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

Abstract

The application discloses a complex equipment operation state prediction method based on multi-order time sequence feature fusion. The method comprises the following steps: firstly, obtaining a time sequence data set after pre-processing operation parameter data of complex equipment; then, calculating multi-order difference value sequences corresponding to each parameter in the time sequence data set, and then constructing a training data set; then, constructing a graph-time domain fusion network, training the graph-time domain fusion network by using the training data set, and obtaining a multi-time sequence scale state prediction model; finally, performing adaptive time sequence window state prediction by using the multi-time sequence scale state prediction model according to actual operation parameter data of the complex equipment, and obtaining corresponding operation state prediction results. The model provided by the application is stable and lightweight in structure, can perform time sequence prediction online during equipment operation, guide equipment operation decision-making, and ensure the stability and safety of complex equipment operation.

Description

TECHNICAL FIELD

[0001] The application relates to a complex equipment operation state prediction method in the field of time series data prediction, in particular to a complex equipment operation state prediction method based on multi-order time series feature fusion. BACKGROUND

[0002] The prediction of the operation state of complex equipment is a typical time series prediction. Through time series prediction, the operation state of equipment at a certain time can be accurately predicted in real time during equipment operation. The time series prediction of the operation state of complex equipment is of great significance to equipment state monitoring, accurate control, efficient decision-making, safety protection and the like. 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 flight, including important flight state parameters such as the east-north coordinates of the aircraft, the attitude angle of the aircraft, the dynamic / static pressure, the Mach number, the airflow angle, the engine thrust, the angle of attack and the sideslip angle, improve the intelligence, autonomy and internal and external coordination capability of the aircraft, optimize the flight driving technology, and have important significance for realizing the stability and safety of flight. In addition, the time series prediction of the operation state of complex equipment is of great significance to the digital transformation of complex equipment.

[0003] The prediction of time series has been a research hotspot of domestic and foreign experts and scholars. Many experts and scholars have made certain progress in the fields of transportation, logistics, power and the like. Common prediction methods include RNN, LSTM, GRU, time domain convolution network (TCN) based on extended causal convolution, Transform and the spatio-temporal graph neural network (STGNNs) proposed in recent years. LSTM and GRU are algorithms proposed for a long time. Due to the recursive structure, the calculation efficiency and the modeling capability of time series data are weak. The TCN model based on parallel convolution operation has higher efficiency, but cannot process the correlation between parameters. The Transform model based on the self-attention mechanism has excellent modeling capability, but has the shortcomings of many parameters and large calculation overhead. The common spatio-temporal graph neural network also has these characteristics of the Transform model.

[0004] Current time series data prediction work focuses on learning long-term dependencies in time series data and focuses on improving algorithm accuracy, without high requirements for algorithm speed. In the scenario of complex equipment operation state prediction, under the premise of a certain prediction accuracy, the efficiency of prediction calculation is also important, and real-time time series prediction is required. SUMMARY

[0005] In order to solve the problems mentioned in the background, the application proposes a complex equipment operation state prediction method based on multi-order time sequence feature fusion. The method proposed by the application fully excavates the time sequence data features, and is used for realizing accurate prediction of multi-parameter operation state in the complex equipment operation process, guiding equipment operation decision and preventing safety hazards caused by adverse decisions. The application can overcome the shortcomings of the existing methods, calculate the multi-order difference value of the time sequence data by using simulation or sensor measured data in the equipment operation process, fully excavate the complex physical relationship between parameters through physical prior knowledge and multi-head attention mechanism, reconstruct the intermediate results, select the sliding window size through dynamic adaptive strategy, intercept the time sequence data with suitable length, learn the time sequence features between parameters through multi-layer convolution based on causal dilated convolution, and perform multi-parameter time sequence prediction. The model is compact as a whole, the data amount is moderate, can be used for real-time prediction in the complex equipment operation process, and the prediction accuracy of multi-parameters is high. It provides a new research idea for fully excavating complex equipment operation data, guiding decision and improving complex equipment operation safety.

[0006] In order to achieve the above purpose, the technical scheme of the application is as follows:

[0007] One, a complex equipment operation state prediction method based on multi-order time sequence feature fusion, characterized by comprising the following steps:

[0008] S1: obtaining a time sequence data set after pre-processing the operation parameter data of the complex equipment;

[0009] S2: calculating the multi-order difference value sequence corresponding to each parameter in the time sequence data set, then establishing the connection relationship between the graph nodes according to the correlation between each time sequence data and the physical prior knowledge and recording it as an adjacency matrix, and aggregating the multi-order difference value sequence corresponding to all parameters in each single time to form a graph node feature, thereby obtaining a training data set;

[0010] S3: constructing a graph-time domain fusion network, training the graph-time domain fusion network by using the training data set, and obtaining a multi-time sequence scale state prediction model;

[0011] S4: according to the actual operation parameter data of the complex equipment, using the multi-time sequence scale state prediction model to perform state prediction of the adaptive time sequence window, and obtaining the corresponding operation state prediction result.

[0012] The pre-processing includes data cleaning and data standardization.

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

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

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

[0016]

[0017] wherein, R MSE is the final loss result, N is the number of parameters, weight is the weight value corresponding to each parameter, is the target time true value of the mth parameter, y m is the target time prediction value of the mth parameter, k1 is the reduction error weight, is the original target time true value of the mth parameter, y’ m is the target time reduction prediction value of the mth parameter.

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

[0019] The S4 is specifically:

[0020] An initial size of a time sequence window is set, an actual time sequence is selected from the actual running parameter data of the complex equipment by using the current time sequence window, the actual time sequence is preprocessed and the corresponding graph node features are generated, and then the graph node features are input into the multi-time scale state prediction model to obtain a corresponding running state prediction result; in the state prediction process, the prediction error of the model is calculated, and the size of the time sequence window is adaptively adjusted according to the prediction error of the model, so as to adjust the actual time sequence, and further obtain a more optimal running state prediction result.

[0021] II. A computer device

[0022] The device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the complex equipment running state prediction method based on multi-order time sequence feature fusion when executing the computer program.

[0023] III. A computer readable storage medium

[0024] The computer program is stored on the medium and is executed by the processor to implement the steps of the complex equipment operation state prediction method based on multi-order time sequence feature fusion.

[0025] Four, a computer program product

[0026] The product includes computer programs / instructions that are executed by the processor to implement the steps of the complex equipment operation state prediction method based on multi-order time sequence feature fusion.

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

[0028] 1) Fully tap the time sequence data generated in the operation process of complex equipment, and the state prediction model solves the problem that traditional TCN cannot process multiple parameters. By using the causal expansion convolution and aggregation of TCN and the self-defined loss function, the time sequence features of multi-dimensional time sequence data are fully learned, and accurate prediction of multi-dimensional and multi-scale parameters is realized. The model stability and modeling ability are better than those of the LSTM and GRU models.

[0029] 2) The graph attention model constructed based on the multi-order features of physical prior knowledge and parameters can fully learn the physical correlation features between parameters, has certain interpretability, and the multi-head attention mechanism also improves the model robustness and stability, so that the prediction result is more accurate.

[0030] 3) The sequential splicing of the graph neural network and the time sequence prediction model makes the data volume of the overall model moderate, and compared with other spatio-temporal graph neural networks or Transform models, the model 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 realized in the operation process of complex equipment, and the overall model has strong generalization ability and adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The flowchart of the complex equipment operation state prediction method based on multi-order time sequence feature fusion of the present application.

[0033] Figure 2 The model structure diagram of the prediction network model based on multi-order time sequence feature fusion.

[0034] Figure 3 The original data and model prediction data comparison curve of part of the test set data in a certain flight state in the embodiment of the present application, including the pitch, yaw and roll angles in the flight process of the aircraft.

[0035] Figure 4 For the original data of part of the test set data in some flight state in the embodiment of the application, the comparison curve of the model prediction data includes east, north and ground coordinates in the flight process of the aircraft. DETAILED DESCRIPTION

[0036] The application will be further described below in combination with the drawings and embodiments:

[0037] The mode of the embodiment is a complex equipment operation state prediction method based on multi-order time sequence feature fusion. In the embodiment, a simulation model of an F-16 aircraft is taken as a specific case of complex equipment to solve the problem that real flight data of the aircraft cannot be publicly used, to mine the space-time features of simulation data, and to realize accurate prediction of multi-parameter simulation data. The overall flowchart is shown in FIG. 1, and the method comprises the following steps: Figure 1

[0038] S1: Simulation obtains a multi-parameter flight process flight state time sequence data set and performs data processing.

[0039] S1.1: Simulation data under different altitudes, different speeds, different operations and other flight conditions are obtained through trimming and a simulation model, multi-scale parameters to be predicted are selected, and are exported as.txt files, including 18 state parameters of north coordinate (npos), east coordinate (epos), altitude (h), roll angle (roll), pitch angle (pitch), yaw angle (yaw), roll angle velocity (p), pitch angle velocity (q), yaw angle velocity (r), airspeed (v), three kinds of overload (nx, ny, nz), dynamic pressure (Q), static pressure (ps), Mach number (mach), angle of attack (alpha), angle of sideslip (beta), and 4 control quantity parameters of engine thrust (T), elevator angle (E), aileron angle (A) and rudder angle (R), simulation time is 100 s, data sampling period is 10 Hz, and the number of flight conditions is 100, thereby generating 100 groups of aircraft operation state data, including aircraft pitch, roll, yaw and other operations.

[0040] S1.2: The data is standardized, the mean and variance are set according to the upper and lower limits of the data for parameters with uneven data distribution, and the mean and variance are calculated according to the time sequence data for data with uniform distribution, so as to scale the features to an overall distribution with a mean of 0 and a variance of 1, thereby improving the convergence speed and stability of the model.

[0041] The data is divided into a training set, a validation set and a test set, and the data in the training set is further divided into multiple batches, each batch being 20 groups of time sequence data, thereby reducing the calculation consumption during training and improving the generalization ability of the model.

[0042] ​S2: Calculate the multi-order difference value sequence corresponding to each parameter in the time series data set. The calculation formula of the multi-order difference value of each parameter is as follows:

[0043]

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

[0045] Each parameter has a corresponding multi-order difference value sequence, and the dth order time series data of the parameter x is represented as is the nth data of the dth order of the parameter x. Then, the connection relationship between the graph nodes is established according to the correlation between the time series data and the physical prior knowledge, and is denoted as an adjacency matrix. For parameters greater than the correlation threshold value 0.8, the value of the adjacency matrix is set to 1, indicating the connectivity between the parameters. For parameters less than the threshold value, the value is set to 0, and the input parameter adjacency matrix value is only set to a one-way value. The dimension of the adjacency matrix is (N, N). The physical prior knowledge determines the order size of the time series characteristics and the association relationship between the parameters. The multi-order difference value sequence corresponding to all parameters in each single moment is aggregated to form a graph node feature, with a dimension of [N, D], wherein N is the number of parameters and D is the order of the parameters, thereby obtaining a training data set;

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

[0047]

[0048] The edge_index adopts the form of a coordinate list to represent a sparse adjacency matrix, wherein the two numbers in each column represent the adjacency matrix value between the corresponding two parameters as 1. The graph node feature with a dimension of (N, D) and the edge_index are taken 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 data set, a multi-time series scale state prediction model is obtained.

[0050] 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 convolution network (TCN) based on a dilated causal convolution, wherein the time domain convolution network based on the dilated causal convolution includes a decoder.

[0051] The number of multi-attention mechanism heads, the number of layers, the aggregation mode and other parameters in the graph neural network model based on the multi-head attention mechanism are selected and represented by the following formula:

[0052]

[0053] where, a ij represents the influence weight between nodes, and the vector is used as aggregation, the weight matrix W is used for vector dimension change, LeakyReLU() is an activation function, and [||] represents that the features are aggregated in a splicing manner. Finally, the result is calculated by softmax to obtain the attention score between nodes. After multiplying the attention score with the features and splicing, the intermediate layer feature of the model is obtained.

[0054]

[0055] where, is the feature value of node i, is the feature value of node j, K is the number of attention heads, and j is a number of nodes connected to node i, is the attention score between nodes calculated in the kth attention head, and σ() is an activation function, W k is the weight matrix of the kth attention head, and || represents the multi-head attention mechanism. [||] represents that the feature values of multiple head parameters are aggregated in a splicing manner. The intermediate layer feature of the model is subjected to a graph aggregation operation again, and finally the output is the graph node feature value at a single time, which sufficiently learns the physical feature correlation between parameters and improves the accuracy of subsequent prediction.

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

[0057]

[0058] In S3, during training, the output single-time parameter feature combination is reconstructed into a multi-parameter feature time series with a dimension of (Batch, T, N), Batch is the batch size, T is the total length of the input time series, and N is the number of parameters. 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, thereby obtaining a feature set at different time scales. Then, the feature sets at different time scales are respectively input into the time-domain convolution network based on the dilated causal convolution, and the time-domain convolution network based on the dilated causal convolution corresponding to different time scales is respectively trained. After training, the graph-time domain fusion network and the time-domain convolution network based on the dilated causal convolution corresponding to all time scales are integrated to form a multi-time scale state prediction model.

[0059] During the training process, the prediction data of the target moment in the predicted time series output by the time domain convolution network based on the expansion causal convolution and the real data of the target moment are used to construct a self-defined loss function for loss evaluation and back propagation, so as to realize model training, verification and testing.

[0060] The size of the multi-size sliding window satisfies the exponential rule, specifically 2 0 ,2 1 ,…,2 k , wherein k is a window scale coefficient, so that the data sequence after {i, 2i, …, ni} can be predicted.

[0061] In the time series convolution model based on causal expansion convolution, appropriate convolution layer and expansion factor are set, and the time t and the parameter time series features before t are fused to avoid gradient explosion or gradient disappearance in the learning process. In this embodiment, the data dimension of the input time convolution network is (batch, 40, 22), the window size of the time series is 40, and the data after t = 2s is used for testing. The time convolution network is composed of two convolution layers and one aggregation layer, the number of hidden layers of the convolution layer is 450, and the number of aggregation layers is 100. The output parameter dimension is (Batch, K, 100), and then a decoder link is used to map the data to n flight state parameters, realizing accurate regression prediction of multi-scale and multi-parameter after time t = 2s.

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

[0063]

[0064] Wherein, R MSE is the final loss result, N is the number of parameters, weight is the weight value corresponding to each parameter, is the target moment true value of the mth parameter, y m is the target moment prediction value of the mth parameter, k1 is the reduction error weight, is the target moment original true value of the mth parameter, y’ m is the target moment reduction prediction value of the mth parameter.

[0065] In the loss function, the loss weight weight of each parameter is constructed according to the magnitude of the original data, so that the multi-parameter can achieve approximate numerical accuracy. And based on the loss function with the set weight, it is still difficult to realize the prediction of the same accuracy of multiple parameters, and it is difficult to fit the dynamic change of the parameter by fixing the weight, so the reduction error in the non-standard state is introduced as as the dynamic scale error in the learning process, so as to realize accurate regression prediction of multiple parameters.

[0066] S4: load the multi-time scale state prediction model into the electronic device and integrate it into the complex equipment. Therefore, the multi-time scale state prediction model can be used to perform adaptive time window state prediction according to the actual operation parameter data of the complex equipment, and the corresponding operation state prediction result can be obtained.

[0067] S4 is specifically:

[0068] The initial size of the time window is set, the actual time sequence is selected from the actual operation parameter data of the complex equipment by using the current time window, the actual time sequence is preprocessed and the corresponding graph node features are generated, and then the actual time sequence is input into the multi-time scale state prediction model to obtain the corresponding operation state prediction result. In the state prediction process, the prediction error of the model is calculated, and the size of the time window is adaptively adjusted according to the prediction error of the model, so as to adjust the actual time sequence, that is, to select the TCN of the corresponding time scale, and then to obtain a better operation state prediction result, thereby improving the overall generalization ability and adaptive ability.

[0069] The model evaluation is performed on the pre-divided test set in this embodiment, and the results are shown in Figs. Figure 3 and Figure 4 The prediction data of some significant flight state parameters after 2s is basically consistent with the target data, and the transformation of multiple different scale parameters is consistent with the change trend of the original simulation value. In summary, the flight state multi-parameter regression prediction model based on multi-order time sequence feature fusion of the present application realizes accurate regression prediction of multiple parameters in the flight process of the aircraft, has the characteristics of fast convergence speed, high prediction accuracy, lightweight model, etc. At the same time, the present application can also be migrated to other complex equipment operation state time prediction scenarios, monitor and assist equipment operation decision-making, and has practical engineering significance.

[0070] The above embodiments are used to explain and illustrate the present application, but not to limit the present application. Any modification and change made to the present application within the spirit and protection scope of the claims of the present application all fall within the protection scope of the present application.

Claims

1. A complex equipment operation state prediction method based on multi-order timing feature fusion, characterized in that, The method comprises the following steps: S1: obtaining a time series data set after pre-processing operation parameter data of complex equipment; S2: calculating a multi-order difference value sequence corresponding to each parameter in the time series data set, then establishing a connection relationship between graph nodes according to the correlation between each time series data and physical prior knowledge and recording it as an adjacency matrix, and aggregating the multi-order difference value sequence corresponding to all parameters in each single time to form a graph node feature, thereby obtaining a training data set; S3: constructing a graph-time domain fusion network, training the graph-time domain fusion network using the training data set, and obtaining a multi-time series scale state prediction model; S4: using the multi-time series scale state prediction model to perform adaptive time series window state prediction according to actual operation parameter data of the complex equipment, and obtaining a corresponding operation state prediction result. 2.The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 1, characterized in that, The pre-processing includes data cleaning and data standardization.

3. The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 1, characterized in that, In S3, the graph-time domain fusion network comprises a graph neural network model based on a multi-head attention mechanism and a time domain convolution network based on a dilated causal convolution, wherein the time domain convolution network based on the dilated causal convolution comprises a decoder.

4. The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 1, characterized in that, 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, thereby obtaining a feature set of different time series scales, and then the feature set of different time series scales is input into the time domain convolution network based on the dilated causal convolution, respectively, to train and obtain the time domain convolution network based on the dilated causal convolution corresponding to different time series scales; After training is completed, the graph-time domain fusion network and the time domain convolution network based on the dilated causal convolution corresponding to all time series scales are integrated to form a multi-time series scale state prediction model.

5. The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 1, characterized in that, In S3, the loss function of the graph-time domain fusion network is an improved mean square error loss, which satisfies the following formula: wherein R MSE is the final loss result, N is the number of parameters, weight is the weight value corresponding to each parameter, is the target time true value of the mth parameter, y m is the target time predicted value of the mth parameter, k1 is the reduction error weight, is the original true value of the mth parameter at the target time, y , m is the reduction predicted value of the mth parameter at the target time.

6. The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 4, characterized in that, The size of the multi-size sliding window satisfies an exponential rule, specifically 2 0 ,2 1 ,…,2 k , where k is a window scale factor.

7. The complex equipment running state prediction method based on multi-order timing feature fusion according to claim 1, characterized in that, S4 is specifically: An initial size of a time series window is set, an actual time series is selected from the actual operation parameter data of the complex equipment using the current time series window, the actual time series is pre-processed and the corresponding graph node feature is generated, and then the graph node feature is input into the multi-time series scale state prediction model to obtain a corresponding operation state prediction result; during the state prediction process, the prediction error of the model is calculated, the size of the time series window is adaptively adjusted according to the prediction error of the model, the actual time series is adjusted, and a more optimal operation state prediction result is obtained.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the complex equipment operation state prediction method based on multi-order time series feature fusion in any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the complex equipment operation state prediction method based on multi-order time series feature fusion in any one of claims 1 to 7.

10. A computer program product comprising computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the complex equipment operation state prediction method based on multi-order time series feature fusion in any one of claims 1 to 7.

Images