporal edges.

4. The method according to claim 1, wherein The introducing of the attention mechanism and the multiplication interaction on the spatiotemporal graph to generate and process the spatiotemporal graph includes: Inputting node features in the spatiotemporal graph into the attention mechanism for transformation processing, and outputting sub-node features; The sub-node features are transformed through the multiplication interaction to obtain transformed sub-node features; Introducing the conversion sub-node features into three types of regularization terms to perform composite regularization term processing to generate a processing result; Attention weight calculation and normalization processing are performed on the processing results to obtain a processing spatiotemporal graph.

5. The method according to claim 4, characterized in that The node features in the spatiotemporal graph are input into the attention mechanism for transformation processing, and the output sub-node features include: The node features are input into the MultKAA module in the attention mechanism, and transformed by the KAN layer using the B-spline basis function as the scoring function to obtain the output sub-node features.

6. The method according to claim 5, characterized in that The performing spatiotemporal feature processing on the processed spatiotemporal graph based on preset processing rules to obtain a global spatiotemporal representation includes: Using a sliding window to traverse the processing space-time graph to generate multiple local space-time subgraphs; Perform graph convolution on the local spatiotemporal subgraph based on the MultKAA module to update node features and generate an updated result; A global spatiotemporal representation is generated based on the updated result.

7. The method according to claim 6, characterized in that Generating a global spatiotemporal representation based on the update result includes: Applying a SAGPOOL graph pooling operation to the updated result to generate a compressed graph structure; Merging the compressed graph structures to generate a global graph; A final graph convolution is performed on the global graph to weight the global spatiotemporal dependencies and generate a global spatiotemporal representation.

8. A device for predicting remaining useful life based on graph neural network and time-frequency analysis, characterized in that: include: A signal data acquisition module, configured to acquire raw sensor signal data of the target sensor; A time series feature generation module, configured to perform feature selection and reorganization on the original sensor signal data to generate sensor time series features; A time series feature conversion module, configured to convert the sensor time series features into a graph structure to obtain a time-space graph; a spatiotemporal data processing module, configured to introduce an attention mechanism and multiplication interaction on the spatiotemporal graph to generate a processed spatiotemporal graph; A global representation generation module, configured to perform spatiotemporal feature processing on the processed spatiotemporal graph based on preset processing rules to obtain a global spatiotemporal representation; A service life prediction module is used to predict the service life of the target sensor based on the global spatiotemporal representation.

9. An electronic device, characterized in that: comprising a processor coupled to a memory; The processor is configured to execute the computer program stored in the memory, so that the electronic device performs the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The method comprises a computer program or an instruction, which, when executed on a computer, causes the computer to execute the method according to any one of claims 1 to 7.

Images