A power station equipment state trend prediction method based on deep learning

Abstract

The application provides a power station equipment state trend prediction method based on deep learning, comprising the following steps: S1, obtaining power station equipment historical operation data and preprocessing; S2, using a deep neural network, constructing a trend early warning model, a temperature time series prediction submodel and a vibration time series prediction submodel based on the power station equipment historical operation data; S3, sequentially outputting a degradation degree prediction value, a temperature trend prediction value and a vibration trend prediction value through the model; S4, establishing a three-layer evaluation system and generating a state evaluation result; S5, generating early warning information according to the state evaluation result. The application constructs a multi-level deep network, realizes multi-dimensional accurate prediction of equipment degradation, temperature and vibration trends, introduces season coding, self-adaptive attention and other technologies to enhance the perception and learning ability under complex working conditions, innovates a three-layer dynamic evaluation system, provides scientific and healthy state evaluation from components to the whole, and provides a more comprehensive and intelligent solution for power station early warning.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring technology, specifically relating to a deep learning-based method for predicting the condition trends of power plant equipment. Background Technology

[0002] As a critical infrastructure of the power system, the operating status of power plant equipment is crucial to the safety and stability of power production. Traditional methods for predicting equipment status trends mainly rely on human experience and simple statistical models, which have technical limitations such as low prediction accuracy, poor adaptability to complex operating conditions, and inaccurate judgment of future development trends.

[0003] With the continuous development of intelligent, large-scale, and complex power equipment, traditional prediction methods are no longer sufficient to accurately depict the evolution patterns and potential risks of equipment condition. Existing technologies still face many challenges in identifying long-term evolution trends of equipment condition, accurately predicting degradation processes, and assessing future conditions in multiple dimensions. There is an urgent need to introduce advanced deep learning technology to improve the scientific rigor and early warning capabilities of equipment condition trend prediction. Summary of the Invention

[0004] The main objective of this invention is to provide a deep learning-based method for predicting the state trends of power plant equipment, addressing the aforementioned problems.

[0005] Therefore, the above-mentioned objective of the present invention is achieved through the following technical solution:

[0006] A deep learning-based method for predicting the state trends of power plant equipment includes the following steps:

[0007] S1. Obtain historical operating data of power plant equipment and preprocess the historical operating data of power plant equipment;

[0008] S2. Using deep neural networks, a trend early warning model and a time-series prediction model are constructed based on historical operating data of power plant equipment. The time-series prediction model includes a temperature time-series prediction sub-model and a vibration time-series prediction sub-model.

[0009] S3. Through the trend early warning model, the temperature time series prediction sub-model and the vibration time series prediction sub-model, the predicted values ​​of deterioration degree, temperature trend and vibration trend are output sequentially.

[0010] S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate state evaluation results based on the predicted values ​​of deterioration, temperature trends, and vibration trends.

[0011] S5. Generate early warning information based on the status evaluation results.

[0012] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions:

[0013] As a preferred technical solution of the present invention: In step S1, the historical operating data of the power station equipment includes operating condition data, vibration data, historical deterioration data, measuring point temperature data, temperature trend data of key components, and vibration trend data of key parts.

[0014] As a preferred technical solution of the present invention: in step S1, the preprocessing includes data cleaning, data standardization and data synchronization.

[0015] As a preferred technical solution of the present invention: step S2, constructing the trend early warning model includes the following sub-steps:

[0016] S211. Obtain the preprocessed working condition data, oscillation data and corresponding historical degradation data, and construct the first training sample set and the first verification sample set.

[0017] S212. Construct a deep long short-term memory neural network, which includes an input layer, a multimodal adaptive dropout layer, a coupled bidirectional LSTM layer, a fully connected layer, and an output layer. The multimodal adaptive dropout layer adaptively adjusts the dropout strategy according to the different feature distributions of the working condition data and the oscillation data. An information interaction mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer.

[0018] S213. The deep long short-term memory neural network is trained using the backpropagation algorithm, the network parameters are updated using an adaptive optimizer, and the gradient values ​​are subject to threshold constraints.

[0019] S214. Calculate the first validation loss value on the first validation sample set. When the decrease of the first validation loss value within a preset number of consecutive rounds is less than the preset degradation threshold, stop training and save the current network parameters.

[0020] S215. Use the trained deep long short-term memory neural network as a trend early warning model.

[0021] As a preferred technical solution of the present invention: Step S2, constructing the temperature time series prediction sub-model includes the following sub-steps:

[0022] S221. Based on the preprocessed working condition data, measuring point temperature data and corresponding key component temperature trend data, seasonal coding information is used as an additional feature to construct a second training sample set and a second verification sample set.

[0023] S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention computation layer, a fully connected layer, and an output layer; wherein, the first GRU layer is used to extract temporal features, the second GRU layer is used to fuse working conditions and seasonal information, and the attention computation layer includes a weight matrix and a scaling factor.

[0024] S223. Input the second training sample set into the gated recurrent neural network, use the RMSprop optimizer and update the network parameters based on the temperature change trend constraint;

[0025] S224. Calculate the second validation loss value on the second validation sample set. When the rate of change of the second validation loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stop training and save the current network parameters.

[0026] S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0027] As a preferred technical solution of the present invention: Step S223, updating the network parameters using the RMSprop optimizer and based on temperature change trend constraints, includes the following sub-steps:

[0028] S2231. Calculate the difference value of the temperature data of the measurement points in the training sample to generate a temperature change trend sequence;

[0029] S2232. Construct a combined loss function based on the temperature change trend sequence. The combined loss function integrates the mean squared error loss term, the trend correlation loss term, and the magnitude penalty term in a weighted manner.

[0030] S2233. Calculate the gradient value of the combined loss function using the RMSprop optimizer, iteratively update the network parameters based on the gradient value, and restrict the gradient norm.

[0031] As a preferred technical solution of the present invention: Step S2, constructing the vibration time series prediction sub-model includes the following sub-steps:

[0032] S231. Obtain working condition data, oscillation data, and corresponding vibration trend data of key parts, and construct a third training sample set and a third verification sample set;

[0033] S232. Construct a multi-scale gated recurrent neural network, which includes an input layer, a multi-scale GRU layer, a feature fusion layer, and an output layer. The multi-scale GRU layer uses sliding windows at different time scales to extract vibration features, and the feature fusion layer integrates multi-scale vibration features with working condition information.

[0034] S233. Input the third training sample set into the multi-scale gated recurrent neural network, and update the network parameters using the Adam optimizer combined with the attention mechanism.

[0035] S234. Calculate the third validation loss value on the third validation sample set. When the fluctuation of the third validation loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stop training and save the current network parameters.

[0036] S235. The trained multi-scale gated recurrent neural network is used as a vibration time series prediction sub-model.

[0037] As a preferred embodiment of the present invention, step S5 further includes the following sub-steps:

[0038] S51. Input the predicted values ​​of degradation, temperature trend, and vibration trend into the three-layer evaluation system;

[0039] S52. In the state quantity evaluation layer, calculate the deduction value of each state quantity based on the deduction system;

[0040] S53. In the component evaluation layer, the evaluation level of the component is determined based on the individual deduction range of the state quantity. The evaluation level includes normal, attention, abnormal and serious.

[0041] S54. At the overall evaluation level, when all component evaluation levels are normal, the overall status is determined to be normal; when there are multiple components with different evaluation levels, the overall status is determined according to the status level progression mechanism.

[0042] As a preferred technical solution of the present invention: in step S5, the early warning information includes the equipment number, evaluation level, key status quantity value, change trend, and early warning time, and generates processing suggestions based on historical maintenance experience.

[0043] Compared with existing technologies, this invention has the following advantages: First, it innovatively constructs a multi-layered deep neural network model, including a trend early warning model, a temperature time series prediction sub-model, and a vibration time series prediction sub-model, achieving multi-dimensional and accurate prediction of equipment degradation, temperature, and vibration trends. Second, by introducing technologies such as seasonal coding, adaptive attention mechanism, and multi-scale feature extraction, it effectively enhances the model's ability to perceive and learn changes in equipment status under complex operating conditions. Third, through an innovatively designed three-layer evaluation system and dynamic weight mechanism, it constructs a scientific and reliable health status evaluation method from state variables and components to the whole system, providing a more comprehensive and intelligent technical solution for power plant equipment status early warning, demonstrating significant technological innovation and practical value. Attached Figure Description

[0044] Figure 1The flowchart shows the deep learning-based power plant equipment status trend prediction method provided by this invention.

[0045] Figure 2 A flowchart for constructing a temperature time series prediction sub-model.

[0046] Figure 3 A flowchart for updating network parameters. Detailed Implementation

[0047] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0048] like Figure 1 As shown, a deep learning-based method for predicting the state trends of power plant equipment includes the following steps:

[0049] S1. Obtain historical operating data of power plant equipment and preprocess the historical operating data of power plant equipment;

[0050] Historical operating data of power plant equipment includes operating condition data, vibration and sway data, historical deterioration data, temperature data at measuring points, temperature trend data of key components, and vibration trend data of key parts.

[0051] Preprocessing includes data cleaning, data standardization, and data synchronization.

[0052] S2. Using deep neural networks, a trend early warning model and a time-series prediction model are constructed based on historical operating data of power plant equipment. The time-series prediction model includes a temperature time-series prediction sub-model and a vibration time-series prediction sub-model.

[0053] S211. Obtain the preprocessed working condition data, oscillation data and corresponding historical degradation data, and construct the first training sample set and the first verification sample set.

[0054] S212. Construct a deep long short-term memory neural network, which includes an input layer, a multimodal adaptive dropout layer, a coupled bidirectional LSTM layer, a fully connected layer, and an output layer. The multimodal adaptive dropout layer adaptively adjusts the dropout strategy according to the different feature distributions of the working condition data and the oscillation data. An information interaction mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer.

[0055] S213. The deep long short-term memory neural network is trained using the backpropagation algorithm, the network parameters are updated using an adaptive optimizer, and the gradient values ​​are subject to threshold constraints.

[0056] S214. Calculate the first validation loss value on the first validation sample set. When the decrease of the first validation loss value within a preset number of consecutive rounds is less than the preset degradation threshold, stop training and save the current network parameters.

[0057] S215. Use the trained deep long short-term memory neural network as a trend early warning model.

[0058] S221. Based on the preprocessed working condition data, measuring point temperature data and corresponding key component temperature trend data, seasonal coding information is used as an additional feature to construct a second training sample set and a second verification sample set.

[0059] S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention computation layer, a fully connected layer, and an output layer; wherein, the first GRU layer is used to extract temporal features, the second GRU layer is used to fuse working conditions and seasonal information, and the attention computation layer includes a weight matrix and a scaling factor.

[0060] S223. Input the second training sample set into the gated recurrent neural network, use the RMSprop optimizer and update the network parameters based on the temperature change trend constraint;

[0061] S2231. Calculate the difference value of the temperature data of the measurement points in the training sample to generate a temperature change trend sequence;

[0062] S2232. Construct a combined loss function based on the temperature change trend sequence. The combined loss function integrates the mean squared error loss term, the trend correlation loss term, and the magnitude penalty term in a weighted manner.

[0063] S2233. Calculate the gradient value of the combined loss function using the RMSprop optimizer, iteratively update the network parameters based on the gradient value, and restrict the gradient norm.

[0064] S224. Calculate the second validation loss value on the second validation sample set. When the rate of change of the second validation loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stop training and save the current network parameters.

[0065] S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0066] S231. Obtain working condition data, oscillation data, and corresponding vibration trend data of key parts, and construct a third training sample set and a third verification sample set;

[0067] S232. Construct a multi-scale gated recurrent neural network, which includes an input layer, a multi-scale GRU layer, a feature fusion layer, and an output layer. The multi-scale GRU layer uses sliding windows at different time scales to extract vibration features, and the feature fusion layer integrates multi-scale vibration features with working condition information.

[0068] S233. Input the third training sample set into the multi-scale gated recurrent neural network, and update the network parameters using the Adam optimizer combined with the attention mechanism.

[0069] S234. Calculate the third validation loss value on the third validation sample set. When the fluctuation of the third validation loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stop training and save the current network parameters.

[0070] S235. The trained multi-scale gated recurrent neural network is used as a vibration time series prediction sub-model.

[0071] S3. Through the trend early warning model, the temperature time series prediction sub-model and the vibration time series prediction sub-model, the predicted values ​​of deterioration degree, temperature trend and vibration trend are output sequentially.

[0072] S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate state evaluation results based on the predicted values ​​of deterioration, temperature trends, and vibration trends.

[0073] S5. Generate early warning information based on the status evaluation results.

[0074] S51. Input the predicted values ​​of degradation, temperature trend, and vibration trend into the three-layer evaluation system;

[0075] S52. In the state quantity evaluation layer, calculate the deduction value of each state quantity based on the deduction system;

[0076] S53. In the component evaluation layer, the evaluation level of the component is determined based on the individual deduction range of the state quantity. The evaluation level includes normal, attention, abnormal and serious.

[0077] S54. At the overall evaluation level, when all component evaluation levels are normal, the overall status is determined to be normal; when there are multiple components with different evaluation levels, the overall status is determined according to the status level progression mechanism.

[0078] The early warning information includes important information such as equipment number, evaluation level, key status value, change trend, and warning time, and generates handling suggestions based on historical maintenance experience.

[0079] Example:

[0080] S1. Obtain historical operating data of power plant equipment and preprocess the historical operating data of power plant equipment.

[0081] Historical operating data for power plant equipment includes operating condition data, vibration and sway data, historical degradation data, temperature data at measuring points, temperature trend data for key components, and vibration trend data for key parts. Operating condition data includes head data, active power data, speed data, and operating mode (pumping / generating). Vibration and sway data includes vibration data and sway data. Temperature trend data for key components includes thrust bearing oil surface temperature trend data, stator coil temperature trend data, water guide bearing bearing temperature trend data, and main transformer oil surface temperature trend data.

[0082] Preprocessing includes three steps: data cleaning, data standardization, and data synchronization. Data cleaning uses outlier identification methods, combined with the 2.5 to 3.5 standard deviation method and confidence interval analysis, to identify and remove outliers. At the same time, missing values ​​are filled using multiple interpolation methods. Data standardization uses the maximum and minimum value standardization method to normalize various types of data to the [0,1] interval. Data synchronization aligns data from different sampling frequencies to the same time series.

[0083] S2. Using deep neural networks, construct trend early warning models and time-series prediction models based on historical operating data of power plant equipment, including the following steps:

[0084] S21. Construct a trend early warning model, including the following steps:

[0085] S211. Obtain the preprocessed operating condition data, oscillation data, and corresponding historical degradation data to construct the first training sample set and the first verification sample set.

[0086] The preprocessed operating condition data and oscillation data were randomly divided into a first training sample set and a first validation sample set at an 8:2 ratio. The operating condition data and oscillation data were used as input data, and the corresponding historical degradation data were used as label data.

[0087] S212. Construct a deep long short-term memory neural network.

[0088] A deep long short-term memory neural network consists of an input layer, a multimodal adaptive dropout layer, a coupled bidirectional LSTM layer, a fully connected layer, and an output layer.

[0089] The number of neurons in the input layer is the same as the feature dimensions of the operating condition data and the oscillation data. The multimodal adaptive Dropout layer is set between the input layer and the coupled bidirectional LSTM layer, and adaptively adjusts the dropout strategy according to the different feature distributions of the operating condition data and the oscillation data. The coupled bidirectional LSTM layer includes forward LSTM units and backward LSTM units. The forward LSTM units process the data in the forward time series to capture the temporal dependencies from history to the present, and the backward LSTM units process the data in the reverse time series to capture the contextual information of the complete time series. An information exchange mechanism is set up between the forward and backward units to capture the temporal dependency features of device degradation through bidirectional processing and information exchange. The fully connected layer contains 32 neurons and uses the ReLU activation function. The output layer outputs the degradation prediction value.

[0090] The adaptive discarding strategy based on the different feature distributions of the operating condition data and the oscillation data includes: calculating the feature distribution differences between the operating condition data and the oscillation data respectively, setting corresponding inactivation rates for different modal data according to the distribution differences, and applying them to the corresponding data for neuron discarding. In addition, the information interaction mechanism is implemented by weighted fusion of the hidden state outputs of the forward LSTM unit and the backward LSTM unit, and the fused features are used as the final output of the coupled bidirectional LSTM layer.

[0091] The multimodal adaptive Dropout layer sets the inactivation rate according to the difference in the characteristic distribution of the working condition data and the oscillation data, realizing differentiated regularization processing for different modal data; the forward and backward units of the coupled bidirectional LSTM layer perform hidden state fusion through information interaction mechanism, which enhances the utilization of bidirectional temporal information; the simplified network structure reduces intermediate layer processing links and reduces computational complexity.

[0092] S213. A deep long short-term memory neural network is trained using the backpropagation algorithm, and an adaptive optimizer is used to update the network parameters, with a threshold limit applied to the gradient values.

[0093] First, the first training sample set is processed in batches, and the loss function value between the predicted output and the actual degradation is calculated using forward propagation. Then, based on the loss function value, the gradient values ​​of the parameters of each layer of the network are calculated using the backpropagation algorithm. Next, the calculated gradient values ​​are thresholded; when the absolute value of the gradient exceeds a preset threshold, it is truncated to a specified range to prevent gradient explosion. Subsequently, an adaptive optimizer is used to update the network parameters based on the processed gradient values. Finally, a learning rate scheduling strategy is used to optimize the training process, and the loss value is recorded during training to monitor the convergence of the network.

[0094] In a preferred embodiment, the batch size can be set to 8-32 samples, the gradient clipping threshold can be set to 1-5, the learning rate of the Adam optimizer can be set to 0.0001-0.01, and the learning rate decay interval can be set to 20-50 rounds.

[0095] S214. Calculate the first validation loss value on the first validation sample set. When the decrease of the first validation loss value within a preset number of consecutive rounds is less than the preset degradation threshold, stop training and save the current network parameters.

[0096] After each training round, the first validation sample set is input into the currently trained deep long short-term memory neural network. The mean squared error between the network output value and the actual degradation value is calculated to obtain the first validation loss value. The trend of the validation loss value is monitored. When the decrease within a preset number of consecutive rounds is less than a preset degradation threshold, the model is considered to have converged and training is stopped. The preset degradation threshold is determined by analyzing the distribution of the rate of change of equipment degradation in historical operating data. At the same time, an early stopping mechanism is set to prevent overfitting. When the validation loss value increases within a consecutive number of rounds, training is terminated and the optimal network parameters are saved.

[0097] In a preferred embodiment, the preset number of consecutive rounds can be set to 5-15 rounds, and the number of early stop rounds can be set to 3-8 rounds.

[0098] S215. Use the trained deep long short-term memory neural network as a trend early warning model.

[0099] The trend warning model is based on a network structure of multimodal adaptive Dropout layer and coupled bidirectional LSTM layer, which can handle the heterogeneous features of working condition data and oscillation data and capture the temporal dependency of equipment degradation. Through gradient threshold limitation and early stopping mechanism, the risks of gradient explosion and overfitting during training are reduced.

[0100] S22. Construct a temperature time series prediction sub-model, such as Figure 2 As shown, it includes the following steps:

[0101] S221. Based on the preprocessed operating condition data, measuring point temperature data, and corresponding key component temperature trend data, seasonal coding information is used as an additional feature to construct a second training sample set and a second verification sample set.

[0102] First, the preprocessed operating condition data and the measured temperature data are aligned chronologically. Then, seasonal coding information is generated based on the data acquisition time, using one-hot coding to convert the four seasons (spring, summer, autumn, and winter) into 4-dimensional binary vectors. Next, the operating condition data, measured temperature data, and seasonal coding information are concatenated along the feature dimension to form an input feature matrix, with the corresponding temperature trend data of key components serving as labels. Finally, the dataset is randomly divided into a second training sample set and a second validation sample set at an 8:2 ratio. The operating condition data and measured temperature data serve as input data, and the corresponding temperature trend data of key components serve as label data.

[0103] S222, Construct a gated recurrent neural network.

[0104] Gated recurrent neural networks consist of an input layer, a dual GRU layer, an attention computation layer, a fully connected layer, and an output layer.

[0105] The number of neurons in the input layer is the same as the dimension of the concatenated features; the first GRU layer contains 128 hidden units and uses the tanh activation function and gating mechanism to extract temporal features; the second GRU layer contains 64 hidden units and uses the tanh activation function and gating mechanism to fuse working conditions and seasonal information; the attention calculation layer uses an adaptive weight initialization method, including a weight matrix and a scaling factor. The weight matrix has a dimension of 64×64, learns the importance of different time steps through trainable parameters and normalizes it using the softmax function. The scaling factor is dynamically adjusted according to the data distribution to regulate the concentration of attention distribution; the fully connected layer uses the ReLU activation function; the number of neurons in the output layer is the same as the dimension of the temperature trend data of key components, and a linear activation function is used to predict the temperature trend.

[0106] S223. Input the second training sample set into the gated recurrent neural network, use the RMSprop optimizer and update the network parameters based on the temperature change trend constraint.

[0107] The network parameter update flowchart is as follows: Figure 3 As shown, the difference values ​​are calculated for the temperature data of the measurement points in the training samples to generate a temperature change trend sequence. Based on the temperature change trend sequence, a combined loss function is constructed. The combined loss function integrates the mean squared error loss term, the trend correlation loss term, and the amplitude penalty term in a weighted manner. The specific formula is as follows:

[0108]

[0109] In the formula, These are the weighting coefficients, which are dynamically adjusted based on the error changes during model training (in this embodiment, they are initially set to 0.6, 0.3, and 0.1). LMAEThe mean squared error loss term represents the squared Euclidean distance between the predicted temperature series and the actual temperature series. LCorr The trend correlation loss term is calculated by subtracting the Pearson correlation coefficient between the predicted trend and the actual trend from 1. LPen The amplitude penalty term represents the second difference value of the predicted temperature sequence. L2 Norm, the specific formula is as follows:

[0110]

[0111] In the formula, y t express t The actual temperature value at that moment. This represents the predicted temperature value at time t. n For sequence length, and These represent the mean values ​​of the actual temperature series and the predicted temperature series, respectively. and These represent the standard deviations of the actual temperature series and the predicted temperature series, respectively.

[0112] The second training sample set is divided into batches of 32 samples each, and the combined loss value for each batch is calculated using forward propagation. Then, the gradient value of the combined loss function is calculated using the RMSprop optimizer, where the learning rate of the optimizer is set to 0.001, the decay rate is set to 0.9, and the numerical stability constant is set to 1e-8. Next, the calculated gradient value is pruned according to its norm, limiting it to the interval [-5, 5]. Finally, the network parameters are iteratively updated based on the pruned gradient value.

[0113] S224. Calculate the second validation loss value on the second validation sample set. When the rate of change of the second validation loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stop training and save the current network parameters.

[0114] The temperature prediction threshold was determined by combining K-fold cross-validation on the training dataset with the experience of domain experts.

[0115] S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0116] The temperature time-series prediction sub-model improves the predictive ability of critical component temperature trends by introducing seasonal coding information, a two-layer GRU structure, and an attention mechanism, combined with a combined loss function based on temperature change trend constraints. This model can utilize seasonal features and historical temperature change patterns to capture the importance of different time steps, improving the temporal continuity and physical consistency of temperature predictions, and providing predictive support for identifying equipment temperature anomalies.

[0117] S23. Construct a vibration time series prediction sub-model.

[0118] S231. Obtain working condition data, oscillation data, and corresponding vibration trend data of key parts to construct a third training sample set and a third verification sample set.

[0119] Stratified random sampling was employed, dividing the collected data into a third training sample set and a third validation sample set at an 8:2 ratio to ensure the representativeness and randomness of the sample sets. Operating condition data and vibration data were used as input data, while vibration trend data of corresponding key components were used as label data.

[0120] S232. Construct a multi-scale gated recurrent neural network.

[0121] Multiscale gated recurrent neural networks consist of an input layer, a multiscale GRU layer, a feature fusion layer, and an output layer.

[0122] The multi-scale GRU layer sets sliding windows with 1, 3, and 5 time steps to capture short-term, medium-term, and long-term features of vibration signals, respectively. Each time scale is equipped with an independent gated recurrent unit. The multi-scale GRU layer extracts vibration features using sliding windows at different time scales and extracts local and global feature patterns of the signals by processing vibration signals at different time scales in parallel. The feature fusion layer integrates multi-scale vibration features with working condition information and adaptively adjusts the importance of features at different scales based on the attention weight mechanism according to the working condition information. The output layer generates vibration trend prediction results and maps the fused multi-scale features to continuous vibration trend prediction values.

[0123] S233. Input the third training sample set into the multi-scale gated recurrent neural network, and use the Adam optimizer combined with the attention mechanism to update the network parameters.

[0124] First, the data from the third training sample set is input into a multi-scale gated recurrent neural network in batches. In each training batch, forward propagation is performed to calculate the predicted output. Then, the mean squared error is used as the loss function to calculate the loss value between the predicted output and the actual vibration trend, and the gradient of the parameters of each layer of the network is calculated through the backpropagation algorithm. The Adam optimizer is used to adaptively update the network parameters based on the calculated gradient information and the attention mechanism, where the attention weights are dynamically adjusted according to the vibration characteristics and working conditions. During the parameter update process, the magnitude of parameter update is controlled by setting the learning rate and regularization term to suppress overfitting and maintain the generalization performance of the model. The above iterative process of forward propagation, loss calculation, backpropagation and parameter update is repeated until the full training of the third training sample set is completed.

[0125] S234. Calculate the third validation loss value on the third validation sample set. When the fluctuation range of the third validation loss value within the preset number of consecutive rounds is less than the vibration prediction threshold, stop training and save the current network parameters.

[0126] The vibration prediction threshold is determined by statistical variance and error analysis of historical vibration data, and is 0.025 in this embodiment.

[0127] S235. The trained multi-scale gated recurrent neural network is used as a vibration time series prediction sub-model.

[0128] The vibration time series prediction sub-model improves the predictive ability of equipment vibration trends by employing a multi-scale GRU layer with sliding windows of different time steps, an attention weight mechanism in the feature fusion layer, and a hierarchical random sampling training method. This model can capture both short-term and long-term characteristics of vibration signals, adjust feature weights based on operating condition information, improve the stability of vibration trend prediction, and provide predictive support for early warning of abnormal equipment vibration.

[0129] S3. Through the trend early warning model, temperature time series prediction sub-model and vibration time series prediction sub-model, the degradation degree prediction value, temperature trend prediction value and vibration trend prediction value are output sequentially.

[0130] S4. Establish a three-tier evaluation system to generate state evaluation results based on the predicted values ​​of deterioration, temperature trends, and vibration trends.

[0131] The three-tiered evaluation system comprises a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer. The process for generating the state evaluation results is as follows: Predicted degradation values, temperature trend values, and vibration trend values ​​are input into the three-tiered evaluation system. In the state quantity evaluation layer, a deduction value for each state quantity is calculated based on a deduction system. The deduction value equals the basic deduction value multiplied by its weight plus the family defect deduction, as shown in the following formula:

[0132]

[0133] In the formula, This is the basic deduction value. For dynamic weight values, Points will be deducted for family defects.

[0134] It should be noted that the basic deduction value is calculated using a piecewise linear function. Based on the deviation of the predicted value from each threshold level (normal, attention, abnormal, severe), the deduction value is calculated proportionally within different intervals. The greater the deviation of the predicted value from the normal range, the higher the deduction value. A reasonable upper limit is set to avoid excessive deductions in extreme cases. Dynamic weight values ​​are determined by analyzing the historical rate of change and predicted trends of the state quantity. Specifically, when the rate of change of the state quantity exceeds a preset threshold or the predicted trend shows a significant deterioration, the weight value of that state quantity is increased accordingly; conversely, when the state quantity remains stable or the predicted trend is positive, the weight value is appropriately decreased. The weight value range is [0.8, 1.2], with an initial weight of 1.0. The weight is updated monthly, calculated based on data within a sliding time window (e.g., 3 months).

[0135] In addition, the family defect penalty is determined based on three dimensions: the frequency of defect occurrence, the severity of the defect, and the speed of its evolution.

[0136] For the frequency of defect occurrence, 0-5 points are deducted when the frequency is less than 5%, 6-10 points are deducted when the frequency is 5%-15%, and 11-15 points are deducted when the frequency is greater than 15%.

[0137] For the severity of the defect, the following deductions apply: minor impact (not affecting normal equipment operation) will incur an additional deduction of 0-2 points; moderate impact (may lead to performance degradation) will incur an additional deduction of 3-4 points; and severe impact (may lead to malfunction and downtime) will incur an additional deduction of 5 points.

[0138] For the speed of defect evolution, no additional points will be deducted. For moderate (several weeks to several months) defects, 1 to 3 points will be deducted. For rapid (several days to several weeks) defects, 3 to 5 points will be deducted.

[0139] The final family defect deduction score is the sum of the deductions from the three dimensions, and cannot exceed the maximum of 20 points. If the score exceeds 20 points, then 20 points will be counted. This scoring standard can be dynamically adjusted based on actual operational data and expert experience.

[0140] The health status of components and the overall health is divided into four evaluation levels: normal, attention, abnormal, and severe. In the component evaluation level, the evaluation level of the component is determined based on the deduction range of individual status quantities, as shown in Table 1.

[0141] Table 1 Evaluation Level Determination Table

[0142]

[0143] To avoid frequent jumps near the critical value, a 5-point buffer mechanism is set up: when the deduction value is within ±2.5 points of the critical value, the evaluation level remains unchanged from the previous evaluation result.

[0144] At the overall evaluation level, when all components are rated as normal, the overall status is determined to be normal; when there are multiple components with different ratings, the overall status is determined according to the status level progression mechanism.

[0145] The status level progression mechanism is as follows: the overall status is comprehensively determined based on the combination of the statuses of each component. When any component is at the severe level, the overall status is directly determined to be severe; when two or more components are at the abnormal level, the overall status is determined to be severe; when only one component is at the abnormal level or three or more components are at the attention level, the overall status is determined to be abnormal; when only one or two components are at the attention level and the other components are at the normal level, the overall status is determined to be attention. In other cases, the rating process follows the "strictest principle," that is, when a situation may correspond to multiple levels, the most severe level is taken as the final rating result.

[0146] This three-tiered evaluation system achieves multi-level evaluation from individual state variables to components and then to the whole system. Through dynamic weighting and state level progression mechanisms, it accurately reflects the health status of equipment at each level, providing a scientific basis for early warning decisions. The system considers both the importance and changing trends of different state variables, as well as the mutual influence between components, ensuring the rationality and reliability of the evaluation results.

[0147] S5. Generate early warning information based on the status evaluation results.

[0148] Based on different overall status evaluation levels, the system automatically generates corresponding early warning messages. For example, when the evaluation level is "Attention," a reminder-level early warning message is generated, suggesting increased monitoring and inspection; when the evaluation level is "Abnormal," a warning-level early warning message is generated, suggesting timely inspection and necessary preventative measures; when the evaluation level is "Severe," an alarm-level early warning message is generated, suggesting immediate shutdown for inspection and maintenance. The early warning message includes important information such as equipment number, evaluation level, key status quantity values, change trends, and warning time, and automatically generates handling suggestions based on historical maintenance experience.

[0149] Practical application verification shows that the method of this invention has significant advantages in power plant equipment condition monitoring, exhibiting excellent prediction accuracy, temperature and vibration prediction precision, equipment anomaly identification capability, and model training efficiency. In practical applications, the system can generate tiered early warning information, enabling maintenance personnel to take timely preventive maintenance measures before equipment failures occur, effectively reducing the risk of unexpected equipment downtime and providing reliable technical support for accurate health status assessment of power plant equipment.

[0150] The technical solution of the present invention has been described in conjunction with the specific experimental procedures shown in the accompanying drawings. However, the scope of protection of the present invention is not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions resulting from such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the state trend of power plant equipment based on deep learning, characterized in that, Includes the following steps: S1. Obtain historical operating data of power plant equipment and preprocess the historical operating data of power plant equipment; S2. Using deep neural networks, a trend early warning model and a time-series prediction model are constructed based on historical operating data of power plant equipment. The time-series prediction model includes a temperature time-series prediction sub-model and a vibration time-series prediction sub-model. S3. Through the trend early warning model, the temperature time series prediction sub-model and the vibration time series prediction sub-model, the predicted values ​​of deterioration degree, temperature trend and vibration trend are output sequentially. S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate state evaluation results based on the predicted values ​​of deterioration, temperature trends, and vibration trends. S5. Generate early warning information based on the status evaluation results; Building a trend early warning model includes the following sub-steps: S211. Obtain the preprocessed working condition data, oscillation data and corresponding historical degradation data, and construct the first training sample set and the first verification sample set. S212. Construct a deep long short-term memory neural network, which includes an input layer, a multimodal adaptive dropout layer, a coupled bidirectional LSTM layer, a fully connected layer, and an output layer. The multimodal adaptive dropout layer adaptively adjusts the dropout strategy according to the different feature distributions of the working condition data and the oscillation data. An information interaction mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer. S213. The deep long short-term memory neural network is trained using the backpropagation algorithm, the network parameters are updated using an adaptive optimizer, and the gradient values ​​are subject to threshold constraints. S214. Calculate the first validation loss value on the first validation sample set. When the decrease of the first validation loss value within a preset number of consecutive rounds is less than the preset degradation threshold, stop training and save the current network parameters. S215. Use the trained deep long short-term memory neural network as a trend early warning model.

2. The method according to claim 1, characterized in that: In step S1, the historical operating data of the power plant equipment includes operating condition data, vibration and sway data, historical deterioration data, temperature data at measuring points, temperature trend data of key components, and vibration trend data of key parts.

3. The method according to claim 1, characterized in that: In step S1, preprocessing includes data cleaning, data standardization, and data synchronization.

4. The method according to claim 1, characterized in that: In step S2, constructing the temperature time series prediction sub-model includes the following sub-steps: S221. Based on the preprocessed working condition data, measuring point temperature data and corresponding key component temperature trend data, seasonal coding information is used as an additional feature to construct a second training sample set and a second verification sample set. S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention computation layer, a fully connected layer, and an output layer; wherein, the first GRU layer is used to extract temporal features, the second GRU layer is used to fuse working conditions and seasonal information, and the attention computation layer includes a weight matrix and a scaling factor. S223. Input the second training sample set into the gated recurrent neural network, use the RMSprop optimizer and update the network parameters based on the temperature change trend constraint; S224. Calculate the second validation loss value on the second validation sample set. When the rate of change of the second validation loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stop training and save the current network parameters. S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

5. The method according to claim 4, characterized in that: Step S223, which involves using the RMSprop optimizer and updating the network parameters based on temperature change trend constraints, includes the following sub-steps: S2231. Calculate the difference value of the temperature data of the measurement points in the training sample to generate a temperature change trend sequence; S2232. Construct a combined loss function based on the temperature change trend sequence. The combined loss function integrates the mean squared error loss term, the trend correlation loss term, and the magnitude penalty term in a weighted manner. S2233. Calculate the gradient value of the combined loss function using the RMSprop optimizer, iteratively update the network parameters based on the gradient value, and restrict the gradient norm.

6. The method according to claim 1, characterized in that: In step S2, constructing the vibration time series prediction sub-model includes the following sub-steps: S231. Obtain working condition data, oscillation data, and corresponding vibration trend data of key parts, and construct a third training sample set and a third verification sample set; S232. Construct a multi-scale gated recurrent neural network, which includes an input layer, a multi-scale GRU layer, a feature fusion layer, and an output layer. The multi-scale GRU layer uses sliding windows at different time scales to extract vibration features, and the feature fusion layer integrates multi-scale vibration features with working condition information. S233. Input the third training sample set into the multi-scale gated recurrent neural network, and update the network parameters using the Adam optimizer combined with the attention mechanism. S234. Calculate the third validation loss value on the third validation sample set. When the fluctuation of the third validation loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stop training and save the current network parameters. S235. The trained multi-scale gated recurrent neural network is used as a vibration time series prediction sub-model.

7. The method according to claim 1, characterized in that: Step S5 also includes the following sub-steps: S51. Input the predicted values ​​of degradation, temperature trend, and vibration trend into the three-layer evaluation system; S52. In the state quantity evaluation layer, calculate the deduction value of each state quantity based on the deduction system; S53. In the component evaluation layer, the evaluation level of the component is determined based on the individual deduction range of the state quantity. The evaluation level includes normal, attention, abnormal and serious. S54. At the overall evaluation level, when all component evaluation levels are normal, the overall status is determined to be normal; when there are multiple components with different evaluation levels, the overall status is determined according to the status level progression mechanism.

8. The method according to claim 1, characterized in that: In step S5, the early warning information includes the equipment number, evaluation level, key status quantity values, change trend, and early warning time, and generates handling suggestions based on historical maintenance experience.

Images