Equipment life prediction method and system

Abstract

The invention discloses an equipment fault prediction method and system, and belongs to the technical field of electric data processing. The method comprises the following steps: inputting state information of each component of equipment into a conditional generative adversarial network to generate time series data, and constructing a plurality of target maps with adjacency relations; time features and local graph features are extracted by using multi-branch convolution and a graph convolution structure of a convolution auto-encoder, and target graph features are obtained through splicing; performing channel weighting through a multi-scale attention module to obtain multi-scale image features; calculating cross attention weights among different target images and fusing the cross attention weights to obtain fused image features; and compressing the potential feature representation as a health index, constructing a health index time sequence, inputting the health index time sequence into a fault prediction model, and obtaining an estimated residual life of the equipment. According to the method, strong noise interference can be effectively filtered out, the problem of feature redundancy is solved, key degradation information is captured robustly, the equipment health state is accurately represented, and the residual life prediction precision is remarkably improved.

Description

Equipment life prediction methods and systems Technical Field

[0001] This application relates to the field of electrical data processing technology, and in particular to a method and system for predicting equipment lifespan. Background Technology

[0002] In the field of autonomous vehicle applications, condition-based maintenance relies on lifespan prediction of critical components. Existing technologies primarily extract features by analyzing monitoring data collected by sensors, such as vibration and temperature. However, due to the complex and variable operating environment of autonomous vehicles, monitoring signals often exhibit strong noise interference, multi-scale characteristics, and feature redundancy. Traditional feature extraction methods often struggle to comprehensively and robustly capture this critical degradation information hidden within complex data streams. This results in insufficient representation of the equipment's health status by the extracted features, failing to clearly map the change process from health to failure, thus severely limiting the accuracy of subsequent remaining lifespan prediction. Summary of the Invention

[0003] This application provides a method and system for predicting equipment lifespan. The technical solution is as follows:

[0004] In a first aspect, embodiments of this application provide a method comprising: inputting state information of various components of a device into a conditional generative adversarial network to generate component time-series data; constructing multiple target graphs with adjacency relationships based on the component time-series data; inputting the target graphs into a convolutional autoencoder; extracting time features at different scales through the multi-branch convolutional structure of the convolutional autoencoder; extracting local graph features at different scales through the graph convolutional structure of the convolutional autoencoder; concatenating the time features and the local graph features to obtain target graph features; inputting the target graph features into a multi-scale attention module; performing channel weighting on the target graph features to obtain multi-scale graph features; calculating cross-attention weights between different target graphs based on the multi-scale graph features; fusing the target graphs based on the cross-attention weights to obtain fused graph features; compressing the fused graph features into a latent feature representation; using the latent feature representation as a health indicator characterizing the degradation state of a component; arranging the health indicators extracted from the same component at different time points in chronological order to form a health indicator time series; and inputting the health indicator time series into a fault prediction model to obtain the estimated remaining lifespan of the device.

[0005] In one possible implementation, constructing multiple target graphs with adjacency relationships based on the component time-series data includes: aligning the component time-series data with corresponding operating condition labels to construct a multimodal time-series data vector; determining at least two neighborhood parameter values ​​using cluster analysis based on the distance between the multimodal time-series data vectors, wherein the neighborhood parameter values ​​characterize the range of the number of neighboring nodes considered by a node when constructing the target graph; and constructing the target graph based on the neighborhood parameter values, wherein each node in the target graph represents the component state at a given time point, and the edges of each target graph are determined by the distance between nodes in the multimodal time-series data vectors under the neighborhood parameter values.

[0006] In one possible implementation, the step of inputting the target image into a convolutional autoencoder (CAE), extracting temporal features at different scales through the multi-branch convolutional structure of the CAE, extracting local graph features at different scales through the graph convolutional structure of the CAE, and concatenating the temporal features and the local graph features to obtain the target image features includes: converting the target image into a node feature sequence; inputting the node feature sequence into the multi-branch convolutional structure of the CAE, extracting temporal features at different scales through multiple convolutional branches with different kernel sizes; inputting the target image into the graph convolutional structure of the CAE, aggregating node neighborhood information and extracting local graph features at different scales through processing by two graph convolutional layers; mapping the temporal features and the local graph features to a preset feature dimension, concatenating them along the channels of the feature dimension to generate concatenated features, and then performing feature dimensionality reduction and compression to obtain the target image features.

[0007] In one possible implementation, the step of inputting the target graph into the graph convolutional structure of the convolutional autoencoder, and aggregating node neighborhood information and extracting local graph features at different scales through processing by two graph convolutional layers, includes: inputting the target graph into the first graph convolutional layer of the graph convolutional structure, aggregating the feature information of the first-order neighbor nodes of the current node based on a preset adjacency matrix, and extracting local node association features; inputting the local node association features into the second graph convolutional layer, aggregating the feature information within the second-order neighbor range of the node, and extracting wide-area topological features of the node; and fusing the local node association features with the wide-area topological features of the node to obtain the local graph features at different scales.

[0008] In one possible implementation, the step of inputting the target map features into a multi-scale attention module and performing channel weighting on the target map features to obtain multi-scale map features includes: dividing the target map features into several sub-feature groups; performing global average pooling and convolution on each sub-feature group to obtain a first sub-feature and a second sub-feature; normalizing all the first sub-features and the second sub-features to obtain a first attention weight matrix and a second attention weight matrix; fusing the first attention weight matrix and the second attention weight matrix to obtain a fused attention weight; and multiplying the fused attention weight by the target map features to obtain the multi-scale map features.

[0009] In one possible implementation, the step of calculating cross-attention weights between different target maps based on the multi-scale map features, and fusing the target maps based on the cross-attention weights to obtain fused map features, includes: selecting multi-scale map features of a first target map as query features, and using multi-scale map features of the remaining target maps as key features; calculating the similarity between the query features and each key feature to obtain cross-attention weights and normalizing the cross-attention weights; using the normalized cross-attention weights to perform a weighted summation of the key features to obtain weighted features; and performing a residual connection between the weighted features and the query features to obtain the fused map features.

[0010] In one possible implementation, compressing the fused graph features into a latent feature representation and using the latent feature representation as a health indicator characterizing the component degradation state includes: extracting features from the fused graph features to obtain semantic features for describing the failure mode; smoothing the semantic features to obtain smoothed features; and vectorizing the smoothed features to obtain a fixed-length latent feature representation as a health indicator characterizing the component degradation state.

[0011] In one possible implementation, the step of inputting the health indicator time series into a fault prediction model to obtain the estimated remaining lifespan of the device includes: preprocessing the health indicator time series to convert it into an input tensor that the fault prediction model can recognize; using the fault prediction model to perform time dependency analysis on the input tensor to extract time series features reflecting degradation trends; inputting the time series features into the regression layer of the fault prediction model for mapping calculation, and outputting the estimated remaining lifespan representing the remaining usage time of the device.

[0012] Secondly, embodiments of this application provide a system including a storage medium and one or more processors; the storage medium stores a computer program, which is invoked by one or more processors to implement any of the above-described device lifetime prediction methods.

[0013] The technical solution provided in this application has at least the following beneficial effects: This application effectively solves the problem that existing technologies are unable to capture key degradation information in complex data by constructing a target graph and combining multi-scale convolution and graph convolution to extract features. It uses a multi-scale attention mechanism to adaptively weight features and combines cross-attention to fuse multi-view information, which significantly enhances the model's ability to perceive degradation trends. The health index obtained by compression can sensitively and stably characterize the equipment status. Combined with a time series prediction model, it greatly improves the accuracy and robustness of fault prediction and realizes accurate decision-making for condition-based maintenance. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments 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.

[0015] Figure 1 is a schematic flowchart of a device life prediction method provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0017] Referring to Figure 1, this application embodiment provides a device lifetime prediction method, which includes the following steps: Step 101: Input the state information of each component of the device into a conditional generative adversarial network to generate component time series data, and construct multiple target graphs with adjacency relationships based on the component time series data.

[0018] In real-world scenarios of autonomous vehicle operation, raw data collected by sensors often suffers from issues such as missing data, inconsistent sampling rates, or severe interference from environmental noise, failing to directly meet the high-precision input data quality requirements for fault prediction. This step first inputs the collected component status information, such as bearing vibration signals and battery voltage data, into a conditional generative adversarial network (GAN). Leveraging the powerful generative distribution capabilities of the GAN, combined with operational condition labels, the raw data is enhanced and repaired to generate high-quality component time-series data that accurately reflects component degradation trends.

[0019] Optionally, the status information of each component of the device can be input into the conditional generative adversarial network to generate component time-series data. This can include: firstly, collecting the original status information of key components in real time through various deployed sensors, aligning the operating condition labels with the status information in time, constructing multimodal input data pairs, and then sending the multimodal input data pairs as conditional inputs into the conditional generative adversarial network. The generator can then learn the data distribution pattern of the components under specific operating conditions.

[0020] In conditional generative adversarial networks (GANs), guided by operating conditions, the network learns and generates time-series data of components that accurately reflect their degradation trends. Simultaneously, a discriminator distinguishes between the generated data and the real data. Through adversarial training between the generator and discriminator, the generative network ultimately generates high-quality time-series data of components that conforms to actual degradation patterns, providing a high-quality input foundation for subsequent graph construction.

[0021] Taking bearings as an example, during vehicle operation, vibration signals are continuously collected by an accelerometer mounted on the bearing housing at a high-frequency sampling rate, such as 25.6 kHz. The 2560 vibration amplitude points collected within a time window, such as 0.1 seconds, are arranged chronologically to form a set of component time-series data reflecting the bearing's current health status. This data contains dynamic degradation information of the bearing as operating conditions change.

[0022] Subsequently, multiple target graphs with adjacency relationships are constructed based on the generated component time-series data. In one possible implementation, this step may include: aligning the component time-series data with corresponding operating condition labels to construct a multimodal time-series data vector; determining at least two neighborhood parameter values ​​using cluster analysis based on the distances between the multimodal time-series data vectors, wherein the neighborhood parameter values ​​characterize the range of the number of neighboring nodes considered by a node when constructing the target graph; and constructing the target graph based on the neighborhood parameter values, wherein each node in the target graph represents the component state at a given time point, and the edges of each target graph are determined by the distances between nodes in the multimodal time-series data vectors under the neighborhood parameter values.

[0023] This step constructs multiple target graphs with different adjacency relationships by building multimodal time-series data vectors and determining multiple neighborhood parameters using cluster analysis. This approach can characterize the evolution of component states from different time scales and topological perspectives, capturing both local fine-grained fluctuations and preserving long-term macroscopic degradation trends. The multi-view graph structure effectively enhances the model's adaptability to complex operating conditions, improves the richness and robustness of key features extracted from time-series data, and lays a solid data structure foundation for subsequent high-precision feature extraction and fault prediction.

[0024] Optionally, in the above steps, the collected raw component time-series data can first be preprocessed. For example, for vibration signals of mechanical components, in order to eliminate noise and highlight degradation features, a Fast Fourier Transform can be performed to convert the signal from the time domain to the frequency domain. Then, the processed feature data is time-aligned with the corresponding operating condition labels, such as rotational speed and load. Through feature concatenation or embedding fusion, a multimodal time-series data vector at time t is constructed. .

[0025] Cluster analysis is used to define the connectivity of the graph structure to accommodate the multi-scale characteristics of the data distribution. First, the distances between multimodal time-series data vectors are calculated. Assume the vectors at two time points i and j are... and The distance is calculated using the Euclidean distance metric, specifically the L2 norm. :

[0026] Based on the calculated distance matrix, clustering algorithms, such as K-Means, are used to analyze the local density or distribution characteristics of the samples, thereby determining the center points of at least two clustering results as neighborhood parameter values ​​k. Smaller k is used to capture closely adjacent local dependencies, while larger k is used to cover a wider time range to capture macro trends.

[0027] Construct multiple target graphs with different topological structures based on determined neighborhood parameters. Each node This represents the component state at a given time point. For each neighborhood parameter value k, an edge set is constructed. The rule is: if node j is located in the set of node i's k nearest neighbors, then an edge is established, and the weight of the edge is... for: Where exp represents the exponential function operation, and γ is a preset hyperparameter. It is the set of the k nearest neighbors of node i, selected based on the neighborhood parameter value k.

[0028] Step 102: Input the target image into the convolutional autoencoder, extract temporal features at different scales through the multi-branch convolutional structure of the convolutional autoencoder, extract local image features at different scales through the graph convolutional structure of the convolutional autoencoder, and concatenate the temporal features and the local image features to obtain the target image features.

[0029] In one possible implementation, this step may include: converting the target graph into a node feature sequence; inputting the node feature sequence into the multi-branch convolutional structure of the convolutional autoencoder; extracting temporal features at different scales through multiple convolutional branches with different kernel sizes; inputting the target graph into the graph convolutional structure of the convolutional autoencoder; aggregating node neighborhood information and extracting local graph features at different scales through processing by two graph convolutional layers; mapping the temporal features and the local graph features to a preset feature dimension; concatenating and splicing them along the channels of the feature dimension to generate concatenated features; and then performing feature dimensionality reduction and compression to obtain the target graph features.

[0030] This scheme extracts temporal features at different scales using a multi-branch convolutional structure, capturing short-term fluctuations and long-term trends in component degradation. Simultaneously, it leverages graph convolutional structures to aggregate neighborhood information, effectively uncovering local topological dependencies between nodes. By cascading and fusing temporal and graph features, complementary multimodal features are achieved. This multi-scale fusion mechanism comprehensively characterizes the complex degradation patterns of components from both the temporal and topological domains, significantly enhancing the expressive power and robustness of features, and providing a high-dimensional basis for accurate subsequent health status assessment and remaining lifespan prediction.

[0031] Optionally, the step of converting the target image into a node feature sequence and inputting the node feature sequence into the multi-branch convolutional structure of the convolutional autoencoder, extracting temporal features at different scales through multiple convolutional branches with different kernel sizes, includes: rearranging the nodes in the target image according to the timestamp index order to construct a node feature sequence that can reflect the evolution of the component state over time; inputting the node feature sequence into parallel multi-branch convolutional structures respectively, and using different-sized temporal convolutional kernels configured in each convolutional branch to perform sliding window convolution operations in the time dimension to obtain output results to characterize the signal change patterns under different time spans; and performing nonlinear activation processing on the output results to extract multi-scale temporal features that can characterize the short-term fluctuations and long-term trends of the component.

[0032] This approach effectively captures short-term fluctuations and long-term degradation trends in component states by rearranging nodes into time series and extracting features in parallel using multi-sized convolutional kernels. This multi-branch temporal convolution mechanism can comprehensively analyze signal patterns from different temporal granularities, significantly enhancing the model's ability to perceive temporal degradation information and providing a rich and in-depth feature foundation for subsequent accurate assessment of component health status.

[0033] Optionally, the step of inputting the target graph into the graph convolutional structure of the convolutional autoencoder, and aggregating node neighborhood information and extracting local graph features at different scales through processing by two graph convolutional layers, includes: inputting the target graph into the first graph convolutional layer of the graph convolutional structure, aggregating the feature information of the first-order neighbor nodes of the current node based on a preset adjacency matrix, and extracting local node association features; inputting the local node association features into the second graph convolutional layer, aggregating the feature information within the second-order neighbor range of the node, and extracting wide-area topological features of the node; and fusing the local node association features with the wide-area topological features of the node to obtain the local graph features at different scales.

[0034] This scheme utilizes a two-layer graph convolutional structure to effectively capture local associations and wide-area topological dependencies between nodes by aggregating first-order and second-order neighbor information. This multi-scale graph feature extraction mechanism can simultaneously preserve microscopic details and global structural information, enhancing the ability to mine spatial relationships within the system, thereby more comprehensively and accurately representing the topological structural features of components.

[0035] Optionally, the step of mapping the temporal features and the local graph features to a preset feature dimension, cascading and concatenating them along the channels of the feature dimension to generate cascaded features, and then performing feature dimensionality reduction and compression to obtain the target graph features may include: inputting the temporal features and the local graph features into two independent fully connected layers respectively, and uniformly mapping them to a preset feature dimension to ensure the consistency of features in the channel space; cascading and concatenating the mapped temporal features and local graph features along the feature channel dimension to fuse temporal information and topological structure information to generate cascaded features containing rich degradation information; and removing redundant information and compressing the feature dimension of the cascaded features through feature dimensionality reduction operations to obtain the target graph features that can comprehensively characterize the health status of the component.

[0036] This scheme achieves feature dimension alignment through fully connected layer mapping, eliminating scale differences between temporal and topological information and ensuring effective fusion. By combining temporal degradation trends with spatial structure dependencies through channel cascading operations, and then removing redundant interference through dimensionality reduction and compression, the final generated target graph features contain comprehensive state information and possess high-dimensional abstract representation capabilities, significantly improving the accuracy and robustness of the model's representation of component health status.

[0037] Optionally, in one specific implementation, in order to capture degradation patterns at different time granularities, a method is set... The sequence represents the rearranged node features, where M is the number of convolutional branches, and the output of the m-th convolutional branch is... as follows:

[0038] Where ReLU is the activation function. The weights are the kernel weights, and * represents the sliding window convolution operation. It is the bias adjustment amount for each convolution branch.

[0039] To uncover topological dependencies between nodes, this scheme employs a two-layer graph convolutional propagation mechanism. Let... Let A be the initial node feature matrix and A be the preset adjacency matrix. The forward propagation formula for the two-layer graph convolution is as follows: The first layer graph convolution is used for local feature extraction: ,in, It is the output feature matrix of the first graph convolutional layer; It is a non-linear activation function; It is an adjacency matrix with self-loops, where The I identity matrix is ​​used to ensure that nodes retain their own feature information when aggregating neighbor information, thus preventing the loss of their own features during multi-layer propagation. yes The corresponding degree matrix; It is the learnable weight matrix of the first graph convolutional layer.

[0040] The second layer of graph convolution is used for wide-area feature extraction: ,in, ϕ is the output feature matrix of the second graph convolutional layer; ϕ is the non-linear activation function. yes The corresponding degree matrix; This is the learnable weight matrix of the second graph convolutional layer. It can characterize the feature information within the second-order neighbor range of an aggregate node.

[0041] Then, by concatenating the two features mentioned above along their feature dimensions, local map features at different scales can be obtained.

[0042] Step 103: Input the target image features into the multi-scale attention module, perform channel weighting on the target image features, and obtain multi-scale image features.

[0043] In one possible implementation, this step may include dividing the target map features into several sub-feature groups. This step divides the high-dimensional feature channels into multiple small blocks, which reduces computational cost compared to processing all channels together, while allowing the attention mechanism to independently mine information in different feature subspaces, increasing the diversity of captured features.

[0044] Global average pooling and convolution are performed on each of the sub-feature groups to obtain the first sub-feature and the second sub-feature. Global average pooling is used to collect global statistical information at all spatial locations within the group to obtain the first sub-feature, which focuses on describing the overall background; convolution captures the local correlations and cross-references between different channels within the group to obtain the second sub-feature, which focuses on describing local details.

[0045] Normalization is performed on all the first and second sub-features to obtain the first attention weight matrix and the second attention weight matrix. After normalization, the abstract feature values ​​are converted into weight coefficients between 0 and 1, generating the first attention weight matrix focusing on global distribution and the second attention weight matrix focusing on local interactions, respectively.

[0046] The first attention weight matrix and the second attention weight matrix are fused to obtain fused attention weights. This step combines the weights from the global perspective with the weights from the local perspective, for example, by adding them element-wise or transforming them after concatenation. Through this fusion, the model can simultaneously take into account the overall importance of features and the specific correlations between channels, thereby obtaining a more comprehensive and robust fused attention weight for subsequent feature selection.

[0047] The fused attention weights are multiplied by the target map features to obtain the multi-scale map features. This step utilizes the generated fused attention weights to perform channel-dimensional weighted multiplication on the original input features. Channel features with high weights are enhanced, for example, by multiplying by a larger coefficient, while channel features with low weights are suppressed, for example, by multiplying by a smaller coefficient. This step filters out noise and irrelevant redundant information from the features, resulting in multi-scale map features that better characterize the health status of the components.

[0048] Step 104: Based on the multi-scale map features, calculate the cross-attention weights between different target maps, and fuse the target maps based on the cross-attention weights to obtain fused map features.

[0049] In one possible implementation, the step may include: selecting multi-scale graph features of a first target graph as query features and multi-scale graph features of the remaining target graphs as key features; calculating the similarity between the query features and each key feature to obtain cross-attention weights and normalizing the cross-attention weights; using the normalized cross-attention weights to perform a weighted summation of the key features to obtain weighted features; and performing a residual connection between the weighted features and the query features to obtain the fused graph features.

[0050] This scheme introduces a cross-attention mechanism, utilizing the similarity calculation between query features and key features to achieve adaptive interaction and fusion of information from different target graph nodes. This scheme effectively captures deep relationships between graph data, aggregating complementary information from other views into the current view, enhancing the richness and completeness of features. Combined with residual connections, it ensures the stability of information transmission and significantly improves the model's comprehensive representation ability of complex multi-graph structures.

[0051] Optionally, in one specific implementation, assuming there are K target images in total, the multi-scale map feature matrix of the k-th target image is selected. The query feature Q is used as the feature set of the remaining K-1 target graphs, which are then used as the key feature set. .

[0052] To measure the relevance between different views, it is necessary to calculate the similarity between query features and each key feature. This scheme captures the correlation in high-dimensional space through the interaction between feature matrices. The cross-attention weights reflect the importance or contribution of other target graphs to the current first target graph. The i-th key feature... Normalized cross-attention weight matrix relative to query feature Q for: Where exp represents the exponential function operation, and T represents the matrix transpose. Let represent the feature matrix of the j-th key.

[0053] Aggregate all key features according to their corresponding weights: ,in, As a weighted feature, This indicates a position-weighted multiplication.

[0054] To prevent feature degradation or loss of original information during deep network processing, a residual connection is performed between the weighted features and the query features. This preserves the original feature information of the first target image itself while supplementing it with complementary information from other views, significantly enhancing the robustness and expressiveness of the features, ultimately resulting in a fused graph feature.

[0055] Step 105: Compress the latent feature representation of the fused graph features, and use the latent feature representation as a health indicator characterizing the degradation state of the component. Arrange the health indicators extracted from the same component at different time points in chronological order to form a health indicator time series. Input the health indicator time series into the fault prediction model to obtain the estimated remaining life of the device.

[0056] In one possible implementation, compressing the latent feature representation of the fused graph features and using the latent feature representation as a health indicator characterizing the component degradation state includes: extracting features from the fused graph features to obtain semantic features for describing the failure mode; smoothing the semantic features to obtain smoothed features; and vectorizing the smoothed features to obtain a fixed-length latent feature representation as a health indicator characterizing the component degradation state.

[0057] This scheme accurately captures fault modes by extracting semantic features and filters out noise and jitter by combining feature smoothing processing, which significantly improves the stability and anti-interference ability of the indicators. The generated fixed-length latent feature representation effectively reduces the data dimensionality and can comprehensively and accurately characterize the real-time health degradation status of components, providing a highly robust data foundation for subsequent predictions.

[0058] Optionally, the fused graph features obtained through multi-view fusion processing are first input into a deep feature extractor. This extractor may contain at least one fully connected layer or convolutional layer for nonlinear mapping and dimensionality reduction of the high-dimensional sparse graph features. Through this process, fault-related information with clear physical meaning can be obtained from complex topological structures and spatiotemporal correlations, removing redundant background noise to obtain semantic features.

[0059] Considering that sensor-collected data often encounters random interference under actual operating conditions, the extracted semantic features may exhibit non-monotonic high-frequency fluctuations. Feature smoothing is then applied to these semantic features. Specifically, signal processing techniques such as moving average algorithms, Gaussian filtering, or low-pass filtering can be used to reduce noise in the feature curves. This operation effectively smooths out local noise while preserving the overall long-term degradation trend of the features.

[0060] The smoothed features are vectorized and encoded, a process that integrates smoothed feature information distributed across different nodes or channels into a one-dimensional vector. By setting a preset dimension size, such as 128 or 256 dimensions, inputs of different times and lengths are mapped to a unified feature space. This fixed-length vector is the latent feature representation, which, as a comprehensive numerical indicator, fully quantifies the current health status of the component and is therefore defined as a health indicator. This indicator not only has a small data volume but also includes degradation information, facilitating real-time monitoring and lifespan prediction.

[0061] In one possible implementation, the step of inputting the health indicator time series into a fault prediction model to obtain the estimated remaining lifespan of the device includes: preprocessing the health indicator time series to convert it into an input tensor that the fault prediction model can recognize; using the fault prediction model to perform time dependency analysis on the input tensor to extract time series features reflecting degradation trends; inputting the time series features into the regression layer of the fault prediction model for mapping calculation, and outputting the estimated remaining lifespan representing the remaining usage time of the device.

[0062] This scheme improves model input quality by generating standard input tensors through preprocessing. It leverages model analysis to identify time dependencies, deeply mining long-term and short-term correlations to accurately capture nonlinear patterns of performance degradation. End-to-end prediction is achieved through regression layer mapping, directly outputting remaining lifetime values, thus enhancing the accuracy of predictions and the timeliness of decision-making.

[0063] Optionally, health indicator data arranged chronologically within a continuous time window can be collected. Considering the differences in numerical dimensions of the original data and the potential for missing values, the time series can be standardized. For example, standardization can map the data to a standard normal distribution with a mean of 0 and a variance of 1 to eliminate the influence of dimensions and accelerate model convergence. Subsequently, a sliding window technique is used to reconstruct the continuous one-dimensional health indicator sequence into three-dimensional sample data, thereby generating the input tensor required for fault prediction models, such as long short-term memory networks or gated recurrent units.

[0064] The constructed input tensor can be fed into the recurrent network layer of a pre-trained fault prediction model. Utilizing the unique gating mechanism of the recurrent unit, key historical information can be filtered and retained at each time step, while irrelevant noise interference is discarded. By performing time-series modeling on the input tensor, the long-term and short-term dependencies and nonlinear dynamic changes of health indicators over time can be automatically captured, thereby extracting deep time-series features that can keenly reflect the rate and state of component performance degradation layer by layer.

[0065] Finally, the extracted high-dimensional time series features are flattened and input into the regression layer at the end of the fault prediction model. This regression layer typically consists of one or more fully connected layers, used to map the abstract high-dimensional time series features to a one-dimensional real number space. Through processing with a preset activation function, the regression layer ultimately calculates a specific numerical value. This value directly represents the length of time the component is expected to continue operating normally under its current health condition, i.e., the estimated remaining lifespan of the equipment.

[0066] On the other hand, embodiments of this application provide a system including a storage medium and one or more processors; the storage medium stores a computer program that is invoked by one or more processors to implement any of the methods described above.

[0067] In an exemplary embodiment, a computer-readable storage medium is also provided, which stores at least one piece of program code that is loaded and executed by a processor to enable a computer to implement any of the methods described above.

[0068] Optionally, the aforementioned computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0069] In an exemplary embodiment, a computer program or computer program product is also provided, which stores at least one computer instruction that is loaded and executed by a processor to enable the computer to implement any of the methods described above.

[0070] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the data protection laws and policies of the country where the application is located, and with authorization from the owner of the relevant device.

[0071] It should be understood that "multiple" as used in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0072] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0073] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting equipment lifespan, characterized in that, The method includes: inputting the state information of each component of the device into a conditional generative adversarial network to generate component time-series data; constructing multiple target graphs with adjacency relationships based on the component time-series data; inputting the target graphs into a convolutional autoencoder, extracting time features at different scales through the multi-branch convolutional structure of the convolutional autoencoder, extracting local graph features at different scales through the graph convolutional structure of the convolutional autoencoder, and concatenating the time features and the local graph features to obtain target graph features; inputting the target graph features into a multi-scale attention module, performing channel weighting on the target graph features to obtain multi-scale graph features; calculating cross-attention weights between different target graphs based on the multi-scale graph features, and fusing the target graphs based on the cross-attention weights to obtain fused graph features; compressing the fused graph features into a latent feature representation, and using the latent feature representation as a health indicator characterizing the degradation state of the component; arranging the health indicators extracted from the same component at different time points in chronological order to form a health indicator time series; and inputting the health indicator time series into a fault prediction model to obtain the estimated remaining lifespan of the device.

2. The method according to claim 1, characterized in that, The step of constructing multiple target graphs with adjacency relationships based on the component time-series data includes: aligning the component time-series data with corresponding operating condition labels to construct multimodal time-series data vectors; determining at least two neighborhood parameter values ​​using cluster analysis based on the distance between the multimodal time-series data vectors, wherein the neighborhood parameter values ​​characterize the range of the number of neighboring nodes considered by a node when constructing the target graph; and constructing the target graphs based on the neighborhood parameter values, wherein each node in the target graph represents the component state at a given time point, and the edges of each target graph are determined by the distance between nodes in the multimodal time-series data vectors under the neighborhood parameter values.

3. The method according to claim 2, characterized in that, The process of inputting the target image into a convolutional autoencoder (CAE), extracting temporal features at different scales through the CAE's multi-branch convolutional structure, extracting local graph features at different scales through the CAE's graph convolutional structure, and concatenating the temporal features and local graph features to obtain the target image features includes: converting the target image into a node feature sequence; inputting the node feature sequence into the CAE's multi-branch convolutional structure, extracting temporal features at different scales through multiple convolutional branches with different kernel sizes; inputting the target image into the CAE's graph convolutional structure, aggregating node neighborhood information through two graph convolutional layers, and extracting local graph features at different scales; mapping the temporal features and local graph features to a preset feature dimension, concatenating them along the channels of the feature dimension to generate concatenated features, and then performing feature dimensionality reduction and compression to obtain the target image features.

4. The method according to claim 3, characterized in that, The process of inputting the target graph into the graph convolutional structure of the convolutional autoencoder, and aggregating node neighborhood information and extracting local graph features at different scales through processing by two graph convolutional layers, includes: inputting the target graph into the first graph convolutional layer of the graph convolutional structure, aggregating the feature information of the first-order neighbor nodes of the current node based on a preset adjacency matrix, and extracting local node association features; inputting the local node association features into the second graph convolutional layer, aggregating the feature information within the second-order neighbor range of the node, and extracting wide-area topological features of the node; and fusing the local node association features with the wide-area topological features of the node to obtain the local graph features at different scales.

5. The method according to any one of claims 1 to 4, characterized in that, The step of inputting the target map features into a multi-scale attention module and performing channel weighting on the target map features to obtain multi-scale map features includes: dividing the target map features into several sub-feature groups; performing global average pooling and convolution on each sub-feature group to obtain a first sub-feature and a second sub-feature; normalizing all the first sub-features and the second sub-features to obtain a first attention weight matrix and a second attention weight matrix; fusing the first attention weight matrix and the second attention weight matrix to obtain a fused attention weight; and multiplying the fused attention weight by the target map features to obtain the multi-scale map features.

6. The method according to claim 5, characterized in that, The step of calculating cross-attention weights between different target maps based on the multi-scale map features, and fusing the target maps based on the cross-attention weights to obtain fused map features, includes: selecting the multi-scale map features of the first target map as query features, and using the multi-scale map features of the remaining target maps as key features; calculating the similarity between the query features and each key feature to obtain cross-attention weights and normalizing the cross-attention weights; using the normalized cross-attention weights to perform a weighted summation of the key features to obtain weighted features; and performing a residual connection between the weighted features and the query features to obtain the fused map features.

7. The method according to claim 6, characterized in that, The step of compressing the fused graph features into a latent feature representation and using the latent feature representation as a health indicator characterizing the component degradation state includes: extracting features from the fused graph features to obtain semantic features for describing the failure mode; smoothing the semantic features to obtain smoothed features; and vectorizing the smoothed features to obtain a fixed-length latent feature representation as a health indicator characterizing the component degradation state.

8. The method according to claim 7, characterized in that, The step of inputting the health indicator time series into the fault prediction model to obtain the estimated remaining lifespan of the equipment includes: preprocessing the health indicator time series into an input tensor that the fault prediction model can recognize; using the fault prediction model to perform time dependency analysis on the input tensor to extract time series features reflecting degradation trends; inputting the time series features into the regression layer of the fault prediction model for mapping calculation, and outputting the estimated remaining lifespan representing the remaining usage time of the equipment.

9. A system comprising a storage medium and one or more processors; said storage medium storing a computer program, characterized in that, The computer program is invoked by one or more processors to implement the device lifetime prediction method according to any one of claims 1 to 8.

Images