Lightweight device fault prediction method based on feature extraction and feature fusion

Abstract

This invention relates to a lightweight equipment fault prediction method based on feature extraction and feature fusion. The method includes: acquiring raw monitoring data from multiple sensors of the equipment; performing K-Means clustering, Z-Score standardization, and remaining life labeling preprocessing on the raw sensor data; evaluating the preprocessed sensor features based on Pearson correlation coefficient, maximum information coefficient, and monotonicity index; determining the weights of each feature evaluation index using the CRITIC weighting method; selecting the optimal subset of sensor features using simulated annealing; and obtaining sensor features using principal component analysis. The features are then input into a lightweight LSTM network for training to obtain a fault prediction model, which predicts the remaining lifespan of the equipment. This invention effectively improves the performance of the equipment remaining lifespan fault prediction model, reduces model training overhead, and is suitable for fault prediction of industrial equipment.

Description

Technical Field

[0001] This invention belongs to the field of predictive maintenance of equipment, and specifically relates to a lightweight equipment fault prediction method based on feature extraction and feature fusion. Background Technology

[0002] Predictive maintenance technology is a technology that uses multiple high-precision sensors to collect performance parameters of key components, then uses data analysis to monitor the health status of the equipment in real time, and uses performance degradation models to assess the remaining service life of the equipment.

[0003] Not all of these sensor data are suitable for training equipment remaining life prediction models, which makes existing predictive maintenance techniques ineffective in predicting equipment remaining life. Summary of the Invention

[0004] To address the shortcomings of the existing technologies, this invention provides a lightweight equipment fault prediction method based on feature extraction and feature fusion.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A lightweight device fault prediction method based on feature extraction and feature fusion is provided, comprising the following steps:

[0007] Step S1: Obtain monitoring data from multiple sensors mounted on the device, and perform K-Means clustering, Z-Score standardization, and remaining lifetime label polyline preprocessing on the multi-sensor monitoring data to obtain sensor preprocessing features;

[0008] Step S2: The sensor preprocessing features in step S1 are evaluated based on three feature indicators: Pearson correlation coefficient, maximum information coefficient (MIC), and monotonicity, to obtain three feature evaluation indicators.

[0009] Step S3: Based on the three feature evaluation indicators in step S2, the weights of each feature evaluation indicator are determined using the CRITIC weighting method.

[0010] Step S4: Based on the weights of each feature evaluation index in step S3, use the simulated annealing algorithm to select the optimal subset of sensor features suitable for the equipment remaining service failure prediction model;

[0011] Step S5: Decentralize all features within the optimal sensor feature subset in step S4, i.e., subtract the average value from each feature vector.

[0012] Step S6: Calculate the covariance matrix of the eigenvectors in step S5;

[0013] Step S7: Perform eigenvalue decomposition on the covariance matrix obtained in step S6 to obtain eigenvalues ​​and eigenvectors;

[0014] Step S8: Sort the feature values ​​in step S7 from largest to smallest, select the two largest ones, and form a feature vector matrix from their corresponding feature vectors.

[0015] Step S9: Transform the optimal subset of sensor features into the new space constructed by the two feature vectors in step S8, thus obtaining the dimensionality-reduced features.

[0016] Step S10: Input the features obtained in step S9 into a lightweight LSTM network model to train a fault prediction model, and then predict the remaining service life of the equipment using the fault prediction model.

[0017] Furthermore, step S1 specifically includes the following:

[0018] Step S11: Acquire monitoring data from the multiple sensors mounted on the device;

[0019] Step S12: Using the K-means clustering method, cluster analysis is performed based on the equipment's operation settings to divide the sensor monitoring data at each time point according to the working status.

[0020] Step S13: Perform Z-Score standardization on the sensor monitoring data under the same working condition;

[0021] Step S14: The remaining life label of the equipment is processed into a piecewise linear representation, which is mathematically represented as a piecewise function. The specific formula is as follows:

[0022]

[0023] Where U is the upper bound of the remaining lifespan, which is determined by the upper limit of the equipment's lifespan;

[0024] Step S15: Obtain the sensor preprocessing features.

[0025] Furthermore, the specific calculation steps for the Maximum Information Coefficient (MIC) index in step S2 are as follows:

[0026] Step S21: Plot a scatter plot with sensor monitoring data X and remaining life tag value Y as coordinate axes. Draw a rows and b columns on the scatter plot along the direction parallel to the coordinate axes to realize the grid division of the scatter plot.

[0027] Step S22: Based on the grid division of the scatter plot in step S21, calculate the maximum mutual information value I(X; Y) under the current division. The mutual information value is calculated by the probability distribution of the scatter points falling into each grid, and the calculation formula is as follows:

[0028]

[0029] Where i and j represent the range of the grid area after being divided into a rows and b columns on the X and Y coordinate axes, respectively;

[0030] Step S23: Normalize the maximum mutual information value in step S22 to obtain the normalized maximum mutual information value;

[0031] Step S24: Select the normalized maximum mutual information value under different partitioning schemes as the MIC result between X and Y. The specific formula is as follows:

[0032]

[0033] Where B = n 0.6 , where n is the data length of X and Y.

[0034] Furthermore, the specific steps of the CRITIC weighting method in step S3 are as follows:

[0035] Step S31: The Min-Max normalization method was used to remove dimensions from all indicators. The specific mathematical formula is as follows:

[0036]

[0037] Step S32: Calculate the contrast of the indicators, that is, calculate the standard deviation of the data for all indicators. The calculation formula is as follows:

[0038]

[0039] Where n is the data length under index j, x ij Let j be the i-th value of index j. The average value of index j;

[0040] Step S33: Calculate the conflict of indicators by measuring the Pearson correlation coefficient between them. The specific calculation formula is shown below:

[0041]

[0042] Where, r ij represents the correlation coefficient between indicator i and indicator j, where p is the number of indicators;

[0043] Step S34: Calculate the information content and perform normalization processing. The specific calculation formula is shown below:

[0044] C j =S j ×R j

[0045]

[0046] Finally, the objective weight W of each evaluation indicator is obtained. j .

[0047] Furthermore, step S4 specifically includes the following:

[0048] Step S41: Select a common residual fault prediction model as the evaluation model for feature subset selection;

[0049] Step S42: Randomly initialize a feature subset, evaluate the feature subset using the residual fault prediction model, use the variance of the prediction results as the internal energy of the current state, and set the initial temperature, termination temperature and temperature decay coefficient.

[0050] Step S43: Randomly perturb the selection of the feature subset, and assign the probability of each feature being selected according to the objective weight obtained in step S3. The greater the weight, the greater the probability of the feature being selected, and the smaller the weight, the greater the probability of the feature being replaced.

[0051] Step S44: Use the residual fault prediction model to evaluate the disturbed feature subset. If the variance of the prediction result decreases, update the feature subset. The temperature is decayed according to the decay coefficient.

[0052] Step S45: Repeat steps S43 and S44 until the temperature decay reaches the set termination temperature, and finally obtain the selected optimal sensor feature subset.

[0053] Furthermore, the construction of the fault prediction model in step S10 specifically includes the following:

[0054] Step S101: Construct a three-layer lightweight LSTM neural network model, including a fully connected input layer, an LSTM neuron layer, and a pooling output layer;

[0055] Step S102: Divide the dimensionality-reduced features obtained in step S9 into a training set, a validation set, and a test set, and input the training set data into the LSTM neural network model for training;

[0056] Step S103: Use the validation set and test set to validate and test the LSTM neural network model, verify the model's performance, and obtain the fault prediction model.

[0057] The beneficial effects of this invention are as follows: First, by combining the changing trends of sensor data with the actual performance degradation process of the device, the remaining lifetime label is processed using a piecewise linear model. After processing, the correlation between the remaining lifetime label and the sensor data is significantly improved. Second, different operating settings of the device determine its working state. Under different working states, the sensor values ​​differ significantly. This difference masks the performance degradation trend of the device itself. In order to extract the changing trend of the remaining lifetime, cluster preprocessing is performed on the device's working states, followed by standardization processing, which can significantly improve the correlation between sensor data and remaining lifetime. Finally, sensor data are evaluated based on Pearson correlation coefficient, maximum information coefficient, and monotonicity index. The Critic weighting method is used to assign weights to the three indices, resulting in the weight of each sensor. The optimal subset of sensor features is then selected using simulated annealing algorithm based on these weights. Finally, principal component analysis is used to reduce the dimensionality of the data, making the output feature dimension suitable for training a lightweight model. This allows for the training of a fault prediction model, which then predicts the remaining lifespan of the equipment. This effectively avoids the problem of overfitting in the subsequent training of the remaining fault prediction model due to redundant features and excessively high data dimensionality, while also reducing the overhead of model training. This achieves the training and deployment of a lightweight fault prediction model. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A detailed flowchart of a lightweight device fault prediction method based on feature extraction and feature fusion provided as an exemplary embodiment;

[0060] Figure 2 A structural diagram of a lightweight device fault prediction device based on feature extraction and feature fusion, provided as an exemplary embodiment. Detailed Implementation

[0061] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0062] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0063] like Figure 1 As shown, a lightweight equipment fault prediction method based on feature extraction and feature fusion is provided, including the following steps:

[0064] Step S1: Obtain monitoring data from multiple sensors mounted on the device, and perform K-Means clustering, Z-Score standardization, and remaining lifetime label polyline preprocessing on the multi-sensor monitoring data to obtain sensor preprocessing features;

[0065] In one embodiment, step S1 specifically includes the following:

[0066] Step S11: Acquire monitoring data from multiple sensors mounted on the device; wherein, the monitoring data from multiple sensors is a time series signal, which generally involves sensing characteristics that indicate key information about the device's operating status, such as temperature, air pressure, degree of deformation, speed, vibration amplitude / frequency, and sound signals.

[0067] Step S12: Using the K-means clustering method, cluster analysis is performed based on the equipment's operation settings to divide the sensor monitoring data at each time point according to the working state, so as to eliminate the differences caused by different working states.

[0068] The main purpose of preprocessing sensor monitoring data and equipment remaining life tags is to alleviate the problem of low correlation between sensor monitoring data and remaining life tags caused by the coexistence of multiple operating states of equipment and the inconsistency of sensor physical dimensions.

[0069] In one embodiment, the specific steps of the K-Means clustering algorithm are as follows:

[0070] Step S121: Randomly select N points from the data sample as the centroids of N clusters, where N is the number of working states of the device;

[0071] Step S122: For each discrete operation setting sample point in the dataset, calculate its distance to each centroid, and assign the discrete point to the cluster to which the centroid with the smallest distance value belongs; use Euclidean distance as the distance metric.

[0072] Step S123: Calculate the geometric center of the cluster and update the centroids of the six clusters;

[0073] Step S124: Repeat steps S122 and S123 until the centroid is fixed and the clustering operation is stopped;

[0074] Step S13: Perform Z-Score standardization on the sensor monitoring data under the same operating condition to eliminate the interference caused by the differences in the physical dimensions and orders of magnitude of the sensors on the feature selection and remaining lifetime prediction process. The specific calculation formula is as follows:

[0075]

[0076] Where μ is the mean of all sensor monitoring data of the sensor, and σ is the standard deviation of all sensor monitoring data;

[0077] Step S14: Sensor data only shows a clear monotonically increasing / decreasing trend in the final stage of the equipment's lifecycle, i.e., close to the time of failure. However, in the short period after the equipment starts operating, the sensor data is relatively stable due to the good equipment condition, and no obvious performance degradation trend is observed. However, the theoretical value of the remaining lifespan is a linear function that monotonically decreases with the increase of the operating cycle, which does not match the sensor change trend in the early stages of equipment operation. To improve the correlation between sensor monitoring data and the equipment's remaining lifespan label, the remaining lifespan label is processed as a piecewise linear function, mathematically represented as a piecewise function, as shown in the following formula:

[0078]

[0079] Where U is the upper bound of the remaining lifespan, which is determined by the upper limit of the equipment's lifespan;

[0080] Step S15: Obtain the sensor preprocessing features.

[0081] Step S2: The sensor preprocessing features in step S1 are evaluated based on three feature indicators: Pearson correlation coefficient, maximum information coefficient (MIC), and monotonicity, to obtain three feature evaluation indicators.

[0082] In one embodiment, the specific formula for calculating the Pearson correlation coefficient in step S2 is as follows:

[0083]

[0084] Where x t Let R be the value monitored by the sensor at time t. t This represents the remaining lifetime tag value of the device at time t. and These are the average values ​​of the sensor monitoring data and the average values ​​of the remaining life tag, respectively.

[0085] In one embodiment, the specific calculation steps for the Maximum Information Coefficient (MIC) index in step S2 are as follows:

[0086] Step S21: Plot a scatter plot with sensor monitoring data X and remaining life tag value Y as coordinate axes. Draw a rows and b columns on the scatter plot along the direction parallel to the coordinate axes to realize the grid division of the scatter plot.

[0087] Step S22: Based on the grid division of the scatter plot in step S21, calculate the maximum mutual information value I(X;Y) under the current division. The mutual information value is calculated by the probability distribution of the scatter points falling into each grid, and the calculation formula is as follows:

[0088]

[0089] Where i and j represent the range of the grid area after being divided into a rows and b columns on the X and Y coordinate axes, respectively;

[0090] Step S23: Normalize the maximum mutual information value in step S22 to obtain the normalized maximum mutual information value;

[0091] Step S24: Select the normalized maximum mutual information value under different partitioning schemes as the MIC result between X and Y. The specific formula is as follows:

[0092]

[0093] Where B = n 0.6 , where n is the data length of X and Y.

[0094] In one embodiment, the specific formula for calculating the monotonicity index in step S2 is as follows:

[0095]

[0096] Where T is the length of the sensor monitoring data sequence, and dx is the derivative of the current sensor monitoring data with respect to the operating cycle; since the sensor data is a discrete sequence, it is replaced by a difference.

[0097] Step S3: Based on the three feature evaluation indicators in step S2, the weights of each feature evaluation indicator are determined using the CRITIC weighting method.

[0098] The CRITIC weighting method assigns appropriate weights to the Pearson correlation coefficient, maximum information coefficient, and monotonicity evaluation index of all sensors, and finally calculates the selection weight for each sensor by weighting.

[0099] weight=α·Pearson+β·MIC+γ·Mono

[0100] The weights α, β, and γ of the three evaluation indicators are determined by the CRITIC weighting method.

[0101] In one embodiment, the specific steps of the CRITIC weighting method in step S3 are as follows:

[0102] Step S31: The Min-Max normalization method was used to remove dimensions from all indicators. The specific mathematical formula is as follows:

[0103]

[0104] Step S32: Calculate the contrast of the indicators, that is, calculate the standard deviation of the data for all indicators. The calculation formula is as follows:

[0105]

[0106] Where n is the data length under index j, x ij Let j be the i-th value of index j. The average value of index j;

[0107] Step S33: Calculate the conflict of indicators by measuring the Pearson correlation coefficient between them. The specific calculation formula is shown below:

[0108]

[0109] Where, r ij represents the correlation coefficient between indicator i and indicator j, where p is the number of indicators;

[0110] Step S34: Calculate the information content and perform normalization processing. The specific calculation formula is shown below:

[0111] C j =S j ×R j

[0112]

[0113] Finally, the objective weight W of each evaluation indicator is obtained. j .

[0114] Step S4: Based on the weights of each feature evaluation index in step S3, use the simulated annealing algorithm to select the optimal subset of sensor features suitable for the equipment remaining service failure prediction model;

[0115] In one embodiment, step S4 specifically includes the following:

[0116] Step S41: Select a common residual fault prediction model as the evaluation model for feature subset selection;

[0117] Step S42: Randomly initialize a feature subset, evaluate the feature subset using the residual fault prediction model, use the variance of the prediction results as the internal energy of the current state, and set the initial temperature, termination temperature and temperature decay coefficient.

[0118] Step S43: Randomly perturb the selection of the feature subset, and assign the probability of each feature being selected according to the objective weight obtained in step S3. The greater the weight, the greater the probability of the feature being selected, and the smaller the weight, the greater the probability of the feature being replaced.

[0119] Step S44: Use the residual fault prediction model to evaluate the disturbed feature subset. If the variance of the prediction result decreases, update the feature subset. The temperature is decayed according to the decay coefficient.

[0120] Step S45: Repeat steps S43 and S44 until the temperature decay reaches the set termination temperature, and finally obtain the selected optimal sensor feature subset.

[0121] Step S5: Decentralize all features within the optimal sensor feature subset in step S4, i.e., subtract the average value from each feature vector.

[0122] Step S6: Calculate the covariance matrix of the eigenvectors in step S5;

[0123] Step S7: Perform eigenvalue decomposition on the covariance matrix obtained in step S6 to obtain eigenvalues ​​and eigenvectors;

[0124] Step S8: Sort the feature values ​​in step S7 from largest to smallest, select the two largest ones, and form a feature vector matrix from their corresponding feature vectors.

[0125] Step S9: Transform the optimal subset of sensor features into the new space constructed by the two feature vectors in step S8, thus obtaining the dimensionality-reduced features.

[0126] Step S10: Input the features obtained in step S9 into a lightweight LSTM network model to train a fault prediction model, and then predict the remaining service life of the equipment using the fault prediction model.

[0127] In one embodiment, the construction of the fault prediction model in step S10 specifically includes the following:

[0128] Step S101: Construct a three-layer lightweight LSTM neural network model, including a fully connected input layer, an LSTM neuron layer, and a pooling output layer;

[0129] Step S102: Divide the dimensionality-reduced features obtained in step S9 into a training set, a validation set, and a test set, and input the training set data into the LSTM neural network model for training;

[0130] Step S103: Use the validation set and test set to validate and test the LSTM neural network model, verify the model's performance, and obtain the fault prediction model.

[0131] This lightweight equipment failure prediction method, based on feature extraction and feature fusion, firstly, performs piecewise linearization on the remaining lifetime label by combining the changing trends of sensor data with the actual performance degradation process of the equipment. This significantly improves the correlation between the remaining lifetime label and the sensor data. Secondly, different operating settings of the equipment determine its working state, and sensor values ​​vary considerably under different working states. This variation masks the performance degradation trend of the equipment itself. To extract the remaining lifetime changing trend, clustering preprocessing is performed on the equipment's working states, followed by standardization, which significantly improves the correlation between sensor data and remaining lifetime. Finally, sensor data are evaluated based on Pearson correlation coefficient, maximum information coefficient, and monotonicity index. The Critic weighting method is used to assign weights to the three indices, resulting in the weight of each sensor. The optimal subset of sensor features is then selected using simulated annealing algorithm based on these weights. Finally, principal component analysis is used to reduce the dimensionality of the data, making the output feature dimension suitable for training a lightweight model. This allows for the training of a fault prediction model, which then predicts the remaining lifespan of the equipment. This effectively avoids the problem of overfitting in the subsequent training of the remaining fault prediction model due to redundant features and excessively high data dimensionality, while also reducing the overhead of model training. This achieves the training and deployment of a lightweight fault prediction model.

[0132] Corresponding to the aforementioned embodiments of the lightweight equipment fault prediction method based on feature extraction and feature fusion, the present invention also provides embodiments of a lightweight equipment fault prediction device based on feature extraction and feature fusion.

[0133] See Figure 2 The present invention provides a lightweight equipment fault prediction device based on feature extraction and feature fusion, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement the lightweight equipment fault prediction method based on feature extraction and feature fusion in the above embodiment.

[0134] The embodiment of the lightweight device fault prediction device based on feature extraction and feature fusion of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2The diagram shown is a hardware structure diagram of any data processing-capable device, including the lightweight equipment fault prediction device based on feature extraction and feature fusion according to the present invention. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0135] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0136] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0137] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the lightweight device fault prediction method based on feature extraction and feature fusion described in the above embodiments.

[0138] The computer-readable storage medium can be an internal storage unit of any data-processing device in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data-processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., mounted on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store computer programs and other programs and data required by any data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0139] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0140] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0141] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of one or more embodiments of this specification. The singular forms “a,” “the,” and “the” used in one or more embodiments of this specification and in the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0142] It should be understood that although the terms first, second, third, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of one or more embodiments of this specification, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "in response to a determination," or "when," or "in the event of a determination."

[0143] The above are merely preferred embodiments of one or more embodiments of this specification and are not intended to limit the scope of one or more embodiments of this specification. Any modifications made within the spirit and principles of one or more embodiments of this specification are permitted.

Claims

1. A lightweight equipment fault prediction method based on feature extraction and feature fusion, characterized in that, Includes the following steps: Step S1: Obtain monitoring data from multiple sensors mounted on the device, and perform K-Means clustering, Z-Score standardization, and remaining lifetime label polyline preprocessing on the multi-sensor monitoring data to obtain sensor preprocessing features; Step S2: The sensor preprocessing features in step S1 are evaluated based on three feature indicators: Pearson correlation coefficient, maximum information coefficient (MIC), and monotonicity, to obtain three feature evaluation indicators. Step S3: Based on the three feature evaluation indicators in step S2, the weights of each feature evaluation indicator are determined using the CRITIC weighting method. Step S4: Based on the weights of each feature evaluation index in step S3, use the simulated annealing algorithm to select the optimal subset of sensor features suitable for the equipment remaining service failure prediction model; Step S5: Decentralize all features within the optimal sensor feature subset in step S4, i.e., subtract the average value from each feature vector. Step S6: Calculate the covariance matrix of the eigenvectors in step S5; Step S7: Perform eigenvalue decomposition on the covariance matrix obtained in step S6 to obtain eigenvalues ​​and eigenvectors; Step S8: Sort the feature values ​​in step S7 from largest to smallest, select the two largest ones, and form a feature vector matrix from their corresponding feature vectors. Step S9: Transform the optimal subset of sensor features into the new space constructed by the two feature vectors in step S8, thus obtaining the dimensionality-reduced features. Step S10: Input the features obtained in step S9 into a lightweight LSTM network model to train a fault prediction model, and then predict the remaining service life of the equipment using the fault prediction model.

2. The lightweight equipment fault prediction method based on feature extraction and feature fusion according to claim 1, characterized in that, Step S1 specifically includes the following: Step S11: Acquire monitoring data from the multiple sensors mounted on the device; Step S12: Using the K-means clustering method, cluster analysis is performed based on the equipment's operation settings to divide the sensor monitoring data at each time point according to the working status. Step S13: Perform Z-Score standardization on the sensor monitoring data under the same working condition; Step S14: The remaining life label of the equipment is processed into a piecewise linear representation, which is mathematically represented as a piecewise function. The specific formula is as follows: Where U is the upper bound of the remaining lifespan, which is determined by the upper limit of the equipment's lifespan; Step S15: Obtain the sensor preprocessing features.

3. The lightweight equipment fault prediction method based on feature extraction and feature fusion according to claim 1, characterized in that, The specific calculation steps for the Maximum Information Coefficient (MIC) in step S2 are as follows: Step S21: Plot a scatter plot with sensor monitoring data X and remaining life tag value Y as coordinate axes. Draw a rows and b columns on the scatter plot along the direction parallel to the coordinate axes to realize the grid division of the scatter plot. Step S22: Based on the grid division of the scatter plot in step S21, calculate the maximum mutual information value I(X; Y) under the current division. The mutual information value is calculated by the probability distribution of the scatter points falling into each grid, and the calculation formula is as follows: Where i and j represent the range of the grid area after being divided into a rows and b columns on the X and Y coordinate axes, respectively; Step S23: Normalize the maximum mutual information value in step S22 to obtain the normalized maximum mutual information value; Step S24: Select the normalized maximum mutual information value under different partitioning schemes as the MIC result between X and Y. The specific formula is as follows: where B = n 0.6 n is the data length of X and Y.

4. The lightweight equipment fault prediction method based on feature extraction and feature fusion according to claim 1, characterized in that, The specific steps of the CRITIC weighting method in step S3 are as follows: Step S31: The Min-Max normalization method was used to remove dimensions from all indicators. The specific mathematical formula is as follows: Step S32: Calculate the contrast of the indicators, that is, calculate the standard deviation of the data for all indicators. The calculation formula is as follows: Where n is the data length under index j, x ij Let j be the i-th value of index j. The average value of index j; Step S33: Calculate the conflict of indicators by measuring the Pearson correlation coefficient between them. The specific calculation formula is shown below: Where, r ij represents the correlation coefficient between indicator i and indicator j, where p is the number of indicators; Step S34: Calculate the information content and perform normalization processing. The specific calculation formula is shown below: C j =S j ×R j Finally, the objective weight W of each evaluation indicator is obtained. j .

5. The lightweight equipment fault prediction method based on feature extraction and feature fusion according to claim 1, characterized in that, Step S4 specifically includes the following: Step S41: Select a common residual fault prediction model as the evaluation model for feature subset selection; Step S42: Randomly initialize a feature subset, evaluate the feature subset using the residual fault prediction model, use the variance of the prediction results as the internal energy of the current state, and set the initial temperature, termination temperature and temperature decay coefficient. Step S43: Randomly perturb the selection of the feature subset, and assign the probability of each feature being selected according to the objective weight obtained in step S3. The greater the weight, the greater the probability of the feature being selected, and the smaller the weight, the greater the probability of the feature being replaced. Step S44: Use the residual fault prediction model to evaluate the disturbed feature subset. If the variance of the prediction result decreases, update the feature subset. The temperature is decayed according to the decay coefficient. Step S45: Repeat steps S43 and S44 until the temperature decay reaches the set termination temperature, and finally obtain the selected optimal sensor feature subset.

6. The lightweight equipment fault prediction method based on feature extraction and feature fusion according to claim 1, characterized in that, The construction of the fault prediction model in step S10 specifically includes the following: Step S101: Construct a three-layer lightweight LSTM neural network model, including a fully connected input layer, an LSTM neuron layer, and a pooling output layer; Step S102: Divide the dimensionality-reduced features obtained in step S9 into a training set, a validation set, and a test set, and input the training set data into the LSTM neural network model for training; Step S103: Use the validation set and test set to validate and test the LSTM neural network model, verify the model's performance, and obtain the fault prediction model.

Images