Early risk warning method and system for molten salt pump of photo-thermal power station based on vibration trend prediction

Abstract

The application discloses a molten salt pump early risk early warning method and system based on vibration trend prediction for a photo-thermal power station, and the method is as follows: collecting multi-channel vibration signals of key parts of the molten salt pump, carrying out preprocessing through denoising filtering, sliding window division and z-score standardization, and generating time sequence samples with unified structure; adopting a causal convolution structure to extract multi-scale time sequence features in parallel, wherein an expansion factor is exponentially increased with the depth of the network, and a residual connection is combined to optimize feature transmission; constructing a vibration trend prediction model, integrating a channel feature enhancement unit, a long-term time sequence modeling unit and a time sequence information aggregation unit; using a mean square error loss function for model training, setting a dynamic threshold based on prediction error statistics; and triggering early warning when the prediction error exceeds the threshold. The application realizes early identification of operation abnormities of the molten salt pump and prolongs the service life of the equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to an early risk warning method and system for molten salt pumps in solar thermal power plants based on vibration trend prediction. Background Technology

[0002] Molten salt pumps in tower-type solar thermal power plants operate continuously under high temperature, high pressure, and highly corrosive molten salt conditions. They are susceptible to various complex factors such as heat load fluctuations, mechanical imbalances, hydraulic and electromagnetic disturbances, and structural aging, leading to abnormal vibration responses. If not detected and addressed in a timely manner, this can cause equipment damage or even operational accidents. Therefore, conducting research on molten salt pump vibration trend prediction is of great significance for achieving fault early warning, extending equipment life, and improving operational economy.

[0003] Traditional vibration monitoring methods mainly rely on threshold judgment, empirical discrimination, or spectrum analysis. While these methods are effective for linear and stationary signals, molten salt pump vibration signals typically exhibit nonlinear, non-stationary, and multi-timescale characteristics under actual operating conditions. Traditional methods struggle to accurately extract key features related to equipment state changes, leading to diagnostic delays, missed diagnoses, and false alarms. Existing data-driven vibration prediction methods, while showing some adaptability to complex operating conditions, still suffer from insufficient multi-timescale feature extraction capabilities, rigid multi-channel data fusion strategies, difficulties in identifying key time-series information, and a lack of adaptability in setting warning thresholds for molten salt pumps. These limitations make it difficult to meet the needs of vibration trend prediction and early risk warning under the high-temperature continuous operation conditions of molten salt pumps. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prediction of vibration signals of molten salt pumps, such as insufficient multi-scale feature extraction capability, limited multi-channel fusion effect, insufficient identification of key time sequence information, and fixed warning threshold. It provides an early risk warning method and system for molten salt pumps in solar thermal power plants based on vibration trend prediction, which improves the accuracy and stability of predicting the vibration state of molten salt pumps under high-temperature continuous operation conditions, and realizes early abnormal trend identification and operation risk warning.

[0005] The technical solution adopted in this invention is:

[0006] An early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction includes the following steps:

[0007] S1: Collect multi-channel vibration signals from key parts of the molten salt pump during operation, and generate vibration time series samples with uniform structure after standardizing and preprocessing the vibration signals.

[0008] S2 extracts multi-scale temporal features from vibration time series by using a causal convolution structure to extract local temporal features at different time scales in parallel to construct a feature matrix.

[0009] S3 divides the dataset into a training set and a test set. The training set is used for model training and threshold construction, while the test set is used to verify model performance and warning accuracy.

[0010] S4. Construct a vibration trend prediction model. The model includes a channel feature enhancement unit, a long-term time series modeling unit, and a time series information aggregation unit. Specifically: the channel feature enhancement unit is used to adaptively weight the multi-scale time series features of the feature matrix, generate global statistical information through compression operations, and learn channel weights through activation operations; the long-term time series modeling unit is used to extract long-term dependent features and selectively retain long-term historical information through a gating mechanism; the time series information aggregation unit is used to dynamically weight the feature vectors of each time step output by the long-term time series modeling unit and identify key time steps through an attention mechanism.

[0011] S5. Train the vibration trend prediction model on the training set, use the mean squared error as the loss function, and calculate the dynamic threshold based on the prediction error of the training set.

[0012] Specifically, by calculating the mean and standard deviation of the prediction error, upper and lower thresholds for early warning judgment are set, i.e., dynamic thresholds.

[0013] S6 uses the trained vibration trend prediction model to predict unknown vibration sequences; when the prediction error exceeds the set dynamic threshold range, it is considered an abnormal trend and triggers an early warning signal, thus realizing early identification and reliable early warning of abnormal operation of molten salt pumps.

[0014] Specifically, the input data includes multi-channel vibration signals (such as vibration velocity and acceleration), and the output data is the vibration trend prediction value. During model inference, the vibration signals are combined with the operating conditions of the molten salt pump (such as temperature and pressure), and the dynamic threshold is calibrated through environmental data.

[0015] Furthermore, the normalization preprocessing in step S1 includes denoising filtering, sliding window segmentation, and normalization operations; the denoising filtering uses a Butterworth bandpass filter with a sliding window size of 60 seconds, and the normalization uses the z-score method, calculated as follows:

[0016] ;

[0017] in, For the j-th feature of the i-th sample, and Let be the mean and standard deviation of the j-th feature of the i-th sample, respectively; The standardized feature value of the j-th feature of the i-th sample.

[0018] Furthermore, step S2 is as follows:

[0019] S21 employs a causal convolutional structure to extract features from vibration time series, extracting multi-scale temporal features in parallel through an exponentially increasing inflation rate; the calculation formula for causal convolution is:

[0020] ;

[0021] in, For dilated convolution operations, M is the kernel size. Let be the filter function of the convolution kernel, and d be the dilation factor. t The value of , where m is the l-th data point in the input sequence data.

[0022] S22 utilizes a residual connection structure to achieve cross-layer information transfer and optimize the feature extraction process; the residual connection calculation formula is as follows:

[0023] ;

[0024] in, The output is a convolutional transformation. For residual connection input;

[0025] S23, align the extracted multi-scale features by time step to construct a feature matrix.

[0026] Furthermore, step S4 specifically includes:

[0027] S41, based on the channel feature enhancement unit, adaptively weights the multi-channel vibration features, generates global statistical information through compression operation, learns channel weights through excitation operation, highlights key channels and suppresses noise;

[0028] S42, based on the long-term temporal modeling unit, performs long-term temporal evolution modeling on the fused features, and selectively retains long-term historical information through a gating mechanism;

[0029] S43, based on the time series information aggregation unit, performs weighted aggregation of long-term time series representations, and dynamically identifies and focuses on key time steps that have a significant impact on the prediction results through the attention mechanism.

[0030] Furthermore, the calculation formula for the compression operation in step S41 is as follows:

[0031] ;

[0032] The formula for calculating the stimulus operation is:

[0033] ;

[0034] The formula for calculating channel reweighting is:

[0035] ;

[0036] in, This is the compressed output of channel c; Let be the feature of channel c at time t, where T is the number of time steps. , δ represents the network weights of the first and second layers, s represents the channel weights, δ represents the ReLU activation function, and σ represents the Sigmoid activation function. The c-th channel feature after reweighting; Let be the weight of the c-th channel; This represents the original c-th channel feature.

[0037] Furthermore, the internal calculation process of the long-term time series modeling unit in step S42 includes:

[0038] Forgotten Gate: ;

[0039] Input Gate: ;

[0040] Cell status update: ;

[0041] Output gate: ;

[0042] Hidden state: ;

[0043] in, The output of the forget gate at time t; The input gate output at time t; Update the cell state output at time t; The output gate outputs at time t; For the input at time t, , Let be the hidden layer states at times t and t-1. , , , These are the weight matrices for the input gate, forget gate, cell state update, and output gate, respectively. , , , Let be the bias vectors for the input gate, forget gate, cell state update, and output gate, respectively; σ is the Sigmoid function; and ⊙ is the element-wise multiplication. The hyperbolic tangent activation function (tanh) is used; the hidden layer dimension is set to 128, and a two-layer structure is adopted.

[0044] Furthermore, in step S43, the time-series information aggregation unit is implemented through a multi-head attention mechanism; the sequence output by the long-term time-series modeling unit is linearly transformed to generate three matrices: query Q, key K, and value V. Attention weights are obtained through attention calculation, and multiple attention heads are run in parallel for weighted aggregation; the formula for generating the query Q, key K, and value V matrices is as follows:

[0045] ;

[0046] in, , , Let X be a learnable linear projection matrix, where X is the output sequence of the long-term time series modeling unit;

[0047] The formula for calculating attention is:

[0048] ;

[0049] in, Let be the dimension of the key vector. This is the scaling factor;

[0050] The multi-head attention mechanism operates H attention heads in parallel, and the calculation formula is as follows:

[0051] ;

[0052] in, For the first The output of each attention head; H is the number of attention heads; This is the output projection matrix. The number of attention heads H is set to 4.

[0053] Furthermore, step S5 specifically includes:

[0054] S51, During the model training phase, the prediction error for each time step is calculated based on a set of known historical data. The prediction error refers to the difference between the model's predicted value and the corresponding actual monitored value.

[0055] ;

[0056] in, It is the actual value. It is a predicted value;

[0057] S52, calculate the mean μ and standard deviation σ for all error samples, and set the upper and lower limits of the dynamic threshold: the dynamic threshold is used to determine whether the current error deviates from the normal range, thereby realizing early warning of the vibration state of the molten salt pump.

[0058] Calculate the mean μ and standard deviation σ for all error samples:

[0059] , ;

[0060] ; ;

[0061] in, Let be the error of the i-th sample; This is the upper limit of the dynamic threshold. This is the lower limit of the dynamic threshold.

[0062] This invention also discloses an early risk warning system for molten salt pumps in solar thermal power plants based on vibration trend prediction, which implements an early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction. The system includes the following modules:

[0063] Vibration data acquisition module: used to acquire multi-channel vibration data of molten salt pumps under different operating conditions in real time;

[0064] Data preprocessing module: Used to receive raw signal data from vibration data acquisition module, and sequentially perform noise reduction filtering, sliding window segmentation and standardization processing to form time series samples with a unified structure;

[0065] Multi-scale feature extraction module: used to extract multi-scale temporal features from the preprocessed vibration time series, and extracts local temporal features at different time scales in parallel through causal convolution structure;

[0066] Model prediction module: used to achieve time series modeling and trend prediction based on channel feature enhancement, long-term time series modeling and time series information aggregation mechanism;

[0067] Trend warning module: It is used to dynamically determine whether there is an abnormal trend based on the error between the model prediction results and the actual observations, trigger warning signals and output maintenance decision suggestions.

[0068] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art:

[0069] 1. This invention introduces a causal convolution structure, enabling the extraction of multi-timescale features from complex molten salt pump nonlinear vibration signals. It captures vibration changes from short-term fluctuations to long-term evolution through exponentially increasing expansion rates, overcoming the limitations of traditional single-timescale modeling and improving prediction accuracy. 2. This invention, through channel feature enhancement units and long-term time-series modeling units, adaptively fuses multi-channel vibration data and identifies key time-step information, fully extracting key temporal features from the vibration sequence and improving the model's sensitivity to vibration trend changes and prediction accuracy. 3. This invention employs a dynamic threshold early warning method based on the statistical characteristics of prediction deviations. It adaptively adjusts the early warning threshold according to actual operating conditions, effectively reducing false alarm and false negative rates, and achieving early identification and reliable early warning of molten salt pump operational anomalies. 4. This invention can accurately predict the vibration trend of molten salt pumps, effectively preventing abnormal equipment vibration, providing a reference for molten salt pump condition-based maintenance, and improving the intelligent operation and maintenance and safety management level of solar thermal power plants. Attached Figure Description

[0070] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0071] Figure 1 This is a flowchart illustrating the early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction, as described in this invention.

[0072] Figure 2 This is a schematic diagram of the causal convolution structure for multi-scale temporal feature extraction in this invention;

[0073] Figure 3 This is a schematic diagram of the residual connection structure of the present invention;

[0074] Figure 4 This is a schematic diagram of the channel feature enhancement unit structure of the present invention;

[0075] Figure 5 This is a schematic diagram of the gating mechanism structure of the long-term timing modeling unit of the present invention;

[0076] Figure 6 This is a schematic diagram of the attention mechanism structure of the temporal information aggregation unit of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0078] like Figures 1 to 6 As shown in one example, this invention discloses an early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction, comprising the following steps:

[0079] S1: Collect multi-channel vibration signals of key parts (such as motor vibration in the Y direction and pump body vibration in the X / Y / Z directions) during the operation of the molten salt pump, and generate vibration time series samples with uniform structure after standardizing and preprocessing the vibration signals.

[0080] S2 extracts multi-scale temporal features from vibration time series by using a causal convolution structure to extract local temporal features at different time scales in parallel to construct a feature matrix.

[0081] S3 divides the dataset into a training set and a test set. The training set is used for model training and threshold construction, while the test set is used to verify model performance and warning accuracy.

[0082] S4. Construct a vibration trend prediction model. The model includes a channel feature enhancement unit, a long-term time series modeling unit, and a time series information aggregation unit. Specifically: the channel feature enhancement unit is used to adaptively weight the multi-scale time series features of the feature matrix, generate global statistical information through compression operations, and learn channel weights through activation operations; the long-term time series modeling unit is used to extract long-term dependent features and selectively retain long-term historical information through a gating mechanism; the time series information aggregation unit is used to dynamically weight the feature vectors of each time step output by the long-term time series modeling unit and identify key time steps through an attention mechanism.

[0083] S5. Train the vibration trend prediction model on the training set, use the mean squared error as the loss function, and calculate the dynamic threshold based on the prediction error of the training set.

[0084] Specifically, by calculating the mean and standard deviation of the prediction error, upper and lower thresholds for early warning judgment are set, i.e., dynamic thresholds.

[0085] S6 uses the trained vibration trend prediction model to predict unknown vibration sequences; when the prediction error exceeds the set dynamic threshold range, it is considered an abnormal trend and triggers an early warning signal, thus realizing early identification and reliable early warning of abnormal operation of molten salt pumps.

[0086] Specifically, the input data includes multi-channel vibration signals (such as vibration velocity and acceleration), and the output data is the vibration trend prediction value. During model inference, the vibration signals are combined with the operating conditions of the molten salt pump (such as temperature and pressure), and the dynamic threshold is calibrated through environmental data.

[0087] Furthermore, the standardization preprocessing in step S1 includes noise reduction filtering, sliding window segmentation, and standardization operations; specifically, the data acquisition and preprocessing steps are as follows:

[0088] S1-1, Vibration sensor deployment: Vibration sensors are deployed at key locations such as the motor vibration in the Y direction and the pump body vibration in the X / Y / Z directions of the molten salt pump. Among them, the motor vibration in the Y direction is highly sensitive to reflecting the operating status of the equipment.

[0089] S1-2, Data Acquisition: Collect operational monitoring data of the molten salt pumps of the Dunhuang 100 MW tower solar thermal power plant as a sample, covering common operating conditions;

[0090] S1-3, Noise Suppression: Noise interference is suppressed by Butterworth bandpass filter, filtering out high-frequency noise and low-frequency drift, and retaining the effective frequency components of the vibration signal;

[0091] S1-4, Prediction Target Construction: The root mean square (RMS) of the vibration signal is calculated over a 60-second sliding window to construct a prediction target for the vibration level over the next 10 minutes. The calculation formula is as follows:

[0092] ;

[0093] In the formula: is the vibration value of the i-th sampling point within the window; N is the number of sampling points within the window.

[0094] Furthermore, a Butterworth bandpass filter is used for denoising, with a sliding window size of 60 seconds. Normalization is performed using the z-score method, calculated as follows:

[0095] ;

[0096] in, For the j-th feature of the i-th sample, and Let be the mean and standard deviation of the j-th feature of the i-th sample, respectively; The standardized feature value of the j-th feature of the i-th sample.

[0097] Furthermore, step S2 is as follows:

[0098] S21 employs a causal convolutional structure to extract features from vibration time series, extracting multi-scale temporal features in parallel through an exponentially increasing inflation rate; the calculation formula for causal convolution is:

[0099] ;

[0100] in, For dilated convolution operations, M is the kernel size. Let be the filter function of the convolution kernel, and d be the dilation factor. For the input sequence at position t The value of , where m is the l-th data point in the input sequence data.

[0101] S22 utilizes a residual connection structure to achieve cross-layer information transfer and optimize the feature extraction process; the residual connection calculation formula is as follows: ;

[0102] in, The output is a convolutional transformation. For residual connection output; For residual connection input;

[0103] S23, align the extracted multi-scale features by time step to construct a feature matrix.

[0104] Furthermore, step S4 specifically includes:

[0105] S41, based on the channel feature enhancement unit, adaptively weights the multi-channel vibration features, generates global statistical information through compression operation, and learns the channel weights through excitation operation;

[0106] S42, based on the long-term temporal modeling unit, performs long-term temporal evolution modeling on the fused features, and selectively retains long-term historical information through a gating mechanism;

[0107] S43, based on the time series information aggregation unit, performs weighted aggregation of long-term time series representations, and dynamically identifies and focuses on key time steps that have a significant impact on the prediction results through the attention mechanism.

[0108] Furthermore, in step S41, the compression operation applies global average pooling to each channel to generate global statistics, calculated using the following formula:

[0109] ;

[0110] The activation operation captures the dependencies between channels through a two-layer mapping network and an activation function. The calculation formula is as follows:

[0111] ;

[0112] The formula for calculating channel reweighting is:

[0113] ;

[0114] in, This is the compressed output of channel c; Let be the feature of channel c at time t, where T is the number of time steps. , δ represents the network weights of the first and second layers, s represents the channel weights, δ represents the ReLU activation function, and σ represents the Sigmoid activation function. The c-th channel feature after reweighting; Let be the weight of the c-th channel; This represents the original c-th channel feature. In one embodiment, the dimensionality reduction ratio r can be set to 16.

[0115] Furthermore, the internal calculation process of the long-term time series modeling unit in step S42 includes:

[0116] Forgotten Gate: ;

[0117] Input Gate: ;

[0118] Cell status update: ;

[0119] Output gate: ;

[0120] Hidden state: ;

[0121] in, The output of the forget gate at time t; The input gate output at time t; Update the cell state output at time t; The output gate outputs at time t; For the input at time t, , Let be the hidden layer states at times t and t-1. , , , These are the weight matrices for the input gate, forget gate, cell state update, and output gate, respectively. , , , Let be the bias vectors for the input gate, forget gate, cell state update, and output gate, respectively; σ is the Sigmoid function; and ⊙ is the element-wise multiplication. The hyperbolic tangent activation function (tanh) is used; the hidden layer dimension is set to 128, and a two-layer structure is adopted.

[0122] Furthermore, in step S43, the time-series information aggregation unit is implemented through a multi-head attention mechanism; the sequence output by the long-term time-series modeling unit is linearly transformed to generate three matrices: query Q, key K, and value V. Attention weights are obtained through attention calculation, and multiple attention heads are run in parallel for weighted aggregation, with the number of attention heads H set to 4; the formula for generating the query Q, key K, and value V matrices is:

[0123] ;

[0124] in, , , Let X be a learnable linear projection matrix, where X is the output sequence of the long-term time series modeling unit;

[0125] The formula for calculating attention is:

[0126] ;

[0127] in, Let be the dimension of the key vector. This is the scaling factor;

[0128] The multi-head attention mechanism operates H attention heads in parallel, and the calculation formula is as follows:

[0129] ;

[0130] in, For the first The output of each attention head; H is the number of attention heads; This is the output projection matrix. In one embodiment, the number of attention heads H can be set to 4.

[0131] Furthermore, step S5 specifically includes:

[0132] S51, During the model training phase, the prediction error for each time step is calculated based on a set of known historical data. The prediction error refers to the difference between the model's predicted value and the corresponding actual monitored value.

[0133] ;

[0134] in, It is the actual value. It is a predicted value;

[0135] S52, calculate the mean μ and standard deviation σ for all error samples, and set the upper and lower limits of the dynamic threshold: the dynamic threshold is used to determine whether the current error deviates from the normal range, thereby realizing early warning of the vibration state of the molten salt pump.

[0136] Calculate the mean μ and standard deviation σ for all error samples:

[0137] , ;

[0138] ; ;

[0139] in, Let be the error of the j-th sample; n is the number of error samples; This is the upper limit of the dynamic threshold. This is the lower limit of the dynamic threshold.

[0140] This invention also discloses an early risk warning system for molten salt pumps in concentrated solar power plants based on vibration trend prediction, which implements an early risk warning method for molten salt pumps in concentrated solar power plants based on vibration trend prediction. The system includes the following modules:

[0141] Vibration data acquisition module: used to acquire multi-channel vibration data of molten salt pumps under different operating conditions in real time;

[0142] Data preprocessing module: Used to receive raw signal data from vibration data acquisition module, and sequentially perform noise reduction filtering, sliding window segmentation and standardization processing to form time series samples with a unified structure;

[0143] Multi-scale feature extraction module: used to extract multi-scale temporal features from the preprocessed vibration time series, and extracts local temporal features at different time scales in parallel through causal convolution structure;

[0144] Model prediction module: used to achieve time series modeling and trend prediction based on channel feature enhancement, long-term time series modeling and time series information aggregation mechanism;

[0145] Trend warning module: It is used to dynamically determine whether there is an abnormal trend based on the error between the model prediction results and the actual observations, trigger warning signals and output maintenance decision suggestions.

[0146] The specific working principle of this invention will be explained in detail below:

[0147] Operational monitoring data of the molten salt pumps at the Dunhuang 100 MW tower solar thermal power plant were collected. Vibration sensors were deployed at key locations such as the motor vibration in the Y-axis and the pump body vibration in the X / Y / Z axes. The collected vibration signals underwent preprocessing, including noise reduction filtering, sliding window segmentation, and standardization, to form time-series samples with a unified structure. Specifically, high-precision vibration sensors were used, and the data sampling frequency was set according to the actual operating conditions. The data included parameters such as vibration velocity and vibration acceleration.

[0148] Multi-scale temporal feature extraction is performed on the preprocessed vibration time series. A causal convolutional structure is employed, using an exponentially increasing dilation rate to extract local temporal features at different time scales in parallel. The multi-scale feature processing architecture is shown in the attached figure. Figure 2 As shown.

[0149] The extracted multi-scale features are input into the channel feature enhancement unit. The channel feature enhancement unit adaptively weights the multi-channel vibration features through compression and excitation operations. The structure of the channel feature enhancement module is shown in the attached figure. Figure 4 As shown.

[0150] The fused features are then input into the long-term time series modeling unit. The long-term time series modeling unit employs a gating mechanism for long-term time series evolution modeling; the structure of the gating mechanism is shown in the attached figure. Figure 5 As shown.

[0151] The output of the long-term time series modeling unit is input into the time series information aggregation unit. The time series information aggregation unit dynamically identifies key time steps through an attention mechanism, as shown in the attached figure. Figure 6 As shown.

[0152] The output layer outputs the predicted value after a nonlinear transformation by a fully connected layer and an activation function. The calculation formula is as follows:

[0153] ;

[0154] In the formula: The predicted vibration value; This is the output of the timing information aggregation unit; These are the output layer weights; This is the output layer bias.

[0155] The process for real-time prediction of molten salt pump vibration signals is as follows: Figure 1 As shown, the collected dataset, after preprocessing and feature extraction, is divided into a training set and a test set in a 7:3 ratio. The training set is used to train the model, and the optimal model parameters are saved. The test set is then imported into the trained model to predict vibration trends. Finally, the predicted values ​​are compared with the actual values ​​to calculate the prediction error. When the prediction error exceeds the dynamic threshold range, an early warning signal is triggered.

[0156] Instance verification analysis:

[0157] The effectiveness of this invention is verified through experimental analysis of vibration data from the molten salt pump of the 100 MW tower solar thermal power plant in Dunhuang. Vibration velocity data in the Y-axis of the molten salt pump motor were collected as the research object, as this channel is highly sensitive to reflecting the equipment's operating status. The raw monitoring data was cleaned and time-synchronized. Noise interference was suppressed using a Butterworth bandpass filter. The root mean square of the vibration signal was calculated using a 60-second sliding window to construct a prediction target for the vibration level in the next 10 minutes. All features were z-score standardized, and the dataset was divided into a 7:3 training set and a test set according to time order.

[0158] The coefficient of determination (R²), mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean square error (RMSE) are used as indicators to evaluate model performance. The closer the R² value is to 1, and the smaller the values ​​of MAPE, MAE, and RMSE, the better the model's predictive performance. The calculation formulas are as follows:

[0159] ;

[0160] ;

[0161] ;

[0162] ;

[0163] Where n is the total number of samples, This represents the actual vibration value of the i-th sample. This represents the predicted value of the i-th sample. This represents the average of the actual values.

[0164] The obtained vibration trend prediction results of the molten salt pump show that the vibration trend values ​​predicted by the model are basically consistent with the actual values, indicating that the method of the present invention can effectively predict the vibration trend of the molten salt pump. In addition, multiple models were selected for comparison to evaluate model performance, and the prediction results are shown in Table 1. As can be seen from the table, the method of the present invention outperforms the comparison models in terms of R², MAPE, MAE, and RMSE. Compared with baseline models such as CNN, GRU, CNN-LSTM, xLSTM, and Transformer, the MAPE index decreases by approximately 32.48%, 28.58%, 8.82%, 17.31%, and 0.49%, respectively, demonstrating the practical application value of the method of the present invention in early risk warning of molten salt pumps in tower solar thermal power plants.

[0165] Table 1 Comparison of Prediction Results

[0166] Model R² MAPE / % MAE / (mm·s⁻¹) RMSE / (mm·s⁻¹) Method of the present invention 0.9182 8.17 0.1357 0.2124 Transformer 0.9060 8.21 0.1465 0.2277 xLSTM 0.9074 9.57 0.1453 0.2261 CNN-LSTM 0.8992 8.96 0.1516 0.2358 GRU 0.8869 11.43 0.1605 0.2498 CNN 0.8551 12.10 0.1817 0.2826

[0167] The present invention adopts the above technical solution and has the following beneficial effects compared with the prior art:

[0168] 1. This invention introduces a causal convolution structure, enabling the extraction of multi-timescale features from complex molten salt pump nonlinear vibration signals. It captures vibration changes from short-term fluctuations to long-term evolution through exponentially increasing expansion rates, overcoming the limitations of traditional single-timescale modeling and improving prediction accuracy. 2. This invention, through channel feature enhancement units and long-term time-series modeling units, adaptively fuses multi-channel vibration data and identifies key time-step information, fully extracting key temporal features from the vibration sequence and improving the model's sensitivity to vibration trend changes and prediction accuracy. 3. This invention employs a dynamic threshold early warning method based on the statistical characteristics of prediction deviations. It adaptively adjusts the early warning threshold according to actual operating conditions, effectively reducing false alarm and false negative rates, and achieving early identification and reliable early warning of molten salt pump operational anomalies. 4. This invention can accurately predict the vibration trend of molten salt pumps, effectively preventing abnormal equipment vibration, providing a reference for molten salt pump condition-based maintenance, and improving the intelligent operation and maintenance and safety management level of solar thermal power plants.

[0169] Obviously, the described embodiments are only a portion, not all, of the embodiments of this application. Without conflict, the embodiments and features described and illustrated herein can be combined with each other. The components of the embodiments of this application generally described and illustrated in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. An early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction, characterized in that: Includes the following steps: S1: Collect multi-channel vibration signals from key parts of the molten salt pump during operation, and generate vibration time series samples with uniform structure after standardizing and preprocessing the vibration signals. S2 extracts multi-scale temporal features from vibration time series by using a causal convolution structure to extract local temporal features at different time scales in parallel to construct a feature matrix. S3 divides the dataset into a training set and a test set. The training set is used for model training and threshold construction, while the test set is used to verify model performance and warning accuracy. S4. Construct a vibration trend prediction model. The model includes a channel feature enhancement unit, a long-term time series modeling unit, and a time series information aggregation unit. Specifically: the channel feature enhancement unit is used to adaptively weight the multi-scale time series features of the feature matrix, generate global statistical information through compression operations, and learn channel weights through activation operations; the long-term time series modeling unit is used to extract long-term dependent features and selectively retain long-term historical information through a gating mechanism; the time series information aggregation unit is used to dynamically weight the feature vectors of each time step output by the long-term time series modeling unit and identify key time steps through an attention mechanism. S5. Train the vibration trend prediction model on the training set, use the mean squared error as the loss function, and calculate the upper and lower thresholds of the dynamic threshold based on the prediction error of the training set. S6 uses the trained vibration trend prediction model to predict unknown vibration sequences; when the prediction error exceeds the set dynamic threshold range, it is considered an abnormal trend and an early warning signal is triggered.

2. The method for early risk warning of molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 1, characterized in that: Step S1, the normalization preprocessing, includes denoising filtering, sliding window segmentation, and normalization. The denoising filtering uses a Butterworth bandpass filter with a sliding window size of 60 seconds. Normalization uses the z-score method, calculated using the following formula: ； in, For the j-th feature of the i-th sample, and Let be the mean and standard deviation of the j-th feature of the i-th sample, respectively; The standardized feature value of the j-th feature of the i-th sample.

3. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 1, characterized in that: The steps in step S2 are as follows: S21 employs a causal convolutional structure to extract features from vibration time series, extracting multi-scale temporal features in parallel through an exponentially increasing inflation rate; the calculation formula for causal convolution is: ； in, For dilated convolution operations, M is the kernel size. Let be the filter function of the convolution kernel, and d be the dilation factor. For the input sequence at position t The value of m is the nth element in the input sequence data. Data points. 4.S22, Cross-layer information transfer is achieved through residual connection structures, optimizing the feature extraction process; the residual connection calculation formula is: ； in, The output is a convolutional transformation. For residual connection output; For residual connection input; S23, align the extracted multi-scale features by time step to construct a feature matrix.

5. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 1, characterized in that: Step S4 specifically includes: S41, based on the channel feature enhancement unit, adaptively weights the multi-channel vibration features, generates global statistical information through compression operation, learns channel weights through excitation operation, highlights key channels and suppresses noise; S42, based on the long-term temporal modeling unit, performs long-term temporal evolution modeling on the fused features, and selectively retains long-term historical information through a gating mechanism; S43, based on the time series information aggregation unit, performs weighted aggregation of long-term time series representations, and dynamically identifies and focuses on key time steps that have a significant impact on the prediction results through the attention mechanism.

6. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 4, characterized in that: The calculation formula for the compression operation in step S41 is: ； The formula for calculating the stimulus operation is: ； The formula for calculating channel reweighting is: ； in, This is the compressed output of channel c; Let be the feature of channel c at time t, where T is the number of time steps. , δ represents the network weights of the first and second layers, s represents the channel weights, δ represents the ReLU activation function, and σ represents the Sigmoid activation function. The c-th channel feature after reweighting; Let be the weight of the c-th channel; This represents the original c-th channel feature.

7. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 4, characterized in that: The internal calculation process of the long-term time series modeling unit in step S42 includes: Forgotten Gate: ; Input Gate: ; Cell status update: ; Output gate: ; Hidden state: ; in, The output of the forget gate at time t; The input gate output at time t; Update the cell state output at time t; The output gate outputs at time t; For the input at time t, , Let be the hidden layer states at times t and t-1. , , , These are the weight matrices for the input gate, forget gate, cell state update, and output gate, respectively. , , , Let be the bias vectors for the input gate, forget gate, cell state update, and output gate, respectively; σ is the Sigmoid function; and ⊙ is the element-wise multiplication. The activation function is hyperbolic tangent; the hidden layer dimension is set to 128, and a two-layer structure is adopted.

8. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 4, characterized in that: In step S43, the time-series information aggregation unit is implemented through a multi-head attention mechanism. The sequence output by the long-term time-series modeling unit undergoes a linear transformation to generate three matrices: query Q, key K, and value V. Attention weights are calculated through attention, and multiple attention heads are run in parallel for weighted aggregation. The formula for generating the query Q, key K, and value V matrices is as follows: ； in, , , Let X be a learnable linear projection matrix, where X is the output sequence of the long-term time series modeling unit; The formula for calculating attention is: ； in, Let be the dimension of the key vector. This is the scaling factor; The multi-head attention mechanism operates H attention heads in parallel, and the calculation formula is as follows: ； in, For the first The output of each attention head; H is the number of attention heads; This is for outputting the projection matrix.

9. The early risk warning method for molten salt pumps in solar thermal power plants based on vibration trend prediction according to claim 1, characterized in that: Step S5 specifically includes: S51, During the model training phase, the prediction error for each time step is calculated based on a set of known historical data. The prediction error refers to the difference between the model's predicted value and the corresponding actual monitored value. ； in, It is the actual value. It is a predicted value; S52, calculate the mean μ and standard deviation σ for all error samples, and set the upper and lower limits of the dynamic threshold: the dynamic threshold is used to determine whether the current error deviates from the normal range, and the calculation formula is as follows: ； ； in, This is the upper limit of the dynamic threshold. This is the lower limit of the dynamic threshold.

10. An early risk warning system for molten salt pumps in solar thermal power plants based on vibration trend prediction, and a method for early risk warning of molten salt pumps in solar thermal power plants based on vibration trend prediction according to any one of claims 1 to 8, characterized in that: The system includes the following modules: Vibration data acquisition module: used to acquire multi-channel vibration data of molten salt pumps under different operating conditions in real time; Data preprocessing module: Used to receive raw signal data from vibration data acquisition module, and sequentially perform noise reduction filtering, sliding window segmentation and standardization processing to form time series samples with a unified structure; Multi-scale feature extraction module: used to extract multi-scale temporal features from the preprocessed vibration time series, and extracts local temporal features at different time scales in parallel through causal convolution structure; Model prediction module: used to achieve time series modeling and trend prediction based on channel feature enhancement, long-term time series modeling and time series information aggregation mechanism; Trend warning module: It is used to dynamically determine whether there is an abnormal trend based on the error between the model prediction results and the actual observations, trigger warning signals and output maintenance decision suggestions.

Images