A device remaining useful life prediction method based on a timing diagram inference

Abstract

This invention discloses a method for predicting the remaining service life of equipment based on time-series graph reasoning. The steps include: 1) constructing a prediction model and acquiring sensor data from sensors of the operating equipment; wherein the prediction model includes a time-series graph construction module, a node association feature learning network, and a remaining service life mapping module; 2) the time-series graph construction module treats each sensor as a node and constructs a time-series graph sequence G based on the sensor data sequence corresponding to each sensor at the current time T0; the graph G in G... j 3) The node association feature learning network performs feature learning based on sensor sensing data and time sequence G to obtain the sensor node feature matrix and sensor association features at each operating time; 4) The remaining service life mapping module predicts the remaining service life T of the equipment by using the sensor node feature matrix and sensor association features at each operating time through regression analysis.

Description

Technical Field

[0001] This invention belongs to the field of computer software technology and relates to a method for inferring the remaining service life of equipment using a time-series graph inference model based on multivariate analysis. It uses a deep learning model to fit a regression function to predict the service life of the equipment. Background Technology

[0002] Remaining useful life (RUL) prediction is a key technology in equipment health management (PHM). Its predictive information can prevent catastrophic failures, effectively ensure the safety and reliability of equipment or systems, and provide guidance for maintenance personnel in maintenance decisions and spare parts ordering, thereby reducing costs and improving efficiency.

[0003] A sequence diagram displays the dynamic collaboration between multiple objects by describing the temporal order in which messages are sent between them. It can represent the sequence of actions in a use case, where each message corresponds to a class operation or a triggering event in a state machine that causes a transition when a use case action is executed.

[0004] Graph Neural Networks (GNNs) are a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation.

[0005] With the development of high technology and the advancement of production processes, a large number of technologically advanced and structurally complex pieces of equipment have been deployed in my country's industrial and defense sectors. However, due to fatigue wear, corrosion, aging, and other factors, the performance degradation of certain components during operation is unavoidable. Once performance degradation leads to equipment failure, it can cause accidents, unnecessary casualties, catastrophic consequences, and significant economic losses. Accurately predicting the remaining service life of equipment in its early stages of performance degradation can provide guidance for preventative maintenance and spare parts ordering by operations and maintenance personnel, significantly reducing unnecessary downtime, preventing accidents, and effectively ensuring equipment safety and reliability while improving task completion rates.

[0006] In existing technologies, Remaining Usage (RUL) prediction techniques can be categorized into mechanistic model-based prediction methods, data-driven prediction methods, and hybrid mechanistic model- and data-driven prediction methods. Mechanism-based RUL prediction methods construct mathematical models by combining the equipment's failure mechanisms with prior knowledge of the equipment, and then estimate the equipment's lifespan in real time using real-time collected monitoring data. With the rapid development of sensor technology, acquiring equipment condition monitoring data has become increasingly easier, providing an opportunity for data-driven remaining usable life prediction methods. These methods do not rely on the mechanistic knowledge of the equipment or system, alleviating the difficulty of modeling complex equipment or systems using mechanistic models. This prediction method is mainly divided into statistical modeling methods and machine learning methods. Because machine learning possesses powerful data processing capabilities and does not require prior knowledge to build degradation mechanism models, using machine learning methods for equipment remaining life prediction has become a research hotspot in industry and academia. Machine learning-based prediction methods can be further divided into traditional machine learning prediction methods and deep learning prediction methods. Deep learning methods can autonomously discover deep data characteristics from large-scale data, automatically extracting degradation features and providing analytical basis for RUL prediction. Convolutional neural networks (CNNs) learn the relationships between different sensing nodes within a unified time period based on combinations of receptive fields of varying sizes, thereby identifying equipment degradation states and predicting remaining lifespan. Recurrent neural networks (RNNs) have significant advantages in processing time-dependent data. To fully exploit the temporal characteristics of equipment condition monitoring data, some researchers have employed traditional RNN models or their extensions for predicting remaining lifespan. However, due to gradient vanishing or exploding, they lack the ability to learn long-term temporal dependencies. To address this issue, some scholars have used LSTM or other improved methods for complex remaining lifespan prediction. Recently, with the introduction of the Transformer network model, it has been widely applied in machine translation, traffic flow prediction, and other fields. Because this model utilizes an attention mechanism to process sequential data, it can better capture long-term dependencies and autonomously identify key time points where equipment states change. Benefiting from the long-distance influence representation capability of the sequence brought by attention, the Transformer network model can fully exploit the spatiotemporal correlations between equipment sensor data, achieving outstanding performance in predicting remaining lifespan.

[0007] For common mechanistic models used for RUL prediction, as the structural complexity of equipment or systems increases and the operating environment of equipment becomes more diverse, the failure modes and failure mechanisms also become more complex. There is a problem of various failure modes being coupled with each other, making it difficult to establish a reasonable mechanistic model and preventing its widespread application in practical engineering.

[0008] In data-driven deep learning methods, convolutional neural networks focus on the correlations between different sensor nodes to discover their degradation features. However, their ability to extract temporal features is insufficient, which is detrimental to tasks involving predicting the remaining service life of equipment, a task with temporal characteristics. Current deep learning models mainly focus on the correlations of individual sensors within local time ranges, without considering the correlation patterns, their impact, and changes over a longer period. In reality, there are certain correlations between the sensing data of different sensor nodes, such as positive correlations, negative correlations, and transitive correlations. These correlations change at different stages of equipment operation, indicating potential performance degradation. Taking aircraft operation as an example, in the early stages, adjacent components can maintain consistent and transitive vibration patterns; however, with the accumulation of operational flights, component aging, and loosening of joints, this consistency gradually weakens, indicating potential equipment failure. Therefore, utilizing the patterns and evolution trends of the correlations between different sensor nodes can also reflect changes in equipment stability, thus providing more sufficient evidence for predicting the remaining service life. Meanwhile, the Transformer model still fails to utilize the state change trends of various components during device operation, specifically in its inability to learn the evolution and changing patterns of relationships between different sensors. This provides motivation for graph neural networks, especially time-series graph neural networks, in predicting remaining useful life.

[0009] In summary, with the increasing complexity of equipment structures and the diversity of operating environments, mechanistic model-based remaining service life prediction methods suffer from poor generalization, high costs, and difficulty in widespread adoption in practical engineering applications. Meanwhile, while data-driven methods can effectively reduce reliance on mechanistic knowledge during modeling and address the challenges of modeling complex equipment, most existing methods only consider the extraction of temporal features from sensor monitoring data, neglecting the patterns and evolution trends of inter-sensor relationships during equipment operation. Therefore, a remaining service life prediction method is needed that can fully exploit both temporal features and the changing patterns of inter-sensor relationships, thereby uncovering deeper information about equipment degradation and improving the accuracy of remaining service life prediction. Summary of the Invention

[0010] To address the problems existing in the prior art, the present invention aims to provide a method for predicting the remaining service life of equipment based on time-series graph reasoning. This invention fully utilizes the coupling relationships and changing trends of sensor operating states to extract the equipment performance degradation information contained therein, enabling accurate prediction of the remaining service life of the equipment. By constructing a sensor node relationship graph within multiple consecutive time slices and performing representation learning based on the continuous graph structure, the spatiotemporal correlations between the states of various components of the equipment are fully discovered, and the changing process of the correlations between different sensor nodes is characterized. This model can predict the node attributes, correlations, and overall information at future moments, thus serving as the analytical basis for predicting the remaining service life.

[0011] The technical solution of this invention is as follows:

[0012] A method for predicting the remaining useful life of equipment based on time sequence diagram reasoning, comprising the following steps:

[0013] 1) Construct a prediction model and acquire sensing data from sensors of the operating equipment; wherein, the prediction model includes a time series graph construction module, a node association feature learning network, and a remaining service life mapping module; a sensor is set on each of the N target locations, and the sensing data sequence corresponding to the i-th sensor is {d i1 ,d i2 ,...,d ij ,...},d ij This represents the sensing data collected by the i-th sensor at the j-th operating time, where i = 1 to N;

[0014] 2) The time sequence diagram construction module treats each sensor as a node and constructs a time sequence diagram based on the sensing data sequence corresponding to each sensor at the current time T0. The j-th graph G in G j ={V j E j} is the timing diagram corresponding to the j-th running time, V j E represents the set of nodes corresponding to the j-th running time. j Represents the set of nodes V j Associations between internal nodes, where e jik ∈E j , representing the j-th runtime node v i and node v j The type of association between nodes; based on node v i The subsequence D consisting of the sensing data collected at the j-th running time and before it. ij and node v k The subsequence D consisting of the sensing data collected at the j-th running time and before it. kj The calculation yields e jik ;

[0015] 3) The node association feature learning network learns features based on sensor sensing data and time series sequence G to obtain sensor node feature matrix and sensor association features at each running time.

[0016] 4) The remaining service life mapping module predicts the remaining service life T of the device by using the sensor node feature matrix and sensor association features obtained at each operating time through regression analysis.

[0017] Furthermore, the node association feature learning network first samples the temporal graph G corresponding to the t-th running time from the graph convolutional network. t Perform feature aggregation to obtain node association features. Relational aggregation feature R′ t Then, the GRU recurrent neural network based on the gating mechanism calculates the node association features. The node association feature H corresponding to the (t-1)th running time t-1 Feature learning is performed to obtain the sensor node feature matrix H corresponding to the t-th running time. t Then, based on the relational feature matrix R corresponding to time t-1... t-1 The relational aggregation feature R′ corresponding to the t-th running time t Obtain sensor-related features R t .

[0018] Furthermore, the recurrent neural network GRU includes an entity-gated recurrent unit and a relation-gated recurrent unit; the entity-gated recurrent unit stores the timing graph G corresponding to the (t-1)th running time. t-1 Sensor data with the last timestamp Copying and with spliced ​​together Then The sensor node feature matrix H corresponding to the (t-1)th running time t-1 The feature matrix H of the sensor nodes at time t is obtained by concatenating the matrix. t The relation-gated loop unit will use the timing diagram G corresponding to the (t-1)th running time. t-1 Sensor data with the last timestamp Copying And compare it with R′ at runtime t-1. t-1 By splicing together, we get R″ t-1 , R″ t-1 The sensor correlation feature R corresponding to time t-1 t-1 The sensor association features R corresponding to the t-th running time are obtained by inputting the recurrent neural network GRU.t .

[0019] Furthermore, in step 4), the remaining service life mapping first transforms the sensor node feature matrix and sensor association features at each operating time into a one-dimensional feature vector X = flatten([H t ;R t Then, the one-dimensional feature vector X is input into a single-layer neural network to predict the remaining service life T of the device.

[0020] Furthermore, the dynamic time normalization algorithm is used to calculate the data subsequence as D. ij and D kj similarity between As a result of e jik .

[0021] Furthermore, for Normalized value, N r The number of intervals divided, i.e., the number of relations. This is for rounding down.

[0022] A server is characterized by comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the methods described above.

[0023] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the above-described method.

[0024] The advantages of this invention are as follows:

[0025] 1. The MSTKG model proposed in this invention combines graph structure and time series data. By capturing the coupling relationship of the operating status of each component of the equipment and its changing trend, and based on the sensor node association characteristics and the time series characteristics of sensing data, it mines the equipment performance degradation information contained therein.

[0026] 2. This invention proposes a graph reasoning-based framework for predicting remaining useful life, which can utilize and combine the correlation between different sensors in the same time period, the correlation between sensors in different time periods, and the correlation implied in the data sensed by the sensors to detect the degradation of equipment performance.

[0027] 3. This invention designs a temporal graph convolutional neural network. Based on the aggregation of sensor node features and sensor correlation features, it introduces original sensor data to expand the semantic meaning of features, performs feature flattening and concatenation operations, and completes the learning of node correlation relationships. This network has smaller errors and stronger robustness in predicting the remaining service life of equipment compared to networks using single feature learning. Attached Figure Description

[0028] Figure 1 This is a diagram of the architecture of the present invention.

[0029] Figure 2 This is a schematic diagram of DTW. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0031] This invention proposes a method for predicting the remaining useful life of equipment based on time-series graph reasoning, in order to address the problem that the accuracy of existing equipment remaining useful life prediction models still has room for improvement. Mechanism-based methods have poor generalization and high cost, making them difficult to promote in practical engineering applications. While data-driven methods can solve the problem of modeling complex equipment, most existing methods only consider the extraction of time-series features from sensor monitoring data, while ignoring the patterns and evolution trends of the correlation between sensors during equipment operation.

[0032] This invention proposes a Multivariate Similarity Temporal Knowledge Graph (MSTKG) inference model based on multivariate analysis to support remaining useful life prediction. The overall structure of the model is as follows: Figure 1 As shown, it mainly consists of three parts: the graph construction model divides the sensor's sensing data into several time slices and constructs a node association graph arranged in time sequence; the node association feature learning module uses a local recurrent encoder to model and represent the spatiotemporal association evolution features of sensor nodes and the original sensing data information; finally, a regression model is used to map the device state features into a continuous space to achieve prediction of remaining service life.

[0033] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0034] S1, acquire the sensing data from the sensors of the operating equipment, and define the input, output and optimization objectives of the model;

[0035] Table 1 provides descriptive information for some of the sensors.

[0036]

[0037] Table 1 shows a description of some of the sensors. The sensing data collected in this embodiment includes vibration amplitude, temperature, and wind speed. Specifically, the operating equipment integrates N sensors in different parts, defined as {S1, S2, ..., S...}. N Each sensor acquires data at a frequency of t0, recording the operating status parameters of its component. All the sensing data forms a dataset {{d}. 11 ,d 12 ,...},{d 21 ,d 22 ,...},...,{d i1 ,d i2 ,...},...,{d N1 ,d N2 ,...}}, where d ij This represents the sensing data collected by the i-th sensor at the j-th operating moment. For example, in an engine system, sensor a collects the engine fan inlet temperature, and sensor b collects the low-pressure compressor outlet temperature; the sensor data is time-series, so there are data from different sensors at different times.

[0038] Furthermore, the model input is defined as a set of perceived data. Where T0 is the current time; the model output is That is, the predicted value of the remaining service life of the equipment at time T0; the model optimization objective is based on the assumptions... Where F() is the prediction model, if the equipment fails after working for T time periods, then at any current time T0, the remaining lifespan of the equipment is denoted as |T-T0|, and the optimization objective is to minimize

[0039] S2, Design a time sequence diagram structure to realize the associated representation of sensor nodes in different time periods;

[0040] The specific timing diagram structure is defined as follows: Given a time series diagram sequence, where each G j ={V j E j} represents the time series diagram corresponding to the j-th time period, V j Let V represent the set of nodes for time period j. j ={v j1 ,...},E j Represents the set of nodes V j Associations between internal nodes, where ejik ∈E j , indicating node v at time j i and v k The relationship type. The node set for each time period consists of the aforementioned N sensors.

[0041] Time series graph models can depict the relationships between nodes at different time periods, thereby modeling their changing trends and providing relevant information for predicting remaining useful life;

[0042] The nodes in a timing diagram are defined as follows: Each node in the timing diagram is defined as a sensor, i.e.:

[0043]

[0044]

[0045] …

[0046]

[0047] For ease of recording, the time period index is omitted and it is simply referred to as {v1,v2,...,v...} N}

[0048] The definition of edges in a sequence graph is as follows: Assuming there is a relationship between any two nodes, this is represented by the edge weight. Specifically, for e... jik That is, during the j-th time period, sensor v i With v k To establish a correlation, first extract a subset of data:

[0049] D ij ={d i(j-1)*P+1 ,d i(j-1)*P+2 ,...,d i(j-1)*P+P}

[0050] D kj ={d k(j-1)*P+1 ,d k(j-1)*P+2 ,...,d k(j-1)*P+P}

[0051] For each pair of data subsequences D ij and D kj The similarity between data subsequences is calculated using a dynamic time-normalization algorithm, which serves as a measure of node association.

[0052]

[0053] DTW is a commonly used sequence distance metric that measures the pattern similarity between two sequences and allows for pattern shifts and stretching. For example... Figure 2As shown, the DTW method adaptively finds the best matching pattern between two sequences. The DTW metric can characterize the pattern correlation between two sequences and is unaffected by time offset, better representing the state relationships between sensor nodes, such as the transmission of vibration in different parts of the device and changes in vibration period. In the specific calculation process, DTW uses dynamic programming to find the closest matching pattern between the two sequences and calculates the distance. The dynamic programming formula is:

[0054]

[0055] in, This represents the DTW distance between time series x[1:j1] and y[1:j2], corresponding to D in the time series diagram. ij With D kj , where abs is the absolute value function. Assuming the total length of time series x is N and the total length of y is M, then the DTW distance between x and y is DTW. N,M This distance is velocity-independent, meaning it is unaffected by time series shape scaling. Subsequently, the DTW distance of the perceived data subsets for all node pairs within the same time period is calculated, and the similarity is normalized. The similarity value range is uniformly divided into N... r Each interval is labeled as follows: For each pair of sensor nodes, the similarity is categorized into corresponding intervals and labeled accordingly, thus transforming continuous similarity into discrete relationships. The normalization calculation and relationship category transformation formula for DTW similarity are as follows:

[0056]

[0057]

[0058] Where, N r The number of intervals divided, i.e., the number of relations. To round down, e jik For time period j, node v i With v k The relationship between them is that max and min represent the maximum and minimum values, respectively.

[0059] Based on the above method, a time series diagram sequence for predicting remaining useful life is constructed. The time series diagram corresponding to j time periods is G. j ={V j E j}; Get a time period t, in Let be the feature attributes of a time period t, s be the head node, and e be the feature attributes of a time period t. tsoLet 'o' be the target node and 't' be the node set V over a time period t. t =[h s,t ,h o,t ], where h s,t It is the entity embedding representation of the head node, h o,t The entity embedding representation of the target node serves as the input to the inference model.

[0060] S3, the node association feature learning network, performs feature learning based on sensor-perceived data and time-series graph structure, mainly including two feature dimensions: sensor node feature matrix H. t Sensor-related features R t .

[0061] A local cyclic encoder is proposed and used. Based on the state correlations of different sensors within the same time period and at different time periods, it learns the changing trends of correlation features through sensor node state representations, providing a basis for predicting remaining service life. Its input is the time series graph sequence constructed in the previous step. and the characteristic attributes of each node in the diagram. A graph convolutional neural network (GCNN) is employed to aggregate information in the time series graph from the perspective of node data correlation. Gated recurrent units are used to evolve the network across multiple time segments, learning the changing relationships between sensor nodes over different time periods. The input to the GCNN is a time series graph G spanning t. t Feature attributes s is the head node of time segment t, e tso Let be the node relationships over time t, and let 'o' be the target node over time t; its corresponding embedding representation is: Represents the entity embedding of the first layer during time period t; r t This represents the relationship embedding over a time period t. Within time period t, the graph convolutional network aggregates multiple relationships and information from multi-hop sensor neighbor nodes using the following formula, outputting a feature representation of the target sensor node over time period t:

[0062]

[0063] Among them, (s,e) tso ,o) represent the head node, relation, and target node, respectively. r represents the entity embedding at layer l in time period t. t Represents the embedding of relationships within the time period t. Here, ψ represents the learnable weight parameters, c0 is the coefficient used for normalization, and its value is equal to the in-degree of entity o. ψ is a one-bit convolution operation, and σ is the ReLU activation function.

[0064] After the above formula is used to aggregate the nodes in time period t through a neural network, the node association feature representation for time period t is obtained.

[0065] Secondly, across multiple time periods, two gated loop units—an entity-gated loop unit and a relation-gated loop unit—are used to learn evolutionary features, characterizing the state changes of sensor nodes during device operation, i.e., the trend of changes in correlation relationships.

[0066] The GRU (Gated Recurrent Neural Network) is based on a gated recurrent neural network. For the entity-gated recurrent unit, the input is the node association features aggregated by the graph neural network over a time interval t. Its specific evolutionary characteristics are represented by the following formula:

[0067]

[0068] H t H t-1 Let be the d-dimensional sensor node feature matrices at time t and t-1, respectively. The node feature representation after GCN aggregation in time period t corresponds to the hidden state at time t, the hidden state at time t-1, and the input at time t in the standard GRU structure, respectively.

[0069] Correspondingly, the evolutionary characteristics of the relation-gated recurrent unit are represented by the following formula:

[0070]

[0071] R t =GRU(R) t-1 ,R′ t )

[0072] Among them, R′ t It is a feature representation of the relation aggregation at time t. It represents all entities connected to relation r at time t. Let d be the sensor association features at time t and time t-1, respectively, corresponding to the hidden state at time t, the hidden state at time t-1, and the input at time t in the standard GRU structure. For a specific relation r, its representation in the current time period t consists of two parts: pooling of the entity representations connected by all relations r in the previous time period, and the global representation of relation r.

[0073] Through the calculations performed by the GCN and GRU modules described above, the final output is the sensor node feature matrix H. t Sensor relationship feature matrix R t .

[0074] Next, based on the time-series graph reasoning, the feature matrix H of the sensor nodes is introduced. tThe sensor-related feature matrix R t Learn to acquire time-series characteristics of data;

[0075] The specific steps are as follows: When the entity and relation-gated recurrent units perform evolutionary learning on entity embeddings and relation embeddings, original sensor data is introduced to expand the feature semantics. The feature joint formula of the entity-gated recurrent unit is improved as follows:

[0076]

[0077]

[0078] in, It is a data time series characteristic. To construct G t-1 The data from the last timestamp in the original sensor data was copied to obtain... [;] represents a concatenation operation; the result after concatenation is... This will be used as the actual input to the physical gated loop unit, and the concatenated result will be... and H t-1 Input GRU, output H t .

[0079] The characteristic joint formula of the relation-gated loop unit is as follows:

[0080]

[0081]

[0082] in, yes R′ obtained after the same copying operation t-1 It is the relation embedding at time t-1. The result after concatenation is... This will be used as the actual input to the relation-gated loop unit, and the concatenated R″ will be... t-1 and R t-1 Input GRU, output R t Based on the above encoding, the original sensing data is incorporated into the feature representation, and the time-varying characteristics are fully learned by the gated recurrent unit to support the prediction of remaining useful life.

[0083] S4, in the remaining service life mapping section, the sensor node feature matrix H is obtained through regression analysis and time series diagram inference. t The feature matrix R of the relationship between the sensor and the sensor t This allows for the prediction of the remaining useful life of the equipment.

[0084] Regression analysis involves two steps: first, concatenating the original features to transform the feature matrix into a one-dimensional eigenvector, which supports subsequent regression analysis. The specific formula is as follows:

[0085] X = flatten([H t ;R t ])

[0086] RUL estimation analysis inputs the one-dimensional feature vector X into a single-layer neural network and outputs the predicted RUL value. The formula is as follows:

[0087]

[0088] in Let be the predicted value of RUL, [;] represent the concatenation operation, W and b are learnable parameters, and σ(·) is the sigmoid function. Furthermore, a normalization operation is implemented on RUL to calculate stability.

[0089] Finally, the loss function for the RUL prediction task is Root Mean Square Error (RMSE), as follows:

[0090]

[0091] In this invention, the model needs to sample time-series data to construct a time-series graph. The sampling window slides within each data unit with a fixed step size, and the RUL of the last timestamp within the window is used as the RUL label for the generated knowledge graph snapshot. Since the system does not degrade in its early stages, system degradation is negligible in its early lifecycle. Referring to existing work, this paper uses a piecewise linear model to label the dataset with RULs, i.e., the maximum upper limit of the RUL is set to 125 (RUL... max =125).

[0092] In summary, this example proposes a method for predicting the remaining useful life of equipment based on a time-series graph inference model using multivariate analysis. This method uses a model with sensing data from various sensor nodes of the equipment as input, estimating the remaining healthy operating time of the equipment from the perspectives of the correlation and evolution of sensor operating states and the changing trends of raw sensor data. The model first divides the sensing data from different sensors into multiple time windows and constructs a node association graph for each window, forming a time-series graph sequence. Then, a graph convolutional neural network is used to learn the sensor node features, and a gated recurrent unit is used to realize the transfer of node states between different time windows. In the prediction part, the decoded time-series graph features are used to predict the corresponding estimated remaining useful life through multi-layer mapping regression analysis. Experimental results based on benchmark data show that the proposed model can improve the prediction accuracy, especially for equipment with complex operating conditions.

[0093] Comparative analysis of effects:

[0094] Representative models based on convolutional neural networks, recurrent neural networks, and encoding / decoding mechanisms were selected as comparison algorithms.

[0095] CNN and CNN-FNN perform convolution operations on sensor data along the time dimension, and use FNN to predict RUL.

[0096] DCNN-FNN replaces the CNN in the aforementioned work with dilated CNN, expanding the receptive field of the convolutional kernel to improve the model's prediction accuracy.

[0097] LSTM-FNN uses LSTM to directly model sensor data, trains the model in an autoregressive manner, and improves the prediction accuracy of RUL through FNN.

[0098] RBM-LSTM-FNN adds a Restricted Boltzmann Machine before the LSTM and pre-trains it in an unsupervised manner, automatically extracting degradation-related features from the unlabeled raw training data.

[0099] The Auto-encoder trains a bidirectional LSTM in an unsupervised manner to model the health index curve, and then combines similarity-based curve matching techniques and zero-center rules to calculate the RUL.

[0100] GCU-Transformer extracts local features through convolutional gating, encodes high-dimensional sensor data and local features using the Transformer's encoding layer, and finally decodes them using a single-layer neural network to obtain the RUL.

[0101] 1. Theory

[0102] Currently, most feature extraction methods for predicting the remaining lifespan of equipment involve extracting features from sensor data in the time or spatial dimensions, ignoring the features related to the relationships between sensors. At the same time, single-dimensional feature extraction may result in incomplete feature representation, which may have a negative impact on subsequent prediction results.

[0103] The model described in this invention takes the sensing data from the sensor nodes of various parts of the device as input. It learns from the perspective of the correlation and evolution of sensor operating states and the changing trends of the original sensor sensing data, rather than being limited to existing methods that only rely on time-series data features or only on the correlation features between sensors. The model of this invention divides the sensing data from different sensors into several time windows and constructs a node correlation graph for each window, forming a time-series graph sequence. A convolutional neural network is used to learn the sensor node correlation features and relationship features, and the data time-series features are added to obtain node feature matrices and relationship feature matrices. After flattening and concatenation operations, the node correlation results are obtained. Through multi-layer mapping regression, RUL prediction is completed.

[0104] 2. Experimental Data

[0105] This experiment uses root mean square error as the performance metric for the RUL prediction task, defined as:

[0106]

[0107] Where N is the number of test units.

[0108] Table 2 shows the root mean square error of RUL prediction.

[0109]

[0110] Table 2 shows the results of the text model MSTKG and the comparative methods for RUL prediction on four subsets of CMAPSS. The results show that MSTKG outperforms the state-of-the-art model on the FD002 and FD004 datasets, with improvements of 0.88 and 1.29 respectively on the two dataset subsets.

[0111] The results show that, similar to autoencoder models, MSTKG is better at modeling data with complex operating conditions. For the most complex FD004 dataset, which has 12 settings: 6 operating conditions and 2 fault modes, the proposed model can not only capture the temporal features of the sensed data but also comprehensively capture the dependencies between features, automatically discovering the correlation changes in the states of various components of the equipment. This expands the information dimensions and provides a more comprehensive representation of the overall operating state of the equipment under complex conditions, achieving optimal performance on the most complex dataset.

[0112] In summary, the model proposed in this invention can fully utilize the correlation and evolution of sensor node states to discover the degradation of different components of the equipment, support the prediction of remaining service life, and its accuracy is better than the existing best solution under complex working conditions.

[0113] In this invention, the model for learning the correlation between sensor nodes at different time periods is GRU, which can be replaced by other time series models such as RNN, LSTM, and Transformer.

[0114] In this invention, the model used to aggregate information in the time series graph is GCN, which can be replaced by other graph models such as GNN, GraphSAGE, and GAT.

[0115] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.

Claims

1. A method for predicting the remaining useful life of equipment based on time-series graph reasoning, comprising the following steps: 1) Construct a prediction model and acquire sensing data from sensors in operating equipment; wherein, the prediction model includes a time series graph construction module, a node association feature learning network, and a remaining service life mapping module; a sensor is set at each of N target locations, and the sensing data sequence corresponding to the i-th sensor is as follows: , Indicates the first The sensor at the first Sensing data collected at each runtime, i=1~N; 2) The time sequence diagram construction module treats each sensor as a node, based on the current time. A time series diagram sequence is constructed from the sensing data sequences corresponding to each sensor. ; The j-th graph For the first Timing diagrams corresponding to each running moment. Indicates the first The set of nodes corresponding to each running time. Represents a set of nodes The relationships between internal nodes, where , indicating the first Each running time node and nodes The relationship type between nodes; based on the nodes No. A subsequence consisting of sensing data collected at each running time and prior to that time. and nodes No. A subsequence consisting of sensing data collected at each running time and prior to that time. Calculations yielded ; 3) The node association feature learning network is based on sensor perception data and time series graph sequences. Feature learning is performed to obtain the sensor node feature matrix and sensor association features at each running time. 4) The remaining service life mapping module predicts the remaining service life T of the device by using the sensor node feature matrix and sensor association features obtained at each operating time through regression analysis.

2. The method according to claim 1, characterized in that, The node association feature learning network first samples the graph convolutional network for the first... Timing diagram corresponding to the running time Perform feature aggregation to obtain node association features. and relational aggregation features Then, the GRU recurrent neural network based on the gating mechanism calculates the node association features. and the -1 Node association characteristics at runtime Perform feature learning to obtain the first Sensor node feature matrix at runtime Then, based on the relational feature matrix corresponding to time t-1... and the Relational aggregation features corresponding to runtime time Obtain sensor-related features .

3. The method according to claim 2, characterized in that, The recurrent neural network (GRU) includes an entity-gated recurrent unit and a relation-gated recurrent unit; the entity-gated recurrent unit stores the timing graph corresponding to the (t-1)th running time. Sensor data with the last timestamp Copying and with spliced ​​together Then The sensor node feature matrix corresponding to the (t-1)th running time By splicing, we can obtain the first Sensor node feature matrix at runtime The relation-gated loop unit will generate the timing diagram corresponding to the (t-1)th running time. Sensor data with the last timestamp Copying and with the first -1 runtime spliced ​​together ,Will Sensor correlation features corresponding to time t-1 Input the recurrent neural network GRU to obtain the first... Sensor correlation features corresponding to the running time .

4. The method according to claim 1, characterized in that, In step 4), the remaining service life mapping first transforms the sensor node feature matrix and sensor association features at each operating time into a one-dimensional feature vector. Then, the one-dimensional feature vector X is input into a single-layer neural network to predict the remaining service life T of the device.

5. The method according to claim 1, 2, or 3, characterized in that, The data subsequence is calculated using the dynamic time normalization algorithm. and Similarity between As a result .

6. The method according to claim 5, characterized in that, , for Normalized value, The number of intervals divided, i.e., the number of relations. This is for rounding down.

7. A server, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing each step of the method of any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images