Signal prediction method and system based on deep learning and multiple signal decomposition

Abstract

The invention provides a signal prediction method and system based on deep learning and multiple signal decomposition, and relates to the technical field of hydroelectric generating set fault diagnosis. Comprising the following steps: acquiring time sequence data of a plurality of different signal channels and preprocessing the time sequence data, decomposing the preprocessed time sequence signal data by adopting EEMD (Ensemble Empirical Mode Decomposition) which can better represent different frequency components of the signal, performing VMD (Variational Modal Decomposition) on a first-order IMF component, further decomposing the first-order IMF component into a plurality of modal components, extracting more detailed modal information, and obtaining more accurate modal information. The method comprises the following steps: firstly, decomposing modal components in a time sequence to improve the prediction precision, then inputting the decomposed modal components into a built LSTM prediction model to identify modes and trends in the time sequence so as to obtain a prediction result, carrying out multi-level signal decomposition by combining EEMD and VMD so as to extract finer features, and then carrying out time sequence prediction through the LSTM model, so that the method has relatively high prediction precision and robustness; and the accuracy of the prediction function is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower unit fault diagnosis, and in particular to a signal prediction method and system based on deep learning and multiple signal decomposition. Background Art

[0002] As the years of operation increase, the performance of hydropower units gradually deteriorates, and the potential factors of failure increase. In order to predict unit failures in advance and avoid unnecessary losses caused by accidental shutdowns, it is necessary to predict the trend of unit changes. At present, trend prediction can be divided into two categories: prediction methods based on physical models and prediction methods based on data-driven: First, the actual operating environment of rotating machinery such as hydropower units is complex and changeable. The physical prediction model constructed from the perspective of operating mechanism considers many factors and relies heavily on expert experience. It has strong uncertainty and its research and application are limited; Second, signal timing prediction from a data-driven perspective is the mainstream method of current hydropower unit trend prediction. Due to the strong nonlinear and non-stationary characteristics of hydropower unit monitoring signals, especially vibration signals, the high-frequency components present in them pose a challenge to the accurate prediction of signals. In order to further improve the prediction accuracy, the "decomposition and integration" theory in signal processing has been gradually applied to trend prediction, which decomposes the original signal into a series of stationary components to enhance the predictability of information.

[0003] In the prior art, deep learning has been used for research, for example: 1. The Baidu trend prediction method for hydropower unit bearings using the regression sliding average model can realize the medium- and short-term prediction of the swing development trend of the unit bearings; 2. The degradation trend prediction method for hydropower units based on the long short-term memory network, including collecting the parameters of the hydropower unit and normalizing them, constructing a matrix using the normalized parameters; establishing a fully connected network model containing the matrix; training and testing the fully connected network model, and establishing a long short-term memory network degradation trend prediction model for each sub-degradation time series; training and testing the long short-term memory network degradation trend prediction model, using the long short-term memory network degradation trend prediction model to predict the degradation trend of the hydropower unit; The long-term memory network predicts the degradation trend of the hydropower unit so as to provide timely warning of abnormal conditions and improve the operation and maintenance accuracy of the hydropower unit; third, the degradation degree prediction method of the hydropower unit based on EEMD and LSTM uses the ensemble empirical mode decomposition to decompose the degradation degree time series into multiple components; each component is reconstructed according to the preset time step to obtain multiple sequence samples; a long short-term memory network degradation trend prediction model is constructed for each sequence sample to obtain the degradation trend prediction component corresponding to each component; all degradation trend prediction components are superimposed to obtain the degradation trend of the hydropower unit to improve the prediction accuracy of the degradation degree of the hydropower unit.

[0004] However, current researches mainly predict the vibration signals of the unit directly. In practice, the vibration signals of the unit are fast variables, and the non-stationary and nonlinear components they contain greatly affect the predictability of the signals, resulting in large signal prediction errors. Moreover, when applying the decomposition and integration theory for signal prediction, the first-order components after decomposition still have more high-frequency components of the original signal, resulting in inaccurate prediction results. Summary of the invention

[0005] The main purpose of the present invention is to provide a signal prediction method based on deep learning and multiple signal decomposition to solve the technical problem that the prediction results are not accurate enough due to the presence of non-stationary and nonlinear components in the current unit vibration signal and the high-frequency components of the original signal after decomposition.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a signal prediction method based on deep learning and multiple signal decomposition, comprising the following steps:

[0007] S1: Acquire the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing;

[0008] S2: Use EEMD to decompose each time series signal data after data preprocessing to obtain multiple IMF components;

[0009] S3: VMD is used to perform secondary modal decomposition on the first-order IMF component to obtain the corresponding multiple modal components;

[0010] S4: Build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and input the integrated modal components into the LSTM prediction model;

[0011] S5: Obtain the prediction result according to the output of the LSTM prediction model.

[0012] In a preferred embodiment, S2 includes:

[0013] S21: Add Gaussian white noise n(t) to the original sign indicator sequence V(t) to obtain a new indicator signal s(t);

[0014] S22: Perform EMD decomposition on the synthetic signal to obtain IMF components and residuals of various orders;

[0015] S23: repeat steps S21 to S22, adding different white noise each time;

[0016] S24: Calculate the mean of all IMFs and residuals of each order as the final IMF components and residuals. The formula is:

[0017]

[0018] In the formula, c i,j (t) is the jth IMF component, n is the total number, i is the component order, and t is time.

[0019] In a preferred embodiment, S3 includes:

[0020] S31: Use Hilbert transform to obtain the modal component u k (t) The corresponding analytical signal z k (t) and the single-sided spectrum, and then z k (t) and the estimated center frequency ω k Mix, adjust the modal component frequency band to the baseband, calculate the gradient square L2 norm to estimate the bandwidth, and construct a constrained variational problem. The formula is:

[0021]

[0022] In the formula, is the partial derivative function, the analytical signal z k (t)=[δ(t)+j / πt]*u k (t), k = 1, 2, ..., K, δ(t) is the unit pulse function;

[0023] S32: Introduce the quadratic penalty term α and the Lagrangian multiplier λ to solve the above constrained variational problem, obtain the augmented Lagrangian function, and obtain the modal component. The formula is:

[0024]

[0025] Where f(t) is the original signal.

[0026] In a preferred embodiment, the S32 includes:

[0027] A1: Initialize u in the augmented Lagrangian function formula k ,ω k and λ, and set the number of iterations n;

[0028] A2: Iterative update u k , according to ω k And λ to find the "saddle point" of the augmented Lagrangian function, the formula is:

[0029]

[0030] A3: Repeat step A2 until the conditions are met

[0031] A4: Obtain K modal components decomposed from the signal sequence.

[0032] In the preferred embodiment, the basic structure of the LSTM prediction model is divided into three layers: input layer, hidden layer, and output layer. The hidden layer controls information transmission by setting a threshold unit or a gate structure.

[0033] The hidden layer mainly includes three gate structures: forget gate, input gate and output gate;

[0034] Among them, the forget gate selects and retains the neuron c at time t-1 t-1 Historical state information, the input gate determines the new input information x of the neuron at time t t The storage of the hidden layer, the output gate controls the output value h of the hidden layer t information output.

[0035] In the preferred solution, the weight threshold is iteratively updated using a back-propagation algorithm and the chain rule over time.

[0036] In the preferred embodiment, in S2, EEMD is used to perform signal decomposition on each component sequence data to obtain IMF modal components, and the correlation between the IMF modal components and the target sequence data is calculated to obtain the IMF modal components whose correlation coefficients are greater than a preset correlation threshold value;

[0037] The correlation coefficient is used as the evaluation index to comprehensively evaluate the signal prediction error, where the root mean square error RMSE, mean absolute error MAE and correlation coefficient R are expressed as:

[0038]

[0039] In the formula, n is the total number, y(t) is the model prediction value, and y′(t) is the actual value. and is the corresponding mean.

[0040] In the preferred solution, data preprocessing includes denoising, arranging the original time series data in order according to a unified time standard, then filling the missing values ​​with the average value, and finally normalizing the data.

[0041] A signal prediction system based on deep learning and multiple signal decomposition, comprising:

[0042] A data acquisition module is used to obtain the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing;

[0043] An EEMD decomposition module is used to decompose each time series signal data after data preprocessing by using EEMD to obtain multiple IMF components;

[0044] A VMD decomposition module is used to perform secondary modal decomposition on the first-order IMF component using VMD to obtain corresponding multiple modal components;

[0045] The prediction model module is used to build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and then input the integrated modal components into the LSTM prediction model;

[0046] The result analysis module is used to obtain the prediction results based on the output of the LSTM prediction model.

[0047] The present invention provides a signal prediction method based on deep learning and multiple signal decomposition, including obtaining time series data of multiple different signal channels and preprocessing them, using EEMD to decompose the preprocessed time series signal data, the decomposed components can better represent the different frequency components of the signal, and the first-order IMF component is subjected to variational mode decomposition (VMD), which is further decomposed into multiple modal components to extract more detailed modal information and improve the prediction accuracy, and then the decomposed modal components are input into the constructed LSTM prediction model to identify the patterns and trends in the time series, so as to obtain the prediction results, and the multi-level signal decomposition is performed by combining EEMD and VMD to extract more refined features, and then the time series prediction is performed through the LSTM model, which has high prediction accuracy and robustness, and greatly improves the accuracy of the prediction function. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0049] Figure 1 is a flow chart of the method of the present invention;

[0050] Figure 2 This is the LSTM neuron structure diagram of the present invention;

[0051] Figure 3 It is a flow chart of the health status prediction model of the hydropower unit of the present invention;

[0052] Figure 4 It is a data division diagram of the peak-to-peak value timing signal of the upper guide swing X direction of the present invention;

[0053] Figure 5 This is the upper guide swing prediction result diagram based on EEMD-LSTM of the present invention;

[0054] Figure 6 It is the prediction error diagram of the upper guide swing EEMD component of the present invention;

[0055] Figure 7 It is the decomposition diagram of the upper guide swing IMF1 and VMD components of the present invention;

[0056] Figure 8This is a comparison chart of the prediction effects of EEMD-VMD-LSTM and EEMD-LSTM of the present invention;

[0057] Fig. 9 It is a diagram of the prediction results after each model of the present invention is trained once. DETAILED DESCRIPTION

[0058] Example 1

[0059] Currently, signal decomposition has been used in the fault prediction of hydropower units to predict the degradation degree of hydropower units, including the following two theories.

[0060] 1. Signal Decomposition Theory

[0061] 1. Empirical Mode Decomposition (EMD) is an effective method that can adaptively analyze nonlinear and non-stationary signals.

[0062] In the process of the unit evolving from normal to fault, the distribution law of the relevant signal changes with time, showing strong non-stationarity. At this time, the index sequence can be analyzed by EMD. The basic idea of ​​EMD is to perform adaptive smoothing on the original signal, and obtain a series of intrinsic mode components (IMFs) by decomposing them step by step. Each intrinsic mode component satisfies: ① The number of extreme points of the intrinsic mode component is equal to the number of zero-crossing points or the relative difference is 1; ② At any data point, the mean of the upper and lower envelopes composed of local maximum and local minimum is 0.

[0063] According to the above definition of IMF, a monitoring signal can be decomposed through the following steps:

[0064] (1) Determine all extreme points, use cubic spline interpolation to connect the maximum and minimum points respectively, and fit the upper and lower envelopes.

[0065] (2) Obtain the mean of the upper and lower envelopes, denoted as, and subtract the mean from the original indicator sequence to obtain the new indicator sequence h 1 (t):

[0066] h 1 (t) = V (t) - m (1)

[0067] (3) Determine whether the new sequence meets the conditions of the intrinsic modal component. If so, then h 1 is the first-order modal component decomposed. Otherwise, for h 1 Continue to repeat steps (1) to (3) to decompose until the conditions of the natural modal component are met. The first-order modal component is recorded as c 1 (t).

[0068] (4) c1 (t) is separated from the original sequence and the residual is obtained:

[0069] r 1 (t) = V(t) - c 1 (t) (2) r 1 (t) Repeat steps (1) to (3) as a new original sequence until the next modal component c is obtained. 2 (t).

[0070] (5) Repeat the above steps. When the remaining sequence is a monotonic sequence, stop the decomposition. The overall process can be expressed as:

[0071]

[0072] Finally, we get 1 (t)~c n (t) a series of IMF components, residual r n The monotonicity of (t) reflects the overall change trend of the original indicator sequence. In summary, the original indicator sequence can be expressed as:

[0073]

[0074] 2. Ensemble Empirical Mode Decomposition (EEMD) improves on the traditional EMD decomposition by adding Gaussian white noise to the original data multiple times, effectively avoiding the modal aliasing phenomenon that exists when using EMD for signal decomposition.

[0075] 2. Deep Learning Theory

[0076] The Long Short Term Memory Neural Network (LSTM) is developed based on the Recurrent Neural Network (RNN). It combines short-term memory with long-term memory through a special "gate" structure, making the network output highly correlated with current and historical inputs. It solves the problem that traditional recurrent neural networks only have short-term memory, and can effectively utilize time series historical information and process long-term correlations of time series.

[0077] 3. Hydropower unit signal prediction integrating deep learning and multiple signal decomposition

[0078] Currently, EEMD combined with deep learning models are used for prediction. EEMD decomposes the unit monitoring signal into a series of IMFs. Among the obtained IMFs, IMF1 is the main cause of the indicator prediction error because it contains the high-frequency components of the signal. Therefore, in order to improve the overall prediction accuracy of the signal, it is necessary to further decompose IMF1 to obtain a more stable signal component, thereby reducing the IMF1 prediction error. However, due to the limitations of EEMD decomposition conditions, it is difficult to effectively obtain a more stable subsequence by continuing to perform EEMD on the IMF components.

[0079] Based on the above research, in order to improve the deficiencies in the prior art, the technical solution of this embodiment is proposed.

[0080] like Figure 1-9 As shown, a signal prediction method based on deep learning and multiple signal decomposition includes the following steps:

[0081] S1: Acquire the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing.

[0082] S2: Use EEMD to decompose each time series signal data after data preprocessing to obtain multiple IMF components.

[0083] S3: VMD is used to perform secondary modal decomposition on the first-order IMF component to obtain the corresponding multiple modal components.

[0084] S4: Build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and then input the integrated modal components into the LSTM prediction model.

[0085] S5: Obtain the prediction result according to the output of the LSTM prediction model.

[0086] This embodiment collects time series data of multiple different signal channels and performs preprocessing, and uses EEMD to decompose the preprocessed time series signal data. These components can better represent the different frequency components of the signal. The first-order IMF component is subjected to variational mode decomposition (VMD) to further decompose it into multiple modal components, extract more detailed modal information, and improve the prediction accuracy. The decomposed modal components are then input into the constructed LSTM prediction model to identify patterns and trends in the time series, thereby obtaining prediction results. EEMD and VMD are combined for multi-level signal decomposition to extract more refined features, and then the LSTM model is used for time series prediction, which has high prediction accuracy and robustness.

[0087] In the preferred solution, the specific step S2 of EEMD decomposing the fault symptom indicator sequence includes:

[0088] S21: Add Gaussian white noise n(t) to the original sign indicator sequence V(t) to obtain a new indicator signal s(t).

[0089] S22: Perform EMD decomposition on the synthesized signal to obtain IMF components and residuals of various orders.

[0090] S23: Repeat steps S21 to S22, adding different white noise each time.

[0091] S24: Calculate the mean of all IMFs and residuals of each order as the final IMF components and residuals. The formula is:

[0092]

[0093] In the formula, c i,j (t) is the jth IMF component, n is the total number, i is the component order, and t is time.

[0094] Through the EEMD algorithm, the signal is decomposed into a series of IMF components at different time scales, and compared with the original signal, the fluctuation of each IMF component is smoother. Theoretically, trend prediction can obtain more accurate prediction results.

[0095] Variational Mode Decomposition (VMD) is based on Wiener filtering, Hilbert transform and frequency mixing theory, and can automatically decompose the input signal V(t) into K modal components with sparse characteristics. Assuming that each modal component has a finite bandwidth with a corresponding center frequency, VMD constructs the modal decomposition problem as a process of solving the optimal solution of the constrained variational problem by minimizing the sum of the estimated bandwidths of the modal components.

[0096] In the preferred embodiment, step S3 comprises:

[0097] S31: Use Hilbert transform to obtain the modal component u k (t) The corresponding analytical signal z k (t) and the single-sided spectrum, and then z k (t) and the estimated center frequency ω k Mix, adjust the modal component frequency band to the baseband, calculate the gradient square L2 norm to estimate the bandwidth, and construct a constrained variational problem. The formula is:

[0098]

[0099] In the formula, is the partial derivative function, the analytical signal z k (t)=[δ(t)+j / πt]*u k(t), k = 1, 2, ..., K, δ(t) is the unit pulse function.

[0100] S32: Introduce the quadratic penalty term α and the Lagrangian multiplier λ to solve the above constrained variational problem, obtain the augmented Lagrangian function, and obtain the modal component. The formula is:

[0101]

[0102] Where f(t) is the original signal.

[0103] In the preferred embodiment, step S32 includes:

[0104] A1: Initialize u in formula (7) k ,ω k and λ, and set the number of iterations n, and use the multiplier alternating direction method to solve the constrained variational problem;

[0105] A2: Iterative update u k , according to ω k And λ to find the "saddle point" of the augmented Lagrangian function. If the result of iteration to the n+1th time is:

[0106]

[0107] A3: Repeat step A2 and iterate equations (8) to (10) until the condition is met.

[0108] A4: At this point, we get K modal components decomposed from the signal sequence.

[0109] This embodiment adopts LSTM neural network.

[0110] In the preferred embodiment, Figure 2 As shown in the figure, the basic structure of the LSTM prediction model is divided into three layers: input layer, hidden layer, and output layer. The hidden layer controls information transmission by setting threshold units or gate structures.

[0111] The hidden layer mainly contains three gate structures: forget gate, input gate and output gate;

[0112] Among them, the forget gate selects and retains the neuron c at time t-1 t-1 Historical state information, the input gate determines the new input information x of the neuron at time t t The storage of the hidden layer, the output gate controls the output value h of the hidden layer t information output.

[0113] like Figure 2 As shown, the forward propagation algorithm of the LSTM network can be obtained as shown in the following formulas (11)-(17).

[0114] f t =sig(W f ·[h t-1 ,x t ]+b f ) (11)

[0115] i t =sig(W i ·[h t-1 ,x t ]+b i ) (12)

[0116]

[0117] o t =sig(W o ·[h t-1 ,x t ]+b o ) (11)

[0118] h t =o t ·tanh(c t ) (16)

[0119] y t =sig(W y ·h t +b y ) (17)

[0120] Among them, f t represents the forget gate, formula i t and represents the input gate, c t is the neuron state update expression, o t and h t is the output gate, y t Calculate the output of the neuron at time t. t is the network input information at time t, h t-1 is the output value of the hidden layer state of the network at time t-1, is the candidate state value of the input gate, c t-1 、c t is the neuron state at different times, sig is the sigmoid activation function, W f , W i , W c , W o , W y is the corresponding weight matrix, b f 、b i 、b c 、b o 、b y is the corresponding threshold vector, yt is the network prediction output at time t.

[0121] In the preferred solution, the weight threshold is iteratively updated using a back-propagation algorithm and the chain rule over time.

[0122] After completing the LSTM forward propagation, the back propagation process begins, which is to expand the LSTM into a deep network in chronological order, and use the back-propagation algorithm (BPTT) and the chain rule to iteratively update the weight threshold until the optimal solution is obtained.

[0123] In the preferred solution, firstly, data preprocessing includes denoising, the original time series data is arranged in order according to a unified time standard, then the missing values ​​are filled with the average value, and finally the data is normalized.

[0124] This embodiment takes advantage of EEMD's advantages in decomposing non-stationary signals, uses EEMD as the main decomposition method for time series signals, and combines the powerful time series prediction capabilities of the LSTM network to construct a hydropower unit health status prediction model process. Figure 3 The following are the steps:

[0125] (1) Multiple signal decomposition

[0126] The unit monitoring signal is decomposed into a series of IMFs using EEMD. Among the obtained IMFs, IMF1 is the main cause of the indicator prediction error because it contains the high-frequency components of the signal. Therefore, in order to improve the overall prediction accuracy of the signal, it is necessary to further decompose IMF1 to obtain a more stable signal component, thereby reducing the IMF1 prediction error. However, due to the limitations of the EEMD decomposition conditions, it is difficult to effectively obtain a more stable subsequence by continuing to perform EEMD on the IMF components. Variational mode decomposition (VMD), as a new complex signal decomposition method, can decompose the signal into stable components with different central frequency components without changing the signal time length, and has attracted much attention in time series prediction. Therefore, VMD is used to perform a secondary decomposition of IMF1 to obtain the corresponding modal components.

[0127] (2) Establishment of signal prediction model

[0128] A signal prediction model based on LSTM is established. Through the signal time series autocorrelation analysis, the LSTM input time step l is determined, so as to construct the input and output matrix of each modal component for network training. The matrix construction form is:

[0129]

[0130] Where Xinput is the input matrix, x(k), k∈N * is the time series value of the signal modal component, m is the number of training samples, X output is the corresponding output matrix.

[0131] (3) Monitoring signal prediction

[0132] Use the trained LSTM network to predict the amplitude of each modal component of the signal at the next moment, superimpose the prediction results, and finally obtain the signal prediction value.

[0133] (4) Example Analysis

[0134] The prediction analysis of the peak-to-peak value monitoring signal of the upper guide swing of the unit in a certain period of time is carried out, and the training set and the test set are divided as follows: Figure 4 shown.

[0135] Perform EEMD operation on the monitoring signal. In terms of the selection of the number of sets and the decomposition noise intensity, as the number of sets increases, the influence of the Gaussian white noise added in the EEMD process on the decomposition effect gradually decreases and tends to be stable, and different studies have shown that the influence of noise intensity on the result error is also relatively slight. Therefore, when using EEMD to process the signal timing sequence, the number of sets is set to 200 and the auxiliary noise intensity is set to 0.4.

[0136] The EEMD decomposition method is combined with LSTM, referred to as EEMD-LSTM, and EEMD-LSTM is used to predict the X-direction peak-to-peak value of the upper guide swing of the unit within a period of time.

[0137] In the preferred solution, the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) are used as evaluation indicators to comprehensively evaluate the signal prediction error.

[0138] The correlation coefficient is used as the evaluation index to comprehensively evaluate the signal prediction error, where the root mean square error RMSE, mean absolute error MAE and correlation coefficient R are expressed as:

[0139]

[0140] Where n is the total number, y(t) is the model predicted value, y′(t) is the actual value, and y(t) and y′(t) are the corresponding means.

[0141] Finally, the prediction result of the upper guide swing after this time period is obtained as follows Figure 5As shown in the figure, it can be seen that the prediction results of the upper guide swing using the EEMD-LSTM model are basically consistent with the actual value, and the signal development trend can be predicted more accurately, and a satisfactory effect has been initially achieved. However, the deviation from the actual value is large in the rising stage of the signal, and the prediction accuracy needs to be further improved.

[0142] Further analysis of the prediction errors of each component is carried out, such as Figure 6 As shown in the figure, among the prediction results of all 8 IMF components, the IMF1 component has the largest prediction error and is the main source of the deviation between the signal prediction value and the actual value. In addition, the MF1 fluctuation frequency is relatively high, representing the high-frequency component of the signal sequence. Therefore, IMF1 can be further decomposed into several components with lower complexity and higher stability. By predicting the components and superimposing the predicted values, the prediction error of IMF1 can be reduced, and the overall prediction accuracy of the upper guide swing can be finally improved.

[0143] like Figure 6 As shown in , in the decomposition of IMF1, since IMF1 already satisfies the intrinsic modal component properties in EEMD, that is, the mean of the envelope composed of local maxima and local minima is 0. Therefore, according to the EEMD principle, it is difficult to effectively obtain a more stable subsequence by continuing to perform EEMD operations on the IMF components. Figure 7 As shown in the figure, variational mode decomposition (VMD) is a new complex signal decomposition method that can decompose the signal into stationary components with different central frequency components without changing the signal time length. Therefore, VMD is used to further decompose IMF1 to obtain the corresponding modal components.

[0144] In terms of the number of decomposition layers, similar to EEMD, when the number of decomposed modes is large, it is easy to generate repeated modes and increase the complexity of model training, resulting in unnecessary computational overhead. Fewer modes are prone to cause under-decomposition of the signal. According to previous studies, in terms of signal processing involving hydropower units, the number of VMD decomposition modes selected is usually 3 to 5. For the upper guide swing signal of the unit in the case, the number of modes is uniformly determined to be 4. The index prediction results after EEMD and VMD decomposition are compared with Figure 6 The prediction results in are compared with Figure 8 Compared with Table 1, it can be seen that the change trend of the signal prediction value after multiple decomposition is more consistent with the actual value, and the error indicators RMSE and MAE are reduced by 28.94% and 23.31% respectively, achieving higher prediction accuracy.

[0145] Table 1 Evaluation of prediction effects of EEMD-VMD-LSTM and EEMD-LSTM

[0146]

[0147] In order to further verify the advantages of the prediction model combining multiple signal decomposition and LSTM (hereinafter referred to as EEMD-VMD-LSTM) in predicting monitoring signals of hydropower units, this embodiment uses the upper guide swing X-direction original sequence-LSTM (control group 1), EMD-LSTM (control group 2), EEMD-ANN (control group 3) and the EEMD-LSTM (control group 4) prediction model in the previous section as comparative experiments. In the control group experiment, the division of the training set and the test set as well as the various parameters of the LSTM network are consistent with those of the experimental group. When using EEMD-ANN for upper guide swing prediction, the ANN network structure is a three-layer structure of input layer-hidden layer-output layer. The parameters are optimized through genetic algorithm, and each model is trained separately 20 times. Finally, the prediction results of each model after a certain training are obtained as follows Fig. 9 As shown. It can be seen that among the several prediction models compared, the prediction results of the constructed EEMD-VMD-LSTM model are more in line with the actual change trend of the upper guide swing. When LSTM is used to directly predict the original signal sequence, the non-stationarity of the symptom indicators in the process of normal to fault evolution causes the prediction results to deviate seriously from the actual changes of the indicators. At the same time, the prediction effect of EEMD-LSTM on the upper guide swing is significantly better than that of EMD-LSTM and EEMD-ANN, which respectively shows the advantages of EEMD over EMD in signal stationarity decomposition and LSTM over traditional neural networks in time series self-learning.

[0148] In general, the overall prediction performance of the proposed EEMD-LSTM model is better than that of the other comparative models.

[0149] Table 2 is an evaluation of the prediction effect of each model. By comparison, it can be seen that directly predicting the original sequence has the worst effect. The EEMD-VMD-LSTM model in this embodiment has the smallest prediction error, and the correlation between the predicted value and the actual value is the highest, and the prediction effect is the best.

[0150] The prediction error of EEMD-LSTM is reduced by 32.1% (RMSE) and 35.69% (MAE) compared with EMD-LSTM, and the error is reduced by 40.19% (RMSE) and 43.94% (MAE) compared with EEMD-ANN.

[0151] Table 2 Evaluation of prediction effects of different models

[0152]

[0153] Example 2

[0154] In combination with Example 1, a signal prediction system based on deep learning and multiple signal decomposition is further described, characterized in that it includes:

[0155] The data acquisition module is used to obtain the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing.

[0156] The EEMD decomposition module is used to decompose each time series signal data after data preprocessing by using EEMD to obtain multiple IMF components.

[0157] The VMD decomposition module is used to perform secondary modal decomposition on the first-order IMF component by using VMD to obtain corresponding multiple modal components.

[0158] The prediction model module is used to build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and then input the integrated modal components into the LSTM prediction model.

[0159] The result analysis module is used to obtain the prediction results based on the output of the LSTM prediction model.

[0160] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A signal prediction method based on deep learning and multiple signal decomposition, characterized in that: The following steps are involved: S1: Acquire the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing; S2: Use EEMD to decompose each time series signal data after data preprocessing to obtain multiple IMF components; S3: VMD is used to perform secondary modal decomposition on the first-order IMF component to obtain the corresponding multiple modal components; S4: Build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and input the integrated modal components into the LSTM prediction model; S5: Obtain the prediction result according to the output of the LSTM prediction model.

2. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: The S2 includes: S21: Add Gaussian white noise n(t) to the original sign indicator sequence V(t) to obtain a new indicator signal s(t); S22: Perform EMD decomposition on the synthetic signal to obtain IMF components and residuals of various orders; S23: repeat steps S21 to S22, adding different white noise each time; S24: Calculate the mean of all IMFs and residuals of each order as the final IMF components and residuals. The formula is: In the formula, c i,j (t) is the jth IMF component, n is the total number, i is the component order, and t is time.

3. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: The S3 includes: S31: Use Hilbert transform to obtain the modal component u k (t) The corresponding analytical signal z k (t) and the single-sided spectrum, and then z k (t) and the estimated center frequency ω k Mix, adjust the modal component frequency band to the baseband, calculate the gradient square L2 norm to estimate the bandwidth, and construct a constrained variational problem. The formula is: In the formula, is the partial derivative function, the analytical signal z k (t)=[δ(t)+j / πt]*u k (t), k = 1, 2, ..., K, δ(t) is the unit pulse function; S32: Introduce the quadratic penalty term α and the Lagrangian multiplier λ to solve the above constrained variational problem, obtain the augmented Lagrangian function, and obtain the modal component. The formula is: Where f(t) is the original signal.

4. The signal prediction method based on deep learning and multiple signal decomposition according to claim 3, characterized in that: The S32 includes: A1: Initialize u in the augmented Lagrangian function formula k ,ω k and λ, and set the number of iterations n; A2: Iterative update u k , according to ω k And λ to find the "saddle point" of the augmented Lagrangian function, the formula is: A3: Repeat step A2 until the conditions are met A4: Obtain K modal components decomposed from the signal sequence.

5. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: The basic structure of the LSTM prediction model is divided into three layers: input layer, hidden layer, and output layer. The hidden layer controls information transmission by setting threshold units or gate structures; The hidden layer mainly includes three gate structures: forget gate, input gate and output gate; Among them, the forget gate selects and retains the neuron c at time t-1 t-1 Historical state information, the input gate determines the new input information x of the neuron at time t t The storage of the hidden layer, the output gate controls the output value h of the hidden layer t information output.

6. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: The weight threshold is iteratively updated using the back-propagation algorithm over time and the chain rule.

7. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: In S2, EEMD is used to perform signal decomposition on each component sequence data to obtain IMF modal components, and the correlation between the IMF modal components and the target sequence data is calculated to obtain the IMF modal components whose correlation coefficients are greater than a preset correlation threshold; The correlation coefficient is used as the evaluation index to comprehensively evaluate the signal prediction error, where the root mean square error RMSE, mean absolute error MAE and correlation coefficient R are expressed as: In the formula, n is the total number, y(t) is the model prediction value, and y′(t) is the actual value. and is the corresponding mean.

8. The signal prediction method based on deep learning and multiple signal decomposition according to claim 1, characterized in that: Data preprocessing includes denoising, arranging the original time series data in order according to a unified time standard, then filling the missing values ​​with the average value, and finally normalizing the data.

9. A signal prediction system based on deep learning and multiple signal decomposition, characterized in that: include: A data acquisition module is used to obtain the time series signal data of multiple different signal channels of the hydropower unit and perform data preprocessing; An EEMD decomposition module is used to decompose each time series signal data after data preprocessing by using EEMD to obtain multiple IMF components; A VMD decomposition module is used to perform secondary modal decomposition on the first-order IMF component using VMD to obtain corresponding multiple modal components; The prediction model module is used to build and train the LSTM prediction model, obtain and integrate the modal components obtained in step S2 and step S3, and then input the integrated modal components into the LSTM prediction model; The result analysis module is used to obtain the prediction results based on the output of the LSTM prediction model.

Images