Complex equipment operation state prediction method based on multi-order time sequence feature fusion

Abstract

The invention discloses a complex equipment operation state prediction method based on multi-order time sequence feature fusion. Comprising the following steps: firstly, preprocessing operation parameter data of complex equipment to obtain a time sequence data set; then calculating a multi-order differential value sequence corresponding to each parameter in the time sequence data set, and further constructing a training data set; constructing a graph-time domain fusion network, and training the graph-time domain fusion network by using the training data set to obtain a state prediction model of multiple time sequence scales; and finally, according to the actual operation parameter data of the complex equipment, performing state prediction of the self-adaptive time sequence window by using the state prediction model of the multiple time sequence scales to obtain a corresponding operation state prediction result. The model provided by the invention is stable and light in structure, and can carry out on-line time sequence prediction in equipment operation, guide the equipment operation decision and ensure the stability and safety of complex equipment operation.

Description

Technical Field

[0001] The present invention relates to a method for predicting the operating state of complex equipment in the field of time series data prediction, and particularly to a method for predicting the operating state of complex equipment based on multi-order time series feature fusion. Background Art

[0002] The prediction of the operating state of complex equipment is a typical time series prediction. Through time series prediction, it is possible to accurately predict the operating state of equipment at a certain time in real time during the operation of the equipment, which is of great significance for multiple aspects such as equipment state monitoring, precise control, efficient decision-making, and safety guarantee. For example, the time series prediction of the flight state during flight can make full use of a large amount of flight control data generated and recorded during the flight, including important flight state parameters such as the northeast ground coordinates of the aircraft, aircraft attitude angles, dynamic / static pressure, Mach number, airflow angle, engine thrust, angle of attack, and sideslip angle, improving the intelligent, autonomous, and internal and external coordination capabilities of the aircraft, optimizing flight driving techniques, and being of great significance for achieving flight stability and safety. In addition, the time series prediction of the operating state of complex equipment is of great significance for the digital transformation of complex equipment.

[0003] The prediction of time series has always been a research hotspot for experts and scholars at home and abroad. Many experts and scholars have made certain progress in multiple fields such as transportation, logistics, and power. Common prediction methods include RNN, LSTM, GRU, the time domain convolutional network (TCN) based on dilated causal convolution, Transformer, and the spatio-temporal graph neural network (STGNNs) proposed in recent years. LSTM and GRU are algorithms proposed for a long time, and their recursive structures make them weak in terms of computational efficiency and the modeling ability of time series data. The TCN model has higher efficiency based on parallel convolutional operations, but it cannot handle the correlation between parameters. The Transformer model has excellent modeling ability based on the self-attention mechanism, but it has the disadvantages of many parameters and large computational overhead. Common spatio-temporal graph neural networks also have these characteristics of the Transformer model.

[0004] Current time series data prediction work mostly focuses on learning the long-term dependencies in time series data and emphasizes improving the algorithm accuracy, with no high requirement for algorithm speed. However, in the scenario of predicting the operating state of complex equipment, on the premise of a certain prediction accuracy, the efficiency of prediction calculation is equally important, and real-time time series prediction needs to be achieved. Summary of the Invention

[0005] To solve the problems mentioned in the background art, the present invention proposes a method for predicting the operating state of complex equipment based on multi-order time series feature fusion. The method proposed by the present invention fully explores the features of time series data and is used to accurately predict the operating states of multiple parameters during the operation of complex equipment, guide equipment operation decisions, and prevent potential safety hazards caused by bad decisions. The present invention can overcome the deficiencies of existing methods, use simulation or sensor measured data during the operation of the equipment, calculate the multi-order difference values of time series data, and combine physical prior knowledge with the multi-head attention mechanism to fully explore the complex physical connections between parameters. After reconstructing the intermediate results, the sliding window size is selected through a dynamic adaptive strategy, and time series data of appropriate length is intercepted. Based on causal dilated convolution, the time series features between parameters are fully learned through multiple layers of convolution to perform time series prediction of multiple parameters. The overall model is compact and the data volume is moderate, which can be used for real-time prediction during the operation of complex equipment, and the prediction accuracy of multiple parameters is high, providing a new research idea for fully exploring the operation data of complex equipment, guiding decisions, and improving the operation safety of complex equipment.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] 1. A method for predicting the operating state of complex equipment based on multi-order time series feature fusion, characterized by including the following steps:

[0008] S1: After preprocessing the operation parameter data of the complex equipment, a time series data set is obtained;

[0009] S2: Calculate the multi-order difference value sequences corresponding to each parameter in the time series data set, then establish the connection relationship between graph nodes according to the correlation between each time series data and physical prior knowledge and record it as an adjacency matrix, and aggregate the multi-order difference value sequences corresponding to all parameters at each single moment to form graph node features, thereby obtaining a training data set;

[0010] S3: Construct a graph-time domain fusion network, and after training the graph-time domain fusion network with the training data set, obtain a multi-time series scale state prediction model;

[0011] S4: According to the actual operation parameter data of the complex equipment, use the multi-time series scale state prediction model to perform state prediction of an adaptive time series window, and obtain the corresponding operating state prediction result.

[0012] The preprocessing includes data cleaning and data standardization.

[0013] In the S3, the graph-time domain fusion network includes a graph neural network model based on the multi-head attention mechanism and a time domain convolutional network based on dilated causal convolution, and a decoder is included in the time domain convolutional network based on dilated causal convolution.

[0014] In S3, during training, a multi-size sliding window is used to extract features from the output of the graph neural network model based on the multi-head attention mechanism, so as to obtain feature sets of different time series scales. Then, the feature sets of different time series scales are respectively input into the time-domain convolutional network based on dilated causal convolution, and the time-domain convolutional network based on dilated causal convolution corresponding to different time series scales is trained respectively. After training, the graph-time domain fusion network and the time-domain convolutional networks based on dilated causal convolution corresponding to all time series scales are integrated to form a multi-time series scale state prediction model.

[0015] In S3, the loss function of the graph-time domain fusion network is an improved mean square error loss, which satisfies the following formula:

[0016]

[0017] where R MSE is the final loss result, N is the number of parameters, weight is the weight value corresponding to each parameter, is the true value at the target time of the m-th parameter, y m is the predicted value at the target time of the m-th parameter, k 1 is the reduction error weight, is the original true value at the target time of the m-th parameter, y’ m is the reduction predicted value at the target time of the m-th parameter.

[0018] The size of the multi-size sliding window satisfies an exponential rule, specifically 2 0 , 2 1 , …, 2 k , where k is the window scale coefficient.

[0019] S4 is specifically as follows:

[0020] Set the initial size of the time series window, use the current time series window to select the actual time series sequence from the actual operation parameter data of the complex equipment, preprocess the actual time series sequence and generate the corresponding graph node features, and then input them into the multi-time series scale state prediction model to obtain the corresponding operation state prediction result; during the state prediction process, calculate the prediction error of the model, adaptively adjust the size of the time series window according to the prediction error of the model, so as to adjust the actual time series sequence, and then obtain a better operation state prediction result.

[0021] II. A computer device

[0022] The device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting the operation state of a complex equipment based on multi-order time series feature fusion are implemented.

[0023] 3. A computer-readable storage medium

[0024] The medium stores a computer program, which, when executed by a processor, implements the steps of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion.

[0025] IV. A COMPUTER PROGRAM PRODUCT

[0026] The product includes a computer program / instruction, which, when executed by a processor, implements the steps of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1) The state prediction model constructed by fully mining the time series data generated during the operation of complex equipment solves the problem that traditional TCN cannot handle multiple parameters. By utilizing the causal expansion convolution and aggregation effect of TCN and the custom loss function, the time series characteristics of multi-dimensional time series data are fully learned to achieve accurate prediction of multi-dimensional and multi-scale parameters. It is superior to LSTM and GRU models in terms of model stability and modeling ability.

[0029] 2) The graph attention model constructed based on physical prior knowledge and multi-order features of parameters can fully learn the physical correlation characteristics between parameters and has a certain degree of interpretability. At the same time, the multi-head attention mechanism also improves the robustness and stability of the model, making the prediction results more accurate.

[0030] 3) The sequential splicing of graph neural networks and time series prediction models makes the data volume of the overall model moderate. Compared with other spatiotemporal graph neural networks or Transform models, it is more lightweight, and the training stability and convergence speed are relatively high.

[0031] 4) Through the integration of multiple input window sizes, real-time adaptive multi-time and multi-scale parameter regression prediction can be achieved during the operation of complex equipment, and the overall model has strong generalization ability and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 The present invention is a flowchart of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion.

[0033] Figure 2 This is a model structure diagram of the prediction network model based on multi-order time series feature fusion.

[0034] Figure 3In a certain flight state of the embodiments of the present invention, the comparison curve of the original data and the model prediction data of part of the test set data, including: pitch, yaw, and roll angles during the flight of the aircraft.

[0035] Figure 4 In a certain flight state of the embodiments of the present invention, the comparison curve of the original data and the model prediction data of part of the test set data, including the east, north, and ground coordinates during the flight of the aircraft. Specific embodiments

[0036] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0037] The method of this embodiment is a method for predicting the operating state of complex equipment based on multi-order time series feature fusion. In this embodiment, the simulation model of the F-16 aircraft is used as a specific case of complex equipment to solve the problem that the real flight data of the aircraft cannot be publicly used, mine the spatio-temporal features of the simulation data, and achieve multi-parameter accurate prediction of the simulation data. The overall flow chart is as Figure 1 shown, and the method includes the following steps:

[0038] S1: Simulate and obtain the time series data set of the flight state during the multi-parameter flight process and perform data processing.

[0039] S1.1: Obtain simulation data under various flight conditions such as different altitudes, different speeds, and different operations through trimming and simulation models, select the multi-scale parameters to be predicted, and export them as.txt files, including: north coordinate (npos), east coordinate (epos), altitude (h), roll angle (roll), pitch angle (pitch), yaw angle (yaw), roll angular velocity (p), pitch angular velocity (q), yaw angular velocity (r), airspeed (v), three overloads (nx, ny, nz), dynamic pressure (Q), static pressure (ps), Mach number (mach), angle of attack (alpha), sideslip angle (beta) 18 state parameters, and engine thrust (T), elevator angle (E), aileron angle (A), rudder angle (R) 4 control parameter, the simulation time is 100s, the data sampling period is 10Hz, the number of flight conditions is 100, and a total of 100 sets of aircraft operating state data are generated, including aircraft pitch, roll, yaw and other operations.

[0040] S1.2: Standardize the data. For parameters with uneven data distribution, set the mean and variance according to the upper and lower limits of the data. For uniformly distributed data, calculate the mean and variance according to the time series data, and scale the features to an overall distribution with a mean of 0 and a variance of 1 to improve the convergence speed and stability of the model.

[0041] Divide the data into a training set, a validation set, and a test set. Then divide the data in the training set into multiple batches, with each batch consisting of 20 groups of time-series data, reducing the computational consumption during training while improving the generalization ability of the model.

[0042] S2: Calculate the multi-order difference value sequences corresponding to each parameter in the time-series dataset. The calculation formula for the multi-order difference value of each parameter is as follows:

[0043]

[0044] Where, is the d-th order and i-th difference value of parameter x, is the (d - 1)-th order and i-th difference value of parameter x, is the (d - 1)-th order and (i - 1)-th difference value of parameter x, and Δt is the time-series interval between the i-th parameter and the (i - 1)-th parameter.

[0045] Each parameter has a corresponding multi-order difference value sequence. The time-series data of the d-th order of parameter x is expressed as is the n-th data of the d-th order of parameter x. Then, based on the correlation between each time-series data and physical prior knowledge, establish the connection relationship between the graph nodes and record it as the adjacency matrix. For parameters with a correlation greater than the threshold of 0.8, the value of its adjacency matrix is set to 1, indicating the connectivity between parameters. For parameters less than the threshold, it is set to 0. The adjacency matrix value of the input parameters is set only with a unidirectional value. The dimension of the adjacency matrix is (N, N). Physical prior knowledge determines the order size of the time-series features and the correlation relationship between parameters. And aggregate the multi-order difference value sequences corresponding to all parameters at each single moment to form graph node features with a dimension of [N, D], where N is the number of parameters and D is the parameter order, thereby obtaining the training dataset;

[0046] The initial adjacency matrix constructed in this embodiment is:

[0047]

[0048] edge_index represents the sparse adjacency matrix in the form of a coordinate list, where the two numbers in each column represent that the adjacency matrix value between the corresponding two parameters is 1. Use the graph node features with a dimension of (N, D) and edge_index as the input of the graph attention model.

[0049] S3: Construct a graph-time domain fusion network. After training the graph-time domain fusion network using the training dataset, obtain a multi-time-scale state prediction model;

[0050] In S3, the graph-time domain fusion network includes a graph neural network model based on the multi-head attention mechanism and a time domain convolutional network (TCN) based on dilated causal convolution, where the time domain convolutional network based on dilated causal convolution contains a decoder.

[0051] Parameters such as the number of multi-attention mechanism heads, the number of layers, and the aggregation method in the graph neural network model selected for the multi-head attention mechanism are represented by the following formula:

[0052]

[0053] Among them, α ij represents the influence weight between nodes, the vector is used for aggregation, the weight matrix W changes the vector dimension, LeakyReLU() is the activation function, and [||] indicates that the features are aggregated in a concatenated manner. Finally, the result is calculated by softmax to obtain the attention scores between nodes, and the attention scores are multiplied by the features and then concatenated to obtain the intermediate layer features of the model.

[0054]

[0055] Among them, is the eigenvalue of node i, is the eigenvalue of node j, K is the number of heads of the attention mechanism, j is several nodes connected to node i, is the attention score between nodes calculated in the k-th attention head of the attention mechanism, σ() is the activation function, W k is the weight matrix of the k-th attention head, || represents the multi-head attention mechanism, and [||] indicates that the eigenvalues of multiple head parameters are aggregated in a concatenated manner. The intermediate layer features of the model are further subjected to a graph aggregation operation, and the final output is the graph node eigenvalue at a single moment, fully learning the physical feature correlations between various parameters and improving the accuracy of subsequent predictions.

[0056] The second layer of the graph attention network adopts a similar structure above, sets the number of heads of the multi-head attention to 1, and obtains the final features in an average manner. The main purpose is to aggregate the graph node features from the intermediate feature layer into the output feature structure.

[0057]

[0058] In S3, during training, the parameter feature combination output at a single moment is reconstructed into a time series of multi-parameter features with a dimension of (Batch, T, N), where Batch is the batch size, T is the total length of the input time series, and N is the number of parameters. A sliding window of multiple sizes is used to extract features from the output of the graph neural network model based on the multi-head attention mechanism, thereby obtaining feature sets of different time series scales. Then, the feature sets of different time series scales are respectively input into the time domain convolutional network based on dilated causal convolution, and the time domain convolutional network based on dilated causal convolution corresponding to different time series scales is respectively trained. After training, the graph-time domain fusion network and the time domain convolutional networks based on dilated causal convolution corresponding to all time series scales are integrated to form a multi-time series scale state prediction model.

[0059] During the training process, a custom loss function is constructed based on the predicted data at the target moment in the predicted time series output by the time domain convolutional network based on dilated causal convolution and the true data at the target moment for loss evaluation and backpropagation, realizing model training, verification, and testing.

[0060] The sizes of the sliding windows of multiple sizes satisfy an exponential rule, specifically 2 0 , 2 1 , …, 2 k , where k is the window scale coefficient, so that the data sequence after the moments {i, 2i, …, ni} can be predicted.

[0061] In the time series convolutional model based on causal dilation convolution, appropriate numbers of convolutional layers and dilation factors are set, and the parameter time series features at time t and before time t are fused to avoid gradient explosion or gradient disappearance during the learning process. In this embodiment, the data dimension input to the time convolutional network is (batch, 40, 22), the window size of the time series is 40, and the data after t = 2s is tested. The time convolutional network consists of two convolutional layers and one aggregation layer. The number of hidden layers in the convolutional layer is 450, and the number of layers in the aggregation layer is 100. The output parameter dimension is (Batch, K, 100), and then through a decoder, the data is mapped to n flight state parameters to achieve accurate regression prediction of multi-scale and multi-parameters at time t = 2s.

[0062] The loss function of the graph-time domain fusion network is an improved mean squared error loss, which satisfies the following formula:

[0063]

[0064] where, R MSE is the final loss result, N is the number of parameters, weight is the weight value corresponding to each parameter, is the true value at the target moment of the m-th parameter, y mis the predicted value of the m-th parameter target moment, k 1 is the reduction error weight, is the original true value of the m-th parameter at the target moment, y’ m is the reduction prediction value of the m-th parameter at the target moment.

[0065] In this loss function, the loss weights weight of each parameter are constructed according to the magnitude of the original data, so that multiple parameters can achieve approximate numerical precision. And it is still difficult to achieve the prediction of the same precision for multiple parameters based on the loss function with set weights. Fixed weights are difficult to fit the dynamic changes of parameters. Therefore, the reduction error in the non-normalized state is introduced, that is as the dynamic scale error in the learning process, so as to achieve accurate regression prediction of multiple parameters.

[0066] S4: Load the state prediction model with multi-temporal scales into the electronic device and integrate it into the complex equipment. Therefore, according to the actual operation parameter data of the complex equipment, the state prediction of the adaptive time series window can be carried out by using the state prediction model with multi-temporal scales, and the corresponding operation state prediction results can be obtained.

[0067] S4 is specifically as follows:

[0068] Set the initial size of the time series window, select the actual time series from the actual operation parameter data of the complex equipment by using the current time series window, preprocess the actual time series and generate the corresponding graph node features, and then input them into the state prediction model with multi-temporal scales to obtain the corresponding operation state prediction results; during the state prediction process, calculate the prediction error of the model, and adaptively adjust the size of the time series window according to the prediction error of the model, so as to adjust the actual time series, that is, select the TCN corresponding to the time series scale, and then obtain better operation state prediction results, so as to improve the overall generalization ability and adaptive ability.

[0069] In this embodiment, the model is evaluated on the pre-divided test set, and the results are as Figure 3 and Figure 4 shown. The predicted data of some significant flight state parameters are basically consistent with the target data after 2s, and the changes of multiple parameters at different scales are consistent with the trends of the original simulation values. In summary, the flight state multi-parameter regression prediction model based on multi-order time series feature fusion of the present invention realizes accurate regression prediction of multiple parameters during the flight process of the aircraft, has the characteristics of fast convergence speed, high prediction accuracy, and lightweight model. At the same time, the present invention can also be migrated to the operation state time series prediction scenarios of other complex equipment to monitor and assist in decision-making for equipment operation, which has practical engineering significance.

[0070] The above embodiments are used to explain the present invention rather than limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A method for predicting the operating status of complex equipment based on multi-order time series feature fusion, characterized in that: The following steps are involved: S1: After preprocessing the operating parameter data of complex equipment, a time series data set is obtained; S2: Calculate the multi-order difference value sequence corresponding to each parameter in the time series data set, then establish the connection relationship between the graph nodes based on the correlation between the time series data and the physical prior knowledge and record it as an adjacency matrix, and aggregate the multi-order difference value sequence corresponding to all parameters at each single moment to form the graph node feature, so as to obtain the training data set; S3: Build a graph-time fusion network, and use the training data set to train the graph-time fusion network to obtain a state prediction model at multiple time scales; S4: According to the actual operating parameter data of the complex equipment, the state prediction model of multiple time series is used to perform state prediction of the adaptive time series window to obtain the corresponding operating state prediction results.

2. According to claim 1, a method for predicting the operating status of complex equipment based on multi-order time series feature fusion is characterized in that: The preprocessing includes data cleaning and data standardization.

3. According to claim 1, a method for predicting the operating status of complex equipment based on multi-order time series feature fusion is characterized in that: In S3, the graph-time domain fusion network includes a graph neural network model based on a multi-head attention mechanism and a time domain convolutional network based on dilated causal convolution, wherein the time domain convolutional network based on dilated causal convolution includes a decoder.

4. According to claim 1, a method for predicting the operating status of complex equipment based on multi-order time series feature fusion is characterized in that: In S3, during training, a sliding window of multiple sizes is used to extract features from the output of the graph neural network model based on the multi-head attention mechanism, thereby obtaining feature sets of different time series scales, and then the feature sets of different time series scales are respectively input into the time domain convolutional network based on dilated causal convolution, and the time domain convolutional networks based on dilated causal convolution corresponding to different time series scales are respectively trained; After training, the graph-time fusion network is integrated with the time convolutional network based on dilated causal convolution corresponding to all time scales to form a state prediction model at multiple time scales.

5. According to claim 1, a method for predicting the operating status of complex equipment based on multi-order time series feature fusion is characterized in that: In S3, the loss function of the graph-time fusion network is an improved mean square error loss, which satisfies the following formula: Among them, R MSE is the final loss result, N is the number of parameters, and weight is the weight value corresponding to each parameter. is the target moment true value of the mth parameter, y m is the predicted value of the mth parameter target moment, k1 is the restoration error weight, is the original true value of the target moment of the mth parameter, y , m Restore the predicted value for the target moment of the mth parameter.

6. The method for predicting the operating status of complex equipment based on multi-order time series feature fusion according to claim 4 is characterized in that: The size of the multi-size sliding window satisfies the exponential rule, specifically 2 0 ,2 1 ,…,2 k , where k is the window scale coefficient.

7. The method for predicting the operating status of complex equipment based on multi-order time series feature fusion according to claim 1 is characterized in that: The S4 is specifically: The initial size of the timing window is set, and the actual timing sequence is selected from the actual operating parameter data of the complex equipment using the current timing window. The actual timing sequence is preprocessed and the corresponding graph node features are generated before being input into the multi-time-scale state prediction model to obtain the corresponding operating state prediction results. In the state prediction process, the prediction error of the model is calculated, and the timing window size is adaptively adjusted according to the prediction error of the model, thereby adjusting the actual timing sequence to obtain a better operating state prediction result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion as described in any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion as described in any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of a method for predicting the operating status of complex equipment based on multi-order time series feature fusion as described in any one of claims 1 to 7 are implemented.

Images