Hydroelectric generating set vibration signal trend prediction method and system

Abstract

The invention discloses a hydroelectric generating set vibration signal trend prediction method and system, and the method comprises the steps: S1, collecting vibration signals of key parts in the operation process of a hydroelectric generating set, and generating a vibration time sequence sample with a unified structure; s2, applying an EEMD (ensemble empirical mode decomposition) algorithm to each section of preprocessed vibration time sequence; s3, dividing the constructed data set into a training set and a test set; s4, constructing a time sequence prediction model based on BiLSTM and an attention mechanism; s5, training the model on the training set, using a mean square error as a loss function, and adopting an Adam optimizer to optimize network parameters; calculating a mean value and a standard deviation based on prediction errors of the training set, and setting upper and lower threshold values of early warning judgment; s6, applying the trained model to a test set, predicting an unknown vibration sequence, and if a prediction error exceeds a set threshold range, regarding that the trend is abnormal; according to the method, the problems of insufficient feature extraction capability, poor modeling effect and low prediction precision in the existing hydroelectric generating set vibration signal prediction are solved.

Description

Technical Field

[0001] This invention relates to the field of hydropower unit operation status monitoring and fault prediction technology, and in particular to a method and system for predicting the trend of vibration signals of hydropower units. Background Technology

[0002] During long-term continuous operation, hydropower units are susceptible to various complex factors such as hydraulic excitation, unbalanced forces, electromagnetic disturbances, and structural aging, leading to abnormal vibration responses. If these abnormalities are not detected and addressed promptly, they may cause equipment damage or even operational accidents. Therefore, conducting research on vibration trend prediction for hydropower units is of great significance for achieving fault early warning, extending equipment lifespan, and improving operational economy.

[0003] Traditional methods for monitoring the vibration of hydropower units mainly rely on threshold judgment, empirical discrimination, or spectrum analysis. These methods are effective in processing linear and stationary signals. However, due to the nonlinear, non-stationary, and highly time-varying characteristics of the vibration signals of the units under actual operating conditions, traditional methods are difficult to accurately extract key features related to changes in equipment status, resulting in problems such as diagnostic delays, missed judgments, and misjudgments. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a method and system for predicting the vibration signal trend of hydropower units. This invention solves the problems of insufficient feature extraction capability, poor modeling effect, and low prediction accuracy in existing hydropower unit vibration signal prediction, improves the accuracy and stability of predicting the vibration state of hydropower units under complex operating conditions, and realizes early abnormal trend identification and operational risk warning.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for predicting the trend of vibration signals of hydropower units, comprising the following steps:

[0006] S1. Collect vibration signals from key components during the operation of the hydropower unit, and use a standardized preprocessing method for the vibration signals, as shown in the following formula, to generate vibration time series samples x′ with uniform structure;

[0007]

[0008] In the formula: μ is the sample mean, and σ is the sample standard deviation;

[0009] S2. Apply the EEMD algorithm to each preprocessed vibration time series to decompose the non-stationary signal into multiple intrinsic mode components (IMFs), and align the multiple IMFs according to time steps to construct a feature matrix, which serves as the input for subsequent models.

[0010] S3. Divide the constructed dataset into a training set and a test set. The training set is used for model training and threshold construction, while the test set is used to verify model performance and warning accuracy.

[0011] S4. Construct a time series prediction model based on BiLSTM and attention mechanism: Take the IMF feature matrix as input, extract the bidirectional time series features of each time step through BiLSTM, and introduce an attention mechanism at the output to weight the feature vectors of each time step;

[0012] S5. Train the model on the training set, using the mean squared error as the loss function, and use the Adam optimizer to optimize the network parameters; calculate the mean and standard deviation of the prediction error based on the training set, and set the upper and lower limit thresholds for early warning judgment.

[0013] S6. Apply the trained model to the test set to predict unknown vibration sequences. If the prediction error exceeds the set threshold range, it is considered an abnormal trend.

[0014] Preferably, in step S1, the standardized preprocessing method for vibration signals includes noise reduction, normalization, and sliding window segmentation.

[0015] Preferably, step S2 specifically includes:

[0016] S21. Add different types of white noise to the original signal to construct multiple noisy signals;

[0017] S22. Perform EMD decomposition on each group of noisy signals to obtain multiple IMF sequences;

[0018] S23. For each modal component, take its average value across all calculations as the final IMF output.

[0019] Preferably, step S21 specifically involves: mixing different white noise n i The signals X(t) are added to the original signal X(t) respectively to construct multiple noisy signals X(t)+n. i (t).

[0020] Preferably, in step S23, the IMF output result is:

[0021]

[0022] In the formula: c j (t) represents the final output signal of the j-th order IMF at time t; M is the total number of noisy signals; c ij (t) is the amplitude of the j-th order IMF component obtained by EEMD decomposition of the i-th group of noisy signals at time t.

[0023] Preferably, step S4 specifically includes:

[0024] S41. Based on BiLSTM, the vibration signal feature matrix is ​​modeled, and combined with the forward and backward information of the time series, the learning ability of long-term dependent features and trend changes is enhanced.

[0025] S42. Given an input sequence X = {x1, x2, ..., x...} T The BiLSTM network consists of two LSTM units operating in opposite directions, performing forward and backward propagation respectively, and outputting the forward hidden state. and backward hidden state The outputs from both directions are concatenated at time step t to obtain the final hidden state output:

[0026] S43. For a given sequence of hidden states output by a BiLSTM, the attention mechanism first calculates the relevance score for each time step, performs normalization processing to obtain the attention weight, and then weights the hidden states of all time steps according to the attention weight to calculate the context representation vector.

[0027] Preferably, in step S42, the internal calculation process of the LSTM unit is as follows:

[0028] f t =σ(W f [h t-1 ,x t ]+b f )

[0029] i t =σ(W i [h t-1 ,x t ]+b i )

[0030] o t =σ(W o [h t-1 ,x t ]+b o )

[0031]

[0032] In the formula: f t It is the forgetting gate vector; σ is the sigmoid activation function; b f Forget gate bias term; W f Here are the weight matrix of the forget gate; the input vector at time step t; and h. t-1 It is the hidden state vector of the previous time step; x t i is the input vector at time step t;t It is the input gate vector; W i b is the weight matrix of the input gate; i It is the bias term of the input gate; t It is the output gate vector; W o It is the weight matrix of the output gate, b o It is the bias term of the output gate; It is the candidate cell state; W c In candidate cell state calculation, it is used to input x at the current time step. t The weight matrix for linear transformation; b c It is the bias term of the candidate cell state; C t It is the unit state vector at time step t; C t-1 It is the unit state vector of the previous time step; ⊙ is the Hadamard product, multiplied element by element.

[0033] Preferably, step S43 specifically includes:

[0034] Given the hidden state sequence h of the BiLSTM output. t The attention mechanism first calculates the relevance score e for each time step t. t That is: e t =v T tanh(Wh t +b), all scores e t Normalization yields attention weights. The hidden states at all time steps are weighted according to the attention weights to calculate the context representation vector. In the formula: v is the learnable weight vector of attention score; h t denoted as the hidden state in the model; tanh(·) is the hyperbolic tangent activation function; W is the weight matrix, b is the bias vector; exp(·) is the exponential function.

[0035] Preferably, step S5 specifically includes:

[0036] S51. During the model training phase, the prediction error at each time step is calculated based on a known set of historical data;

[0037] S52. Calculate the mean and standard deviation of all error samples, and set upper and lower limit thresholds to determine whether the current error deviates from the normal range, thereby realizing early warning of the vibration status of hydropower units.

[0038] Preferably, step S51 specifically includes:

[0039] During the model training phase, the prediction error ε at each time step is calculated based on a set of known historical data. t ,Right now: In the formula: y t It is the actual value. This is a predicted value.

[0040] Preferably, step S52 specifically includes:

[0041] Calculate the mean μ and standard deviation σ for all error samples, i.e.: Set upper limit threshold T up With lower threshold T low This is used to determine whether the current error deviates from the normal range, i.e.: T up =μ+3σ,T low =μ-3σ.

[0042] This invention also discloses a hydropower unit vibration signal trend prediction system, which executes the above-described hydropower unit vibration signal trend prediction method and includes the following modules:

[0043] Vibration data acquisition module: Used to collect vibration data of hydropower units under different operating conditions in real time;

[0044] Data preprocessing module: Used to receive raw signal data from vibration data acquisition module, and sequentially perform signal noise reduction and sliding window segmentation. The sliding window is set with a length of L and a step size of S to segment the signal sequence and form a time series sample with a uniform structure.

[0045] Feature extraction module: used to add white noise to the preprocessed vibration time series and perform signal decomposition using the EEMD method;

[0046] Model prediction module: used for time series modeling and trend prediction based on bidirectional long short-term memory networks and attention mechanisms;

[0047] Trend warning module: used to dynamically determine whether there are trend anomalies based on the error between the model prediction results and the actual observed values.

[0048] Beneficial effects of this invention:

[0049] This invention addresses the problems of insufficient feature extraction, poor modeling effect, and low prediction accuracy in existing hydropower unit vibration signal prediction by combining signal decomposition technology, deep temporal modeling network, and attention mechanism. It improves the accuracy and stability of predicting the vibration state of hydropower units under complex operating conditions, and enables early identification of abnormal trends and warning of operational risks.

[0050] This invention utilizes ensemble empirical mode decomposition to decompose the original vibration signal and extract different frequency components; then, it constructs a bidirectional long short-term memory neural network to perform time-series modeling on the decomposed signal, capturing the bidirectional dynamic characteristics of the vibration signal; finally, it introduces an attention mechanism to assign weights to the features, improving the model's sensitivity to changes in vibration trends and its prediction accuracy, thus achieving more accurate and efficient vibration prediction and health status assessment under complex working conditions. Attached Figure Description

[0051] Figure 1 A flowchart of a method for predicting the vibration trend of a hydroelectric generator unit provided in an embodiment of the present invention;

[0052] Figure 2 This is a diagram of the EEMD network architecture provided in an embodiment of the present invention;

[0053] Figure 3 A diagram of the BiLSTM network structure provided in an embodiment of the present invention;

[0054] Figure 4 The vibration signal EEMD decomposition results provided in the embodiments of the present invention;

[0055] Figure 5 The prediction results of EEMD-BiLSTM-Attention provided in the embodiments of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0057] Example 1: As Figure 1 As shown, a method for predicting the vibration trend of a hydropower unit includes the following steps:

[0058] Step 1: Collect vibration signals of key components during the operation of the hydropower unit, and perform sliding window segmentation preprocessing on the original vibration signals to obtain vibration time series samples with uniform structure;

[0059] Step 2: Add different white noise levels with an amplitude of 0.1 times the signal standard deviation to each segment of the preprocessed vibration time series, constructing a total of 100 noisy signals. Then, use the EEMD algorithm to decompose each noisy signal into 7 IMF components. For each modal component IMF, take the average of the 100 decomposed IMFs of the same order as the final IMF output result. The calculation formula is: k = 1, 2, ..., 7, align these IMF components according to time steps to construct a feature matrix, forming the model input feature set;

[0060] Step 3: Divide the constructed IMF feature matrix data into training and test sets in a 7:3 ratio;

[0061] Step 4: Construct a vibration trend prediction model based on a combination of a Bidirectional Long Short-Term Memory (BiLSTM) network and an attention mechanism. The specific network structure is as follows: the BiLSTM network has 128 hidden layer nodes and two layers. The BiLSTM network extracts the bidirectional temporal features of the vibration signal in both forward and backward directions. An attention mechanism is introduced after the output layer of the BiLSTM, with the dimension of the attention mechanism set to be the same as the number of hidden layer nodes, i.e., 128 dimensions. For the hidden state at each time step, its relevance score is calculated using the attention mechanism. After normalization using the softmax function, the attention weights for each time step are obtained. Finally, the hidden states at all time steps are weighted and summed based on the attention weights to output the final prediction result.

[0062] Step 5: Train the constructed BiLSTM-Attention model on the training set. During training, mean squared error is used as the loss function, and the Adam optimization algorithm is used to optimize the network parameters. Then, based on the statistical characteristics of the prediction error, the mean and standard deviation of the prediction error are calculated, and then the upper and lower bounds of the prediction error, T, are calculated. up =μ+3σ,T low =μ-3σ;

[0063] Step 6: Input the test set data into the trained prediction model. If the error of the prediction result exceeds the set dynamic threshold range, it is determined that there is an abnormal trend in the current vibration sequence, thereby realizing early warning of the vibration status of the hydropower unit.

[0064] Example 2:

[0065] Based on the above method, this embodiment provides a hydropower unit vibration trend prediction system, which mainly includes:

[0066] Vibration Data Acquisition Module: This module is responsible for real-time acquisition of vibration data of the hydropower unit under different operating conditions. The system uses a high-precision CRY2301 noise sensor with a sampling frequency set to 10240Hz. The data includes the upper guide, lower guide, water guide, thrust swing, etc. of the hydropower unit.

[0067] Data preprocessing module: This module receives raw signal data from the vibration data acquisition module and performs signal noise reduction and sliding window segmentation in sequence. The sliding window is set with a length of L and a step size of S to segment the signal sequence and form a time series sample with a uniform structure.

[0068] Feature extraction module: This module adds multiple sets of white noise with an amplitude of 0.1 times the standard deviation to the preprocessed vibration time series and performs signal decomposition using the EEMD method. Each signal is decomposed into 6 IMFs and 1 residual term. For each IMF, the average of all noise-added results is taken to form the final feature input. The final IMF feature matrix, constructed from k = 1, 2, ..., 6, serves as the input to the neural network model.

[0069] Model Prediction Module: This module utilizes a bidirectional long short-term memory network and an attention mechanism to achieve time series modeling and trend prediction. The network structure consists of two layers of BiLSTM, with 128 hidden units in each layer. The output is connected to an attention mechanism of the same dimension. This module supports adaptive adjustment of network parameters based on the complexity of the actual vibration signal and the required prediction accuracy. It outputs the predicted value after weighted integration of the hidden states at each time step.

[0070] Trend Early Warning Module: This module dynamically determines whether there are trend anomalies based on the error between the model's predictions and actual observations. During the model training phase, the error discrimination threshold range is determined by statistically analyzing the mean μ and standard deviation σ of the training set error: Upper threshold = μ + 3σ, Lower threshold = μ - 3σ. Upper threshold T up =μ+3σ lower limit threshold T low =μ-3σ. When the prediction error exceeds this range, it is considered that a potential abnormal trend has occurred, and the system triggers an early warning signal to ensure the safe and stable operation of the hydropower unit.

[0071] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for predicting the vibration signal trend of a hydroelectric generator unit, characterized in that, Includes the following steps: S1. Collect vibration signals from key components during the operation of the hydropower unit, and use a standardized preprocessing method for the vibration signals, as shown in the following formula, to generate vibration time series samples x′ with uniform structure; In the formula: μ is the sample mean, and σ is the sample standard deviation; S2. Apply the EEMD algorithm to each preprocessed vibration time series to decompose the non-stationary signal into multiple intrinsic mode components (IMFs), and align the multiple IMFs according to time steps to construct a feature matrix, which serves as the input for subsequent models. S3. Divide the constructed dataset into a training set and a test set. The training set is used for model training and threshold construction, while the test set is used to verify model performance and warning accuracy. S4. Construct a time series prediction model based on BiLSTM and attention mechanism: Take the IMF feature matrix as input, extract the bidirectional time series features of each time step through BiLSTM, and introduce an attention mechanism at the output to weight the feature vectors of each time step; S5. Train the model on the training set, using the mean squared error as the loss function, and use the Adam optimizer to optimize the network parameters; calculate the mean and standard deviation of the prediction error based on the training set, and set the upper and lower limit thresholds for early warning judgment. S6. Apply the trained model to the test set to predict unknown vibration sequences. If the prediction error exceeds the set threshold range, it is considered an abnormal trend.

2. The method for predicting the vibration signal trend of a hydropower unit according to claim 1, characterized in that, In step S1, the standardized preprocessing method for vibration signals includes noise reduction, normalization, and sliding window segmentation.

3. The method for predicting the vibration signal trend of a hydropower unit according to claim 1, characterized in that, Step S2 specifically includes: S21. Add different types of white noise to the original signal to construct multiple noisy signals; S22. Perform EMD decomposition on each group of noisy signals to obtain multiple IMF sequences; S23. For each modal component, take its average value across all calculations as the final IMF output.

4. The method for predicting the vibration signal trend of a hydropower unit according to claim 3, characterized in that, Step S21 specifically involves: processing different white noise n i The signals X(t) are added to the original signal X(t) respectively to construct multiple noisy signals X(t)+n. i (t).

5. The method for predicting the vibration signal trend of a hydropower unit according to claim 3, characterized in that, In step S23, the IMF output result is: In the formula: c j (t) represents the final output signal of the j-th order IMF at time t; M is the total number of noisy signals; c ij (t) is the amplitude of the j-th order IMF component obtained by EEMD decomposition of the i-th group of noisy signals at time t.

6. The method for predicting the vibration signal trend of a hydropower unit according to claim 1, characterized in that, Step S4 specifically includes: S41. Based on BiLSTM, the vibration signal feature matrix is ​​modeled, and combined with the forward and backward information of the time series, the learning ability of long-term dependent features and trend changes is enhanced. S42. Given an input sequence X = {x1, x2, ..., x...} T The BiLSTM network consists of two LSTM units operating in opposite directions, performing forward and backward propagation respectively, and outputting the forward hidden state. and backward hidden state The outputs from both directions are concatenated at time step t to obtain the final hidden state output: S43. For a given sequence of hidden states output by a BiLSTM, the attention mechanism first calculates the relevance score for each time step, performs normalization processing to obtain the attention weight, and then weights the hidden states of all time steps according to the attention weight to calculate the context representation vector.

7. The method for predicting the vibration signal trend of a hydropower unit according to claim 6, characterized in that, In step S42, the internal calculation process of the LSTM unit is shown in the following formula: f t =σ(W f [h t-1 ,x t ]+b f ) i t =σ(W i [h t-1 ,x t ]+b i ) the t =σ(W o [h t-1 ,x t ]+b o ) In the formula: f t It is the forgetting gate vector; σ is the sigmoid activation function; b f Forget gate bias term; W f Here are the weight matrix of the forget gate; the input vector at time step t; and h. t-1 It is the hidden state vector of the previous time step; x t i is the input vector at time step t; t It is the input gate vector; W i This is the weight matrix of the input gate; b i It is the bias term of the input gate; o t It is the output gate vector; W o It is the weight matrix of the output gate, b o It is the bias term of the output gate; It is the candidate cell state; W c In candidate cell state calculation, it is used to input x at the current time step. t The weight matrix for linear transformation; b c It is the bias term of the candidate cell state; C t It is the unit state vector at time step t; C t-1 It is the unit state vector of the previous time step; ⊙ is the Hadamard product, multiplied element by element.

8. The method for predicting the vibration signal trend of a hydropower unit according to claim 6, characterized in that, Step S43 specifically involves: Given the hidden state sequence h of the BiLSTM output. t The attention mechanism first calculates the relevance score e for each time step t. t That is: e t =v T tanh(Wh t +b), all scores e t Normalization yields attention weights. The hidden states at all time steps are weighted according to the attention weights to calculate the context representation vector. In the formula: v is the learnable weight vector of attention score; h t denoted as the hidden state in the model; tanh(·) is the hyperbolic tangent activation function; W is the weight matrix, b is the bias vector; exp(·) is the exponential function.

9. The method for predicting the vibration signal trend of a hydropower unit according to claim 1, characterized in that, Step S5 specifically includes: S51. During the model training phase, the prediction error at each time step is calculated based on a known set of historical data; S52. Calculate the mean and standard deviation of all error samples, and set upper and lower limit thresholds to determine whether the current error deviates from the normal range, thereby realizing early warning of the vibration status of hydropower units.

10. The method for predicting the vibration signal trend of a hydropower unit according to claim 9, characterized in that, Step S51 specifically involves: During the model training phase, the prediction error ε at each time step is calculated based on a set of known historical data. t ,Right now: In the formula: y t It is the actual value. This is a predicted value.

11. The method for predicting the vibration signal trend of a hydropower unit according to claim 10, characterized in that, Step S52 specifically involves: Calculate the mean μ and standard deviation σ for all error samples, i.e.: Set upper limit threshold T up With lower threshold T low This is used to determine whether the current error deviates from the normal range, i.e.: T up =μ+3σ,T low =μ-3σ.

12. A hydropower unit vibration signal trend prediction system, executing the hydropower unit vibration signal trend prediction method according to any one of claims 1 to 11, characterized in that: Includes the following modules: Vibration data acquisition module: Used to collect vibration data of hydropower units under different operating conditions in real time; Data preprocessing module: Used to receive raw signal data from vibration data acquisition module, and sequentially perform signal noise reduction and sliding window segmentation. The sliding window is set with a length of L and a step size of S to segment the signal sequence and form a time series sample with a uniform structure. Feature extraction module: used to add white noise to the preprocessed vibration time series and perform signal decomposition using the EEMD method; Model prediction module: used for time series modeling and trend prediction based on bidirectional long short-term memory networks and attention mechanisms; Trend warning module: used to dynamically determine whether there are trend anomalies based on the error between the model prediction results and the actual observed values.

Images