Method and model for predicting residual service life of equipment based on fact enhancement

Abstract

The invention provides an equipment remaining service life prediction method and model based on fact enhancement, and the method comprises the steps: receiving sensor data of equipment, dividing and setting a sliding time window scale of the sensor data, and selecting the sensor data with significant fluctuation to form a feature matrix; measuring the correlation between the sensors and constructing a time-space map; dividing real correlation and false correlation through the initialized screening parameters; after the false correlation is eliminated, inputting the updated time-null graph into the graph convolutional network to generate sensor features containing real correlation information, and training a model by adopting a root-mean-square error loss function; the screening parameters are updated based on model training backhaul gradient information; further mining real correlation and updating the time-null map to generate accurate sensor features until a root mean square error loss value infinitely tends to 0; and evaluating the remaining service life of the equipment by using the prediction model. Accurate association learning is realized by eliminating false association and synthesizing real association.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment performance prediction, and in particular relates to a method and model for predicting the remaining useful life of equipment based on fact enhancement. Background Art

[0002] In the field of equipment remaining useful life prediction, early technologies primarily relied on deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs). These models can automatically extract essential features related to remaining useful life from time series data monitored by sensors. CNNs, with their excellent feature extraction capabilities in image processing, are applied to one-dimensional time series data to identify local spatiotemporal features, while RNNs, due to their suitability for processing sequential data, are used to capture temporal dependencies. In particular, variants such as gated recurrent units (GRUs) and long short-term memory (LSTM) networks have been widely used in this field due to their advantages in alleviating the vanishing gradient problem.

[0003] Current technologies construct fully connected graphs to fully exploit the spatial and temporal dependencies between sensors. While these technologies demonstrate strong performance in RLS prediction, they overlook the potential impact of spurious correlations within the graph structure. True correlations between sensors, where similar data distributions at different timestamps indicate a strong association between two sensors, are inherently related. However, due to the influence of long time series, the spatiotemporal dependencies between these strongly correlated sensors cannot be accurately established, resulting in the model learning a high number of spurious correlations. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and model for predicting the remaining useful life of equipment based on fact enhancement, aiming to solve the problem that existing prediction schemes have many false correlations.

[0005] The present invention is implemented as follows: a method for predicting the remaining useful life of equipment based on fact enhancement, the method comprising the following steps:

[0006] Step S1: Receive sensor data from the device, divide and set the sliding time window scale of the sensor data according to the number of times the device is used, select Z sensor data with significant fluctuations to form a feature matrix, use the Transformer encoder to extract features from the sensor data from the feature matrix and output the extracted features;

[0007] Step S2: After feature extraction is completed, the correlation between sensors is measured and a time-space graph is constructed to output the correlation between sensors;

[0008] Step S3: divide the true correlation and false correlation by the initialized screening parameter ∈, update and output the constructed time-space graph;

[0009] Step S4: Input the updated spatiotemporal graph into the graph convolutional network to generate sensor features containing real correlation information, output predictions, train the model using the root mean square error loss function, and update the screening parameters ∈ in step S3 through gradient backpropagation;

[0010] Step S5: Based on the updated screening parameter ∈, mine the true correlation and update the time-space graph, repeating step S4 until the root mean square error loss value approaches 0 infinitely;

[0011] Step S6: Use the trained prediction model to predict the real-time RUL of each device based on the test data set, and use the root mean square error to evaluate the remaining service life of the device.

[0012] A further technical solution of the present invention is: step S2 includes the following steps:

[0013] Step S21: After feature extraction is completed, the time-space graph construction phase begins. The time-space graph G is composed of graph nodes N and the correlation matrix E between nodes. Graph nodes N are all sensor features under all time sliding windows, expressed as where n i,t represents the i-th sensor feature under the t-th time sliding window;

[0014] Step S22: Based on the sensor feature matrix, a dynamic time warping network is used to calculate and measure the temporal correlation between sensor features to obtain a correlation matrix c. The cosine similarity algorithm is used to measure the spatial correlation between all sensors in the five time sliding windows, that is, the adjacency matrix s. The correlation between sensor features is measured by multiplying the adjacency matrix s with the correlation matrix c. The formula is as follows:

[0015] e ij,t =c·s

[0016] The correlation between all sensors under different time sliding windows constitutes the correlation matrix E, which is expressed as where e ij,t Represents the correlation characteristics between the i-th and j-th sensors under the t-time sliding window;

[0017] Step S23: Construct a time-space graph G = (N, E) of sensor features based on the sensor features N and the calculated correlation matrix E. If e ij,t =W, which means that in the process of constructing the time-space graph, the i-th sensor feature and the j-th sensor feature in the t-time window are connected and the weight value W is marked.

[0018] A further technical solution of the present invention is: step S3 includes the following steps:

[0019] Step S31: Using a predefined threshold ε, the correlation features are divided into true correlations and false correlations. The formula is as follows:

[0020]

[0021] Where A represents the set of activated features that approximate the true correlation, and R represents the set of remaining features that are regarded as false correlations. Represents the absolute value of the gradient returned by the model, A set representing the association relationship between sensors;

[0022] Step S32: Eliminate the false associations of the division and update the correlation matrix E and the time-space graph G=(N, E) according to step S23.

[0023] A further technical solution of the present invention is: step S4 includes the following steps:

[0024] Step S41: Based on the time-space graph G = (N, E), using the sensor features with real correlation generated in step S5 and the predicted RUL value, calculate the root mean square error based on the RUL value predicted by the model and the RUL value annotated in the dataset:

[0025]

[0026] RUL i Indicates the component lifespan annotated in the dataset, represents the component life predicted by the model;

[0027] Step S42: Use the root mean square error as the loss function to train the model, update the threshold ε through the gradient information returned by the training, and then divide the true correlation and false correlation from the remaining feature set to further explore the true correlation features. The gradient information calculation formula is:

[0028]

[0029] in Represents the association relationship between sensors, and f represents the loss function L RMSE , Denotes the loss function L RMSE The gradient of The derivative of the parameter, express The partial derivative of the parameter, u represents the remaining useful life feature of the dataset annotation, represents the remaining useful life characteristics predicted by the model;

[0030] Step S43: Through iterative training, the model will gradually extract true correlation features from the remaining feature set R and put them into the activated feature set A, thereby obtaining complete true correlation features.

[0031] A further technical solution of the present invention is: step S5 includes the following steps:

[0032] Step S51: Use a graph convolutional network to integrate the true dependency between sensors into sensor features. For the sensor feature set N and the true correlation feature set E between sensors, a more accurate sensor feature h is obtained after learning the association relationship through the graph convolutional network. The formula is as follows:

[0033]

[0034] in represents the sensor features after learning the true association, represents the initial sensor features input into the model, σ(.) represents the activation function, N represents the nodes in the graph convolution model, e represents the trainable hyperparameter, and W l represents the weight matrix of the lth layer, b l Represents the bias of the lth layer, l represents the number of layers of the network model;

[0035] Step S52: After updating the sensor features, a multi-layer perceptron is used to convert the stacked multi-layer sensor features into representations. Finally, the sensor representations are input into a single-layer neural network to output the RUL prediction value.

[0036] Another object of the present invention is to provide a prediction model for a method for predicting the remaining useful life of equipment based on facts, comprising:

[0037] The data preprocessing module receives sensor data from the device, divides and sets the sliding time window size of the sensor data according to the number of times the device is used, selects Z sensor data with significant fluctuations to form a feature matrix, uses the Transformer encoder to extract features from the sensor data from the feature matrix, and outputs the extracted features;

[0038] The time-space graph construction module is used to measure the correlation between sensors and construct a time-space graph after feature extraction is completed, and output the correlation between sensors;

[0039] Association learning module: used to divide true correlation and false correlation by initializing the screening parameter ∈, and update and output the constructed time-space graph;

[0040] The prediction module is used to input the updated spatiotemporal graph into the graph convolutional network to generate sensor features containing real correlation information, output predictions, use the root mean square error loss function to train the model, update the screening parameter ∈ through gradient backpropagation, and based on the updated screening parameter ∈, mine the real correlation and update the spatiotemporal graph. Repeat step S4 until the root mean square error loss value tends to 0 infinitely.

[0041] A further technical solution of the present invention is: the time-space graph construction module is specifically used to:

[0042] After feature extraction is completed, the time-space graph construction phase begins. The time-space graph G consists of graph nodes N and the correlation matrix E between nodes. Graph nodes N are all sensor features under all time sliding windows, expressed as where n i,t represents the i-th sensor feature under the t-th time sliding window;

[0043] According to the sensor feature matrix, the dynamic time warping network is used to calculate and measure the temporal correlation between sensor features to obtain the correlation matrix c. The cosine similarity algorithm is used to measure the spatial correlation between all sensors under five time sliding windows, that is, the adjacency matrix s. The correlation between sensor features is measured by the dot product of the adjacency matrix s and the correlation matrix c. The formula is as follows:

[0044] e ij,t =c·s

[0045] The correlation between all sensors under different time sliding windows constitutes the correlation matrix E, which is expressed as where e ij,t Represents the correlation characteristics between the i-th and j-th sensors under the t-time sliding window;

[0046] Based on the sensor features N and the calculated correlation matrix E, a time-space graph of sensor features G = (N, E) is constructed. If e ij,t =W, which means that in the process of constructing the time-space graph, the i-th sensor feature and the j-th sensor feature in the t-time window are connected and the weight value W is marked.

[0047] A further technical solution of the present invention is that the association learning module includes a correlation selection module, which divides the correlation features into true correlations and false correlations.

[0048] The correlation features are divided into true correlation and false correlation by a predefined threshold ε, as follows:

[0049]

[0050] Where A represents the set of activated features that approximate the true correlation, and R represents the set of remaining features that are regarded as false correlations. Represents the absolute value of the gradient returned by the model, Represents the set of association relationships between sensors; removes the false associations of the partition and updates the correlation matrix E and the time-space graph G = (N, E).

[0051] A further technical solution of the present invention is that the association learning module includes a self-extraction module to further explore potential real correlations.

[0052] Based on the spatiotemporal graph G = (N, E), the generated sensor features with real correlation and the predicted RUL values ​​are used to calculate the root mean square error (RMSE) between the RUL values ​​predicted by the model and the RUL values ​​annotated in the dataset:

[0053]

[0054] RUL i Indicates the component lifespan annotated in the dataset, Represents the component lifespan predicted by the model; the root mean square error is used as the loss function to train the model, and the threshold ε is updated through the gradient information returned by the training. Then, the true correlation and false correlation are divided from the remaining feature set to further explore the true correlation features. The gradient information calculation formula is:

[0055]

[0056] in Represents the association relationship between sensors, and f represents the loss function L RMSE , Denotes the loss function L RMSE The gradient of The derivative of the parameter, express The partial derivative of the parameter, u represents the remaining useful life feature of the dataset annotation, represents the remaining useful life characteristics predicted by the model;

[0057] Through iterative training, the model will gradually extract the true correlation features from the remaining feature set R and put them into the activation feature set A, thereby obtaining complete true correlation features.

[0058] The beneficial effects of the present invention are as follows: starting from a causal perspective, the present invention introduces a new paradigm for the prediction of the remaining useful life of equipment, constructs a causal inference scheme, makes the association learning process more explainable, and achieves accurate association learning by eliminating false associations and synthesizing true associations. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the main flow chart of the prediction method of the present invention;

[0060] Figure 2 It is a flow chart of the prediction model of the present invention. DETAILED DESCRIPTION

[0061] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0062] The present invention provides a method and model for predicting the remaining useful life of equipment based on fact enhancement to eliminate the false correlations caused by the time span during the model learning process. The key idea is to adaptively identify causal factors and interference factors while fully integrating the real correlations to optimize the prediction performance. Specifically, correlation selection is first proposed, that is, selecting causal factors and interference factors from candidate correlations as real correlations and false correlations, respectively. However, due to the influence of time factors, current prediction models often judge real correlations as false correlations. In order to solve this problem, a self-extraction module is proposed to encourage the model to extract real correlations from false correlations as completely as possible, and eliminate false correlations while integrating real correlations.

[0063] In fields such as aerospace, industrial, and military applications, degraded equipment performance can lead to catastrophic failure. Therefore, diagnostics and health management (PHM) is crucial for maintaining mechanical systems in optimal operating condition. Remaining useful life prediction (RSV) is a foundational technology for fault diagnosis and health management. It can predict the duration until equipment ceases to function effectively as designed, thereby reducing the risk of sudden system failure. Consequently, RVS prediction of mechanical equipment is widely used in aerospace, industrial, and military applications.

[0064] In the aerospace sector, this technology can be integrated into real-time monitoring systems on aircraft and spacecraft to continuously assess the health of critical components such as engine turbine blades and landing gear systems. By accurately predicting remaining useful life, this technology can guide ground maintenance teams in preventive maintenance and optimize spare parts inventory management, thereby avoiding emergency groundings or delays due to component failures. In spacecraft missions in particular, this technology can help extend mission lifespans and maximize scientific output and economic benefits.

[0065] In industrial applications, this technology can combine sensor data from production lines to analyze and predict the remaining useful life of equipment such as CNC machine tools, industrial robots, and compressors in real time. This information allows companies to effectively develop maintenance plans, reduce unplanned downtime, and improve overall production efficiency. Furthermore, it facilitates the transition from reactive maintenance to data-driven proactive maintenance, thereby extending equipment life and reducing maintenance costs.

[0066] In military applications, this technology can be embedded in the maintenance systems of military equipment such as tanks and fighter jets, enabling real-time monitoring of key performance indicators. Through predictive analysis, technicians can proactively identify potential failures that could impact the equipment's combat effectiveness, enabling targeted maintenance during non-combat windows to ensure high equipment availability. This not only improves the reliability of various types of equipment but also significantly reduces combat readiness vulnerabilities and logistical support pressures caused by unexpected failures.

[0067] Data preprocessing: First, the data preprocessing module receives sensor data from the equipment. Based on the number of engine flight cycles, the sensor data is partitioned into a sliding time window of 50. Z sensor data points with significant fluctuations are selected. For example, in aerospace technology, 14 sensor data points can be selected: fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan inlet pressure, bypass duct total pressure, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow and P30 ratio, corrected fan speed, and corrected core engine speed. This data is then used to construct a 14x50 engine feature matrix. The model then uses a Transformer encoder to extract features from the sensor data. The data preprocessing module selects the sensor data most relevant to the equipment degradation process and extracts the corresponding sensor features.

[0068] Time-space graph construction phase: After feature extraction is completed, the time-space graph construction phase begins. The time-space graph G consists of graph nodes N and the correlation matrix E between nodes. Graph nodes N represent all sensor features under all time sliding windows, expressed as:

[0069]

[0070] where n i,t represents the i-th sensor feature under the t-th time sliding window.

[0071] Based on the existing sensor feature matrix of 5*14*50 dimensions, we first use a dynamic time warping (DTW) network to calculate and measure the temporal correlation between sensor features, resulting in a correlation matrix c of (5*14)*(5*14) dimensions. We then use the cosine similarity algorithm to measure the spatial correlation between all sensors within five temporal sliding windows, i.e., the adjacency matrix s of (5*14)*(5*14) dimensions. Finally, we measure the correlation between sensor features by taking the dot product of the adjacency matrix s and the correlation matrix c, as follows:

[0072] e ij,t =c·s

[0073] The correlation between all sensors under different time sliding windows constitutes the correlation matrix E, which is expressed as:

[0074]

[0075] where e ij,t Represents the correlation feature between the i-th and j-th sensors in the t-time sliding window.

[0076] Finally, a spatiotemporal graph G = (N, E) of sensor features is constructed based on the sensor features N and the calculated correlation matrix E. If e ij,t = 0.84, indicating that during the construction of the spatiotemporal graph, the i-th sensor feature and the j-th sensor feature in time window t are connected and labeled with a weight of 0.84. The spatiotemporal graph construction helps the model learn the associations between sensor features from both temporal and spatial perspectives, enhancing sensor feature representation.

[0077] Fact-enhanced association learning: After constructing the spatiotemporal graph, the model obtains correlation features between all sensors. However, due to temporal factors, some true associations are measured as false associations. Therefore, fact-enhanced association learning is proposed to capture potential true associations from false ones. The association learning module aims to effectively capture potential true associations and eliminate false ones, ensuring that the model accurately reflects the true interactions between sensors.

[0078] Fact-enhanced association learning includes a relevance selection module and a self-extraction module. The relevance selection module first divides the relevance features into true and false relevance. Specifically, the model calculates the root mean square error (RMSE) using sensor features and relevance features. The RMS error is used as the loss function for model training. The relevance features are divided using the gradient information returned by training. The gradient information calculation formula is as follows:

[0079]

[0080] RUL iIndicates the component lifespan annotated in the dataset, represents the component life predicted by the model, Represents the association relationship between sensors, and f represents the loss function L RMSE , Denotes the loss function L RMSE The gradient of The derivative of the parameter, express The partial derivative of the parameter, u represents the remaining useful life feature of the dataset annotation, Represents the remaining useful life characteristics predicted by the model.

[0081] The correlation features are divided into true correlation and false correlation by a predefined threshold ε, as follows:

[0082]

[0083] Where A represents the set of activated features that approximate the true correlation, and R represents the set of remaining features that are regarded as false correlations. Represents the absolute value of the gradient returned by the model, A collection representing the association relationships between sensors.

[0084] Due to time constraints, some true correlations can be mistakenly classified as false correlations. Therefore, a self-extraction module is used to further explore potential true correlations. Specifically, during training, true correlation features in the activated feature set A are first removed. By discarding the most important dependencies, the model is forced to extract true correlation features from the remaining feature set R. This means that the root mean square error (RMSE) is recalculated using sensor features and correlation features from the remaining feature set to obtain new gradient information, which is then used to further explore true correlation features from the remaining feature set.

[0085] Finally, the model extracts the true correlation features from the remaining feature set R and puts them into the activated feature set A to obtain the complete true correlation features.

[0086] Remaining useful life prediction: Based on true correlation features, the model uses a graph convolutional network (GCN) to incorporate the true dependencies between sensors into sensor features, thereby generating accurate sensor characteristic representations. Specifically, for a sensor feature set N and a true correlation feature set E between sensors, the GCN network learns the association relationship to obtain a more accurate sensor feature h, as shown below:

[0087]

[0088] in represents the sensor features after learning the true association, represents the initial sensor features input into the model, σ(.) represents the activation function, N represents the nodes in the graph convolution model, e represents the trainable hyperparameter, and W l represents the weight matrix of the lth layer, b l "(L)" represents the bias of layer l, where l represents the number of layers in the network model. After updating the sensor features, a multilayer perceptron (MLP) is used to stack multiple layers of sensor features and transform the learned sensor features into representations. Finally, the sensor representations are input into a single-layer neural network, which outputs the RUL prediction value.

[0089] Through these steps, this technical route models the real correlation between sensor sequences in time and space, and eliminates the influence of false correlation, making the remaining service life prediction of mechanical equipment more accurate and reliable.

[0090] The present invention is described below with reference to an embodiment.

[0091] The specific process of the equipment remaining useful life prediction model based on fact enhancement is as follows: First, the turbofan engine degradation monitoring data set published by the National Aeronautics and Space Administration (NASA) is smoothed and minimum-maximum normalized. Then, the 14 sensor data most relevant to the engine degradation process are screened out (fan inlet temperature, low-pressure compressor outlet temperature, high-pressure compressor outlet temperature, low-pressure turbine outlet temperature, fan inlet pressure, duct total pressure, high-pressure compressor outlet total pressure, uncorrected fan speed, uncorrected core engine speed, engine pressure ratio, high-pressure compressor outlet static pressure, fuel flow and P30 ratio, fan corrected speed, core engine corrected speed). Then, the monitoring data are grouped and processed using a sliding time window. Finally, the transformer is used to extract features from the grouped data and output the extracted features.

[0092] After data preprocessing, the sensor,features in the training set are input into the prediction model.,The correlation between sensors is measured and a,time-space graph is constructed through the time-space graph,,and the correlation between sensors is output;

[0093] The correlation features between sensors are input into the fact-enhanced association learning module, and the real correlation and false correlation are divided by the initialized screening parameter ∈, and the constructed time-space graph is updated and output;

[0094] The updated spatiotemporal graph is fed into the GCN network to generate sensor features containing real correlation information and output the predicted value. Finally, the model is trained using the root mean square error loss function, and the updated screening parameters ∈ are passed back through the gradient;

[0095] Based on the updated screening parameter ∈, the true correlation is mined and the time-space graph is updated. The previous step is repeated until the root mean square error loss value approaches 0 infinitely, and the training is stopped.

[0096] During the model testing phase, the trained prediction model is used to predict the real-time RUL of each aircraft engine based on the test dataset, and the root mean square error is used to evaluate it.

[0097] From a causal perspective, this paper introduces a new paradigm for equipment remaining useful life prediction and constructs a causal inference scheme to make the association learning process more explainable, achieving accurate association learning by eliminating false associations and synthesizing true associations.

[0098] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the remaining useful life of equipment based on fact enhancement, characterized in that: The method comprises the following steps: Step S1: Receive sensor data from the device, divide and set the sliding time window scale of the sensor data according to the number of times the device is used, select Z sensor data with significant fluctuations to form a feature matrix, use the Transformer encoder to extract features from the sensor data from the feature matrix and output the extracted features; Step S2: After feature extraction is completed, the correlation between sensors is measured and a time-space graph is constructed to output the correlation between sensors; Step S3: divide the true correlation and false correlation by the initialized screening parameter ∈, update and output the constructed time-space graph; Step S4: Input the updated spatiotemporal graph into the graph convolutional network to generate sensor features containing real correlation information, output predictions, train the model using the root mean square error loss function, and update the screening parameters ∈ in step S3 through gradient backpropagation; Step S5: Based on the updated screening parameter ∈, mine the true correlation and update the time-space graph, repeating step S4 until the root mean square error loss value approaches 0 infinitely; Step S6: Use the trained prediction model to predict the real-time RUL of each device based on the test data set, and use the root mean square error to evaluate the remaining service life of the device.

2. The method for predicting the remaining useful life of equipment based on fact enhancement according to claim 1, characterized in that: The step S2 comprises the following steps: Step S21: After feature extraction is completed, the time-space graph construction phase begins. The time-space graph G is composed of graph nodes N and the correlation matrix E between nodes. Graph nodes N are all sensor features under all time sliding windows, expressed as where n i,t represents the i-th sensor feature under the t-th time sliding window; Step S22: Based on the sensor feature matrix, a dynamic time warping network is used to calculate and measure the temporal correlation between sensor features to obtain a correlation matrix c. The cosine similarity algorithm is used to measure the spatial correlation between all sensors in the five time sliding windows, that is, the adjacency matrix s. The correlation between sensor features is measured by multiplying the adjacency matrix s with the correlation matrix c. The formula is as follows: e ij,t =c·s The correlation between all sensors under different time sliding windows constitutes the correlation matrix E, which is expressed as where e ij,t Represents the correlation characteristics between the i-th and j-th sensors under the t-time sliding window; Step S23: Construct a time-space graph G = (N, E) of sensor features based on the sensor features N and the calculated correlation matrix E. If e ij,t =W, which means that in the process of constructing the time-space graph, the i-th sensor feature and the j-th sensor feature in the t-time window are connected, and the weight value is marked as W.

3. The method for predicting the remaining useful life of equipment based on fact enhancement according to claim 2, characterized in that: The step S3 comprises the following steps: Step S31: Using a predefined threshold ε, the correlation features are divided into true correlations and false correlations. The formula is as follows: Where A represents the set of activated features that approximate the true correlation, and R represents the set of remaining features that are regarded as false correlations. Represents the absolute value of the gradient returned by the model, A set representing the association relationship between sensors; Step S32: Eliminate the false associations of the division and update the correlation matrix E and the time-space graph G=(N, E) according to step S23.

4. The method for predicting the remaining useful life of equipment based on fact enhancement according to claim 1, characterized in that: The step S4 comprises the following steps: Step S41: Based on the time-space graph G = (N, E), using the sensor features with real correlation generated in step S5 and the predicted RUL value, calculate the root mean square error based on the RUL value predicted by the model and the RUL value annotated in the dataset: RUL i Indicates the component lifespan annotated in the dataset, represents the component life predicted by the model; Step S42: Use the root mean square error as the loss function to train the model, update the threshold ε through the gradient information returned by the training, and then divide the true correlation and false correlation from the remaining feature set to further explore the true correlation features. The gradient information calculation formula is: in Represents the association relationship between sensors, and f represents the loss function L RMSE , Denotes the loss function L RMSE The gradient of The derivative of the parameter, express The partial derivative of the parameter, u represents the remaining useful life feature of the dataset annotation, represents the remaining useful life characteristics predicted by the model; Step S43: Through iterative training, the model will gradually extract true correlation features from the remaining feature set R and put them into the activated feature set A, thereby obtaining complete true correlation features.

5. The method for predicting the remaining useful life of equipment based on fact enhancement according to claim 1, characterized in that: The step S5 comprises the following steps: Step S51: Use a graph convolutional network to integrate the true dependency between sensors into sensor features. For the sensor feature set N and the true correlation feature set E between sensors, a more accurate sensor feature h is obtained after learning the association relationship through the graph convolutional network. The formula is as follows: in represents the sensor features after learning the true association, represents the initial sensor features input into the model, σ(.) represents the activation function, N represents the nodes in the graph convolution model, e represents the trainable hyperparameter, and W l represents the weight matrix of the lth layer, b l Represents the bias of the lth layer, l represents the number of layers of the network model; Step S52: After updating the sensor features, a multi-layer perceptron is used to convert the stacked multi-layer sensor features into representations. Finally, the sensor representations are input into a single-layer neural network to output the RUL prediction value.

6. A prediction model for implementing the method for predicting the remaining useful life of equipment based on fact enhancement according to any one of claims 1 to 5, characterized in that: include: The data preprocessing module receives sensor data from the device, divides and sets the sliding time window size of the sensor data according to the number of times the device is used, selects Z sensor data with significant fluctuations to form a feature matrix, uses the Transformer encoder to extract features from the sensor data from the feature matrix, and outputs the extracted features; The time-space graph construction module is used to measure the correlation between sensors and construct a time-space graph after feature extraction is completed, and output the correlation between sensors; Association learning module: used to divide true correlation and false correlation by initializing the screening parameter ∈, and update and output the constructed time-space graph; The prediction module is used to input the updated spatiotemporal graph into the graph convolutional network to generate sensor features containing real correlation information, output predictions, use the root mean square error loss function to train the model, update the screening parameter ∈ through gradient backpropagation, and based on the updated screening parameter ∈, mine the real correlation and update the spatiotemporal graph. Repeat step S4 until the root mean square error loss value tends to 0 infinitely.

7. The equipment remaining useful life prediction model based on fact enhancement according to claim 6 is characterized in that: The space-time graph construction module is specifically used to: After feature extraction is completed, the time-space graph construction phase begins. The time-space graph G consists of graph nodes N and the correlation matrix E between nodes. Graph nodes N are all sensor features under all time sliding windows, expressed as where n i,t represents the i-th sensor feature under the t-th time sliding window; According to the sensor feature matrix, the dynamic time warping network is used to calculate and measure the temporal correlation between sensor features to obtain the correlation matrix c. The cosine similarity algorithm is used to measure the spatial correlation between all sensors under five time sliding windows, that is, the adjacency matrix s. The correlation between sensor features is measured by the dot product of the adjacency matrix s and the correlation matrix c. The formula is as follows: e ij,t =c·s The correlation between all sensors under different time sliding windows constitutes the correlation matrix E, which is expressed as where e ij,t Represents the correlation characteristics between the i-th and j-th sensors under the t-time sliding window; Based on the sensor features N and the calculated correlation matrix E, a time-space graph of sensor features G = (N, E) is constructed. If e ij,t =W, which means that in the process of constructing the time-space graph, the i-th sensor feature and the j-th sensor feature in the t-time window are connected and the weight value W is marked.

8. The equipment remaining useful life prediction model based on fact enhancement according to claim 6 is characterized in that: The association learning module includes a correlation selection module, which divides the correlation features into true correlations and false correlations. The correlation features are divided into true correlations and false correlations by a predefined threshold ε, and the formula is as follows: Where A represents the set of activated features that approximate the true correlation, and R represents the set of remaining features that are regarded as false correlations. Represents the absolute value of the gradient returned by the model, Represents the set of association relationships between sensors; removes the false associations of the partition and updates the correlation matrix E and the time-space graph G = (N, E).

9. The equipment remaining useful life prediction model based on fact enhancement according to claim 8, characterized in that: The association learning module includes a self-extraction module to further explore potential real correlations. Based on the time-space graph G = (N, E), the generated sensor features with real correlations and the predicted RUL values ​​are used to calculate the root mean square error based on the RUL values ​​predicted by the model and the RUL values ​​annotated in the dataset: RUL i Indicates the component lifespan annotated in the dataset, Represents the component lifespan predicted by the model; the root mean square error is used as the loss function to train the model, and the threshold ε is updated through the gradient information returned by the training. Then, the true correlation and false correlation are divided from the remaining feature set to further explore the true correlation features. The gradient information calculation formula is: in Represents the association relationship between sensors, and f represents the loss function L RMSE , Denotes the loss function L RMSE The gradient of The derivative of the parameter, express The partial derivative of the parameter, u represents the remaining useful life feature of the dataset annotation, represents the remaining useful life characteristics predicted by the model; Through iterative training, the model will gradually extract the true correlation features from the remaining feature set R and put them into the activation feature set A, thereby obtaining complete true correlation features.

Images