A device component residual life prediction method, device, apparatus and storage medium

Abstract

The application discloses a kind of equipment component remaining life prediction method, device, equipment and storage medium, and is based on mixed fixed weight and improved GAT-BiMamba model to realize industrial equipment component remaining useful life and health state prediction.Adopting the technical scheme of the application, the degradation characteristics implied in the vibration signal of industrial equipment component during operation are efficiently and accurately extracted to realize the rapid and accurate prediction of the remaining useful life of industrial equipment component.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of failure prediction and health management of industrial equipment components, and particularly relates to a device component residual life prediction method, device, equipment and storage medium. BACKGROUND

[0002] In the industrial internet field which attaches importance to reliability and production safety, the residual service life prediction of industrial equipment components has become an important part of the field of device failure prediction and health management (PHM). In the modern industrial intelligent production process, the running components of the industrial equipment will produce cracks, fractures, corrosion and other hidden dangers due to high temperature and high pressure, high salt and humidity, and metal fatigue, material loss and other factors caused by high-intensity operation in a harsh operating environment, which will cause the service performance of the components to gradually degrade, eventually fail, and completely lose the service performance, thereby causing the equipment to run poorly and reducing the production efficiency. If the core industrial components fail, there may even be a risk of a chain failure of the equipment components and the overall damage of the production equipment, which threatens the safety of production operation.

[0003] In order to monitor the health status of the equipment and ensure the safety of the production operation of the equipment, it is necessary to predict the failure and manage the health of the key components of the industrial equipment. In the field of failure prediction and health management, evaluating the health status of the equipment, accurately predicting the residual service life and timely warning the failure risk are the key implementation goals of the field. There are two implementable technical routes in the field: one is a physical model-based method, which predicts the residual service life by constructing a physical failure model, using failure mechanism constraints and combining mathematical operations. Although it can efficiently reason, it has the problems of difficulty in complex model construction, unclear failure mechanism, dependence on expert knowledge, large amount of calculation parameters, high data precision dependence and poor universality. The other is a data-driven machine learning and deep learning method, which automatically extracts features in a large amount of data by constructing a deep learning model, has good generalization and is suitable for complex systems, but also has the problems of large data volume requirement, poor interpretability, slow reasoning speed and long-term dependence. SUMMARY

[0004] In order to solve the problems existing in the prior art, the application provides a device component residual service life prediction method, device, equipment and storage medium, which efficiently and accurately extracts the degradation features implied in the vibration signals of the industrial equipment components during operation, so as to realize the rapid and accurate prediction of the residual service life of the industrial equipment components.

[0005] To achieve the above object, the application provides the following solutions.

[0006] A device component residual service life prediction method, comprising:

[0007] Collect and preprocess horizontal and vertical vibration signal data during operation of an industrial equipment component;

[0008] Extract time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data, and label the life label;

[0009] Calculate the evaluation index related to the life label of the time domain, frequency domain and time-frequency domain features, determine the index weight using the improved mixed fixed weight method, select the high correlation features to form a data set and divide it into a training set, a validation set and a test set;

[0010] Build an encoder-decoder architecture model; wherein the encoder-decoder architecture model comprises: 2-layer encoder multilayer neural network module and 1-layer decoder multilayer neural network module; the encoder multilayer neural network module is composed of an improved graph attention network layer and a BiMamba module; the decoder multilayer neural network module is composed of a multilayer perceptron MLP layer and a multi-scale convolution layer;

[0011] Train the encoder-decoder architecture model according to the training data set;

[0012] Use the trained encoder-decoder architecture model to perform residual useful life prediction on the test data set, and output the residual useful life prediction value of each step.

[0013] The application also provides a device component residual life prediction device, comprising:

[0014] The first processing unit is used for collecting and preprocessing horizontal and vertical vibration signal data during operation of an industrial equipment component;

[0015] The second processing unit is used for extracting time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data, and labeling the life label;

[0016] The third processing unit is used for calculating the evaluation index related to the life label of the time domain, frequency domain and time-frequency domain features, determining the index weight using the improved mixed fixed weight method, selecting the high correlation features to form a data set and dividing it into a training set, a validation set and a test set;

[0017] The fourth processing unit is used for building an encoder-decoder architecture model; wherein the encoder-decoder architecture model comprises: 2-layer encoder multilayer neural network module and 1-layer decoder multilayer neural network module; the encoder multilayer neural network module is composed of an improved graph attention network layer and a BiMamba module; the decoder multilayer neural network module is composed of a multilayer perceptron MLP layer and a multi-scale convolution layer;

[0018] The fifth processing unit is used for training the encoder-decoder architecture model according to the training data set;

[0019] The sixth processing unit is configured to use the trained encoder-decoder architecture model to perform the remaining useful life prediction on a test data set, and output a remaining useful life prediction value at each step.

[0020] The application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program which is run by the processor, and the computer program performs the device component remaining life prediction method when run by the processor.

[0021] The application further provides a storage medium, wherein the storage medium stores a computer program which performs the device component remaining life prediction method when run.

[0022] Compared with the prior art, the application has the following beneficial effects:

[0023] 1. Efficient feature information mining capability, the application mines the performance degradation information in the running process by calculating the time domain, frequency domain and time-frequency domain features of the collected industrial equipment component vibration detection signal data, and then uses the improved subjective and objective mixed weight method to confirm the comprehensive score of the time domain, frequency domain and time-frequency domain features, selects the features with high correlation with the life label, reduces the influence of redundant features, reduces the consumption of computing resources, and improves the model calculation efficiency.

[0024] 2. Three-dimensional feature relationship modeling capability, the application converts the screened vibration feature data set into complex three-dimensional non-Euclidean data, and uses an improved graph attention network layer to extract the relationship between features to clearly show the mutual influence between different features, and then models the complex changes in the performance degradation process in a three-dimensional manner, which significantly improves the accuracy of the prediction method.

[0025] 3. Strong long sequence time sequence dependence extraction capability, the BiMamba model used in the application accurately models the super-long time sequence by introducing the HiPPO technology, greatly alleviating the long-term dependence problem, and its unique parallel operation mechanism can extract time sequence features while cooperating with the hardware perception algorithm without damaging the cycle operation process, optimizing the time complexity, greatly improving the operation efficiency and time sequence feature extraction performance, and more accurately depicting the overall law of non-monotonic degradation by extracting the time sequence relationship in both directions in the time dimension, extracting from multiple angles of the whole and the part to help improve the final prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0027] Figure 1 The flow chart of the equipment component residual life prediction method of the embodiment of the present application;

[0028] Figure 2 The structure diagram of the encoder multi-layer neural network module;

[0029] Figure 3 The structure diagram of the improved graph attention network of the present application

[0030] Figure 4 The structure diagram of the decoder multi-layer neural network module. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0032] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0033] Embodiment 1

[0034] As shown in the following table, the present application provides a device component residual life prediction method, comprising: Figure 1

[0035] Step S1, collecting and preprocessing the horizontal and vertical vibration signal data in the operation process of the industrial equipment component;

[0036] Step S2, extracting the time domain, frequency domain and time-frequency domain features of the preprocessed horizontal and vertical vibration signal data, and labeling the life label;

[0037] Step S3, calculating the evaluation index related to the life label of the time domain, frequency domain and time-frequency domain features, determining the index weight using the improved mixed fixed weight method, screening out the high correlation features to form a data set and dividing it into a training set, a validation set and a test set;

[0038] Step S4, constructing an encoder-decoder architecture model (improved GAT-BiMamba model);

[0039] Step S5, constructing an encoder multi-layer neural network module;

[0040] Step S6, constructing a decoder multi-layer neural network module;

[0041] ​Step S7, training the improved GAT-BiMamba model according to the training data set;

[0042] Step S8, using the trained improved GAT-BiMamba model to make remaining useful life prediction on the test data set, and outputting the remaining useful life prediction value of each step.

[0043] As an embodiment of the present application, in step S1, in order to remove the influence of running noise, it is necessary to use a variational mode decomposition method to filter the noise of the vibration monitoring signal data in the running process, and set the decomposition layer number K=8.

[0044] As an embodiment of the present application, in step S2, in order to mine as much information as possible from the denoised vibration signal data, the time domain features, frequency domain features and time-frequency domain features in the above denoised vibration signal data are calculated and extracted. The time domain features include peak amplitude, peak-to-peak value, peak-to-peak ratio, root mean square value, absolute mean value, standard deviation, skewness, kurtosis, margin factor, waveform factor and pulse factor; the frequency domain features include amplitude mean value, amplitude sample standard deviation, amplitude skewness, amplitude kurtosis, weighted average frequency, weighted frequency standard deviation, weighted root mean square frequency, frequency coefficient of variation and frequency root mean square; the time-frequency domain features include approximate entropy, sample entropy, permutation entropy and wavelet entropy; the RMS feature in the above time domain features is extracted, the degradation path of the equipment component is confirmed, and the remaining useful life label of each time step is calculated; the above time domain, frequency domain, time-frequency domain features and remaining useful life label are combined to construct a data set; the above data set is divided into a training data set, a validation data set and a test data set. Through the extraction of multi-domain features, the implicit information of the denoised vibration signal is greatly mined, which provides an important basis for the extraction and prediction of the subsequent improved GAT-BiMamba model.

[0045] As an embodiment of the present application, in step S3, in order to improve the information density of the data set, reduce the influence of redundant features and noise, reduce the model complexity, reduce the calculation resources and improve the calculation efficiency, it is necessary to use four evaluation indexes to comprehensively evaluate the features. The evaluation indexes should include the Pearson correlation coefficient, the Spearman correlation coefficient, the monotonicity coefficient and the cross-correlation coefficient, which correspond to the linear correlation, the nonlinear correlation, the monotonicity and the cross-correlation between the extracted features and the life label respectively. The four index values obtained are obtained by linear weighting, and the index weight is determined by an improved subjective and objective mixed weighting method. The improved subjective and objective mixed weighting method needs to calculate the subjective weight and the objective weight by using the subjective weighting method and the objective weighting method respectively. The subjective weighting method uses the analytic hierarchy process to construct the judgment matrix of the four indexes to calculate the subjective weight, and ensures the logic correctness through consistency check, and then evaluates the subjective reliability Rs based on formula 1 through the consistency ratio CR; the objective weighting method uses the entropy weight method, and determines the objective weight based on the information entropy of the features, and then evaluates the objective reliability Ro based on formula 2 through the entropy of the entropy weight method. After obtaining the subjective weight and the objective weight, the subjective and objective weight proportion is calculated based on formula 3 through the subjective reliability and the objective reliability. The maximum weight of the four evaluation indexes is confirmed based on the subjective and objective weight proportion, the subjective weight and the objective weight. After obtaining the comprehensive evaluation value of the features, the features with high evaluation value are selected to form the data set, and then the training data set, the validation data set and the test data set are divided.

[0046] (Formula 1)

[0047] (Formula 2)

[0048] (Formula 3)

[0049] Wherein, CR is the consistency ratio of the subjective judgment matrix in the analytic hierarchy process, which is used to represent the logical consistency degree of the judgment matrix; Rs is the subjective reliability; is the information entropy of the objective weight determined by the entropy weight method; Ro is the objective reliability; is the subjective and objective weight proportion.

[0050] As an embodiment of the present application, in step S4, the constructed encoder-decoder architecture needs to include 2 layers of encoder multi-layer neural network modules and 1 layer of decoder multi-layer neural network modules.

[0051] As an embodiment of the present application, in step S5, the constructed encoder multi-layer neural network module needs to include an improved graph attention network layer and a BiMamba module, and the encoder structure is as follows: Figure 2The training data set is converted into a non-Euclidean data form, the feature variable is used as a node, the node similarity is used as an initial edge weight, a full connection graph is constructed, and then the constructed graph data is input into the improved graph attention network layer; the graph attention network layer includes a graph attention network layer GATv2, an adjacent attention layer, a residual attention layer, a normalization layer and a dropout layer, and the improved graph attention network layer structure is as shown in Figure 3 When the training data set in the non-Euclidean form is input into the improved graph attention network layer, the adjacent attention layer is input first, the Gaussian kernel weight between nodes is calculated, and the Gaussian kernel decay coefficient is set to 1.0; the Top-K sparsification operation is performed on the graph data based on the calculated Gaussian kernel weight to realize the regularization processing of the graph structure, which is used for focusing on key associations and reducing the risk of oversmoothing, and K is set to 10; after the sparsified graph data is stabilized by normalization, it is further enhanced by the dropout layer with a coefficient of 0.2 to improve generalization, and the obtained graph node data, graph edge index data and graph edge weight data are input into the graph attention layer; the graph attention layer performs graph attention calculation based on the received three kinds of graph data to obtain the relationship data between graph nodes; the relationship data between graph nodes is input into the residual attention layer, if the layer belongs to the first encoder multi-layer neural network module, the above graph data is directly projected into the feature relationship set in the Euclidean data form, saved as historical data and output to the BiMamba module; if the layer belongs to the second encoder multi-layer neural network module, the attention score is calculated after the current layer relationship data between graph nodes is combined with the historical data, and the relationship data between graph nodes in different layers is weighted and fused using a gating mechanism, and then projected into the feature relationship set in the Euclidean data form for output, which can effectively alleviate the overfitting phenomenon. The above feature relationship set is input into the BiMamba module to extract the time sequence dependence relationship of the running component feature, the BiMamba model is composed of two parallel Mamba models in forward and reverse order, the Mamba model dimension is set to 256, the state parameter is set to 32, the convolution kernel size is 2, and the expansion is 1 layer of zero padding, the forward and reverse time sequence dependence relationships are extracted respectively, then the reverse calculation result is reversed and spliced and fused with the forward calculation result, to obtain the time sequence dependence relationship of the output vibration signal data. The improved graph attention network in the encoder module is closely combined with the BiMamba network, extracts the feature relationship and time sequence dependence relationship and organically fuses them, which effectively improves the feature extraction performance of the encoder module for vibration signal data in a complex environment.

[0052] As an embodiment of the application, in step S6, the constructed decoder multi-layer neural network module needs to include a multi-layer perceptron MLP layer and a multi-scale convolution layer, and the decoder structure is as shown in Figure 4The time sequence dependent data outputted in the above encoder neural network is inputted into a multi-layer perception MLP layer, firstly passes through a full connection layer, the mapping dimension is 1 / 2 of the original channel number, then uses a GELU activation function for nonlinear activation, after the activation, the data passes through a dropout function with a coefficient of 0.2 for random inactivation to improve the generalization, finally uses a full connection layer to map the data into single-channel time sequence prediction data; the above single-channel time sequence prediction data is inputted into a multi-scale convolution layer, the convolution kernel size is respectively set to 3, 5 and 7, the time sequence relationship of different receptive fields is extracted through different size convolution kernels, and finally the time sequence prediction data is fused into single-step prediction data and outputted. In this way, the data of different channels and different time steps is effectively fused from the channel dimension and the time dimension, as a prediction head to accurately output the remaining useful life prediction value, effectively reduces the information loss of feature fusion, and improves the prediction accuracy.

[0053] As an embodiment of the present application, in step S7, the life label in the training data set needs to be removed in advance before training, the time label of the training data set is processed as an additional 3 time dimensions, the data width parameter is set to 96 unit time steps, the upper limit of the training round is 40 rounds, the early stop patience upper limit is 3 rounds, the life label is combined with the processed training data set and input into the model for training. The training data needs to pass through 2 layers of encoding layer neural network model to extract the relationship between features and time sequence dependence, and then pass through 1 layer of decoding layer neural network model to fuse data features and output predicted values. During the training process, the mean square error loss value (MSE) is calculated according to the above predicted value and the life label, and then the model parameters are updated by using the Adam optimizer for back propagation according to the mean square error loss value, the learning rate of the Adam optimizer is set to 1e-4, and the weight decay is set to 2e-4, which helps the model converge faster and more stably; after completing one round of training, the verification stage is entered, the model training parameters need to be frozen, the remaining useful life is predicted using the verification data set, and the verification data mean square error loss is calculated according to the verification data set predicted value and the life label; after completing a complete verification process, the verification data mean square error loss needs to be recorded and compared with the historical lowest verification data mean square error loss, if the current loss is lower than the historical lowest loss, the current loss is set as the historical lowest loss, the current model parameters are saved as the optimal parameters, and the patience value is set to zero, if the current loss is higher than the historical lowest loss, the patience value is increased by a certain value; after completing the loss comparison of the verification data set, the patience value needs to be judged according to the patience value, if the patience value is higher than the early stop patience upper limit, the training is stopped and the prediction stage is entered, if the patience value is lower than the early stop patience upper limit, the next round of training is entered, and the training gradient is updated according to the mean square error loss of the current round of verification data set; when the patience value is higher than the patience upper limit or the number of training rounds is higher than the set number of rounds, the prediction stage is entered, and the optimal model parameters are used to construct the above-mentioned encoder-decoder architecture multi-layer neural network model to predict the remaining useful life of the test data set.

[0054] Embodiment 2

[0055] The present application also provides a device component remaining life prediction device, comprising:

[0056] The first processing unit is used for collecting and preprocessing the horizontal and vertical vibration signal data during the operation of the industrial device component;

[0057] The second processing unit is used for extracting the time domain, frequency domain and time-frequency domain features of the preprocessed horizontal and vertical vibration signal data, and labeling the life label;

[0058] The third processing unit is configured to calculate evaluation indexes of the time domain, frequency domain and time-frequency domain features related to the life label, determine the index weight by using an improved mixed weight method, screen out a high correlation feature group to form a data set, and divide the data set into a training set, a validation set and a test set;

[0059] The fourth processing unit is configured to construct an encoder-decoder architecture model, wherein the encoder-decoder architecture model comprises a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module is composed of an improved graph attention network layer and a BiMamba module; and the decoder multi-layer neural network module is composed of a multi-layer perceptron (MLP) layer and a multi-scale convolution layer.

[0060] The fifth processing unit is configured to train the encoder-decoder architecture model according to the training data set.

[0061] The sixth processing unit is configured to use the trained encoder-decoder architecture model to perform the remaining useful life prediction on the test data set, and output the remaining useful life prediction value of each step.

[0062] Embodiment 3

[0063] The application further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program which is run by the processor, and the computer program performs the device component remaining life prediction method when being run by the processor.

[0064] Embodiment 4

[0065] The application further provides a storage medium, wherein the storage medium stores a computer program which performs the device component remaining life prediction method when being run.

[0066] The above-described embodiments only describe the preferred modes of the application, and do not limit the scope of the application, and various modifications and improvements to the technical solutions of the application made by those skilled in the art without departing from the design spirit of the application shall fall within the protection scope of the claims of the application.

Claims

1. A method for predicting the remaining life of equipment components, characterized in that, The method comprises the following steps: Collect and preprocess horizontal and vertical vibration signal data during operation of an industrial equipment component; Extract time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data, and label the life label; Calculate the evaluation index related to the life label of the time domain, frequency domain and time-frequency domain features, determine the index weight using an improved mixed weight method, screen out high correlation features to form a data set and divide it into a training set, a validation set and a test set; Build an encoder-decoder architecture model; wherein the encoder-decoder architecture model comprises: a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module is composed of an improved graph attention network layer and a BiMamba module; the decoder multi-layer neural network module is composed of a multi-layer perceptron MLP layer and a multi-scale convolution layer; Train the encoder-decoder architecture model according to the training data set; Use the trained encoder-decoder architecture model to predict the remaining useful life on the test data set, and output the remaining useful life prediction value of each step; The encoder converts the training data set into a non-Euclidean data form, uses the feature variable as the node, uses the node similarity as the initial edge weight, constructs a fully connected graph, and then inputs the constructed graph data into the improved graph attention network layer; the graph attention network layer includes a graph attention network layer GATv2, an adjacent attention layer, a residual attention layer, a normalization layer and a dropout layer; when the non-Euclidean form training data set is input into the improved graph attention network layer, the adjacent attention layer is input first, and the Gaussian kernel weight between the nodes is calculated; based on the calculated Gaussian kernel weight, the Top-K sparsification operation is performed on the graph data to realize the regularization processing of the graph structure; after the sparsified graph data is stabilized by normalization, it is input into the graph attention layer through the dropout layer with a coefficient of 0.2 to enhance the generalization; the graph node data, the graph edge index data and the graph edge weight data obtained after processing are input into the graph attention layer; the graph attention layer performs graph attention calculation based on the received three kinds of graph data to obtain the relationship data between the graph nodes; the relationship data between the graph nodes is input into the residual attention layer; if the layer belongs to the first encoder multilayer neural network module, the above graph data is directly projected into the Euclidean data form of the feature relationship set, saved as historical data and output to the BiMamba module; if the layer belongs to the second encoder multilayer neural network module, the current layer graph node relationship data and the historical data are merged to calculate the attention score, and the graph node relationship data between different layers is weighted and fused using a gating mechanism, and then projected into the Euclidean data form of the feature relationship set and output; this operation can effectively alleviate the overfitting phenomenon; the feature relationship set is input into the BiMamba module to extract the time sequence dependence relationship of the running component feature; the BiMamba model is composed of two parallel Mamba models in forward and reverse order, the Mamba model dimension is set to 256, the state parameter is set to 32, the convolution kernel size is 2, and the expansion is 1 layer of zero padding; after the forward and reverse time sequence dependence relationships are extracted respectively, the reverse calculation result is inverted and spliced and fused with the forward calculation result to obtain the time sequence dependence relationship of the output vibration signal data; The decoder multilayer neural network module inputs the time sequence dependence relationship data output from the encoder neural network into the multilayer perceptron MLP layer; first, the data is mapped to 1 / 2 of the original channel number through the full connection layer 1, then the data is nonlinearly activated through the GELU activation function, then the data is randomly inactivated through the dropout function with a coefficient of 0.2 to improve the generalization, and finally the data is mapped to single-channel time sequence prediction data through the full connection layer; the single-channel time sequence prediction data is input into the multi-scale convolution layer, the convolution kernel size is set to 3, 5 and 7 respectively, the time sequence relationship of different receptive fields is extracted through different size convolution kernels, and finally the time sequence prediction data is fused into single-step prediction data and output.

2. A device component residual life prediction apparatus for implementing the device component residual life prediction method according to claim 1, characterized by, The first processing unit is configured to collect and preprocess horizontal and vertical vibration signal data during the operation of the industrial equipment component. ​ The second processing unit is configured to extract time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data, and label a life label; The third processing unit is configured to calculate evaluation indexes of the time domain, frequency domain and time-frequency domain features related to the life label, determine index weights using an improved mixed weighting method, screen out high correlation features to form a data set, and divide the data set into a training set, a validation set and a test set; The fourth processing unit is configured to construct an encoder-decoder architecture model; wherein the encoder-decoder architecture model comprises: a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module is composed of an improved graph attention network layer and a BiMamba module; the decoder multi-layer neural network module is composed of a multi-layer perceptron MLP layer and a multi-scale convolution layer; The fifth processing unit is configured to train the encoder-decoder architecture model according to the training data set; The sixth processing unit is configured to use the trained encoder-decoder architecture model to perform remaining useful life prediction on the test data set, and output a remaining useful life prediction value for each step.

3. An electronic device, comprising: Comprise: A memory and a processor, the memory has a computer program stored thereon, which is run by the processor, and the computer program performs the device component remaining life prediction method of claim 1 when being run by the processor.

4. A storage medium, characterized by The storage medium has a computer program stored thereon, which performs the device component remaining life prediction method of claim 1 when being run.

Images