Equipment fault prediction method based on multi-source sensing data and space-time neural network

Abstract

The invention discloses an equipment fault prediction method based on multi-source sensing data and a space-time neural network. The state and fault prediction of a complex industrial production system needs to monitor more correlative process variables at the same time. According to the method, dimension reduction processing is not carried out on collected data, information in the data cannot be lost, the collected data of a plurality of monitoring variables are synchronously sampled and normalized respectively, the data are segmented according to the equal duration and then assembled into a time sequence two-dimensional array data set, the row of each two-dimensional array represents the monitoring variable, and the row of each two-dimensional array represents the monitoring variable; the column represents a sampling time. And learning correlation change features and time sequence features of each monitoring variable in the data set by using a space-time prediction neural network, and training a neural network model. And predicting data of each monitoring variable in the future by using the trained space-time prediction neural network model. And according to the predicted data, whether an abnormality or a fault occurs in the future equipment operation and the occurrence time are judged, so that a basis is provided for overhaul and maintenance plans.

Description

Technical Field

[0001] The present invention relates to a method for monitoring the health status of industrial production process equipment and predicting faults, and in particular to a method for predicting equipment faults based on multi-source sensor data and spatiotemporal neural networks. Background Art

[0002] Complex industrial production systems, such as process industries, have very complex production processes, and their measurement systems contain many process variables. Process variables usually contain a variety of physical quantities, such as temperature, flow, speed, displacement, angle, pressure, vibration, noise, etc. Moreover, process variables are interrelated and coupled with each other. Changes in any variable may cause changes in other variables, making the cause-effect relationship complicated, increasing the difficulty and complexity of operating status and fault prediction, and making it difficult to establish an accurate mathematical model. It is usually necessary to monitor data containing multiple variables and collect a large amount of data for analysis to find patterns. Multivariate data sets provide rich information, but also increase the difficulty of data analysis. Due to the large number of measurement variables and the correlation between variables, the first task of fault diagnosis is usually data dimensionality reduction. However, data dimensionality reduction will always lose some implicit information in the data to a greater or lesser extent, and sometimes the lost information may be important information.

[0003] The principal component analysis (PCA) method is a multivariate statistical analysis method widely used in modeling and fault detection in process industries. It projects the multivariate high-dimensional data space into a relatively independent low-dimensional space to obtain the orthogonal projection axis that maximizes the data variance, so as to eliminate data correlation and reduce dimensionality. By constructing a set of new variables to reduce the dimension of the original data space, the main change information is extracted from the mapping space for statistical analysis and fault diagnosis.

[0004] The data collected by process industries are generally highly correlated, and each type of data is unevenly distributed. Although the principal components extracted by the PCA method can retain most of the data variation characteristics, these principal components can only capture the overall structure of the process data and do not consider the inherent local information between different samples. In nonlinear situations, small features in the data may also contain important information. The PCA method often cannot achieve satisfactory results for nonlinear problems that arise in engineering. Moreover, PCA is a static algorithm that assumes that the process is time-invariant, that is, the relationship between variables does not change with time, and does not consider the dynamic characteristics of process industry systems with time-varying characteristics. Most process industries have slow-changing characteristics. If the normal operating condition model established earlier is used for fault detection, it is easy to cause misjudgment.

[0005] For multi-spatiotemporal time series data collected in process industries, existing data modeling methods do not effectively utilize multi-source sensor data and historical operation data.

[0006] In order to solve the aforementioned problems, the present invention proposes a method which does not require dimensionality reduction processing on the data, and therefore does not lose the correlation and fine feature information in the data, while being able to extract the mutually coupled time series features in the data of multiple process physical variables, for predicting the operating status and failure of production process equipment, as well as the time when the failure may occur. Summary of the invention

[0007] Purpose of the invention: The difficulty in predicting the operating status and faults of complex industrial production systems lies in the problem of feature extraction of multiple process variable data that change over time and have correlation. Feature extraction from multiple monitoring variable data collected by multi-source sensors is the premise and basis for predicting future operating status and faults. The present invention proposes a method that does not require dimensionality reduction processing of the collected data, does not lose correlation information in the data, and can accurately extract mutually coupled time series features from multiple monitoring variable data, which is used to predict the operating status of equipment, potential faults, and the time when faults occur in industrial production processes.

[0008] Technical solution: The data of monitoring variables collected by multiple sensors are sampled at the same sampling frequency, normalized respectively, and segmented according to equal time lengths. The segmented data are connected in the row direction (vertical axis) to form a time-synchronized two-dimensional array. Each row in the vertical axis direction represents a monitoring variable, and the column direction (horizontal axis) represents the sampling time. The value of the array represents the amplitude of each monitoring variable changing over time. Normalization cannot be performed on all data to avoid drowning the characteristics of small variables, but should be performed separately on the data of each monitoring variable. The above processing is performed on the collected data to obtain a two-dimensional array data set of time series. The spatiotemporal prediction neural network model can well extract the characteristics of the two dimensions of time and space. Applying the spatiotemporal prediction neural network to the time series two-dimensional array data set of the above multiple monitoring variables can extract the characteristics of the monitoring variable dimension and the time series dimension. It can not only extract the characteristics of the correlation changes between variables, but also extract the time series characteristics reflecting the trend of variable changes. In order to train the spatiotemporal prediction neural network model, the above time series two-dimensional array data set is transformed in dimension and divided into a training set, a validation set, and a test set. The training set and validation set are used to train the spatiotemporal prediction neural network model, and the test set is used to evaluate the trained model. The deployed spatiotemporal prediction neural network model can be used to predict the operating status and faults of equipment in the future. An array or a sequence of multiple arrays of past time is input to the deployed spatiotemporal prediction neural network model. The neural network model can predict an array or a sequence of multiple arrays of future time. The prediction result represents the change pattern of each monitoring variable in the future. Based on the prediction results, it is possible to determine the operating status of the equipment in the future, whether a fault will occur, and the time when the fault will occur. The main steps include: (1) Data collection: Use sensors and other measurement methods to collect data on various monitoring variables related to the production process and equipment operation status, such as temperature, flow, speed, displacement, angle, pressure, vibration, noise, concentration, etc.; (2) Data preprocessing: In order to ensure that the data of each monitoring variable are synchronized in time, the collected data are resampled at the same sampling frequency. The data of each variable are normalized and segmented according to equal time length. Then, the data of each monitoring variable that has undergone the above processing are connected in the row direction (vertical axis) to form a time-synchronized two-dimensional array. The rows of the two-dimensional array represent the monitoring variables, the columns of the array (horizontal axis) represent the sampling time, and the values ​​of the array represent the amplitude of each variable over time. Each array contains the law of the change of each monitoring variable over time and the characteristics of the correlation between the implicit variables. After the sampling data is processed by the above method, a time series two-dimensional array data set is obtained. At this time, the time series two-dimensional array data set cannot be directly used for the training of the spatiotemporal prediction neural network model. It is also necessary to expand the dimension of the time series two-dimensional array data set according to the input and output dimensions of the spatiotemporal prediction neural network model, and add one dimension to the two-dimensional array to become a three-dimensional array. The dimension of the array becomes (row, column, channel) or (channel, row, column). Then, the sliding window is used to slide the time series three-dimensional array data set to generate a new data set for neural network model training. The length of the sliding window is determined by the length of the input and output array sequences of the neural network model. The newly generated data set contains sample data and label data, and the sample data and label data are respectively composed of an array or a sequence of multiple arrays. The newly generated data set is divided into a training set, a validation set, and a test set. The training set and validation set are used to train the spatiotemporal prediction neural network model, and the test set is used to evaluate the trained model; (3) Training and deploying spatiotemporal prediction neural network models: Use the spatiotemporal prediction neural network model to learn the characteristics of the interrelated changes and change trends of the variables in the training data set, that is, the characteristics of the monitoring variables and time series. The spatiotemporal prediction neural network combines the advantages of convolutional neural networks (CNN) and long short-term memory neural networks (LSTM). One implementation method is to simply combine the CNN architecture with the LSTM architecture, first use convolution operations to learn the spatial features in the input data, and then use the LSTM layer to learn the time series features of the input data, that is, the CNN-LSTM network architecture. Another implementation method is to deeply combine CNN and LSTM to form a convolutional long short-term memory network (ConvLSTM). The ConvLSTM neural network can simultaneously extract the features of the spatial and temporal dimensions of the input data. In addition, there are other network architectures that are improved on the basis of these two basic implementation methods. Use the training set and validation set processed in step (2) to train the spatiotemporal prediction neural network model, and use the test set to evaluate the trained model. The trained spatiotemporal prediction neural network model can be deployed and used for prediction; (4) Predicting operating status and faults: Input a three-dimensional array or a sequence of multiple three-dimensional arrays of the past time into the deployed spatiotemporal prediction neural network model. The output of the model is a three-dimensional array or a sequence of multiple three-dimensional arrays predicted for a period of time in the future. The prediction results represent the change patterns of each monitoring variable over time in the future. Denormalizing and visualizing the predicted data can clearly show the changes of each monitoring variable in the future. Based on the prediction results, it can be further determined whether the operating status of the equipment is abnormal, whether a fault will occur, and the time when the abnormality or fault occurs within the predicted time range. The occurrence of a fault is generally a slow-changing process, which can be judged by observing and analyzing the change trend of the predicted data and whether it exceeds the set threshold. Alternatively, under the condition that historical fault feature data is available, the predicted data can be compared and analyzed with the fault feature data for judgment.

[0009] Beneficial effects: (1) The method proposed in the present invention can fuse the monitoring variable data of various physical quantities collected by various sensors, and can more fully extract the correlation characteristics between various variables; (2) When fusing the data of multiple variables, the original collected data is not subjected to dimensionality reduction processing, and the information of the correlation between the various monitoring variables and the small feature data in the original data is not lost; (3) The use of spatiotemporal prediction neural network can better learn the correlation change characteristics and time series characteristics between the various monitoring variables in the data, and thus can more accurately predict the production process and equipment operation status in the future; (4) The prediction result includes the change amplitude and time information of each monitoring variable, which is convenient for visual observation and analysis of the changes in the future production process and equipment operation status, and judge whether anomalies or failures will occur, and when they will occur; (5) The prediction time range can be flexibly set according to the periodic characteristics of the production process and the actual needs of inspection and maintenance, and prediction schemes of different time lengths of short, medium and long can be adopted to adapt to the different needs of control management and planning, and help reduce the losses caused by abnormal downtime. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a flow chart of the method of the present invention.

[0012] Figure 2 It is a schematic diagram of the input, output and prediction of a spatiotemporal prediction neural network model.

[0013] Figure 3 This is a typical structural diagram of the ConvLSTM spatiotemporal prediction neural network model.

[0014] Figure 4 This is a typical structural diagram of the Conv-LSTM spatiotemporal prediction neural network model. Implementation

[0015] The technical solution of the present invention is further described below in conjunction with an embodiment. The device fault prediction method based on multi-source sensor data and spatiotemporal neural network described in this embodiment comprises the following steps: (1) Data collection: Use various sensors to collect data on various monitoring variables related to the production process and equipment operation status, such as temperature, flow, speed, displacement, angle, pressure, vibration, noise, concentration, composition, etc.; (2) Data preprocessing: The collected data are resampled at the same sampling frequency to ensure that the data of all monitoring variables are synchronized in time. The data of each variable are normalized and segmented according to equal time lengths, for example, the data of each variable are segmented with a fixed length of 24 hours. Then, the data of each monitoring variable that has undergone the above processing are connected in the row direction (vertical axis) to form a two-dimensional array. The rows of the array represent the variables, and the columns of the array (horizontal axis) represent the sampling time. Each two-dimensional array contains the changing rules of each monitoring variable within 24 hours and the characteristics of the correlation between the implicit variables. After the above processing method, a time series two-dimensional array data set is obtained. According to the input and output dimensions of the spatiotemporal prediction neural network model, the dimension of the time series two-dimensional array data set is expanded, and the two-dimensional array is increased by one dimension to become a three-dimensional array. The dimension of the three-dimensional array is (row, column, channel) or (channel, row, column). Then, according to the length of the array sequence of the input and output of the neural network model, the data set is slid using a sliding window to generate a new data set that can be used for neural network model training. For example, we want to use the sequence of 10 arrays in the past to predict the sequence of 10 arrays in the future, and the length of the sliding window is 10. Slide the sliding window along the time axis of the data set with a step size of 1 to generate a new data set. Each sample in the new data set contains a sequence of 10 three-dimensional arrays, and the corresponding label is also a sequence of 10 three-dimensional arrays. Divide the newly generated data set into training set, validation set and test set; (3) Training and deploying spatiotemporal prediction neural network models: Taking the convolutional long short-term memory (ConvLSTM) spatiotemporal prediction neural network as an example. The basic structure of the model consists of a two-dimensional convolutional long short-term memory layer (ConvLSTM2D), a normalization layer (BatchNormalization), and a three-dimensional convolution layer (Conv3D). The ConvLSTM2D layer is used to extract the characteristics of the interrelated changes of each variable in the input sample data, as well as the time series characteristics. The Conv3D layer is used as the output layer to format the output. The number of layers, number of filters, size of the convolution kernel, learning rate, activation function, loss function, and optimizer selection of the ConvLSTM2D, BatchNormalization, and Conv3D layers in the model can be determined by adjusting parameters. When enough training data samples are collected, the spatiotemporal prediction neural network model can be trained. The model is trained offline. The model is trained with a training set and a validation set, and finally the model is evaluated with a test set. The trained model can be saved and deployed for prediction; (4) Predicting operating status and failures: Input a sequence of 10 three-dimensional arrays of past time to the deployed spatiotemporal prediction neural network model, and the model will output a predicted sequence of 10 three-dimensional arrays. The input sequence and output sequence of the model are continuous in time. The length of time represented by each three-dimensional array is 24 hours. The sequence of 10 three-dimensional arrays predicted by the neural network model represents the operating status of the equipment in the next 240 hours. Denormalizing and visualizing the predicted data can clearly show the changes in various monitoring variables in the future. Based on the predicted data, it can be further determined whether the equipment will have an abnormality or failure within the predicted time range, and the time when the abnormality or failure will occur.

[0016] The above-described embodiment is only a specific implementation example of the present invention, and is not intended to limit the scope of implementation of the present invention. All changes made based on the principle of the present invention should be included in the protection scope of the present invention.

Claims

1. Equipment failure prediction method based on multi-source sensor data and spatiotemporal neural network, characterized by: Includes steps: (1) Collect data on the changes in physical variables related to the production process and equipment operating status over time as monitoring variable data; (2) The data of the collected monitoring variables are synchronously sampled, normalized separately, and then divided into segments of equal time length and assembled into a time series two-dimensional array data set. The rows of each array in the data set represent the monitoring variables, the columns represent the sampling time, and the values ​​represent the amplitude of the monitoring variables changing over time. The time series two-dimensional array contains the correlation change characteristics of each monitoring variable and the time series change trend characteristics; (3) Use the spatiotemporal prediction neural network to extract the features of the monitoring variable dimension and the time series dimension in the time series array, train and deploy the spatiotemporal prediction neural network model; (4) Use the deployed spatiotemporal prediction neural network model to predict the data of equipment operation status monitoring variables in the future, and determine whether a fault will occur and the time when the fault will occur based on the predicted data.

2. The equipment failure prediction method based on multi-source sensor data and spatiotemporal neural network according to claim 1 is characterized in that: In the step (2), according to the dimensional requirements of the input and output of the spatiotemporal prediction neural network model, the two-dimensional array is dimensionally expanded and transformed into a three-dimensional array, and the dimension of the three-dimensional array is (row, column, channel) or (channel, row, column).

3. The equipment failure prediction method based on multi-source sensor data and spatiotemporal neural network according to claim 1 is characterized in that: In the step (2), according to the length requirements of the array sequences of the input and output of the spatiotemporal prediction neural network model, a sliding operation is performed on the time series three-dimensional array data set using a sliding window to generate a data set for training the spatiotemporal prediction neural network model.

4. The equipment failure prediction method based on multi-source sensor data and spatiotemporal neural network according to claim 1 is characterized in that: In the step (3), one implementation method of the spatiotemporal prediction neural network model is to use a convolutional long short-term memory network (ConvLSTM) to simultaneously learn the correlation change characteristics and time series characteristics between the monitoring variables in the input data, that is, the ConvLSTM network architecture; another implementation method is to first use a convolutional network (CNN) to learn the correlation change characteristics between the monitoring variables in the input data, and then use a long short-term memory network (LSTM) to learn the time series characteristics of the input data, that is, the CNN-LSTM network architecture.

5. The equipment failure prediction method based on multi-source sensor data and spatiotemporal neural network according to claim 1 is characterized in that: In step (4), the prediction result of the spatiotemporal prediction neural network model is a time series array. The predicted data is denormalized and visualized to clearly display the changes of various monitoring variables in the future. Based on the predicted data, it is determined whether the equipment will fail within the predicted time range and the specific time when the failure will occur.

Images