Power station equipment state trend prediction method based on deep learning

Abstract

The invention provides a power station equipment state trend prediction method based on deep learning, and the method comprises the following steps: S1, obtaining the historical operation data of power station equipment, and carrying out the preprocessing; s2, constructing a trend early warning model, a temperature time sequence prediction sub-model and a vibration time sequence prediction sub-model based on the historical operation data of the power station equipment by using a deep neural network; 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; and S5, generating early warning information according to a state evaluation result. According to the method, a multi-level deep network is constructed, and multi-dimensional accurate prediction of equipment degradation, temperature and vibration trend is realized; technologies of seasonal coding, adaptive attention and the like are introduced, so that the perception and learning capabilities under complex working conditions are enhanced; a three-layer dynamic evaluation system is innovated, scientific health state evaluation from parts to the whole is provided, and a more comprehensive and intelligent solution is provided for early warning of a power station.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment status monitoring, and specifically relates to a power station equipment status trend prediction method based on deep learning. Background Art

[0002] As critical infrastructure in power systems, the operational status of power plant equipment is crucial to the safety and stability of power production. Traditional methods for predicting equipment status trends rely primarily on manual experience and simple statistical models. These methods suffer from technical limitations such as low prediction accuracy, poor adaptability to complex operating conditions, and inaccurate predictions of future development trends.

[0003] As power equipment continues to become more intelligent, larger, and more complex, traditional forecasting methods are struggling to accurately capture the evolution of equipment status and potential risks. Existing technologies still face numerous challenges in identifying long-term trends in equipment status, accurately predicting degradation processes, and assessing multi-dimensional future status. Advanced deep learning techniques are urgently needed to enhance the scientific nature and early warning capabilities of equipment status trend forecasting. Summary of the Invention

[0004] The main purpose of the present invention is to provide a power plant equipment status trend prediction method based on deep learning to address the above-mentioned problems.

[0005] To this end, the above-mentioned purpose of the present invention is achieved through the following technical solutions: A method for predicting power plant equipment status trends based on deep learning, comprising the following steps: S1. Obtain historical operation data of power station equipment and preprocess the historical operation data of power station equipment; S2. Using deep neural networks, a trend warning model and a time series prediction model are constructed based on the 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. Outputting the degradation degree prediction value, temperature trend prediction value, and vibration trend prediction value in sequence through the trend warning model, temperature time series prediction sub-model, and vibration time series prediction sub-model; S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate a state evaluation result based on the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value; S5. Generate early warning information based on the status evaluation results.

[0006] While adopting the above technical solutions, the present invention may also adopt or combine the following technical solutions: As a preferred technical solution of the present invention: in step S1, the historical operation data of the power station equipment includes operating condition data, vibration data, historical degradation data, measuring point temperature data, key component temperature trend data, and key part vibration trend data.

[0007] As a preferred technical solution of the present invention: in step S1, preprocessing includes data cleaning, data standardization and data synchronization.

[0008] As a preferred technical solution of the present invention: in step S2, constructing a trend warning model includes the following sub-steps: S211, obtaining pre-processed working condition data, vibration data, and corresponding historical degradation data, and constructing a first training sample set and a first verification sample set; S212. Construct a deep long short-term memory neural network, the deep long short-term memory neural network including 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 vibration data; an information exchange mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer; S213, using a back propagation algorithm to train the deep long short-term memory neural network, using an adaptive optimizer to update network parameters, and setting a threshold limit on the gradient value; S214, calculating a first validation loss value on the first validation sample set, and when the first validation loss value decreases by less than a preset degradation threshold within a preset number of consecutive rounds, stopping training and saving the current network parameters; S215. Use the trained deep long short-term memory neural network as a trend warning model.

[0009] As a preferred technical solution of the present invention: in step S2, constructing the temperature time series prediction sub-model includes the following sub-steps: S221, based on the pre-processed operating condition data, the measuring point temperature data, and the corresponding key component temperature trend data, the seasonal coding information is used as an additional feature to construct a second training sample set and a second validation sample set; S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention calculation layer, a fully connected layer, and an output layer; wherein the first GRU layer is used to extract time series features, the second GRU layer is used to integrate working condition and seasonal information, and the attention calculation layer includes a weight matrix and a scaling factor; S223, inputting the second training sample set into the gated recurrent neural network, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint; S224, calculating a second verification loss value on the second verification sample set, and when the rate of change of the second verification loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stopping training and saving the current network parameters; S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0010] As a preferred technical solution of the present invention: in step S223, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint includes the following sub-steps: S2231, calculating differential values ​​for the temperature data of the measuring 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, where the combined loss function fuses the mean square error loss term, the trend correlation loss term, and the amplitude penalty term in a weighted manner; S2233. Calculate the gradient value of the combined loss function through the RMSprop optimizer, iteratively update the network parameters according to the gradient value, and limit the gradient norm.

[0011] As a preferred technical solution of the present invention: in step S2, constructing the vibration time series prediction sub-model includes the following sub-steps: S231, obtaining working condition data, vibration data, and corresponding key parts vibration trend data, and constructing 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; wherein the multi-scale GRU layer uses sliding windows of different time scales to extract vibration features, and the feature fusion layer integrates the multi-scale vibration features and working condition information; S233, inputting the third training sample set into the multi-scale gated recurrent neural network, and updating the network parameters using the Adam optimizer combined with the attention mechanism; S234, calculating a third verification loss value on the third verification sample set, and when the fluctuation amplitude of the third verification loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stopping training and saving the current network parameters; S235. Use the trained multi-scale gated recurrent neural network as a vibration time series prediction sub-model.

[0012] As a preferred technical solution of the present invention: step S5 further includes the following sub-steps: S51, inputting the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value into the three-tier evaluation system; S52. At the state quantity evaluation layer, the deduction value of each state quantity is calculated based on the deduction system; S53. At the component evaluation layer, the component evaluation level is determined based on the individual deduction intervals of the status quantity. The evaluation levels include normal, caution, abnormal, and severe. S54. At the overall evaluation level, when the evaluation levels of all components 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 progressive mechanism.

[0013] As a preferred technical solution of the present invention: in step S5, the warning information includes the equipment number, evaluation level, key state quantity value, change trend, warning time, and generates processing suggestions based on historical maintenance experience.

[0014] Compared with the existing technology, the present invention has the following beneficial effects: the present invention innovatively constructs a multi-level deep neural network model, including a trend warning model, a temperature time series prediction sub-model and a vibration time series prediction sub-model, to achieve multi-dimensional and accurate prediction of equipment degradation, temperature and vibration trends; secondly, the introduction of seasonal coding, adaptive attention mechanism and multi-scale feature extraction and other technologies effectively enhances the model's perception and learning ability of equipment status changes under complex working conditions; thirdly, through the innovatively designed three-tier evaluation system and dynamic weight mechanism, a scientific and reliable health status evaluation method from state quantity, components to the whole is constructed, providing a more comprehensive and intelligent technical solution for power station equipment status warning, which has significant technical innovation and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a flow chart of the power plant equipment status trend prediction method based on deep learning provided by the present invention.

[0016] Figure 2 Flowchart built for the temperature time series prediction submodel.

[0017] Figure 3 Flowchart for network parameter update. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, a power plant equipment status trend prediction method based on deep learning specifically includes the following steps: S1. Obtain historical operation data of power station equipment and preprocess the historical operation data of power station equipment; The historical operation data of power station equipment includes operating condition data, vibration data, historical degradation data, measuring point temperature data, key component temperature trend data, and key part vibration trend data.

[0020] Preprocessing includes data cleaning, data standardization and data synchronization.

[0021] S2. Using deep neural networks, a trend warning model and a time series prediction model are constructed based on the 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. S211, obtaining pre-processed working condition data, vibration data, and corresponding historical degradation data, and constructing a first training sample set and a first verification sample set; S212. Construct a deep long short-term memory neural network, the deep long short-term memory neural network including 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 vibration data; an information exchange mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer; S213, using a back propagation algorithm to train the deep long short-term memory neural network, using an adaptive optimizer to update network parameters, and setting a threshold limit on the gradient value; S214, calculating a first validation loss value on the first validation sample set, and when the first validation loss value decreases by less than a preset degradation threshold within a preset number of consecutive rounds, stopping training and saving the current network parameters; S215. Use the trained deep long short-term memory neural network as a trend warning model.

[0022] S221, based on the pre-processed operating condition data, the measuring point temperature data, and the corresponding key component temperature trend data, the seasonal coding information is used as an additional feature to construct a second training sample set and a second validation sample set; S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention calculation layer, a fully connected layer, and an output layer; wherein the first GRU layer is used to extract time series features, the second GRU layer is used to integrate working condition and seasonal information, and the attention calculation layer includes a weight matrix and a scaling factor; S223, inputting the second training sample set into the gated recurrent neural network, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint; S2231, calculating differential values ​​for the temperature data of the measuring 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, where the combined loss function fuses the mean square error loss term, the trend correlation loss term, and the amplitude penalty term in a weighted manner; S2233. Calculate the gradient value of the combined loss function through the RMSprop optimizer, iteratively update the network parameters according to the gradient value, and limit the gradient norm.

[0023] S224, calculating a second verification loss value on the second verification sample set, and when the rate of change of the second verification loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stopping training and saving the current network parameters; S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0024] S231, obtaining working condition data, vibration data, and corresponding key parts vibration trend data, and constructing 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; wherein the multi-scale GRU layer uses sliding windows of different time scales to extract vibration features, and the feature fusion layer integrates the multi-scale vibration features and working condition information; S233, inputting the third training sample set into the multi-scale gated recurrent neural network, and updating the network parameters using the Adam optimizer combined with the attention mechanism; S234, calculating a third verification loss value on the third verification sample set, and when the fluctuation amplitude of the third verification loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stopping training and saving the current network parameters; S235. Use the trained multi-scale gated recurrent neural network as a vibration time series prediction sub-model.

[0025] S3. Outputting the degradation degree prediction value, temperature trend prediction value, and vibration trend prediction value in sequence through the trend warning model, temperature time series prediction sub-model, and vibration time series prediction sub-model; S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate a state evaluation result based on the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value; S5. Generate early warning information based on the status evaluation results.

[0026] S51, inputting the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value into the three-tier evaluation system; S52. At the state quantity evaluation layer, the deduction value of each state quantity is calculated based on the deduction system; S53. At the component evaluation layer, the component evaluation level is determined based on the individual deduction intervals of the status quantity. The evaluation levels include normal, caution, abnormal, and severe. S54. At the overall evaluation level, when the evaluation levels of all components 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 progressive mechanism.

[0027] The early warning information includes important information such as equipment number, evaluation level, key status value, change trend, early warning time, etc., and generates processing suggestions based on historical maintenance experience.

[0028] Example: S1. Obtain historical operation data of power station equipment and preprocess the historical operation data of power station equipment.

[0029] Historical operating data for power plant equipment includes operating condition data, runout data, historical degradation data, temperature data at measurement points, temperature trends for key components, and vibration trends for key locations. Operating condition data includes head data, active power data, speed data, and operating mode (pumping / generating mode). Runout data includes vibration data and swing data. Key component temperature trends include thrust bearing oil level temperature trends, stator coil temperature trends, water guide bearing bearing temperature trends, and main transformer oil level temperature trends.

[0030] Preprocessing includes three steps: data cleaning, data standardization, and data synchronization. Data cleaning uses the outlier identification method, combined with the 2.5~3.5 times standard deviation method and confidence interval analysis to identify and eliminate outliers, and uses multiple interpolation methods to fill missing values. Data standardization uses the maximum and minimum value standardization method to normalize all types of data to the [0,1] interval. Data synchronization aligns data with different sampling frequencies to the same time series.

[0031] S2. Using a deep neural network, a trend warning model and a time series prediction model are constructed based on the historical operating data of power plant equipment, including the following steps: S21. Constructing a trend warning model, including the following steps: S211 , obtaining pre-processed operating condition data, vibration data, and corresponding historical degradation data, and constructing a first training sample set and a first verification sample set.

[0032] The preprocessed operating condition data and vibration data are randomly divided into the first training sample set and the first validation sample set in a ratio of 8:2. The operating condition data and vibration data are used as input data, and the corresponding historical degradation data are used as label data.

[0033] S212. Construct a deep long short-term memory neural network.

[0034] The 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.

[0035] The number of neurons in the input layer is the same as the characteristic dimension of the working condition data and the vibration data; the multimodal adaptive Dropout layer is set between the input layer and the coupled bidirectional LSTM layer, and the dropout strategy is adaptively adjusted according to the different characteristic distributions of the working condition data and the vibration data; the coupled bidirectional LSTM layer includes a forward LSTM unit and a backward LSTM unit. The forward LSTM unit processes data forward according to the time series to capture the temporal dependency from history to the present, and the backward LSTM unit processes data backward according to the time series to capture the contextual information of the complete time series. An information interaction mechanism is set between the forward and backward units to capture the temporal dependency characteristics of equipment degradation through bidirectional processing and information interaction; the fully connected layer contains 32 neurons and adopts the ReLU activation function; the output layer outputs the predicted value of the degradation degree.

[0036] Adaptively adjusting the dropout strategy based on the different feature distributions of the operating and vibration data involves calculating the differences in feature distributions for the operating and vibration data, setting corresponding deactivation rates for the different modal data based on these differences, and applying these rates to the corresponding data for neuron dropout. Furthermore, the information exchange mechanism is implemented by weighted fusion of the hidden state outputs of the forward and backward LSTM units. The fused features serve as the final output of the coupled bidirectional LSTM layer.

[0037] The multimodal adaptive Dropout layer sets the deactivation rate according to the differences in the characteristic distributions of the working condition data and the vibration data, achieving differentiated regularization processing for different modal data; the forward and backward units of the coupled bidirectional LSTM layer perform hidden state fusion through an information interaction mechanism, enhancing the utilization of bidirectional temporal information; the simplified network structure reduces the intermediate layer processing links and reduces the computational complexity.

[0038] S213. Use the back-propagation algorithm to train the deep long short-term memory neural network, use the adaptive optimizer to update the network parameters, and set a threshold limit on the gradient value.

[0039] First, the first training sample set is processed in batches, and forward propagation is used to calculate the loss function between the predicted output and the actual degradation. Then, based on the loss function value, the gradient values ​​of the network parameters at each layer are calculated using the backpropagation algorithm. The calculated gradient values ​​are then 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 network's convergence.

[0040] 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.

[0041] S214. Calculate a first verification loss value on the first verification sample set. When the first verification loss value decreases by less than a preset degradation threshold within a preset number of consecutive rounds, stop training and save the current network parameters.

[0042] After each round of training, 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 and the actual degradation value is calculated to obtain the first validation loss value. The validation loss value is monitored for its changing trend. When the decrease within a preset number of rounds is less than a preset degradation threshold, the model is considered converged and training is terminated. The preset degradation threshold is determined by analyzing the distribution of the rate of change of equipment degradation in historical operating data. An early stopping mechanism is also implemented to prevent overfitting. When the validation loss value increases within a consecutive number of rounds, training is terminated and the optimal network parameters are back-tested and saved.

[0043] 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.

[0044] S215. Use the trained deep long short-term memory neural network as a trend warning model.

[0045] The trend warning model is based on a network structure consisting of a multimodal adaptive Dropout layer and a coupled bidirectional LSTM layer. It can process the heterogeneous features of operating condition data and vibration data and capture the temporal dependencies of equipment degradation. Through gradient threshold limiting and early stopping mechanisms, it reduces the risks of gradient explosion and overfitting during training.

[0046] S22, build a temperature time series prediction sub-model, such as Figure 2 As shown, the following steps are included: S221 : Based on the pre-processed operating condition data, the measuring point temperature data, and the corresponding key component temperature trend data, the season coding information is used as an additional feature to construct a second training sample set and a second validation sample set.

[0047] First, the preprocessed operating condition data and the measuring point temperature data were aligned in chronological order. Then, seasonal coding information was generated based on the data collection time, and the four seasons (spring, summer, autumn, and winter) were converted into 4-dimensional binary vectors using a one-hot encoding method. Next, the operating condition data, measuring point temperature data, and seasonal coding information were concatenated along the feature dimension to form the input feature matrix, with the temperature trend data of key components at the corresponding time as the label. Finally, the dataset was randomly divided into a second training sample set and a second validation sample set in an 8:2 ratio. The operating condition data and measuring point temperature data served as input data, and the corresponding key component temperature trend data served as the label data.

[0048] S222. Construct a gated recurrent neural network.

[0049] The gated recurrent neural network consists of an input layer, a dual GRU layer, an attention calculation layer, a fully connected layer, and an output layer.

[0050] The number of neurons in the input layer is the same as the feature dimension after splicing; the first GRU layer contains 128 hidden units, and uses the tanh activation function and gating mechanism to extract time series 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 adopts an adaptive weight initialization method, including a weight matrix and a scaling factor. The weight matrix dimension is 64×64, and the importance of different time steps is learned through trainable parameters and normalized by the softmax function. The scaling factor is dynamically adjusted according to the data distribution to adjust 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 key component temperature trend data, and a linear activation function is used to predict the temperature trend.

[0051] S223 , inputting the second training sample set into the gated recurrent neural network, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint.

[0052] The network parameter update flow chart is as follows: Figure 3 As shown in the figure, the differential value of the temperature data of the measuring points in the training sample is calculated 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 square error loss term, the trend correlation loss term, and the amplitude penalty term in a weighted manner. The specific formula is as follows:

[0053] Where, is the weight coefficient, which dynamically adjusts each weight according to the error changes during model training (in this embodiment, it is initially set to 0.6, 0.3, and 0.1); LMAE is the mean square error loss term, which represents the square of the Euclidean distance between the predicted temperature series and the true temperature series; LCorr is the trend correlation loss term, which is calculated by subtracting the Pearson correlation coefficient between the predicted trend and the true trend from 1; LPen is the amplitude penalty term, which represents the second-order difference value of the predicted temperature series L2 Norm, the specific formula is as follows:

[0054] Where, y t express t The actual temperature value at the moment, represents the predicted temperature value at time t, n is the sequence length, and represent the mean of the actual temperature series and the predicted temperature series, and Represent the standard deviation of the actual temperature series and the predicted temperature series respectively.

[0055] The second training sample set is divided into batches of 32 samples, and the combined loss value of 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. Then, the calculated gradient value is norm clipped to limit it to the interval [-5, 5]. Finally, the network parameters are iteratively updated according to the clipped gradient value.

[0056] S224. Calculate a second verification loss value on the second verification sample set. When the rate of change of the second verification loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stop training and save the current network parameters.

[0057] The temperature prediction threshold is determined by performing K-fold cross validation on the training dataset and combining the experience of domain experts.

[0058] S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

[0059] The temperature time series prediction sub-model improves its ability to predict temperature trends of key components by incorporating seasonal encoding information, a two-layer GRU structure, and an attention mechanism, combined with a combined loss function based on temperature trend constraints. This model leverages seasonal characteristics and historical temperature variation 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.

[0060] S23. Construct a vibration time series prediction sub-model.

[0061] S231. Acquire operating condition data, runout data, and corresponding key parts vibration trend data to construct a third training sample set and a third verification sample set.

[0062] Using stratified random sampling, the collected data was divided into a third training sample set and a third validation sample set at a ratio of 8:2 to ensure the representativeness and randomness of the sample sets. Operating condition data and vibration data were used as input data, and the corresponding key parts vibration trend data was used as label data.

[0063] S232. Construct a multi-scale gated recurrent neural network.

[0064] The multi-scale gated recurrent neural network includes an input layer, a multi-scale GRU layer, a feature fusion layer, and an output layer.

[0065] The multi-scale GRU layer sets sliding windows of 1, 3, and 5 time steps to capture the short-term, medium-term, and long-term features of the vibration signal, respectively. Each time scale is equipped with an independent gated recurrent unit. The multi-scale GRU layer uses sliding windows of different time scales to extract vibration features. By processing vibration signals of different time scales in parallel, it extracts the local and global characteristic patterns of the signal. The feature fusion layer integrates multi-scale vibration features with operating condition information. Based on the attention weight mechanism, it adaptively adjusts the importance of features of different scales according to the operating condition information. The output layer generates vibration trend prediction results, mapping the fused multi-scale features into continuous vibration trend prediction values.

[0066] 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.

[0067] First, the data of the third training sample set are input into the multi-scale gated recurrent neural network in batches. In each training batch, forward propagation is performed to calculate the predicted output. Subsequently, the mean square error is used as the loss function to calculate the loss value between the predicted output and the actual vibration trend, and the gradients of the parameters of each layer of the network are calculated through the back propagation algorithm. The Adam optimizer is used to adaptively update the network parameters based on the calculated gradient information combined with the attention mechanism, where the attention weights are dynamically adjusted according to the vibration characteristics and working condition information. During the parameter update process, the learning rate and regularization terms are set to control the parameter update amplitude, suppress overfitting, and maintain the generalization performance of the model. The above iterative process of forward propagation, loss calculation, back propagation and parameter update is repeated until the full training of the third training sample set is completed.

[0068] S234. Calculate a third verification loss value on the third verification sample set. When the fluctuation amplitude of the third verification loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stop training and save the current network parameters.

[0069] The vibration prediction threshold is determined by statistical variance and error analysis of historical vibration data, and is 0.025 in this embodiment.

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

[0071] The vibration time series prediction sub-model enhances its ability to predict equipment vibration trends through the use of sliding windows with varying time steps in the multi-scale GRU layer, an attention weighting mechanism in the feature fusion layer, and a stratified random sampling training approach. This model captures both short-term and long-term characteristics of vibration signals and adjusts feature weights based on operating conditions, improving the stability of vibration trend prediction and providing predictive support for early warning of abnormal equipment vibration.

[0072] S3. Output the degradation degree prediction value, temperature trend prediction value and vibration trend prediction value in sequence through the trend warning model, temperature time series prediction sub-model and vibration time series prediction sub-model.

[0073] S4. Establish a three-tier evaluation system to generate status evaluation results based on the degradation degree prediction value, temperature trend prediction value, and vibration trend prediction value.

[0074] The three-tier evaluation system includes a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer. The process of generating the state evaluation results is as follows: the degradation degree prediction value, temperature trend prediction value, and vibration trend prediction value are input into the three-tier evaluation system; in the state quantity evaluation layer, the deduction value of each state quantity is calculated based on the deduction system. The deduction value is equal to the basic deduction value multiplied by the weight plus the family defect deduction value. The specific formula is as follows:

[0075] Where, is the basic deduction value, is the dynamic weight value, Points are deducted for family defects.

[0076] It should be noted that the basic penalty is calculated using a piecewise linear function. Deductions are calculated proportionally within different intervals based on the degree of deviation between the predicted value and each level's threshold (normal, caution, abnormal, and severe). The further the predicted value deviates from the normal range, the higher the penalty. A reasonable upper limit is set to avoid excessive penalty in extreme cases. Dynamic weights are determined by analyzing the historical rate of change of the state variable and the predicted trend. Specifically, when the rate of change of the state variable exceeds a preset threshold or the predicted trend shows a significant deterioration, the weight of the state variable is increased accordingly. Conversely, when the state variable remains stable or the predicted trend improves, the weight is appropriately reduced. The weight range is [0.8, 1.2], with initial weights set to 1.0. Weights are updated monthly and calculated based on data within a sliding time window (e.g., 3 months).

[0077] In addition, the family defect deduction points are determined based on three dimensions: the frequency of defect occurrence, the degree of harm, and the speed of evolution: For the frequency of defect occurrence, the deduction is 0-5 points when the frequency is less than 5%, 6-10 points when the frequency is 5%-15%, and 11-15 points when the frequency is greater than 15%. For the degree of damage caused by defects, an additional deduction of 0 to 2 points will be made for minor impact (not affecting normal equipment operation), 3 to 4 points for moderate impact (may cause performance degradation), and 5 points for severe impact (may cause equipment downtime). For the speed of defect evolution, no additional points will be deducted if it is slow (more than several months), 1 to 3 points will be deducted if it is medium (several weeks to several months), and 3 to 5 points will be deducted if it is fast (several days to several weeks).

[0078] The final deduction for family defects is the sum of the deductions from the three dimensions, and shall not exceed the upper limit of 20 points. Any points exceeding 20 will be deducted from the upper limit of 20. This scoring standard can be dynamically adjusted based on actual operating data and expert experience.

[0079] The component and overall health status are divided into four evaluation levels: normal, caution, abnormal, and severe. At the component evaluation level, the component evaluation level is determined based on the single deduction interval of the status quantity, as shown in Table 1.

[0080] Table 1 Evaluation grade determination table

[0081] In order to avoid frequent jumps near the critical value, a 5-point buffer mechanism is set: when the deduction value is within the range of ±2.5 points of the critical value, the evaluation level remains unchanged from the previous evaluation result.

[0082] At the overall evaluation level, when the evaluation levels of all components 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 progressive mechanism.

[0083] The status grading mechanism is as follows: the overall status is determined based on the combination of the status of each component. If any component is rated severe, the overall status is directly rated severe. If two or more components are rated abnormal, the overall status is rated severe. If only one component is rated abnormal or three or more components are rated caution, the overall status is rated abnormal. If only one or two components are rated caution and the other components are rated normal, the overall status is rated caution. In other cases, the rating process follows the "strict principle" that means when a situation may correspond to multiple levels, the more severe level is used as the final rating.

[0084] This three-tiered evaluation system enables multi-level assessments, from single status variables to components and then to the entire system. Through a dynamic weighting mechanism and a progressive status level progression mechanism, it accurately reflects the health status of each level of equipment, providing a scientific basis for early warning decisions. This evaluation system considers the importance and changing trends of different status variables, while also accounting for the mutual influence between components, ensuring the rationality and reliability of the evaluation results.

[0085] S5. Generate early warning information based on the status evaluation results.

[0086] The system automatically generates warning messages of corresponding levels based on the overall status evaluation level. For example, if the evaluation level is "Caution," a reminder-level warning message is generated, advising increased monitoring and inspections; if the evaluation level is "Abnormal," a warning-level warning message is generated, advising prompt inspection and necessary preventive measures; and if the evaluation level is "Critical," an alarm-level warning message is generated, advising immediate shutdown, inspection, and repair. The warning message includes important information such as the equipment number, evaluation level, key status value, trend, and warning time, and automatically generates actionable suggestions based on historical maintenance experience.

[0087] Practical applications have demonstrated that the proposed method offers significant advantages in monitoring the condition of power plant equipment, demonstrating excellent prediction accuracy, temperature and vibration prediction precision, equipment anomaly identification, and model training efficiency. In practical applications, the system generates graded early warning information, enabling maintenance personnel to take preventive maintenance measures before equipment failures occur. This effectively reduces the risk of unplanned equipment downtime and provides reliable technical support for accurate health assessment of power plant equipment.

[0088] Thus far, the technical solutions of the present invention have been described in conjunction with the specific experimental processes 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 may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A power plant equipment status trend prediction method based on deep learning, characterized in that: The steps include: S1. Obtain historical operation data of power station equipment and preprocess the historical operation data of power station equipment; S2. Using deep neural networks, a trend warning model and a time series prediction model are constructed based on the 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. Outputting the degradation degree prediction value, temperature trend prediction value, and vibration trend prediction value in sequence through the trend warning model, temperature time series prediction sub-model, and vibration time series prediction sub-model; S4. Establish a three-layer evaluation system including a state quantity evaluation layer, a component evaluation layer, and an overall evaluation layer, and generate a state evaluation result based on the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value; S5. Generate early warning information based on the status evaluation results.

2. The method according to claim 1, wherein: In step S1 , the historical operation data of the power station equipment includes operating condition data, vibration data, historical degradation data, measurement point temperature data, key component temperature trend data, and key part vibration trend data.

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

4. The method according to claim 1, wherein: In step S2, constructing a trend warning model includes the following sub-steps: S211, obtaining pre-processed working condition data, vibration data, and corresponding historical degradation data, and constructing a first training sample set and a first verification sample set; S212. Construct a deep long short-term memory neural network, the deep long short-term memory neural network including 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 vibration data; an information exchange mechanism is set between the forward LSTM unit and the backward LSTM unit of the coupled bidirectional LSTM layer; S213, using a back propagation algorithm to train the deep long short-term memory neural network, using an adaptive optimizer to update network parameters, and setting a threshold limit on the gradient value; S214, calculating a first validation loss value on the first validation sample set, and when the first validation loss value decreases by less than a preset degradation threshold within a preset number of consecutive rounds, stopping training and saving the current network parameters; S215. Use the trained deep long short-term memory neural network as a trend warning model.

5. The method according to claim 1, wherein: In step S2, building a temperature time series prediction sub-model includes the following sub-steps: S221, based on the pre-processed operating condition data, the measuring point temperature data, and the corresponding key component temperature trend data, the seasonal coding information is used as an additional feature to construct a second training sample set and a second validation sample set; S222. Construct a gated recurrent neural network, including an input layer, a dual GRU layer, an attention calculation layer, a fully connected layer, and an output layer; wherein the first GRU layer is used to extract time series features, the second GRU layer is used to integrate working condition and seasonal information, and the attention calculation layer includes a weight matrix and a scaling factor; S223, inputting the second training sample set into the gated recurrent neural network, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint; S224, calculating a second verification loss value on the second verification sample set, and when the rate of change of the second verification loss value within a preset number of consecutive rounds is less than the temperature prediction threshold, stopping training and saving the current network parameters; S225. Use the trained gated recurrent neural network as a temperature time series prediction sub-model.

6. The method according to claim 5, characterized in that: In step S223, using the RMSprop optimizer and updating the network parameters based on the temperature change trend constraint includes the following sub-steps: S2231, calculating differential values ​​for the temperature data of the measuring 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, where the combined loss function fuses the mean square error loss term, the trend correlation loss term, and the amplitude penalty term in a weighted manner; S2233. Calculate the gradient value of the combined loss function through the RMSprop optimizer, iteratively update the network parameters according to the gradient value, and limit the gradient norm.

7. The method according to claim 1, wherein: In step S2, constructing the vibration time series prediction sub-model includes the following sub-steps: S231, obtaining working condition data, vibration data, and corresponding key parts vibration trend data, and constructing 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; wherein the multi-scale GRU layer uses sliding windows of different time scales to extract vibration features, and the feature fusion layer integrates the multi-scale vibration features and working condition information; S233, inputting the third training sample set into the multi-scale gated recurrent neural network, and updating the network parameters using the Adam optimizer combined with the attention mechanism; S234, calculating a third verification loss value on the third verification sample set, and when the fluctuation amplitude of the third verification loss value within a preset number of consecutive rounds is less than the vibration prediction threshold, stopping training and saving the current network parameters; S235. Use the trained multi-scale gated recurrent neural network as a vibration time series prediction sub-model.

8. The method according to claim 1, wherein: Step S5 also includes the following sub-steps: S51, inputting the degradation degree prediction value, the temperature trend prediction value, and the vibration trend prediction value into the three-tier evaluation system; S52. At the state quantity evaluation layer, the deduction value of each state quantity is calculated based on the deduction system; S53. At the component evaluation layer, the component evaluation level is determined based on the individual deduction intervals of the status quantity. The evaluation levels include normal, caution, abnormal, and severe. S54. At the overall evaluation level, when the evaluation levels of all components 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 progressive mechanism.

9. The method according to claim 1, wherein: In step S5, the warning information includes the equipment number, evaluation level, key state quantity value, change trend, warning time, and generates processing suggestions based on historical maintenance experience.

Images