Complex equipment fault diagnosis system and diagnosis method based on multivariate data fusion

Abstract

The present invention provides a complex equipment fault diagnosis system and method based on multivariate data fusion, and relates to the technical field of equipment fault diagnosis. The diagnostic method periodically collects the operation data of complex equipment by installing multiple sensors to ensure the comprehensiveness and integrity of the data; constructs time series samples by using data preprocessing methods to ensure the high quality of the data; uses convolutional neural networks to extract high-dimensional features, and uses principal component analysis to reduce data dimensionality; and uses a specific LSTM network to train time series samples offline, while retaining the original data information to the greatest extent, reducing the complexity of the model, enhancing the learning efficiency of time dynamic characteristics, and improving the accuracy of multi-classification fault diagnosis; finally, the model is deployed on the edge computing device, and the upper computer software receives the sensor data in real time through Ethernet and loads the model for real-time reasoning, so as to realize real-time and accurate fault diagnosis and remote monitoring of complex equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault diagnosis, and in particular to a complex equipment fault diagnosis system and method based on multivariate data fusion. Background Art

[0002] Complex equipment has complex systems, numerous devices, and strong coupling. Its data has the characteristics of huge data volume, changes over time, strong correlation between sensor data, and many types of faults, which leads to many technical problems such as sudden faults, unclear mechanisms of fault occurrence and propagation, and difficulty in troubleshooting. The maintenance measures for complex equipment generally adopt the strategy of periodic inspection or post-maintenance. Among them, periodic fault maintenance will greatly waste manpower and financial resources, and it is easy to have excessive maintenance. The newly replaced parts may not match the original system; although post-maintenance reduces costs, it affects the use of product functions. Such maintenance methods can no longer meet the use needs of modern equipment.

[0003] The health management technology that has emerged in recent years is based on a thorough study of the equipment failure mechanism and failure laws, and can analyze the failure location and failure time in advance. It has the advantages of small logistics scale, high degree of automation, and good economic affordability.

[0004] Among the existing health management technologies, patent CN202110512033.4 discloses a "complex equipment health management method and system based on digital twins". This technical solution simulates the digital twin of the complex equipment entity to be tested to obtain maintenance task planning information and future health status prediction information for the complex equipment entity to be tested. This method can realize information interaction between the digital twin and complex equipment to a certain extent, making it convenient to determine the healthy operating status of complex equipment in real time. However, domestic digital twin technology is still in its infancy, and its modeling process is complex and difficult. There are inevitably differences between the digital twin model and the actual product, such as external environment, operating conditions, interface relationships, etc., which makes the simulation results prone to errors and difficult to implement in actual engineering.

[0005] Patent CN202010488395.X discloses "a complex equipment fault diagnosis method based on multi-sensor fusion". This technology uses multi-sensor data stacking and convolutional neural network methods to perform data fusion and multi-classification fault diagnosis. This method can improve the accuracy of complex equipment fault diagnosis, but it does not perform feature dimensionality reduction operations on big data, resulting in high model complexity and low diagnostic efficiency. In addition, the convolutional neural network has poor processing effect on time series data and cannot capture the long-term time dependency in the data, which is not suitable for this research object.

[0006] Patent CN202211649995.5 discloses "a complex equipment fault diagnosis system based on deep learning". This technology uses a variational autoencoder and a jump-connected bidirectional gate recurrent network to process time series data, reduces the dimension of high-order feature data, and enhances the ability to extract timing information. However, the autoencoder requires a lot of experimental debugging to determine the size and number of hidden layers, and usually does not consider the time correlation of the time series, which is not suitable for this research object.

[0007] Patent 202310334054.0 discloses "a coal mill fault warning method based on PCA and LSTM fusion algorithm", which uses principal component analysis to effectively reduce the dimension of many variables that affect the safe operation of the coal mill, converting the original high-dimensional related variables into low-dimensional unrelated variables, reducing the time and space complexity of the subsequent deep learning algorithm; by constructing a fault warning model based on long short-term memory neural network, deep feature extraction is performed on the relationship before and after the operation data, and the coal mill fault-related parameters are accurately and effectively predicted based on the associated variables, realizing the monitoring of the real-time operation status of the coal mill and fault warning. However, PCA cannot effectively capture the nonlinear characteristics in the data, especially when the data distribution is complex and contains highly nonlinear relationships. It lacks the ability to extract spatial features, which is not conducive to distinguishing different fault modes and may affect the accuracy of fault prediction.

[0008] In view of this, it is necessary to design an improved complex equipment fault diagnosis system and diagnosis method based on multivariate data fusion to solve the above problems. Summary of the invention

[0009] The purpose of the present invention is to provide a complex equipment fault diagnosis system and a diagnosis method based on multivariate data fusion.

[0010] To achieve the above-mentioned purpose of the invention, the present invention provides a complex equipment fault diagnosis system based on multivariate data fusion, which includes a data acquisition and transmission module, a data preprocessing module, a model training module, a model deployment and reasoning module, and a remote monitoring module;

[0011] The data acquisition and transmission module is used to comprehensively collect multi-sensor data of complex equipment;

[0012] The data preprocessing module is electrically connected to the data acquisition and transmission module, and is used to preprocess the multivariate sensor data and construct a data set, which is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to monitor the model performance during the training process;

[0013] The model training module is electrically connected to the data preprocessing module, uses a convolutional neural network to extract high-dimensional features from the preprocessed multivariate sensor data, uses principal component analysis to reduce the data dimension, uses an LSTM network to offline train the reduced-dimensional time series samples, and finally uses an Adam learning rate adaptive algorithm to optimize the model.

[0014] As a further improvement of the present invention, the model deployment and reasoning module is used to deploy the model on the edge computing device, receive sensor data in real time through Ethernet and load the model for real-time reasoning, so as to realize real-time and accurate multi-classification online fault diagnosis of complex equipment;

[0015] The remote monitoring module is electrically connected to the model deployment and reasoning module. After receiving the online fault diagnosis result in real time through the network, the current fault is highlighted in real time on the human-computer interaction interface and the maintenance personnel are reminded through a buzzer.

[0016] To achieve the above-mentioned object of the invention, the present invention also provides a complex equipment fault diagnosis method based on multivariate data fusion, which performs diagnosis based on the above-mentioned complex equipment fault diagnosis system based on multivariate data fusion, and comprises the following steps:

[0017] S1, collection of multi-sensor data: collection of multi-sensor data of key parts of complex equipment;

[0018] S2, data preprocessing: denoising, normalizing and missing value processing of multivariate sensor data, then constructing time series samples, and finally labeling and dividing the data set to obtain preprocessed multivariate data;

[0019] S3, model construction and training: Convolutional neural network is used to extract features from preprocessed multivariate data, and then principal component analysis is used to reduce data dimension to reduce computational complexity and overfitting risk. LSTM network is used to train the reduced-dimensional time series data offline, and the model is trained and optimized.

[0020] S4, model deployment and reasoning: export the trained model into a suitable format and deploy it in the edge computing device, load the model for real-time reasoning, and realize multi-classification online fault diagnosis of complex equipment;

[0021] S5, remote monitoring: After receiving the online fault diagnosis results in real time through the network, the current fault is highlighted in real time on the human-computer interaction interface, and the maintenance personnel are reminded through the buzzer to realize remote monitoring.

[0022] As a further improvement of the present invention, the specific process of data dimensionality reduction processing in step S3 is:

[0023] S32-1, standardization processing: standardize the feature data processed by the convolutional neural network , so that the mean of each feature is 0 and the variance is 1; where, is the original eigenvalue, is the standardized eigenvalue, It is a feature The mean of It is a feature The standard deviation of

[0024] S32-2, calculate the covariance matrix: the standardized data matrix is , then the covariance matrix The calculation formula is , where n is the sample size;

[0025] S32-3, Calculating Eigenvalues ​​and Eigenvectors: Solving the Covariance Matrix The eigenvalues ​​and eigenvectors of are given by ,in, is the eigenvalue to be solved, is the feature vector, is the covariance matrix;

[0026] S32-4, select the main components: first calculate the The contribution rate of the eigenvalue Then sort the eigenvalue contribution rates from large to small, and select the top ones with a cumulative contribution rate of 85%. features as principal components, forming a eigenvector matrix ;

[0027] S32-5, projection data: the standardized data matrix Projection onto the selected eigenvector matrix On the above, we get the data matrix after dimensionality reduction .

[0028] As a further improvement of the present invention, the construction method of the LSTM network in step S3 is: designing a multi-input and multi-output LSTM model, including an input layer, an LSTM layer, a Dropout layer, a fully connected layer, and an output layer; the specific process is:

[0029] S33-1, input layer: using the data matrix after multivariate data fusion As input data;

[0030] S33-2, LSTM layer: contains 128 units, used to learn the time series data in the input layer; each LSTM unit The update formula of the moment is as follows, where: is at the time step Input, is the time step The hidden state of is the time step The cell state, is the current cell state, and are the weight matrix and bias vector respectively, yes sigmoid Activation Function , tanh is the hyperbolic tangent function ;

[0031] ;

[0032] S33-3, Dropout layer: Set the dropout ratio to reduce overfitting during network training;

[0033] S33-4, fully connected layer: maps the output of LSTM to the target output space, with the same number of nodes as the number of output categories;

[0034] S33-5, output layer: uses SoftMax activation function to output the fault type of the equipment; Represents the model for The predicted value of each category, Indicates that it belongs to The probability value of class failure, Represents the total number of fault categories, and the formula is as follows:

[0035] .

[0036] As a further improvement of the present invention, the specific process of model training and optimization in step S3 is:

[0037] S34-1, Calculate the loss function: Use the multi-classification cross entropy function as the loss function ,in, is the total number of samples, is the total number of fault categories, It is Whether the sample belongs to the category The real situation, The model predicts samples belong to the category The probability of is as follows:

[0038] ;

[0039] S34-2, model optimization: The Adam learning rate adaptive algorithm is used to optimize the model; Indicates the current iteration number, is The gradient of time, Represents the learning model, is the gradient is the exponential moving average, is the exponential moving average of the bungalow gradient, is the learning rate, and is the decay rate, set to 0.9 and 0.999, for , the formula is as follows:

[0040] ;

[0041] S34-3, Model Evaluation: Accuracy Accuracy As an evaluation metric, it is used to measure the accuracy of classification;

[0042] S34-4, model training and saving: Execute training, dynamically adjust the training process according to the performance of the model on the validation set, stop training and save the model when the performance of the validation set does not improve within the set period.

[0043] As a further improvement of the present invention, the specific process of model deployment and reasoning in step S4 is: exporting the trained model into a suitable format and deploying it in the edge computing device, and then encapsulating the process of loading the model file and real-time data on the edge computing device into an API service, so that it can receive requests and return diagnostic results;

[0044] During real-time diagnosis, the host computer software receives periodic data from various sensors via Ethernet and performs preprocessing. It then infers the real-time data by calling the API interface, outputs the fault diagnosis results, and realizes online fault diagnosis of complex equipment.

[0045] As a further improvement of the present invention, the specific process of data preprocessing in step S2 is:

[0046] S2-1, denoising: using moving average filtering method, ,in is the filtered data, is the original data, N is the window size;

[0047] S2-2, Normalization: Using the standard deviation method, ,in, is the standardized data, is the original data, is the mean of the original data, is the standard deviation of the original data;

[0048] S2-3, missing value processing: linear interpolation method is used, ,in, is the interpolated data, and is the nearest known data point before and after the missing point, and is the corresponding time point;

[0049] S2-4, construct time series samples: divide the two-dimensional matrix into multiple time series samples according to the time window and step size, where the time window is 1 hour and the step size is 1;

[0050] S2-5, data labeling: Combined with expert experience, label each time series sample with a corresponding fault label y, where normal operation is 0, fault type is 1, and fault types are 2~n;

[0051] S2-6, divide the data set: randomly construct a data set with the processed time series samples and the corresponding labels, and divide it into a training set and a validation set. The training set is used to train the model, and the validation set is used to monitor the model performance during the training process.

[0052] As a further improvement of the present invention, the specific process of using the convolutional neural network to perform feature extraction processing in step S3 is:

[0053] S31, build a convolutional neural network: CNN consists of an input layer, three convolutional layers and three pooling layers;

[0054] Among them, the number of convolution kernels in the first convolution layer is 32, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer is 128. The convolution kernel sizes are all 3, the zero-filling step size is all 1, the activation function uses the ReLU function, and the pooling layer uses the maximum pooling method to effectively extract the high-dimensional features of the time series data set.

[0055] As a further improvement of the present invention, the multi-sensor data includes but is not limited to one or more combinations of flow sensor data, pressure sensor data, voltage sensor data, current sensor data, temperature, humidity and air pressure sensors.

[0056] The beneficial effects of the present invention are:

[0057] The complex equipment fault diagnosis system and method based on multivariate data fusion provided by the present invention have the following technical advantages:

[0058] 1. Multi-sensor data fusion: By installing multiple sensors to collect complex equipment operation data in real time and comprehensively utilizing multi-source (multivariate) data, the fault diagnosis system is able to acquire status information of complex equipment in different dimensions, ensuring the accuracy, completeness and comprehensiveness of input data and improving the reliability of model training and reasoning.

[0059] 2. Efficient feature extraction and dimensionality reduction: Using convolutional neural networks (CNN) for high-dimensional feature extraction can automatically learn complex features in the data. Combining principal component analysis (PCA) for data dimensionality reduction can retain the main features, reduce the dimensions of the data, and improve the computational efficiency and generalization ability of the model. Using long short-term memory networks (LSTM) to process the time series of the reduced features can capture the time series characteristics of the data. The combination of CNN+PCA+LSTM can significantly improve the accuracy and robustness of fault diagnosis.

[0060] 3. Real-time reasoning and remote monitoring: After the host computer software receives multi-sensor data through Ethernet, the API service of the edge computing device loads the model and online data, and returns the fault diagnosis results in real time, realizing online fault alarm and remote monitoring, and notifying maintenance personnel to carry out maintenance in time, improving the timeliness and effectiveness of equipment maintenance. Edge computing devices are used for online diagnosis and monitoring, and edge computing devices are characterized by light and small size, strong computing power, and high real-time performance.

[0061] 4. Modular design and scalability: The system architecture is clear and has good stability and scalability through modular design. It can adapt to different application scenarios and needs and facilitate subsequent system upgrades and functional expansions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a system architecture diagram of the complex equipment fault diagnosis system based on multivariate data fusion provided by the present invention.

[0063] Figure 2 A schematic flow chart of the complex equipment fault diagnosis method based on multivariate data fusion provided by the present invention.

[0064] Figure 3 Schematic diagram of LSTM internal units in the complex equipment fault diagnosis system based on multivariate data fusion provided by the present invention. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the scheme of the present invention are shown in the drawings, while other details that are not closely related to the present invention are omitted.

[0067] In addition, it should be noted that the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or apparatus.

[0068] See also Figure 1 As shown, the present invention provides a complex equipment fault diagnosis system based on multivariate data fusion, including a data acquisition and transmission module, a data preprocessing module, a model training module, a model deployment and reasoning module, and a remote monitoring module.

[0069] The data acquisition and transmission module: installs multiple sensors (such as flow sensors, temperature sensors, pressure sensors, etc.) at key positions of complex equipment, periodically collects operating data of the complex equipment, and transmits the sensor data to the host computer software via Ethernet.

[0070] The data preprocessing module performs data preprocessing on the sensor data.

[0071] The model training module: the host computer software inputs the preprocessed sensor data into the CNN model, extracts the feature vector, and then uses PCA to reduce the dimension of the feature vector to reduce the data dimension. Finally, the reduced feature is input into the LSTM network for time series processing to complete model training and optimization.

[0072] The model deployment and reasoning module: The edge computing device deploys the trained model, receives real-time sensor data from the host computer software, performs reasoning in real time, outputs fault diagnosis results, and realizes online diagnosis.

[0073] The remote monitoring module: After the host computer software receives the diagnosis results of the model deployment and reasoning module in real time through the network, it highlights the current fault in real time on the human-computer interaction interface and reminds the maintenance personnel through the buzzer.

[0074] Example 1

[0075] See also Figure 2 As shown, Embodiment 1 of the present invention provides a complex equipment fault diagnosis method based on multivariate data fusion, which uses the complex equipment fault diagnosis system based on multivariate data fusion for diagnosis, and specifically includes the following steps:

[0076] S1, data acquisition and transmission module: install various sensors such as flow, pressure, voltage, current, temperature and humidity at key parts of complex equipment, collect signals during the operation of complex equipment, complete A / D conversion and filtering and denoising of the signal, report the numerical data to the host computer at a cycle of 1s, and store it in the database as data to be trained. At this time, each sensor data is a one-dimensional vector, and the multi-sensor data constitutes a two-dimensional matrix.

[0077] S2, data preprocessing module: denoise, normalize and process missing values ​​for each one-dimensional vector to eliminate the influence of noise and outliers, then construct time series samples, and finally label the data and divide the data set.

[0078] Denoising: Using moving average filtering method, ,in is the filtered data, is the original data, and N is the window size.

[0079] Normalization: Using the standard deviation method, ,in, is the standardized data, is the original data, is the mean of the original data, is the standard deviation of the original data.

[0080] Missing value processing: linear interpolation is used. ,in, is the interpolated data, and is the nearest known data point before and after the missing point, and is the corresponding time point.

[0081] Construct time series samples: Divide the two-dimensional matrix into multiple time series samples according to the time window and step size, where the time window is 1 hour and the step size is 1.

[0082] Data labeling: Based on expert experience, each time series sample is labeled with a corresponding fault label y (normal operation 0, fault type 1, fault type 2, ..., fault type n).

[0083] Divide the data set: Randomly construct a data set with the processed time series samples and the corresponding labels, and divide it into a training set (70%) and a validation set (30%). The training set is used to train the model, and the validation set is used to monitor the model performance during the training process.

[0084] S3, model training module: Convolutional neural network (CNN) is used to extract features from preprocessed multivariate data, and then principal component analysis (PCA) is used to reduce data dimension to reduce computational complexity and overfitting risk. LSTM network is used to process the reduced-dimensional time series data offline, and model training optimization is performed.

[0085] S31, Feature Extraction: CNN consists of an input layer, three convolutional layers, and three pooling layers. The number of convolutional kernels in the first convolutional layer is 32, the number of convolutional kernels in the second convolutional layer is 64, and the number of convolutional kernels in the third convolutional layer is 128. The convolutional kernel size is 3, the zero padding step is 1, the activation function uses the ReLU function, and the pooling layer uses the maximum pooling method. After CNN processing, the high-dimensional features of the time series data set can be effectively extracted.

[0086] S32, data dimensionality reduction: consists of five steps: standardization, calculation of covariance matrix, calculation of eigenvalues ​​and eigenvectors, selection of main components, and projection of data. After PCA processing, the extracted high-dimensional features can be reduced in dimensionality and the main components can be retained.

[0087] S32-1, standardization processing: standardize the feature data processed by CNN , so that the mean of each feature is 0 and the variance is 1. Among them, is the original eigenvalue, is the standardized eigenvalue, It is a feature The mean of It is a feature The standard deviation of .

[0088] S32-2, calculate the covariance matrix: the standardized data matrix is , then the covariance matrix The calculation formula is , where n is the number of samples.

[0089] S32-3, calculate eigenvalues ​​and eigenvectors: that is, solve the covariance matrix The eigenvalues ​​and eigenvectors of , and the characteristic equation are ,in, is the eigenvalue to be solved, is the feature vector, is the covariance matrix.

[0090] S32-4, select the main components: first calculate the The contribution rate of the eigenvalue Then sort the eigenvalue contribution rates from large to small, and select the top ones with a cumulative contribution rate of 85%. features as principal components, forming a eigenvector matrix .

[0091] S32-5, projection data: the standardized data matrix Projection onto the selected eigenvector matrix On the above, we get the data matrix after dimensionality reduction .

[0092] S33, build LSTM network: see Figure 3 As shown in the figure, a multi-input and multi-output LSTM model is designed, including input layer, LSTM layer, Dropout layer, fully connected layer, and output layer.

[0093] S33-1, input layer: using the data matrix after multivariate data fusion as input data.

[0094] S33-2, LSTM layer: contains 128 units, which are used to learn the time series data in the input layer. Each LSTM unit The update formula of the moment is as follows, where: is at the time step Input, is the time step The hidden state of is the time step The cell state, is the current cell state, and are the weight matrix and bias vector respectively, yes sigmoid Activation Function , tanh is the hyperbolic tangent function ;

[0095] .

[0096] S33-3, Dropout layer: Set the dropout ratio to 0.2 to reduce overfitting during network training.

[0097] S33-4, fully connected layer: maps the output of LSTM to the target output space, with the same number of nodes as the number of output categories.

[0098] S33-5, output layer: uses SoftMax activation function to output the fault type of the equipment. Represents the model for The predicted value of each category, Indicates that it belongs to The probability value of class failure, Indicates the total number of fault categories.

[0099] .

[0100] S34, model training and optimization: After the model structure is determined, the model training is carried out next.

[0101] S34-1, Calculate the loss function: Use the multi-classification cross entropy function as the loss function ,in, is the total number of samples, is the total number of fault categories, It is Whether the sample belongs to the category The real situation, The model predicts samples belong to the category probability.

[0102] .

[0103] S34-2, model optimization: Adopt Adam learning rate adaptive algorithm to optimize the model. Indicates the current iteration number, is The gradient of time, Represents the learning model, is the gradient is the exponential moving average, is the exponential moving average of the bungalow gradient, is the learning rate, and is the decay rate, set to 0.9 and 0.999, for ;

[0104] .

[0105] S34-3, Model Evaluation: Accuracy Accuracy As an evaluation metric, it is used to measure the accuracy of classification.

[0106] S34-4, model training and saving: Execute training and dynamically adjust the training process according to the performance of the model on the validation set. That is, when the performance of the validation set does not improve within the set period, stop training and save the model.

[0107] S4, model deployment and reasoning module: export the trained model into a suitable format and deploy it in the edge computing device, and then encapsulate the process of loading the model file and real-time data on the edge computing device into an API service, so that it can receive requests and return diagnostic results. During real-time diagnosis, the host computer software receives periodic data from each sensor through Ethernet and pre-processes it, and then infers the real-time data by calling the API interface, outputs the fault diagnosis results, and realizes online fault diagnosis of complex equipment.

[0108] S5, remote monitoring module: After the host computer software receives the diagnosis results of the model deployment and reasoning module in real time through the network, it highlights the current fault in real time on the human-computer interaction interface and reminds the maintenance personnel through the buzzer.

[0109] In summary, the present invention provides a complex equipment fault diagnosis system and method based on multivariate data fusion, which is mainly used for multi-classification fault diagnosis of complex equipment, and relates to the field of equipment fault diagnosis. The diagnostic method periodically collects the operation data of complex equipment by installing multiple sensors to ensure the comprehensiveness and integrity of the data; constructs time series samples using data preprocessing methods to ensure the high quality of the data; uses convolutional neural networks for high-dimensional feature extraction, uses principal component analysis for data dimensionality reduction, and uses a specific LSTM network to train time series samples offline, while retaining the original data information to the greatest extent, reducing the complexity of the model, enhancing the learning efficiency of time dynamic characteristics, and improving the accuracy of multi-classification fault diagnosis; finally, the model is deployed on the edge computing device, and the host computer software receives the sensor data in real time through Ethernet and loads the model for real-time reasoning, so as to realize real-time and accurate fault diagnosis and remote monitoring of complex equipment.

[0110] The above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A complex equipment fault diagnosis system based on multivariate data fusion, characterized by: It includes data collection and transmission module, data preprocessing module, model training module, model deployment and reasoning module, and remote monitoring module; The data acquisition and transmission module installs flow sensors, pressure sensors, voltage sensors, current sensors, and temperature and humidity sensors at key locations of complex equipment to comprehensively collect multi-sensor data of complex equipment; each sensor data is a one-dimensional vector, and multi-sensor data constitutes a two-dimensional matrix; The data preprocessing module is electrically connected to the data acquisition and transmission module, and is used to preprocess the multivariate sensor data and construct a data set, which is divided into a training set and a validation set. The training set is used to train the model, and the validation set is used to monitor the model performance during the training process; The model training module is electrically connected to the data preprocessing module, and adopts a CNN+PCA+LSTM combined method to perform efficient feature extraction and dimensionality reduction processing, adopts a convolutional neural network CNN to extract high-dimensional features from the preprocessed multivariate sensor data, uses principal component analysis PCA to reduce the data dimension, adopts a long short-term memory network LSTM network to offline train the time series samples after dimensionality reduction, and finally adopts an Adam learning rate adaptive algorithm to optimize the model; The model deployment and reasoning module is used to deploy the model on the edge computing device, receive sensor data in real time through Ethernet and load the model for real-time reasoning, so as to realize real-time and accurate multi-classification online fault diagnosis of complex equipment; The remote monitoring module is electrically connected to the model deployment and reasoning module, and after receiving the online fault diagnosis result in real time through the network, the current fault is highlighted in real time on the human-computer interaction interface, and the maintenance personnel are reminded through a buzzer; The specific process of feature extraction using convolutional neural network is as follows: S31, build a convolutional neural network: CNN consists of an input layer, three convolutional layers and three pooling layers; Among them, the number of convolution kernels in the first convolution layer is 32, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer is 128. The convolution kernel size is 3, the zero padding step is 1, the activation function uses the ReLU function, and the pooling layer uses the maximum pooling method to effectively extract the high-dimensional features of the time series data set; The specific process of using principal component analysis PCA to reduce data dimension is as follows: S32-1, standardization processing: standardize the feature data processed by the convolutional neural network , so that the mean of each feature is 0 and the variance is 1; among them, is the original eigenvalue, is the standardized eigenvalue, It is a feature The mean of It is a feature The standard deviation of S32-2, calculate the covariance matrix: the standardized data matrix is , then the covariance matrix The calculation formula is , where n is the sample size; S32-3, Calculating Eigenvalues ​​and Eigenvectors: Solving the Covariance Matrix The eigenvalues ​​and eigenvectors of are given by ,in, is the eigenvalue to be solved, is the feature vector, is the covariance matrix; S32-4, select the main components: first calculate the The contribution rate of the eigenvalue Then sort the eigenvalue contribution rates from large to small, and select the top ones with a cumulative contribution rate of 85%. features as principal components, forming a eigenvector matrix ; S32-5, projection data: the standardized data matrix Projection onto the selected eigenvector matrix On the above, we get the data matrix after dimensionality reduction ; The construction method of the long short-term memory network LSTM network is to design a multi-input and multi-output LSTM model, including input layer, LSTM layer, Dropout layer, fully connected layer, and output layer; the specific process is: S33-1, input layer: using the data matrix after multivariate data fusion As input data; S33-2, LSTM layer: contains 128 units, used to learn the time series data in the input layer; each LSTM unit The update formula of the moment is as follows, where: is at the time step Input, is the time step The hidden state of is the time step The cell state, is the current cell state, and are the weight matrix and bias vector respectively, yes sigmoid Activation Function , tanh is the hyperbolic tangent function ; ； S33-3, Dropout layer: Set the dropout ratio to reduce overfitting during network training; S33-4, fully connected layer: maps the output of LSTM to the target output space, with the same number of nodes as the number of output categories; S33-5, output layer: uses SoftMax activation function to output the fault type of the equipment; Represents the model for The predicted value of each category, Indicates that it belongs to The probability value of class failure, Represents the total number of fault categories, and the formula is as follows: 。 2. A complex equipment fault diagnosis method based on multivariate data fusion, characterized by: The complex equipment fault diagnosis system based on multivariate data fusion according to claim 1 is used for diagnosis, comprising the following steps: S1, Collection of multi-sensor data: Collect multi-sensor data of key parts of complex equipment; S2, data preprocessing: denoising, normalizing and missing value processing of multivariate sensor data, then constructing time series samples, and finally labeling and dividing the data set to obtain preprocessed multivariate data; S3, model construction and training: Convolutional neural network is used to extract features from preprocessed multivariate data, and then principal component analysis is used to reduce data dimension to reduce computational complexity and overfitting risk. LSTM network is used to train the reduced-dimensional time series data offline, and the model is trained and optimized. S4, model deployment and reasoning: export the trained model into a suitable format and deploy it in the edge computing device, load the model for real-time reasoning, and realize multi-classification online fault diagnosis of complex equipment; S5, remote monitoring: After receiving the online fault diagnosis results in real time through the network, the current fault is highlighted in real time on the human-computer interaction interface, and the maintenance personnel are reminded through the buzzer to realize remote monitoring; The specific process of feature extraction using convolutional neural network is as follows: S31, build a convolutional neural network: CNN consists of an input layer, three convolutional layers and three pooling layers; Among them, the number of convolution kernels in the first convolution layer is 32, the number of convolution kernels in the second convolution layer is 64, and the number of convolution kernels in the third convolution layer is 128. The convolution kernel size is 3, the zero padding step is 1, the activation function uses the ReLU function, and the pooling layer uses the maximum pooling method to effectively extract the high-dimensional features of the time series data set; The specific process of using principal component analysis PCA to reduce data dimension is as follows: S32-1, standardization processing: standardize the feature data processed by the convolutional neural network , so that the mean of each feature is 0 and the variance is 1; where, is the original eigenvalue, is the standardized eigenvalue, It is a feature The mean of It is a feature The standard deviation of S32-2, calculate the covariance matrix: the standardized data matrix is , then the covariance matrix The calculation formula is , where n is the sample size; S32-3, Calculating Eigenvalues ​​and Eigenvectors: Solving the Covariance Matrix The eigenvalues ​​and eigenvectors of are given by ,in, is the eigenvalue to be solved, is the feature vector, is the covariance matrix; S32-4, select the main components: first calculate the The contribution rate of the eigenvalue Then sort the eigenvalue contribution rates from large to small, and select the top ones with a cumulative contribution rate of 85%. features as principal components, forming a eigenvector matrix ; S32-5, projection data: the standardized data matrix Projection onto the selected eigenvector matrix On the above, we get the data matrix after dimensionality reduction ; The construction method of the long short-term memory network LSTM network is to design a multi-input and multi-output LSTM model, including input layer, LSTM layer, Dropout layer, fully connected layer, and output layer; the specific process is: S33-1, input layer: using the data matrix after multivariate data fusion As input data; S33-2, LSTM layer: contains 128 units, used to learn the time series data in the input layer; each LSTM unit The update formula of the moment is as follows, where: is at the time step Input, is the time step The hidden state of is the time step The cell state, is the current cell state, and are the weight matrix and bias vector respectively, yes sigmoid Activation Function , tanh is the hyperbolic tangent function ; ； S33-3, Dropout layer: Set the dropout ratio to reduce overfitting during network training; S33-4, fully connected layer: maps the output of LSTM to the target output space, with the same number of nodes as the number of output categories; S33-5, output layer: uses SoftMax activation function to output the fault type of the equipment; Represents the model for The predicted value of each category, Indicates that it belongs to The probability value of class failure, Represents the total number of fault categories, and the formula is as follows: 。 3. The complex equipment fault diagnosis method based on multivariate data fusion according to claim 2 is characterized in that: The specific process of model training and optimization in step S3 is: S34-1, Calculate the loss function: Use the multi-classification cross entropy function as the loss function ,in, is the total number of samples, is the total number of fault categories, It is Whether the sample belongs to the category The real situation, The model predicts samples belong to the category The probability is as follows: ； S34-2, model optimization: The Adam learning rate adaptive algorithm is used to optimize the model; Indicates the current iteration number, is The gradient of time, Represents the learning model, is the gradient is the exponential moving average, is the exponential moving average of the bungalow gradient, is the learning rate, and is the decay rate, set to 0.9 and 0.999, for , the formula is as follows: ； S34-3, Model Evaluation: Accuracy Accuracy As an evaluation metric, it is used to measure the accuracy of classification; S34-4, model training and saving: Execute training, dynamically adjust the training process according to the performance of the model on the validation set, stop training and save the model when the performance of the validation set does not improve within the set period.

4. The complex equipment fault diagnosis method based on multivariate data fusion according to claim 2 is characterized in that: The specific process of model deployment and reasoning in step S4 is: exporting the trained model into a suitable format and deploying it in the edge computing device, and then encapsulating the process of loading the model file and real-time data on the edge computing device into an API service so that it can receive requests and return diagnostic results; During real-time diagnosis, the host computer software receives periodic data from various sensors via Ethernet and performs preprocessing. It then infers the real-time data by calling the API interface, outputs the fault diagnosis results, and realizes online fault diagnosis of complex equipment.

5. The complex equipment fault diagnosis method based on multivariate data fusion according to claim 2 is characterized in that: The specific process of data preprocessing in step S2 is: S2-1, denoising: using moving average filtering method, ,in is the filtered data, is the original data, N is the window size; S2-2, Normalization: Using the standard deviation method, ,in, is the standardized data, is the original data, is the mean of the original data, is the standard deviation of the original data; S2-3, missing value processing: linear interpolation method is used, ,in, is the interpolated data, and is the nearest known data point before and after the missing point, and is the corresponding time point; S2-4, construct time series samples: divide the two-dimensional matrix into multiple time series samples according to the time window and step size, where the time window is 1 hour and the step size is 1; S2-5, data labeling: Combined with expert experience, label each time series sample with a corresponding fault label y, where normal operation is 0, fault type is 1, and fault types are 2~n; S2-6, divide the data set: randomly construct a data set with the processed time series samples and the corresponding labels, and divide it into a training set and a validation set. The training set is used to train the model, and the validation set is used to monitor the model performance during the training process.

6. The complex equipment fault diagnosis method based on multivariate data fusion according to claim 2 is characterized in that: The multi-sensor data includes but is not limited to one or more combinations of flow sensor data, pressure sensor data, voltage sensor data, current sensor data, and temperature, humidity and air pressure sensors.

Images