Equipment fault diagnosis method and system based on missing data
By combining LSTM and random forest models with sparse principal component analysis and the K-nearest neighbor method for gray-level correlation coefficients, the problem of fault diagnosis with missing data in complex equipment systems was solved, achieving efficient and accurate fault diagnosis and real-time monitoring, and improving equipment operation and maintenance efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI AEROSPACE VEHICLE RES INST
- Filing Date
- 2025-11-25
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies struggle to balance the accuracy of missing data interpolation with the ability to model complex time series data when dealing with missing data in complex equipment systems. This results in insufficient accuracy and robustness in fault diagnosis, especially when the data volume is large, changes over time, and sensor data are highly correlated.
A two-layer model is constructed using LSTM and random forest. Combined with data preprocessing, sparse principal component analysis and the gray-level correlation coefficient K-nearest neighbor method, the model achieves accurate data filling and fault diagnosis through a collaborative architecture of data collection, server and client.
It improves the accuracy and robustness of fault diagnosis for complex equipment, ensures data integrity and time consistency, and realizes fully automated processing from data acquisition to fault diagnosis, reducing unplanned downtime and improving operation and maintenance efficiency.
Smart Images
Figure CN121901903A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of industrial automation and fault diagnosis technology, and more specifically, relates to a method and system for equipment fault diagnosis based on missing data. Background Technology
[0002] Complex equipment systems are complex due to their high complexity and strong coupling between numerous devices, resulting in a large amount of data that changes significantly over time. Furthermore, there are strong correlations between the data from various sensors, which poses many challenges in fault diagnosis, especially when data is missing.
[0003] To address the problem of equipment fault diagnosis in the presence of missing data, existing methods typically involve simple deletion or ignoring of missing items, or imputation using zero, mean, or median values. While these methods are simple to operate, they can significantly reduce the available data in the dataset, distort the distribution characteristics of the dataset, and alter underlying features, which is detrimental to subsequent analysis. Furthermore, when analyzing time series data, they may disrupt the temporal correlation of data items, leading to significant biases and making subsequent analysis more difficult.
[0004] The conventional approach uses random forest regression to impute missing data and combines it with an XGBoost model for transformer fault diagnosis. This method performs well in handling incomplete data and can be combined with different data sources. However, by only using a regression model to impute missing data, it fails to consider the complex relationships between missing data and other data features. This may lead to biases in the imputation results and affect the accuracy of the final fault diagnosis.
[0005] There are also methods that construct tree-like models and use Manhattan distance to find the nearest neighboring node data to repair missing data, solving the problem of missing transformer oil chromatographic data. This method improves the accuracy of fault diagnosis to some extent, but it relies on specific oil chromatographic data, and the repair method depends on adjacent node data, making it susceptible to uneven data distribution. This method has limitations in other broader fault diagnosis scenarios, especially when the data dimension is high and the missing patterns are complex, making it difficult to guarantee the consistency and accuracy of its repair effect. Therefore, there is currently no publicly available technology that can simultaneously achieve both the accuracy of missing data imputation and the ability to model complex time series data, while realizing a fully deployable equipment fault diagnosis system.
[0006] In view of this, it is necessary to propose a fault diagnosis method for missing data, especially for missing data of multivariate time series, so that it can be applied to complex equipment with huge data volume, time variation, strong correlation of data from various sensors, and many types of faults. It is also necessary to fully explore the correlation between missing data and non-missing data, establish an effective time series data imputation model, complete the imputation of missing data, and achieve efficient and accurate fault diagnosis. Summary of the Invention
[0007] This invention aims to provide a method and system for diagnosing equipment faults based on missing data. By filling in missing values through data preprocessing, constructing a two-layer model using LSTM and random forest, and combining a collaborative architecture of acquisition end, server end, and client end, it can accurately extract time-series features and diagnose faults, thereby improving the accuracy and robustness of fault diagnosis for complex equipment and ensuring reliable equipment operation.
[0008] In view of the above-mentioned defects or improvement needs of the existing technology, as a first aspect of the present invention, the present invention provides an equipment fault diagnosis method based on missing data, comprising: S1. Complete the cleaning of the raw data; S2. The missing values of the data are initially imputed by the mean imputation method to form a new data sample and complete the data normalization; then the missing values are initially estimated by the sparse principal component analysis method to obtain the preliminary estimated values of the missing values; finally, the K nearest neighbor method of gray-level correlation coefficient is used to iteratively optimize and obtain the final estimated values of the relevant missing values. S3. The LSTM method was selected and a preliminary diagnostic model was trained based on multidimensional time series historical data. The model was then optimized using the random forest method. The LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification. S4. Deploy the trained model on the server side; collect data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmit the data to the server after completing the data preprocessing through the data processing methods in S1 and S2; the server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client; the client is configured to view the equipment status and fault diagnosis results in real time.
[0009] Furthermore, the sparse principal component analysis method in S2 is specifically as follows: The objective function for the SPCA model is established as follows: , in, It is the data matrix after data normalization. It is a low-dimensional principal component matrix. It is a projection matrix. It is the sparse regularization coefficient. It is the Frobenius norm of the data matrix. It is the first The L1 norm of each principal component vector; The SPCA method was used to make a preliminary estimate of missing values, and the missing values were then... The corresponding estimated value is used as the initial interpolation value, as shown in the following formula: , in, This represents the final output, corresponding to the [number]th [item] in the data matrix. line, number The estimated column position is used to fill in missing parts of the original data.
[0010] Furthermore, the nearest neighbor method for the gray-level correlation coefficient K in S2 includes: constructing the gray-level correlation coefficient to calculate the overall gray-level similarity; Construct a gray-level correlation coefficient matrix and calculate the gray-level similarity between data samples. Let the first... Each sample is a reference sequence. , No. Each sample is a comparison sequence. Then the first The formula for calculating the correlation coefficient of each dimension is as follows; , in, Indicates the first In each dimension and The gray-scale correlation coefficient; Represents the resolution coefficient; Next, calculate the overall gray-level similarity, which is to take the average of all dimensions to obtain the overall gray-level similarity between the two data samples. The formula is as follows: , in, This represents the final calculation result, i.e., the first... The first sample and the first The overall gray-level similarity value of each sample.
[0011] Furthermore, the K-nearest neighbor method for gray-level correlation coefficients in S2 also includes GKNN imputation of missing values, specifically: For each sample with missing values, select the sample with the highest gray-level similarity. For each complete sample, construct a neighborhood set. Then, weighted interpolation is performed; the interpolation formula is as follows: , in, It is the first The first sample Estimates of the missing features; It is the first The neighbor in the The true value of the dimension; It is a sample With sample Gray-scale similarity.
[0012] Furthermore, the LSTM model in S2 includes a data matrix obtained by fusing multivariate data. The input layer serves as the input data, the LSTM layer is configured to learn the time series data in the input layer, the Dropout layer is configured to reduce overfitting during network training, and the output layer is configured to output the fault types of the equipment. The Dropout layer is set to a dropout ratio of no less than 0.2; the output layer uses the SoftMax activation function.
[0013] Furthermore, in the LSTM model of S2, each LSTM unit is... The specific formula for updating time is as follows: , in, In time step Input; It is a time step The hidden state; It is a time step The hidden state; It is a time step cellular state; It is a candidate cell state; It is a time step cellular state; and These are the weight matrix and the bias vector, where the subscripts are... These represent the weight matrix and bias vector corresponding to the forget gate, input gate, output gate, and candidate cell state, respectively. yes Activation function , It is the tangent function of a hyperbola. ; It is the output of the forget gate; It is an input gate output; It is the input gate output.
[0014] Furthermore, the specific process of optimizing the model using the random forest method in S2 is as follows: After the LSTM model is trained, a random forest method is used for optimization. This method improves diagnostic accuracy through a voting process involving multiple decision trees, as shown in the following formula: , in, It is the first The prediction function for each tree. It's the weight.
[0015] Furthermore, the training process of the LSTM model in S2 is as follows: The LSTM model is trained using a historical dataset, and the weights are updated using backpropagation. The multi-class cross-entropy function is used as the loss function. To assess the error between the predicted and actual values, the following methods are used: , in, It is the total number of samples. It is the total number of fault categories. It is the first Does each sample belong to a category? The true situation The model predicts the first Each sample belongs to category The probability of.
[0016] As a second aspect of the present invention, an equipment fault diagnosis system based on missing data is also provided, including... The data cleaning unit is used to clean the raw data. The missing value processing unit is used to initially fill in the missing values of the data using the mean imputation method, forming new data samples and completing data normalization; then, it uses the sparse principal component analysis method to perform initial estimation of the missing values, obtaining preliminary estimates of the missing values; finally, it uses the gray-level correlation coefficient K nearest neighbor method to iteratively optimize and obtain the final estimates of the relevant missing values. The model training unit is used to select the LSTM method and train a preliminary diagnostic model based on multidimensional time series historical data, and optimize the model using the random forest method; the LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification; The model deployment service unit is used to deploy the model trained on the server side; it collects data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmits the data to the server after completing the data preprocessing through the data processing methods in S1 and S2; the server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client; the client is configured to view the equipment status and fault diagnosis results in real time.
[0017] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor of any step of the equipment fault diagnosis method based on missing data.
[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The equipment fault diagnosis method based on missing data of the present invention maintains consistency in sensor data from different sources through standardized format processing, eliminates duplicate records to reduce data redundancy, and removes outliers to eliminate noise interference. This series of operations solves the problems of chaotic data format, redundancy, and noise interference in industrial fields, providing standardized and clean data for subsequent processing. It avoids diagnostic bias caused by problems with the quality of the original data, thus building a solid data foundation for equipment fault diagnosis, ensuring that subsequent analysis is based on a reliable data foundation, and improving the initial quality of the entire diagnostic process.
[0019] 2. The equipment fault diagnosis method based on missing data of the present invention maintains the basic structure of the data by using mean imputation, and then unifies the data to the same dimension range to eliminate the scale influence between different features. Next, a dimensionality reduction method is used to initially estimate the missing values, reducing data dimensionality while retaining key features. Finally, by calculating the similarity between samples, similar samples are selected to perform weighted imputation of missing values. This multi-level missing value processing method effectively solves the problem of data loss caused by environmental interference during equipment operation, restores data integrity and temporal consistency, allows the data to more realistically reflect the equipment's operating status, provides complete data support for accurate fault diagnosis, and avoids diagnostic errors caused by data loss.
[0020] 3. The equipment fault diagnosis method based on missing data of this invention utilizes a special network to extract time-series features from equipment operation data, captures the patterns of parameter changes over time, and then optimizes the classification results through a multi-decision-tree voting method, improving the model's accuracy and robustness. After deploying the trained model, real-time collected data, after cleaning and missing value repair, is input into the model to quickly output fault probabilities. Clients can view the status and diagnostic results in real time. This system achieves fully automated processing from data acquisition to fault diagnosis, solving the problem of real-time monitoring of industrial equipment faults. It enables maintenance personnel to promptly grasp the equipment status, detect potential faults in advance, reduce unplanned downtime, and improve equipment operation and maintenance efficiency, which is of significant practical importance for ensuring the continuity and stability of industrial production. Attached Figure Description
[0021] Figure 1 This is a flowchart of an equipment fault diagnosis method based on missing data according to an embodiment of the present invention; Figure 2This is a schematic diagram of the overall architecture of an embodiment of the present invention; Figure 3 This is a schematic diagram of the missing value filling process according to an embodiment of the present invention; Figure 4 This is a system unit diagram of an embodiment of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0023] Example 1 Please refer to Figure 1 This embodiment 1 provides a method for diagnosing equipment faults based on missing data, including: S1. Complete the cleaning of the raw data; S2. The missing values of the data are initially imputed by the mean imputation method to form a new data sample and complete the data normalization; then the missing values are initially estimated by the sparse principal component analysis method to obtain the preliminary estimated values of the missing values; finally, the K nearest neighbor method of gray-level correlation coefficient is used to iteratively optimize and obtain the final estimated values of the relevant missing values. S3. The LSTM method was selected and a preliminary diagnostic model was trained based on multidimensional time series historical data. The model was then optimized using the random forest method. The LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification. S4. Deploy the trained model on the server side; collect data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmit the data to the server after completing the data preprocessing through the data processing methods in S1 and S2; the server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client; the client is configured to view the equipment status and fault diagnosis results in real time.
[0024] This embodiment 1 further elaborates on the above steps.
[0025] (1) Data Cleaning Form Data preprocessing is performed on the original dataset to remove erroneous, duplicate, and noisy data. This includes three aspects: data standardization, handling duplicate data, and outlier detection. Data standardization: Convert all collected data to the same format to ensure consistency of data such as timestamps and sensor outputs. The formula is as follows, where, It is standardized data. It is the raw data. and These are the mean and standard deviation of the original data, respectively.
[0026] , Handling duplicate data: Compare the data in each row to see if it is a duplicate of the previous row. If it is a duplicate, delete the row.
[0027] Outlier detection: Outlier detection is performed using the standard deviation method, with a threshold set to 3 times the standard deviation. If any data point exceeds the threshold, it is deleted.
[0028] (2) Handling missing values Please refer to Figure 3 The mean imputation method is used to initially fill in the missing values in the data, forming a new data sample; the formula is as follows: , in, This is the result of filling in missing values. It is the total number of valid data for this feature. It is a valid data value.
[0029] Then, data normalization is performed to ensure that the data values of different features are within the same dimensional range [0,1], so as to prevent features of different scales from affecting the model's performance; the formula is as follows: , in, It is normalized data. This is a data sample with the mean filled in. and These are the minimum and maximum values of the feature, respectively.
[0030] Then, by using the sparse principal component analysis (SPCA) method to initialize the estimation of missing values, a preliminary estimate of the missing values is obtained, and the data dimensionality is reduced.
[0031] The objective function for the SPCA model is established as follows: , in, It is the data matrix after data normalization. It is a low-dimensional principal component matrix. It is a projection matrix. It is the sparse regularization coefficient. It is the Frobenius norm of the data matrix. It is the first The L1 norm of each principal component vector; The SPCA method was used to make a preliminary estimate of missing values, and the missing values were then... The corresponding estimated value is used as the initial interpolation value, as shown in the following formula: , in, This represents the final output, corresponding to the [number]th [item] in the data matrix. line (number) (sample number) Column (number) The estimated value of the location of each feature is used to fill in the missing parts of the original data.
[0032] Furthermore, the Gray-Level Correlation Coefficient K-Nearest Neighbor (GKNN) method is used for iterative optimization to obtain the final estimate of the relevant missing values. This includes three aspects: constructing the gray-level correlation coefficient, calculating the overall gray-level similarity, and using GKNN to imput missing values.
[0033] Construct a gray-level correlation coefficient matrix and calculate the gray-level similarity between data samples. Let the first... Each sample is a reference sequence. , No. Each sample is a comparison sequence. Then the first The formula for calculating the correlation coefficient of each dimension is as follows; , in, Indicates the first In each dimension and The gray-scale correlation coefficient; The resolution coefficient is 0.5 in the preferred embodiment; Next, calculate the overall gray-level similarity, which is to take the average of all dimensions to obtain the overall gray-level similarity between the two data samples. The formula is as follows: , in, This represents the final calculation result, i.e., the first... The first sample and the first The overall gray-level similarity value of each sample.
[0034] In a preferred embodiment, the GKNN imputation of missing values is specifically as follows: For each sample with missing values, select the sample with the highest gray-level similarity. For each complete sample, construct a neighborhood set. Then, weighted interpolation is performed; the interpolation formula is as follows: , in, It is the first The first sample Estimates of the missing features; It is the first The neighbor in the The true value of the dimension; It is a sample With sample Gray-scale similarity.
[0035] (3) Model training Model computation includes two steps: model training and real-time diagnosis. In model training, an initial diagnostic model is trained using the LSTM method based on multi-dimensional time-series historical data, and the model is optimized using the Random Forest method. This process comprises four steps: constructing the LSTM model, determining the training process, selecting the optimization method, and completing model optimization.
[0036] Construct an LSTM model; design a multi-input multi-output LSTM model, including an input layer, an LSTM layer, a Dropout layer, and an output layer.
[0037] The input layer uses a data matrix obtained after multivariate data fusion. As input data; The LSTM layer contains 128 units and is used to learn the time series data from the input layer. Each LSTM unit... The specific formula for updating time is as follows: , in, In time step Input; It is a time step The hidden state; It is a time step The hidden state; It is a time step cellular state; It is a candidate cell state; It is a time step cellular state; and These are the weight matrix and the bias vector, where the subscripts are... These represent the weight matrix and bias vector corresponding to the forget gate, input gate, output gate, and candidate cell state, respectively. yes Activation function , It is the tangent function of a hyperbola. ; It is the output of the forget gate; It is an input gate output; It is the input gate output.
[0038] The dropout ratio is set to 0.2 in the Dropout layer to reduce overfitting during network training.
[0039] The output layer uses the SoftMax activation function to output the fault type of the equipment. The representation model for the th Predicted values for each category, Indicates belonging to the first The probability value of a type of failure. This indicates the total number of fault categories.
[0040] , Training process: The LSTM model is trained using a historical dataset, and the weights are updated through backpropagation. A multi-class cross-entropy function is used as the loss function. To assess the error between the predicted and actual values, the following methods are used: , in, It is the total number of samples. It is the total number of fault categories. It is the first Does each sample belong to a category? The true situation The model predicts the first Each sample belongs to category The probability of.
[0041] Optimization method: The Adam optimizer is used to minimize the loss function. , , , , , , in, Indicates the current iteration number. Is Gradient at time, Representation learning model, The gradient is the exponential moving average. It is the exponential moving average of the gradient of bungalows. It's the learning rate. and This refers to the attenuation rate, which, in a preferred embodiment, is set to 0.9 and 0.999, respectively. for .
[0042] Model optimization; After the LSTM model is trained, a random forest method is used for optimization. This method improves diagnostic accuracy through a voting process involving multiple decision trees, as shown in the following formula: , in, It is the first The prediction function for each tree. It's the weight.
[0043] In real-time diagnostics, real-time diagnosis is performed based on newly acquired data and trained models. The system determines the operating status and fault type of the equipment based on the matching of real-time data with historical data.
[0044] Real-time data input; Data from sensors of complex equipment systems is collected in real time via the acquisition terminal, and the data is processed in the same way as the training data (including data standardization, handling duplicate data, and outlier detection) before being transmitted to the server.
[0045] Model inference and fault diagnosis: A trained LSTM+random forest model is used to predict real-time data and output the system's fault probability. The formula is as follows, where... The input data at the current moment, It is the fault diagnosis probability output by the model.
[0046] , Alarm triggered; if fault diagnosis results If a fault or potential risk is detected, the fault diagnosis system will trigger an alarm and transmit the alarm to the client.
[0047] (4) Model deployment service In a preferred embodiment, please refer to Figure 2 The above methods were deployed and applied; the equipment fault diagnosis method based on missing data was implemented by combining the three parts of the data acquisition end, server end and client end. The acquisition unit consists of three modules: a data acquisition module, a communication connection module, and a data processing module.
[0048] The data acquisition module collects various operational data of the equipment through multiple information acquisition devices (i.e., sensors), including time-series data such as angular error, force accuracy, current, voltage, support force, tilt angle, temperature, and speed. These data are characterized by strong time sequence and multi-dimensionality, and are often missing or damaged in practical applications.
[0049] The communication connection module is responsible for enabling data communication between the data acquisition module, the data processing module, and the server. The communication method is Ethernet communication to ensure stable communication and real-time data transmission.
[0050] The data processing module is responsible for preprocessing and imputing missing values in the collected time-series data. Specifically, for imputation, a combination of sparse principal component analysis (SPCA) and gray-level correlation coefficient K-nearest neighbor (GKNN) is used to perform local similarity interpolation on the missing data.
[0051] The server-side primarily stores and analyzes the data collected from the acquisition end, and provides data services for client applications to access and call. It consists of a database, model computation, and application services. The database stores the collected data according to a defined format and provides interfaces for model computation and application services. Model computation uses historical data from the database to train deep learning methods, deriving a deep learning model suitable for the current dataset, and then uses new real-time data and the deep learning model to infer equipment fault diagnosis results. The application services further modularize and encapsulate the data from the database and the results output by the model, providing data access services for various client business operations.
[0052] The client receives fault diagnosis results and real-time equipment operating status from the server, providing a user interface for real-time viewing of equipment status and fault diagnosis results. After logging into the system interface, the system displays real-time operating data, fault diagnosis results, and maintenance suggestions through data, reports, and icons. When the diagnostic results determine that the equipment has a fault or potential risk, the system automatically issues an alarm and suggests maintenance or inspection.
[0053] Example 2 Please refer to Figure 4 This embodiment 2 provides a method for diagnosing equipment faults based on missing data. The data cleaning unit is used to clean the raw data. The missing value processing unit is used to initially fill in the missing values of the data using the mean imputation method, forming new data samples and completing data normalization; then, it uses the sparse principal component analysis method to perform initial estimation of the missing values, obtaining preliminary estimates of the missing values; finally, it uses the gray-level correlation coefficient K nearest neighbor method to iteratively optimize and obtain the final estimates of the relevant missing values. The model training unit is used to select the LSTM method and train a preliminary diagnostic model based on multidimensional time series historical data, and optimize the model using the random forest method; the LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification; The model deployment service unit is used to deploy the model trained on the server side; it collects data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmits the data to the server after completing the data preprocessing through the data processing methods in S1 and S2; the server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client; the client is configured to view the equipment status and fault diagnosis results in real time.
[0054] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement any step of an equipment fault diagnosis method based on missing data.
[0055] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0056] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.
[0057] Those skilled in the art will readily understand that the above description is merely 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 within the scope of protection of the present invention.
Claims
1. A method for diagnosing equipment faults based on missing data, characterized in that, include: S1. Complete the cleaning of the raw data; S2. Use the mean imputation method to initially fill in the missing values of the data, form a new data sample, and complete the data normalization; Then, the missing values are initially estimated by sparse principal component analysis to obtain preliminary estimates of the missing values; finally, the K-nearest neighbor method of gray-level correlation coefficient is used for iterative optimization to obtain the final estimates of the relevant missing values. S3. The LSTM method was selected and a preliminary diagnostic model was trained based on multidimensional time series historical data. The model was then optimized using the random forest method. The LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification. S4. Deploy the trained model on the server side; collect data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmit the data to the server after completing the data preprocessing through the data processing methods in S1 and S2; the server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client; the client is configured to view the equipment status and fault diagnosis results in real time.
2. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, The sparse principal component analysis method in S2 is specifically as follows: The objective function for the SPCA model is established as follows: ， in, It is the data matrix after data normalization. It is a low-dimensional principal component matrix. It is a projection matrix. It is the sparse regularization coefficient. It is the Frobenius norm of the data matrix. It is the first The L1 norm of each principal component vector; The SPCA method was used to make a preliminary estimate of missing values, and the missing values were then... The corresponding estimated value is used as the initial interpolation value, as shown in the following formula: ， in, This represents the final output, corresponding to the [number]th [item] in the data matrix. line, number The estimated column position is used to fill in missing parts of the original data.
3. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, The K-nearest neighbor method for gray-level correlation coefficient in S2 includes: constructing gray-level correlation coefficients to calculate overall gray-level similarity; Construct a gray-level correlation coefficient matrix and calculate the gray-level similarity between data samples. Let the first... Each sample is a reference sequence. , No. Each sample is a comparison sequence. Then the first The formula for calculating the correlation coefficient of each dimension is as follows; ， in, Indicates the first In each dimension and The gray-scale correlation coefficient; Represents the resolution coefficient; Next, calculate the overall gray-level similarity, which is to take the average of all dimensions to obtain the overall gray-level similarity between the two data samples. The formula is as follows: ， in, This represents the final calculation result, i.e., the first... The first sample and the first The overall gray-level similarity value of each sample.
4. The equipment fault diagnosis method based on missing data according to claim 3, characterized in that, The K-nearest neighbor method for gray-level correlation coefficients in S2 also includes GKNN imputation of missing values, specifically: For each sample with missing values, select the sample with the highest gray-level similarity. For each complete sample, construct a neighborhood set. Then, weighted interpolation is performed; the interpolation formula is as follows: ， in, It is the first The first sample Estimates of the missing features; It is the first The neighbor in the The true value of the dimension; It is a sample With sample Gray-scale similarity.
5. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, The LSTM model in S2 includes a data matrix obtained by multivariate data fusion. The input layer serves as the input data, the LSTM layer is configured to learn the time series data in the input layer, the Dropout layer is configured to reduce overfitting during network training, and the output layer is configured to output the fault types of the equipment. The Dropout layer is set to a dropout ratio of no less than 0.2; the output layer uses the SoftMax activation function.
6. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, In the LSTM model of S2, each LSTM unit is... The specific formula for updating time is as follows: ， in, In time step Input; It is a time step The hidden state; It is a time step The hidden state; It is a time step cellular state; It is the candidate cell state; It is a time step cellular state; and These are the weight matrix and the bias vector, where the subscripts are... These represent the weight matrix and bias vector corresponding to the forget gate, input gate, output gate, and candidate cell state, respectively. yes Activation function , It is the tangent function of a hyperbola. ; It is the output of the forget gate; It is an input gate output; It is the input gate output.
7. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, The specific process of optimizing the model using the random forest method in S2 is as follows: After the LSTM model is trained, a random forest method is used for optimization. This method improves diagnostic accuracy through a voting process involving multiple decision trees, as shown in the following formula: ， in, It is the first The prediction function for each tree. It's the weight.
8. The equipment fault diagnosis method based on missing data according to claim 1, characterized in that, The training process of the LSTM model in S2 is as follows: The LSTM model is trained using a historical dataset, and the weights are updated via backpropagation. A multi-class cross-entropy function is used as the loss function. To assess the error between the predicted and actual values, the following methods are used: ， in, It is the total number of samples. It is the total number of fault categories. It is the first Does each sample belong to a category? The true situation The model predicts the first Each sample belongs to category The probability of.
9. An equipment fault diagnosis system based on missing data, characterized in that, include The data cleaning unit is used to clean the raw data. The missing value processing unit is used to initially fill in the missing values of the data using the mean imputation method, forming new data samples and completing data normalization. Then, the missing values are initially estimated by sparse principal component analysis to obtain preliminary estimates of the missing values; finally, the K-nearest neighbor method of gray-level correlation coefficient is used for iterative optimization to obtain the final estimates of the relevant missing values. The model training unit is used to select the LSTM method and train a preliminary diagnostic model based on multidimensional time series historical data, and optimize the model using the random forest method; the LSTM method extracts dynamic time series features from the running data and uses its output as a feature vector to input into the random forest for fault classification; The model deployment service unit is used to deploy the model trained on the server. It collects data from the sensors of the complex equipment system in real time through the acquisition terminal, and transmits the data to the server after preprocessing through the data cleaning unit and the missing value handling unit. The server then performs model inference on the data, outputs the failure probability of the system, and summarizes the results and transmits them to the client. The client is configured to view the equipment status and fault diagnosis results in real time.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8: a method for diagnosing equipment faults based on missing data.