Hydroelectric generating set mechanical vibration trend prediction method, device, equipment and medium

Abstract

The invention belongs to the technical field of hydroelectric generating set vibration prediction, and particularly relates to a hydroelectric generating set mechanical vibration trend prediction method and device, equipment and a medium, and the method comprises the steps: processing a vibration signal of a hydroelectric generating set through employing a variational mode decomposition algorithm; training the deep belief network by using the training set to obtain a trained deep belief network model; using a sparrow search algorithm to optimize parameters of the deep belief network model to obtain a vibration trend prediction model; and collecting a vibration signal of the hydroelectric generating set, processing the vibration signal by using a variational mode decomposition algorithm, inputting the processed vibration signal into the vibration trend prediction model to obtain a vibration prediction value, and taking the obtained vibration prediction value as an input of next prediction to realize multi-step prediction of the vibration trend of the hydroelectric generating set. By means of multi-step prediction, operation and maintenance personnel can know the vibration trend of the hydroelectric generating set in a period of time in the future in advance, potential fault hidden dangers can be found in time, and the operation safety and stability of the hydroelectric generating set are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of hydroelectric unit vibration prediction, and in particular relates to a method, device, equipment and medium for predicting the mechanical vibration trend of a hydroelectric unit. Background Art

[0002] As the core component of a hydropower station, the normal operation of a hydropower unit is of great significance to the safety and stability of the power grid. However, the hydropower unit will generate vibrations during operation. This continuous vibration will reduce the operational stability of the equipment, thereby increasing the possibility of equipment failure. As the types of unit failures are extremely complex and the fault information is difficult to capture, the traditional maintenance methods of hydropower units and subsequent fault diagnosis can no longer meet the requirements for safe operation of hydropower units.

[0003] The vibration signal of a hydropower unit can reflect the fault information of the unit. However, the nonlinearity and non-stationarity of the vibration signal make it difficult to accurately predict the vibration signal. Traditional trend prediction methods usually consider monitoring signals of a single dimension, which easily ignores the mapping relationship between the operating status of the equipment and the monitoring signal. Therefore, traditional prediction methods can no longer effectively predict the vibration trend changes of hydropower units.

[0004] At present, the prediction of vibration trend of hydropower units is mostly carried out by selecting multi-dimensional monitoring signals associated with the equipment status for analysis, extracting effective and accurate performance indicators, and then building a prediction model to obtain the vibration trend change law. These prediction models of unit vibration trend generally have the problem of difficult setting of hyperparameters, and the existing prediction models only consider single-step prediction and cannot predict the vibration trend changes of hydropower units in the future for a long time. Summary of the invention

[0005] In view of the problems that the existing vibration trend prediction model has the difficulty in setting hyperparameters and cannot predict the vibration trend changes of the hydropower unit for a long time in the future, the present invention provides a method, device, equipment and medium for predicting the mechanical vibration trend of a hydropower unit.

[0006] In a first aspect, the technical solution of the present invention provides a method for predicting the mechanical vibration trend of a hydropower unit, comprising the following steps: The vibration signal of the hydropower unit is processed using the variational mode decomposition algorithm and the processed data is divided into a training set and a validation set; The deep belief network is trained with the training set, and the trained deep belief network is verified with the verification set to obtain a trained deep belief network model; The sparrow search algorithm is used to optimize the parameters of the deep belief network model to obtain the vibration trend prediction model; The vibration signal of the hydropower unit is collected and processed using the variational mode decomposition algorithm, and then input into the vibration trend prediction model to obtain the vibration prediction value. The obtained vibration prediction value is used as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

[0007] As a further limitation of the technical solution of the present invention, the steps of processing the vibration signal of the hydropower unit using the variational mode decomposition algorithm and dividing the processed data into a training set and a validation set include: Collect the mechanical vibration signal of the hydropower unit, and perform normalization after denoising the collected original vibration signal; Determine the modal parameters and perform VMD decomposition on the normalized vibration signal according to the determined parameters; Perform feature extraction on each modal component obtained by decomposition; The features extracted from all modal components are combined into a feature vector to generate a data set; The dataset is divided into a training set and a validation set according to the set ratio.

[0008] As a further limitation of the technical solution of the present invention, in the step of determining the modal parameters and performing VMD decomposition on the normalized vibration signal according to the determined parameters, the VMD decomposition obtains: Modal Components

[0009] Center frequency .

[0010] As a further limitation of the technical solution of the present invention, the steps of training the deep belief network with a training set and verifying the trained deep belief network with a verification set to obtain a trained deep belief network model include: The number of neurons in the input layer of the deep belief network is determined according to the dimension of the characteristic vector of the vibration signal after the variational mode decomposition process, and the number of neurons in the output layer is determined according to the prediction task, thereby determining the structure of the deep belief network. Randomly initialize the weight matrix and bias vector for each connection between the layers of the deep belief network; Use the contrastive divergence algorithm to pre-train each RBM in the deep belief network; Combine the pre-trained RBMs layer by layer into a complete pre-trained deep belief network; Input the training set into the pre-trained deep belief network, obtain the predicted output through forward propagation, and calculate the mean square error between the predicted output and the actual label as the loss function; According to the calculated loss function, the back propagation algorithm is used to update the weights and biases of the entire network to obtain a fine-tuned deep belief network; Input the data of the validation set into the fine-tuned deep belief network, and obtain the prediction output of the validation set through forward propagation; The evaluation indicators are used to evaluate the performance of the model on the validation set. When the performance on the validation set meets the requirements, a trained deep belief network model is obtained.

[0011] As a further limitation of the technical solution of the present invention, the step of using the contrastive divergence algorithm to pre-train each RBM in the deep belief network includes: The data of the training set is input into the visible layer, and mapped to the hidden layer by calculating the activation probability of the neurons in the hidden layer. The visible layer is reconstructed from the hidden layer by calculating the activation probability of the neurons in the visible layer to update the weights and biases to achieve pre-training of RBM.

[0012] As a further limitation of the technical solution of the present invention, the step of optimizing the parameters of the deep belief network model by using the sparrow search algorithm to obtain the vibration trend prediction model includes: Initialize the sparrow population and define the fitness function. Each sparrow individual represents a set of parameters of the deep belief network, including the weights and biases of all layers. Through iterative updating of the sparrow search algorithm, the sparrow individual with the smallest fitness value is selected, and its corresponding parameters are the parameters of the optimized deep belief network; The optimized parameters are used to construct the final deep belief network model, i.e., the vibration trend prediction model.

[0013] As a further limitation of the technical solution of the present invention, in the step of iterative updating through the sparrow search algorithm, the follower position update formula is:

[0014] Where: Indicates the worst position of an individual in the population at the current moment; Represents the best position in the population after one iteration; A is a 1-row A matrix of columns whose elements are either 1 or -1; ;when At that time, Only the follower needs to fly to other places to find food; otherwise, it means that the follower has found the optimal foraging location; The update formula of the sentinel position is:

[0015] Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and They represent the highest and lowest fitness values ​​of individuals in the current population respectively.

[0016] In a second aspect, the technical solution of the present invention also provides a device for predicting the mechanical vibration trend of a hydropower unit, including a variational mode decomposition processing module, a training module, an optimization module and a multi-step prediction execution module; A variational mode decomposition processing module is used to process the vibration signal of the hydropower unit using a variational mode decomposition algorithm and divide the processed data into a training set and a validation set; The training module is used to train the deep belief network with the training set and verify the trained deep belief network with the verification set to obtain a trained deep belief network model; An optimization module, used to optimize the parameters of the deep belief network model using a sparrow search algorithm to obtain a vibration trend prediction model; The multi-step prediction execution module is used to collect the vibration signal of the hydropower unit and process it using the variational mode decomposition algorithm, then input it into the vibration trend prediction model to obtain the vibration prediction value, and use the obtained vibration prediction value as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

[0017] As a further limitation of the technical solution of the present invention, the variational mode decomposition processing module includes a data acquisition preprocessing unit, a variational mode decomposition unit, a feature extraction unit, a data set generation unit, and a data set division unit; A data acquisition preprocessing unit is used to collect the mechanical vibration signal of the hydropower unit and perform normalization after denoising the collected original vibration signal; A variational mode decomposition unit, used to determine the mode parameters and perform VMD decomposition on the normalized vibration signal according to the determined parameters; A feature extraction unit, used for extracting features from each modal component obtained by decomposition; A data set generating unit, used for combining the features extracted from all modal components into a feature vector to generate a data set; The data set division unit is used to divide the data set into a training set and a validation set according to a set ratio.

[0018] As a further limitation of the technical solution of the present invention, in the step of determining the modal parameters and performing VMD decomposition on the normalized vibration signal according to the determined parameters, the VMD decomposition obtains: Modal Components

[0019] Center frequency .

[0020] As a further limitation of the technical solution of the present invention, the training module includes a network structure determination unit, an initialization unit, a pre-training unit, a pre-trained network generation unit, a training unit, a verification unit and a model evaluation unit; A network structure determination unit, used to determine the number of neurons in the input layer of the deep belief network according to the dimension of the characteristic vector of the vibration signal after the variational mode decomposition process, determine the number of neurons in the output layer according to the prediction task, and further determine the structure of the deep belief network; An initialization unit, used to randomly initialize the weight matrix and bias vector for the connection between each layer of the deep belief network; A pre-training unit for pre-training each RBM in a deep belief network using the contrastive divergence algorithm; The pre-trained network generation unit is used to combine the pre-trained RBMs layer by layer into a complete pre-trained deep belief network; The training unit is used to input the training set into the pre-trained deep belief network, obtain the predicted output through forward propagation, and calculate the mean square error between the predicted output and the actual label as the loss function. According to the calculated loss function, the back propagation algorithm is used to update the weights and biases of the entire network to obtain a fine-tuned deep belief network. The verification unit is used to input the data of the verification set into the fine-tuned deep belief network and obtain the prediction output of the verification set through forward propagation; The model evaluation unit is used to evaluate the performance of the model on the validation set using evaluation indicators. When the performance on the validation set meets the requirements, a trained deep belief network model is obtained.

[0021] As a further limitation of the technical solution of the present invention, the pre-training unit is specifically used to input the data of the training set into the visible layer, and map it to the hidden layer by calculating the activation probability of the neurons in the hidden layer, and reconstruct the visible layer from the hidden layer by calculating the activation probability of the neurons in the visible layer, so as to update the weights and biases to realize the pre-training of RBM.

[0022] As a further limitation of the technical solution of the present invention, the optimization module includes an initialization setting unit, a parameter optimization unit and a model building unit; The initialization setting unit is used to initialize the sparrow population and define the fitness function. Each sparrow individual represents a set of parameters of the deep belief network, including the weights and biases of all layers; The parameter optimization unit is used to iteratively update the sparrow search algorithm and select the sparrow individual with the smallest fitness value. The corresponding parameters are the parameters of the optimized deep belief network. The model building unit is used to build a final deep confidence network model, i.e., a vibration trend prediction model, using the optimized parameters.

[0023] As a further limitation of the technical solution of the present invention, in the step of iterative updating through the sparrow search algorithm, the follower position update formula is:

[0024] Where: Indicates the worst position of an individual in the population at the current moment; Represents the best position in the population after one iteration; A is a 1-row A matrix of columns whose elements are either 1 or -1; ;when At that time, Only the follower needs to fly to other places to find food; otherwise, it means that the follower has found the optimal foraging location; The update formula of the sentinel position is:

[0025] Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and They represent the highest and lowest fitness values ​​of individuals in the current population respectively.

[0026] In a third aspect, the technical solution of the present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores computer program instructions executable by the at least one processor, and the computer program instructions are executed by the at least one processor so that the at least one processor can execute the method for predicting the mechanical vibration trend of a hydropower unit as described in the first aspect.

[0027] In a fourth aspect, the technical solution of the present invention further provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the method for predicting the mechanical vibration trend of a hydropower unit as described in the first aspect.

[0028] The beneficial effects of the technical solution of the present invention are as follows: by using the variational mode decomposition algorithm to process the vibration signal of the hydropower unit, the complex vibration signal can be decomposed into multiple intrinsic mode functions with different characteristic scales, the key information in the signal can be effectively extracted, and noise interference can be removed, so that the data quality used for subsequent model training is higher. Compared with directly using the original vibration signal, the availability and reliability of the data are greatly improved.

[0029] The processed data is divided into a training set and a validation set. Data division helps the model learn data features better during the training process. At the same time, the training effect is evaluated through the validation set to prevent overfitting of the model, enhance the generalization ability of the model, and enable the model to better adapt to the vibration of the hydropower unit under different working conditions. The deep belief network is used for training. The deep belief network has a strong nonlinear modeling ability and can mine the complex relationships hidden in the vibration signal, so as to make a more accurate prediction of the vibration trend of the hydropower unit. Through sufficient training of the training set, the model can continuously adjust its own parameters to optimize the prediction performance. The sparrow search algorithm is used to optimize the parameters of the deep belief network model, which can find the optimal solution among many parameter combinations and avoid the model from falling into the local optimum, so that the vibration trend prediction model can run with a better parameter configuration and improve the accuracy and reliability of the prediction.

[0030] In the actual prediction stage, the vibration prediction value is used as the input for the next prediction, realizing the multi-step prediction of the vibration trend of the hydropower unit. This multi-step prediction mechanism allows operation and maintenance personnel to understand the vibration trend of the hydropower unit in the future in advance, discover potential fault hazards in time, take corresponding maintenance measures in advance, effectively reduce the equipment failure rate, improve the operating safety and stability of the hydropower unit, reduce downtime, reduce maintenance costs, and provide strong guarantee for the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solution of the present invention, the accompanying drawings required for use in the description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0032] Figure 1 A schematic flow chart of a method provided in an embodiment of the present invention.

[0033] Figure 2 Flowchart of the deep belief network training process.

[0034] Figure 3 A diagram showing single-step prediction using the method provided by the present invention.

[0035] Figure 4 A schematic block diagram of a device provided by an embodiment of the present invention.

[0036] Figure 5 Comparison chart of predictions from different models. DETAILED DESCRIPTION

[0037] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in this specific embodiment. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0038] like Figure 1 As shown, an embodiment of the present invention provides a method for predicting the mechanical vibration trend of a hydropower unit, comprising the following steps: S1: Use the variational mode decomposition algorithm to process the vibration signal of the hydropower unit and divide the processed data into a training set and a validation set; Variational mode decomposition (VMD) can decompose complex vibration signals into multiple modal components, which may contain information of different frequencies and time scales, and help to extract more effective features. By reasonably dividing the training set and the validation set, data support can be provided for subsequent model training and validation.

[0039] S2: Use the training set to train the deep belief network, and use the verification set to verify the trained deep belief network to obtain a trained deep belief network model; First, DBN is trained using a training set generated using the variational mode decomposition algorithm VMD, so that the model can learn the features and patterns in the data. Then the trained model is verified using the validation set to evaluate the generalization ability of the model and avoid overfitting.

[0040] S3: The parameters of the deep belief network model are optimized using the sparrow search algorithm to obtain the vibration trend prediction model; After the initial training and verification of the DBN model, the performance of the model can be further improved by optimizing the model parameters using the sparrow search algorithm SSA. SSA can find a better combination of parameters in the hope of obtaining a more accurate vibration trend prediction model, namely the VMD-SSA-DBN model. S4: The vibration signal of the hydropower unit is collected and processed by the variational mode decomposition algorithm, and then input into the vibration trend prediction model to obtain the vibration prediction value, and the obtained vibration prediction value is used as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

[0041] After building the optimized vibration trend prediction model, new vibration signals of the hydropower unit are collected, processed by VMD and input into the model to obtain the vibration prediction value. This is the normal operation of using the model for prediction. The value obtained by the single-step prediction is used as the input for the next prediction, which can realize the multi-step prediction of the vibration trend of the hydropower unit, so as to better grasp the dynamic change process of the unit vibration.

[0042] In some embodiments, the steps of processing the vibration signal of the hydropower unit using the variational mode decomposition algorithm and dividing the processed data into a training set and a validation set include: S11: collecting mechanical vibration signals of the hydropower unit, and performing normalization processing on the collected original vibration signals after denoising; In this step, use accelerometers and other equipment to collect vibration signals at key parts of the hydropower unit (such as bearing seats, housings, etc.). Ensure that the sensors are firmly installed and the sampling frequency of the acquisition equipment is set reasonably. Generally, according to the unit vibration frequency range and Nyquist sampling theorem, the sampling frequency should be higher than twice the highest frequency of the vibration signal. For example, for a vibration signal with a maximum frequency of 1000Hz, the sampling frequency can be set to 2000Hz or higher.

[0043] The collected original vibration signal is denoised. Filtering techniques such as Butterworth filter and wavelet threshold denoising can be used. For example, if there is high-frequency electromagnetic interference noise in the vibration signal, a low-pass Butterworth filter can be used to remove noise components above a certain frequency (such as 1500Hz). At the same time, check whether there are abnormal values ​​in the signal (such as maximum or minimum values ​​caused by sensor failure or electromagnetic pulse). If there are abnormal values, mean substitution, linear interpolation and other methods can be used to correct them.

[0044] In order to process vibration signals of different amplitude ranges at the same scale, the cleaned signals are normalized. Common normalization methods include minimum-maximum normalization, that is, for signals The normalized signal ,in, , mapping the signal range to the interval .

[0045] S12: determining modal parameters and performing VMD decomposition on the normalized vibration signal according to the determined parameters; Select an appropriate modal number K. You can preliminarily estimate the modal number by observing the spectrum of the original vibration signal. If there are several obvious peaks in the spectrum, they may correspond to several major vibration modes, which can be used to determine the value of K. At the same time, set the penalty factor , which is used to control the bandwidth of each mode after decomposition, and is usually between 1000 and 10000.

[0046] According to the determined parameters K and , perform VMD decomposition on the normalized vibration signal. The goal of VMD is to decompose the original signal into K modal components, each of which has its center frequency By constructing and solving the variational problem, each modal component and its center frequency are continuously updated iteratively so that the original signal can be accurately represented by these modal components.

[0047] In actual calculations, the alternating direction method of multipliers (ADMM) is often used to solve this variational problem. After multiple iterations, until the convergence condition is met (such as the change of each modal component between two adjacent iterations is less than the set threshold), K modal components are obtained.

[0048] S13: extracting features from each modal component obtained by decomposition; Extract features from each modal component obtained by decomposition. Common features include the energy, amplitude, frequency, etc. of the modal component. For example, for the modal component , its energy (t1 and t2 are time intervals), the amplitude can be expressed by calculating the maximum energy or root mean square value within a certain time range, and the frequency characteristics can be determined by analyzing the spectrum peak position after performing fast Fourier transform (FFT) on the modal components.

[0049] S14: Combining the features extracted from all modal components into a feature vector to generate a data set; Combine the features extracted from all modal components into a feature vector. For example, if there are K modal components and each modal component extracts n features, then we can get a dimensional feature vectors. Arrange these feature vectors in chronological order.

[0050] S15: Divide the dataset into a training set and a validation set according to the set ratio.

[0051] Divide the training set and validation set according to a certain ratio. Usually, 70% - 80% of the data can be used as the training set, and the remaining data can be used as the validation set. For example, for N feature vectors arranged in chronological order, select the first The first Nm feature vectors are used as the training set, and the last Nm feature vectors are used as the validation set. At the same time, it is important to ensure that the data distribution of the training set and the validation set are similar, that is, they can both represent the vibration characteristics of the hydropower unit under different operating conditions.

[0052] It should be noted that variational mode decomposition (VMD) is based on Wiener filtering and Hilbert transform in a variational framework using alternating direction multipliers to solve the variational model, decomposing the signal into several quasi-orthogonal mode functions with finite bandwidth and center frequency, then estimating the center frequency and transferring the spectrum to the baseband, and finally obtaining the bandwidth of each frequency band by demodulating the signal.

[0053] The VMD algorithm targets real-valued input signals Split into a discrete number of sub-signals (modes) . Assume that each mode has a center frequency The surrounding is tight, that is, the frequency components contained in this mode are all within Nearby, It is determined as the decomposition progresses.

[0054] To evaluate the modal bandwidth, the Hilbert transform method can be used, and the resulting constrained variational formula is as follows: (2-1) In the formula, represents the number of modal components; is the kth modal component; is the center frequency of the kth modal component; represents the derivative with respect to time; is the unit pulse signal function; Indicates convolution calculation. ; are the shorthand symbols for all modes and their center frequencies respectively; is the sum of all modes.

[0055] Solving the above equation with the augmented Lagrange function yields the following equation: (2-2) in, represents the quadratic penalty factor, represents the derivative with respect to time; is the unit pulse signal function, is the kth modal component, represents the Lagrangian operator, is the initial signal.

[0056] Solving the above equations, we get the mode The solution formula is as follows: (2-3) In the formula, is the center frequency, for The Fourier transform of for The Fourier transform of for The Fourier transform of represents the quadratic penalty factor, is the number of iterations. The corresponding center frequency The iteration formula is: (2-4) In the formula, is the center frequency, is the number of iterations, for The Fourier transform of .

[0057] In some embodiments, the step of training the deep belief network with a training set and verifying the trained deep belief network with a verification set to obtain a trained deep belief network model includes: S21: Determine the number of neurons in the input layer of the deep belief network according to the dimension of the characteristic vector of the vibration signal after the variational mode decomposition processing, determine the number of neurons in the output layer according to the prediction task, and then determine the structure of the deep belief network; In this step, the structure of the deep belief network (DBN) is clarified, including the number of neurons in the input layer, hidden layer, and output layer. For example, assuming that the input layer receives the vibration signal features after variational mode decomposition, the feature dimension is n, and the number of neurons in the input layer is n. The number of hidden layers and the number of neurons in each layer are determined based on experience or experiments. For example, if two hidden layers are set, the number of neurons in the first hidden layer is , the number of neurons in the second hidden layer is , the output layer is set to 1 neuron according to the prediction task (such as predicting vibration amplitude).

[0058] S22: Randomly initialize the weight matrix and bias vector for each connection between the layers of the deep belief network; For from Layer to +1 layer connection, weight matrix The size is , and They are Layer and +1 layer number of neurons, bias vector The size is .

[0059] S23: Use the contrastive divergence algorithm to pre-train each RBM in the deep belief network; In this step, the data of the training set is input into the visible layer, and mapped to the hidden layer by calculating the activation probability of the neurons in the hidden layer. The visible layer is reconstructed from the hidden layer by calculating the activation probability of the neurons in the visible layer to update the weights and biases to realize the pre-training of RBM.

[0060] It should be noted that the visible layer neurons and hidden layer neurons The activation probability is:

[0061]

[0062] Where: Visible layer neurons Status; The neurons in the hidden layer Status; For neurons , The connection weight between them; and For neurons , The bias.

[0063] S24: Combine the pre-trained RBMs layer by layer into a complete pre-trained deep belief network; S25: Input the training set into the pre-trained deep belief network, obtain the predicted output through forward propagation, and calculate the mean square error between the predicted output and the actual label as the loss function; S26: Based on the calculated loss function, the back propagation algorithm is used to update the weights and biases of the entire network to obtain a fine-tuned deep belief network; S27: Input the data of the validation set into the fine-tuned deep belief network, and obtain the prediction output of the validation set through forward propagation; S28: Use evaluation indicators to evaluate the performance of the model on the validation set. When the performance on the validation set meets the requirements, a trained deep belief network model is obtained.

[0064] The BP classifier classifies the features extracted by the last layer of RBM, compares the obtained results with the actual results, and calculates the corresponding errors. Then, in order to obtain the global optimal solution of DBN, the network parameters are fine-tuned using the back error propagation algorithm. This process is the final stage of the entire DBN training. Through supervised learning, DBN can learn more discriminative feature representations, thereby improving classification performance. This training method that combines unsupervised pre-training and supervised fine-tuning can make full use of unlabeled data for feature learning, and further optimize through labeled data, thereby improving the generalization ability and classification performance of the model.

[0065] The training process of DBN is as follows Figure 2 As shown. First, many RBMs continuously train the data set and then determine the network parameters. Then, the previous RBM is input into the visible layer of the next RBM, and the weights and biases of the neurons are determined according to the corresponding algorithm, and the network parameters are continuously reconstructed, updated, and trained. This cycle completes the supervised fine-tuning of the DBN network parameters. In other words, DBN training is essentially RBM training. By determining the local optimal solution through the first-layer training, the BP algorithm will not fall into the local optimal solution due to the random initialization of the parameters when adjusting the parameters. The training time can be greatly shortened and the efficiency can be significantly improved.

[0066] In some embodiments, the step of optimizing the parameters of the deep belief network model using the sparrow search algorithm to obtain the vibration trend prediction model includes: S31: Initialize the sparrow population and define the fitness function. Each sparrow individual represents a set of parameters of the deep belief network, including the weights and biases of all layers. S32: Through the iterative update of the sparrow search algorithm, the sparrow individual with the smallest fitness value is selected, and its corresponding parameters are the parameters of the optimized deep belief network; S33: Use the optimized parameters to build the final deep belief network model, i.e., the vibration trend prediction model.

[0067] Set the number of sparrows in the population, for example, to 50. Determine the proportion of discoverers in the population, for example, to 0.2. Set the proportion of vigilants in the population, for example, to 0.15. Set the maximum number of iterations of the algorithm, for example, to 200.

[0068] In this step, each sparrow individual represents a set of DBN parameters (weights and biases). The initial population is randomly generated, and the parameter value of each individual is within a certain range, for example, the weight is between [-1, 1] and the bias is between [-0.1, 0.1]. For a network with n input layer neurons, The first hidden layer neurons, The DBN has a second hidden layer of neurons and 1 output layer neuron. The parameter dimension of a sparrow individual is , corresponding to the weights and bias numbers from the input layer to the first hidden layer, from the first hidden layer to the second hidden layer, and from the second hidden layer to the output layer. The prediction error of DBN on the training set is used as the fitness function of the sparrow search algorithm. The commonly used error metric is the mean square error (MSE). (in is the input feature vector, is the corresponding actual vibration value), the predicted output of DBN is , then the MSE calculation formula is: ,Here, M is the number of training data. The goal of the fitness function is to find the DBN parameter combination that minimizes MSE, that is, to find the optimal sparrow individual.

[0069] The discoverer's position is updated as follows: (2-5) In the formula, Indicates the number of iterations; and They refer to the sparrow’s serial number and the sparrow’s position dimension respectively; is a random number between 0 and 1 ( cannot be 0); is a random number that follows a standard normal distribution; For 1 row A matrix with all 1 columns; Indicates the warning value; Indicates a safe value. When , it indicates that the surrounding environment is safe and the finder can search for food in the surrounding area; when When the signal is received, it indicates that there is danger around. In this way, the sentinel will immediately send out a warning signal, and the population will quickly fly to a safe area after receiving the signal. The follower position update is as follows: (2-6) Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and They represent the highest and lowest fitness values ​​of individuals in the current population respectively.

[0070] The sentinel position is updated as follows: (2-7) Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and Respectively represent the highest and lowest fitness values ​​of individuals in the current population. When , it means that the sparrow is at the outermost edge of the population and needs to move to the center of the population; when This means that the sparrow in the best position has discovered the danger of this place and needs to go to a safe place.

[0071] Specifically, VMD is used to decompose the training set and test set of the original vibration signal into modal components of different frequencies, and the input and output values ​​of the training set and test set are constructed based on these modal components. That is, the 480 sets of upper frame x-direction vibration data signals are decomposed into actual values ​​using equations (2-2) and (2-3), and then divided into training set and test set in a ratio of 2:1. At this time, the input values ​​of the training set and test set are obtained. Each input value of the training set is used as the corresponding input value of the softmax activation function, according to equations (2-12)-(2-14) and Figure 2 After training, the output value of the test set and the trained DBN model are obtained. The obtained DBN model then selects the best prediction position according to the sparrow search algorithm (2-6)-(2-7). The single-step prediction result is shown in the figure below. Figure 3 As shown in the figure, the predicted values ​​of the improved algorithm after adding the sparrow search algorithm have a high degree of overlap with the actual values, and the trends are basically consistent, which verifies the superiority of the algorithm.

[0072] In order to make the operation of hydropower units safer and more stable, it is necessary to carry out multi-step prediction research on the vibration trend of hydropower units.

[0073] The experimental data still uses 480 sets of upper frame x-direction vibration data of a hydropower station hydroelectric unit. At the same time, all prediction experiments are consistent with single-step prediction. The main method of multi-step prediction is to use the output of the previous step as the input of the next step, that is, the input of the test set of the second step is obtained by the output value of the training set predicted by the single step, that is, the input of the test set of the third step is obtained by the output value of the training set predicted by the second step, and so on. The subsequent prediction method remains the same as before.

[0074] The VMD-SSA-DBN model of this application is compared with the existing VMD-DBN, VMD-ELM, ELM, and DBN models. The prediction comparison chart of different models is as follows: Figure 5 shown.

[0075] Depend on Figure 5 It can be seen that single models like DBN and ELM are similar to the true value only in trend, but the amplitude of vibration is far from the true value; combined models like VMD-DBN and VMD-ELM are similar to the true value in trend, and the overall amplitude is well matched to the true value, but most of the endpoints are still far from the true value, and some endpoints still have aliasing; in contrast, the results predicted by the VMD-SSA-DBN model of the present application are not only close to the true value in trend, but also the overall amplitude and endpoints can be well matched to the true value. The model of the present application can better fit the true value. The specific prediction indicators are shown in Table 1, which preliminarily verifies that the model of the present application has good prediction performance.

[0076] Table 1: Prediction indicator data table

[0077] In order to quantitatively analyze the prediction results of different models, three indicators, namely root mean square error (RMSE), mean absolute error (MAE), and mean relative percentage error (MAPE), are introduced to analyze the prediction performance of the model. The specific results are shown in Table 1. Compared with the models of VMD-DBN, VMD-ElM, DBN, and ELM, the RMSE index of the VMD-SSA-DBN model decreased by 0.0265, 0.0926, 1.93414, and 1.64194, respectively, the MAE index decreased by 0.02567, 0.00906, 1.62957, and 1.29257, respectively, and the MAPE index decreased by 0.01603%, 0.00572%, 1.01973%, and 0.80693%, respectively. It is not difficult to see from the three indicators that the VMD-SSA-DBN model has the best prediction effect on the upper frame X-direction vibration signal data.

[0078] like Figure 4As shown, an embodiment of the present invention further provides a device for predicting the mechanical vibration trend of a hydropower unit, comprising a variational mode decomposition processing module, a training module, an optimization module and a multi-step prediction execution module; A variational mode decomposition processing module is used to process the vibration signal of the hydropower unit using a variational mode decomposition algorithm and divide the processed data into a training set and a validation set; The training module is used to train the deep belief network with the training set and verify the trained deep belief network with the verification set to obtain a trained deep belief network model; An optimization module, used to optimize the parameters of the deep belief network model using a sparrow search algorithm to obtain a vibration trend prediction model; The multi-step prediction execution module is used to collect the vibration signal of the hydropower unit and process it using the variational mode decomposition algorithm, then input it into the vibration trend prediction model to obtain the vibration prediction value, and use the obtained vibration prediction value as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

[0079] In some embodiments, the variational mode decomposition processing module includes a data acquisition preprocessing unit, a variational mode decomposition unit, a feature extraction unit, a data set generation unit, and a data set partitioning unit; A data acquisition preprocessing unit is used to collect the mechanical vibration signal of the hydropower unit and perform normalization after denoising the collected original vibration signal; A variational mode decomposition unit, used for determining the mode parameters and performing VMD decomposition on the normalized vibration signal according to the determined parameters; A feature extraction unit, used for extracting features from each modal component obtained by decomposition; A data set generating unit, used for combining the features extracted from all modal components into a feature vector to generate a data set; The data set division unit is used to divide the data set into a training set and a validation set according to a set ratio.

[0080] In some embodiments, in the step of determining modal parameters and performing VMD decomposition on the normalized vibration signal according to the determined parameters, VMD decomposition obtains: Modal Components

[0081] Center frequency .

[0082] In some embodiments, the training module includes a network structure determination unit, an initialization unit, a pre-training unit, a pre-trained network generation unit, a training unit, a verification unit, and a model evaluation unit; A network structure determination unit, used to determine the number of neurons in the input layer of the deep belief network according to the dimension of the characteristic vector of the vibration signal after the variational mode decomposition process, determine the number of neurons in the output layer according to the prediction task, and further determine the structure of the deep belief network; An initialization unit, used to randomly initialize the weight matrix and bias vector for the connection between each layer of the deep belief network; A pre-training unit for pre-training each RBM in a deep belief network using the contrastive divergence algorithm; The pre-trained network generation unit is used to combine the pre-trained RBMs layer by layer into a complete pre-trained deep belief network; The training unit is used to input the training set into the pre-trained deep belief network, obtain the predicted output through forward propagation, and calculate the mean square error between the predicted output and the actual label as the loss function. According to the calculated loss function, the back propagation algorithm is used to update the weights and biases of the entire network to obtain a fine-tuned deep belief network. The verification unit is used to input the data of the verification set into the fine-tuned deep belief network and obtain the prediction output of the verification set through forward propagation; The model evaluation unit is used to evaluate the performance of the model on the validation set using evaluation indicators. When the performance on the validation set meets the requirements, a trained deep belief network model is obtained.

[0083] In some embodiments, the pre-training unit is specifically used to input the data of the training set into the visible layer, and map it to the hidden layer by calculating the activation probability of the neurons in the hidden layer, and reconstruct the visible layer from the hidden layer by calculating the activation probability of the neurons in the visible layer, so as to update the weights and biases to realize the pre-training of RBM.

[0084] In some embodiments, the optimization module includes an initialization setting unit, a parameter optimization unit, and a model building unit; The initialization setting unit is used to initialize the sparrow population and define the fitness function. Each sparrow individual represents a set of parameters of the deep belief network, including the weights and biases of all layers; The parameter optimization unit is used to iteratively update the sparrow search algorithm and select the sparrow individual with the smallest fitness value. The corresponding parameters are the parameters of the optimized deep belief network. The model building unit is used to build a final deep confidence network model, i.e., a vibration trend prediction model, using the optimized parameters.

[0085] In the iterative update step of the sparrow search algorithm, the follower position update formula is:

[0086] Where: Indicates the worst position of an individual in the population at the current moment; Represents the best position in the population after one iteration; A is a 1-row A matrix of columns whose elements are either 1 or -1; ;when At that time, Only the follower needs to fly to other places to find food; otherwise, it means that the follower has found the optimal foraging location; The update formula of the sentinel position is:

[0087] Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and They represent the highest and lowest fitness values ​​of individuals in the current population respectively.

[0088] The embodiment of the present invention also provides an electronic device, which includes: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus. The communication bus can be used for information transmission between the electronic device and the sensor. The processor can call the logic instructions in the memory to execute the following method: S1: Process the vibration signal of the hydropower unit using the variational mode decomposition algorithm and divide the processed data into a training set and a verification set; S2: Train the deep belief network with the training set and verify the trained deep belief network with the verification set to obtain a trained deep belief network model; S3: Use the sparrow search algorithm to optimize the parameters of the deep belief network model to obtain a vibration trend prediction model; S4: Collect the vibration signal of the hydropower unit and use the variational mode decomposition algorithm to process it and then input it into the vibration trend prediction model to obtain a vibration prediction value, and use the obtained vibration prediction value as the input of the next prediction to realize multi-step prediction of the vibration trend of the hydropower unit.

[0089] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0090] An embodiment of the present invention provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute the method provided by the above method embodiment, for example, including: S1: using a variational mode decomposition algorithm to process the vibration signal of a hydropower unit and dividing the processed data into a training set and a verification set; S2: using the training set to train a deep belief network, and using the verification set to verify the trained deep belief network to obtain a trained deep belief network model; S3: using a sparrow search algorithm to optimize the parameters of the deep belief network model to obtain a vibration trend prediction model; S4: collecting the vibration signal of the hydropower unit and using the variational mode decomposition algorithm to process it, and then inputting it into the vibration trend prediction model to obtain a vibration prediction value, and using the obtained vibration prediction value as the input for the next prediction to achieve multi-step prediction of the vibration trend of the hydropower unit.

[0091] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0092] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the mechanical vibration trend of a hydropower unit, characterized in that: The steps include: The vibration signal of the hydropower unit is processed using the variational mode decomposition algorithm and the processed data is divided into a training set and a validation set; The deep belief network is trained with the training set, and the trained deep belief network is verified with the verification set to obtain a trained deep belief network model; The sparrow search algorithm is used to optimize the parameters of the deep belief network model to obtain the vibration trend prediction model; The vibration signal of the hydropower unit is collected and processed using the variational mode decomposition algorithm, and then input into the vibration trend prediction model to obtain the vibration prediction value. The obtained vibration prediction value is used as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

2. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 1, characterized in that: The steps of using the variational mode decomposition algorithm to process the vibration signal of the hydropower unit and dividing the processed data into a training set and a validation set include: Collect the mechanical vibration signal of the hydropower unit, and perform normalization after denoising the collected original vibration signal; Determine the modal parameters and perform VMD decomposition on the normalized vibration signal according to the determined parameters; Perform feature extraction on each modal component obtained by decomposition; The dataset is divided into a training set and a validation set according to the set ratio.

3. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 2, characterized in that: The deep belief network is trained with the training set, and the trained deep belief network is verified with the verification set. The steps of obtaining the trained deep belief network model include: The number of neurons in the input layer of the deep belief network is determined according to the dimension of the characteristic vector of the vibration signal after the variational mode decomposition process, and the number of neurons in the output layer is determined according to the prediction task, thereby determining the structure of the deep belief network. Randomly initialize the weight matrix and bias vector for each connection between the layers of the deep belief network; Use the contrastive divergence algorithm to pre-train each RBM in the deep belief network; Combine the pre-trained RBMs layer by layer into a complete pre-trained deep belief network; Input the training set into the pre-trained deep belief network, obtain the predicted output through forward propagation, and calculate the mean square error between the predicted output and the actual label as the loss function; According to the calculated loss function, the back propagation algorithm is used to update the weights and biases of the entire network to obtain a fine-tuned deep belief network; Input the data of the validation set into the fine-tuned deep belief network, and obtain the prediction output of the validation set through forward propagation; The evaluation indicators are used to evaluate the performance of the model on the validation set. When the performance on the validation set meets the requirements, a trained deep belief network model is obtained.

4. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 3, characterized in that: The steps to pre-train each RBM in a deep belief network using the contrastive divergence algorithm include: The data of the training set is input into the visible layer, and mapped to the hidden layer by calculating the activation probability of the neurons in the hidden layer. The visible layer is reconstructed from the hidden layer by calculating the activation probability of the neurons in the visible layer to update the weights and biases to achieve pre-training of RBM.

5. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 4, characterized in that: Visible layer neurons and hidden layer neurons The activation probabilities are: Where: Visible layer neurons Status; The neurons in the hidden layer Status; For neurons , The connection weight between them; and For neurons , The bias.

6. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 5, characterized in that: The steps of using the sparrow search algorithm to optimize the parameters of the deep belief network model to obtain the vibration trend prediction model include: Initialize the sparrow population and define the fitness function. Each sparrow individual represents a set of parameters of the deep belief network, including the weights and biases of all layers. Through iterative updating of the sparrow search algorithm, the sparrow individual with the smallest fitness value is selected, and its corresponding parameters are the parameters of the optimized deep belief network; The optimized parameters are used to construct the final deep belief network model, i.e., the vibration trend prediction model.

7. The method for predicting the mechanical vibration trend of a hydroelectric unit according to claim 6, characterized in that: In the iterative update step of the sparrow search algorithm, the follower position update formula is: Where: Indicates the worst position of an individual in the population at the current moment; Represents the best position in the population after one iteration; A is a 1-row A matrix of columns whose elements are either 1 or -1; ;when At that time, Only the follower needs to fly to other places to find food; otherwise, it means that the follower has found the optimal foraging location; The update formula of the sentinel position is: Where: Indicates passing The best position found by the sentinel after iterations; and are all random numbers, the former follows a normal distribution, and the latter The value between the two is used to determine the length of each step; is a constant that approaches 0 but is not 0; express The fitness value of the sparrow individual after iterations, and They represent the highest and lowest fitness values ​​of individuals in the current population respectively.

8. A device for predicting the mechanical vibration trend of a hydroelectric unit, characterized in that: It includes variational mode decomposition processing module, training module, optimization module and multi-step prediction execution module; A variational mode decomposition processing module is used to process the vibration signal of the hydropower unit using a variational mode decomposition algorithm and divide the processed data into a training set and a validation set; The training module is used to train the deep belief network with the training set and verify the trained deep belief network with the verification set to obtain a trained deep belief network model; An optimization module, used to optimize the parameters of the deep belief network model using a sparrow search algorithm to obtain a vibration trend prediction model; The multi-step prediction execution module is used to collect the vibration signal of the hydropower unit and process it using the variational mode decomposition algorithm, then input it into the vibration trend prediction model to obtain the vibration prediction value, and use the obtained vibration prediction value as the input of the next prediction to realize the multi-step prediction of the vibration trend of the hydropower unit.

9. An electronic device, characterized in that: The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores computer program instructions executable by the at least one processor, and the computer program instructions are executed by the at least one processor so that the at least one processor can execute the method for predicting the mechanical vibration trend of a hydropower unit as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the method for predicting the mechanical vibration trend of a hydropower unit as described in any one of claims 1 to 7.

Images