Method for identifying abnormal mode of thermal energy storage system

Abstract

The invention provides a method for identifying an abnormal mode of a thermal energy storage system, and belongs to the technical field of thermal energy storage systems.The method comprises the steps that temperature distribution data, pressure fluctuation data and heat flux density data in the operation process of the thermal energy storage system are collected, and multi-scale wavelet decomposition is conducted on the temperature data to extract high-frequency components and low-frequency components; performing time-domain and frequency-domain analysis on the pressure data to extract peak frequency and amplitude attenuation coefficient, performing gradient calculation and topological analysis on the heat flow data to extract singular point distribution characteristics, and fusing and normalizing the three types of characteristics to establish a system operation state characteristic library; and applying the trained model to a system operation state feature library to extract deep abnormal features and output an abnormal mode classification result and a confidence score, and when the confidence score exceeds a preset threshold, determining that the state is an abnormal state, thereby solving the technical problem of insufficient recognition accuracy of the abnormal mode of the thermal energy storage system.

Description

Technical Field

[0001] This invention belongs to the technical field of thermal energy storage systems, and more specifically, relates to a method for identifying abnormal patterns in thermal energy storage systems. Background Technology

[0002] Thermal energy storage systems, as an important technology for energy storage, are widely used in areas such as peak shaving and valley filling of electricity, district heating, and industrial waste heat recovery. Traditional anomaly pattern recognition methods mainly rely on monitoring single physical parameters or rule-based threshold judgments. These methods detect system anomalies by setting predefined rules such as upper and lower limits for temperature, pressure alarm values, and abnormal flow ranges. An alarm is triggered when monitored parameters exceed preset ranges. This approach has become a mature application model in large-scale energy storage power stations, district heating systems, and industrial thermal energy management platforms. Some advanced systems even employ single-feature classifiers based on machine learning. However, traditional parameter monitoring and shallow machine learning methods cannot effectively extract these deep-level anomaly discrimination features, resulting in insufficient accuracy in recognizing anomalies such as localized overheating, pressure oscillations, and heat flow concentration. This is especially true when training data is limited or anomaly pattern features are similar, making existing methods more prone to misjudgment or missed detection. In other words, existing technologies suffer from insufficient accuracy in anomaly pattern recognition for thermal energy storage systems. Summary of the Invention

[0003] In view of this, the present invention provides a method for anomaly pattern recognition in thermal energy storage systems, which can solve the technical problem of insufficient accuracy in anomaly pattern recognition in existing technologies.

[0004] This invention is implemented as follows: A method for anomaly pattern recognition in a thermal energy storage system includes the following steps: collecting system temperature distribution data, pressure fluctuation data, and heat flux density data; performing multi-scale wavelet decomposition on the temperature distribution data to establish a temperature feature vector set; analyzing the pressure fluctuation data to establish a pressure feature vector set; processing the heat flux density data to establish a heat flux feature vector set; fusing the temperature, pressure, and heat flux feature vector sets to establish a system operating state feature library; extracting historical anomaly pattern samples from similar thermal energy storage system data sources, aligning them with domain features, and establishing a transfer learning source dataset; and constructing a mathematical simulation of the thermal energy storage system. The model is injected with abnormal perturbation parameters to generate a simulated abnormal sample set. Data augmentation processing is then performed to obtain an extended training sample set. The transfer learning source dataset is mixed with the extended training sample set to establish a hybrid training dataset. An abnormal pattern recognition network model is constructed and trained using the hybrid training dataset. The system operating state feature library is input into the trained abnormal pattern recognition network model, which outputs abnormal pattern classification results and abnormal confidence scores. When the abnormal confidence score is greater than a preset abnormal threshold, an abnormal type identifier is extracted. Based on the abnormal type identifier, the system control parameter set is obtained, and the operating parameters of the thermal energy storage system are adjusted. The core of this approach lies in using a consistency regularization framework based on multi-view learning for training.

[0005] In the step of establishing the temperature feature vector set, high-frequency and low-frequency components of the temperature distribution data are extracted. The high-frequency components reflect the rapid change characteristics of temperature, while the low-frequency components reflect the slow change trend of temperature. The multi-scale wavelet decomposition refers to performing wavelet transform on the temperature distribution data at different scales. By utilizing the localization characteristics of wavelet basis functions in the time and frequency domains, the temperature signal is decomposed into high-frequency and low-frequency components of different frequencies.

[0006] In the step of establishing the pressure feature vector set, time-domain and frequency-domain analysis is performed on the pressure fluctuation data to extract the peak frequency and amplitude attenuation coefficient of the pressure fluctuation data. The peak frequency refers to the frequency value corresponding to the maximum amplitude in the spectrum after the pressure fluctuation data is Fourier transformed. The amplitude attenuation coefficient refers to the attenuation rate of the amplitude of the pressure fluctuation data with time in the time domain, which is obtained by calculating the ratio of the peak amplitudes of continuous waves and taking the logarithm.

[0007] In the step of establishing the heat flux feature vector set, gradient calculation is performed on the heat flux density data to obtain the heat flux density gradient field. Topological analysis is then performed on the heat flux density gradient field to extract the singular point distribution characteristics. The topological analysis refers to the geometric structure analysis of the heat flux density gradient field to identify saddle points, maxima, and minima in the heat flux density gradient field. The singular point distribution characteristics include the spatial coordinates of the singular points, the number of singular points, and the connection relationships between the singular points.

[0008] In the step of establishing the system operation status feature library, the temperature feature vector set, pressure feature vector set, and heat flow feature vector set are fused to obtain a fused feature vector, and the fused feature vector is normalized. The feature fusion refers to aligning the temperature feature vector set, pressure feature vector set, and heat flow feature vector set according to the timestamp and then concatenating the vectors to form a fused feature vector containing multi-physics information. The normalization process adopts the maximum and minimum value normalization method, which maps each element of the fused feature vector to the interval [0, 1].

[0009] The data source of the similar thermal energy storage system refers to the historical operation database of other thermal energy storage systems that are similar to the current thermal energy storage system in terms of structural type, capacity level and operating conditions. The historical operation database stores the normal operation data and the marked abnormal operation data recorded by the similar thermal energy storage system in the past.

[0010] In the step of establishing the transfer learning source dataset, the historical abnormal pattern samples are aligned with domain features. The domain feature alignment refers to calculating the distribution difference between the historical abnormal pattern samples and the current thermal energy storage system operation data in the feature space by similar thermal energy storage system data sources, using the maximum mean difference criterion to measure the distance between the two data distributions, and mapping the feature distribution of the historical abnormal pattern samples to a feature space similar to that of the current thermal energy storage system through feature transformation.

[0011] In the step of generating the simulation anomaly sample set, abnormal disturbance parameters are injected into the mathematical simulation model of the thermal energy storage system. The mathematical simulation model of the thermal energy storage system refers to the mathematical description model of the thermal energy storage system established based on the thermodynamic equation and the heat transfer equation. The mathematical simulation model of the thermal energy storage system simulates the temperature field distribution, pressure field distribution and heat flow field distribution of the thermal energy storage system under different operating conditions through numerical solution methods. The abnormal disturbance parameters include the leakage rate parameter of the heat storage medium, the scale thickness parameter of the heat exchanger, the efficiency reduction rate parameter of the circulating pump, and the aging coefficient parameter of the insulation layer.

[0012] The data augmentation process includes random noise injection, time series slice recombination, and feature dimension shuffling on the simulated abnormal sample set. Random noise injection refers to superimposing random noise that follows a normal distribution on the feature values ​​of the simulated abnormal samples. Time series slice recombination refers to randomly recombining the time series data after dividing it according to a time window. Feature dimension shuffling refers to randomly shuffling the dimensional order of the feature vectors.

[0013] The structure of the anomaly pattern recognition network model is as follows: the input layer receives the fused feature vector, the first convolutional layer extracts local features, the second convolutional layer extracts global features, the multi-head attention mechanism layer performs weighted aggregation of features, the recurrent neural network layer captures temporal dependencies, and the fully connected layer outputs the anomaly pattern classification result and anomaly confidence score.

[0014] Specifically, after outputting the abnormal mode classification results and the abnormal confidence score, when the abnormal confidence score is greater than the preset abnormal threshold, the current system operating state is determined to be an abnormal state; in the step of obtaining the system control parameter set according to the abnormal type identifier, the abnormal response strategy library is queried to adjust the operating temperature range, circulation flow parameters and heat exchange power parameters of the thermal energy storage system.

[0015] The number of attention heads in the multi-head attention mechanism layer is determined based on the number of dimensions of the fused feature vector, the number of samples in the system operating state feature library, and the operating temperature range of the thermal energy storage system. When the number of dimensions of the fused feature vector ∈ [1, 64), it is set to 4 attention heads; when the number of dimensions of the fused feature vector ∈ [64, 128), it is set to 8 attention heads; and when the number of dimensions of the fused feature vector ≥ 128, it is set to 16 attention heads.

[0016] In the step of training using a hybrid training dataset, the anomaly pattern recognition network model is trained using a consistency regularization framework based on multi-view learning. Multiple viewpoint samples are generated from the same fused feature vector using different data augmentation methods, including feature perturbation, time window shifting, and scale transformation. Through multi-view consistency constraints, the anomaly pattern recognition network model is forced to learn essential anomaly features that are independent of viewpoint transformation.

[0017] The steps of using the consistency regularization framework based on multi-view learning are as follows: apply feature perturbation to each fused feature vector in the training set to generate view 1 samples, apply time window translation to generate view 2 samples, apply scale transformation to generate view 3 samples, and input view 1 samples, view 2 samples and view 3 samples into the abnormal pattern recognition network model to obtain prediction probability distribution 1, prediction probability distribution 2 and prediction probability distribution 3 respectively.

[0018] Specifically, after summing the three Jensen-Shannon divergences (JS divergences) as the consistency loss, the consistency loss and the classification cross-entropy loss are weighted and summed with weight coefficients of 0.3 and 0.7 to obtain the total loss function. The parameters of the anomaly pattern recognition network model are updated by minimizing the total loss function.

[0019] The fine-tuning training refers to further training the anomaly pattern recognition network model using the training set in the mixed training dataset on the basis of the pre-trained model, adjusting the model parameters to adapt to the anomaly pattern characteristics of the target thermal energy storage system, and the learning rate used in the fine-tuning training is less than the learning rate used in the pre-training.

[0020] This invention achieves high-accuracy identification of abnormal patterns in thermal energy storage systems by constructing a multi-physics feature fusion mechanism combined with a consistency regularization training framework based on multi-view learning. The method first uses multi-scale wavelet decomposition to extract high-frequency and low-frequency components of temperature distribution data, extracts the peak frequency and amplitude attenuation coefficient of pressure fluctuations through time-domain and frequency-domain analysis, and uses topological analysis to extract the singular point distribution characteristics of the heat flux density gradient field. The three types of features are fused to form a comprehensive feature representation containing multi-physics information. This multi-physics feature fusion breaks through the limitations of traditional single-parameter monitoring and can comprehensively capture the coupling change law of temperature, pressure and heat flux density, providing richer and more complete feature information for abnormal pattern identification. Compared to the limitations of traditional methods that rely solely on single-parameter threshold judgments or shallow machine learning, this invention extracts local and global features through convolutional layers in a deep neural network structure, weights and aggregates these features using a multi-head attention mechanism, and captures temporal dependencies through recurrent neural network layers. This enables the extraction of more abstract and essential deep anomaly features from the fused feature vectors. Consequently, the anomaly pattern recognition network maintains high accuracy in recognizing various anomalies, including abnormal temperature distribution, abnormal pressure fluctuations, abnormal heat flux density, as well as leakage of the thermal storage medium, scaling of the heat exchanger, decreased efficiency of the circulating pump, and aging of the insulation layer. This solves the technical problem of insufficient accuracy in anomaly pattern recognition of thermal energy storage systems mentioned in the background art. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a spatial distribution map of deep anomaly features in Example 2.

[0023] Figure 3 This is a topological diagram of the heat flux density gradient field in Example 2. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] like Figure 1 The diagram shown is a flowchart of a method for anomaly pattern recognition in a thermal energy storage system provided by the present invention. This method includes the following steps: S01. During the operation of the thermal energy storage system, collect system temperature distribution data, pressure fluctuation data and heat flux density data, perform multi-scale wavelet decomposition on the temperature distribution data, extract the high-frequency components and low-frequency components of the temperature distribution data, and establish a temperature feature vector set. S02. Perform time-domain and frequency-domain analysis on the pressure fluctuation data, extract the peak frequency and amplitude attenuation coefficient of the pressure fluctuation data, and establish a pressure feature vector set; S03. Perform gradient calculation on the heat flux density data to obtain the heat flux density gradient field, perform topological analysis on the heat flux density gradient field, extract the singular point distribution characteristics of the heat flux density gradient field, and establish a heat flux feature vector set. S04. The temperature feature vector set, the pressure feature vector set, and the heat flow feature vector set are fused to obtain a fused feature vector. The fused feature vector is then normalized to establish a system operating status feature library. S05. Extract historical abnormal pattern samples from similar thermal energy storage system data sources, perform domain feature alignment on the historical abnormal pattern samples, and establish the aligned historical abnormal pattern samples as a transfer learning source dataset. S06. Construct a mathematical simulation model of the thermal energy storage system, inject different types of abnormal disturbance parameters into the mathematical simulation model of the thermal energy storage system, generate a simulation abnormal sample set, perform data augmentation processing on the simulation abnormal sample set, and obtain an extended training sample set. S07. Mix the transfer learning source dataset with the extended training sample set to establish a mixed training dataset. Divide the mixed training dataset into a training set and a validation set in a 7:3 ratio. S08. Construct an anomaly pattern recognition network model, pre-train and fine-tune the anomaly pattern recognition network model using the training set, and verify the training process using the validation set to obtain the trained anomaly pattern recognition network model. S09. Input the system operation status feature library into the trained anomaly pattern recognition network model, extract deep anomaly features through the trained anomaly pattern recognition network model, and output anomaly pattern classification results and anomaly confidence scores. S10. Calculate the difference between the anomaly confidence score and the preset anomaly threshold. When the difference is greater than 0, determine that the current system operation state is an abnormal state and extract the anomaly type identifier corresponding to the anomaly pattern classification result. S11. Query the anomaly response strategy library according to the anomaly type identifier, obtain the corresponding system control parameter set, and adjust the working temperature range, circulation flow parameters and heat exchange power parameters of the thermal energy storage system so that the anomaly confidence score is reduced to below the preset anomaly threshold.

[0026] The multi-scale wavelet decomposition refers to performing wavelet transform on temperature distribution data at different scales. By leveraging the localization characteristics of wavelet basis functions in the time and frequency domains, the temperature signal is decomposed into high-frequency and low-frequency components with different frequencies. The high-frequency components reflect the rapid temperature change characteristics, while the low-frequency components reflect the slow temperature change trend. A temperature feature vector set is constructed by extracting wavelet coefficients at different scales.

[0027] The peak frequency refers to the frequency value corresponding to the maximum amplitude in the spectrum after the pressure fluctuation data is Fourier transformed. The amplitude attenuation coefficient refers to the attenuation rate of the amplitude of the pressure fluctuation data over time in the time domain, which is obtained by calculating the ratio of the peak amplitudes of consecutive waves and taking the logarithm.

[0028] The topology analysis refers to the geometric structure analysis of the heat flux density gradient field, identifying saddle points, maxima, and minima in the heat flux density gradient field. The distribution characteristics of the singular points include the spatial coordinates of the singular points, the number of singular points, and the connection relationships between the singular points.

[0029] The feature fusion refers to aligning the temperature feature vector set, pressure feature vector set, and heat flow feature vector set according to the timestamp and then concatenating the vectors to form a fused feature vector containing multi-physics information. The normalization process adopts the maximum-minimum value normalization method to map each element of the fused feature vector to the interval [0, 1].

[0030] The system operation status feature library refers to a data set formed by organizing the normalized fused feature vectors according to the time series, which is used to store the operation status features of the thermal energy storage system at different times.

[0031] The similar thermal energy storage system data source refers to the historical operation database of other thermal energy storage systems that are similar to the current thermal energy storage system in terms of structural type, capacity level and operating conditions. The historical operation database stores the normal operation data and the marked abnormal operation data recorded by the similar thermal energy storage systems in the past.

[0032] The aforementioned domain feature alignment refers to calculating the distribution difference between historical anomaly pattern samples from similar thermal energy storage system data sources and current thermal energy storage system operating data in the feature space, using the maximum mean difference criterion to measure the distance between the two data distributions, and mapping the feature distribution of the historical anomaly pattern samples to a feature space similar to that of the current thermal energy storage system through feature transformation, thereby reducing the data distribution offset between different systems.

[0033] The transfer learning source dataset refers to a set of historical abnormal pattern samples aligned with domain features. It is used to transfer the abnormal pattern recognition knowledge learned by the source system to the target system through pre-training when the training data of the target thermal energy storage system is insufficient, thereby accelerating the convergence of the target system model and improving the generalization ability of the model.

[0034] The mathematical simulation model of the thermal energy storage system refers to a mathematical description model of the thermal energy storage system established based on thermodynamic equations and heat transfer equations. The mathematical simulation model of the thermal energy storage system simulates the temperature field distribution, pressure field distribution, and heat flow field distribution of the thermal energy storage system under different operating conditions through numerical solution methods. The abnormal disturbance parameters include the leakage rate parameter of the thermal storage medium, the fouling thickness parameter of the heat exchanger, the efficiency reduction rate parameter of the circulating pump, and the aging coefficient parameter of the insulation layer. By setting different ranges of abnormal disturbance parameter values ​​in the mathematical simulation model of the thermal energy storage system, simulation operation data corresponding to different abnormal modes are generated.

[0035] The data augmentation process includes injecting random noise, recombining time series slices, and shuffling feature dimensions into the simulated abnormal sample set to generate diverse training samples, increase the coverage of training data, and enhance the model's adaptability to data changes. The random noise injection refers to superimposing normally distributed random noise onto the feature values ​​of the simulated abnormal samples. The time series slice recombining refers to randomly recombining the time series data after dividing it according to a time window. The feature dimension shuffling refers to randomly shuffling the dimensional order of the feature vectors.

[0036] The structure of the anomaly pattern recognition network model is as follows: the input layer receives the fused feature vector, the first convolutional layer extracts local features, the second convolutional layer extracts global features, the multi-head attention mechanism layer performs weighted aggregation of features, the recurrent neural network layer captures temporal dependencies, and the fully connected layer outputs the anomaly pattern classification result and anomaly confidence score. The number of attention heads in the multi-head attention mechanism layer is determined according to the dimension of the fused feature vector, the number of samples in the system operating state feature library, and the operating temperature range of the thermal energy storage system. When the dimension of the fused feature vector is in the interval [1, 64), it is set to 4 attention heads; when the dimension of the fused feature vector is in the interval [64, 128), it is set to 8 attention heads; and when the dimension of the fused feature vector is ≥128, it is set to 16 attention heads.

[0037] The aforementioned consistency regularization framework based on multi-view learning refers to constructing multiple observation representations from different perspectives or transformations of the same data. Consistency loss constrains the consistency of predictions from different perspectives, strengthening the model's invariant feature learning to data transformations. This enables the anomaly pattern recognition network model to extract more stable discriminative features when facing similar anomaly patterns, reducing the model's sensitivity to noise from a single perspective. The steps of using the consistency regularization framework based on multi-view learning are as follows: Apply feature perturbation to each fused feature vector in the training set to generate perspective 1 samples; apply time window translation to generate perspective 2 samples; apply scale transformation to generate perspective 3 samples; and then... Angle 1 sample, viewpoint 2 sample, and viewpoint 3 sample are input into the anomaly pattern recognition network model to obtain predicted probability distribution 1, predicted probability distribution 2, and predicted probability distribution 3, respectively. The Jensen-Shannon divergence between predicted probability distribution 1 and predicted probability distribution 2, the Jensen-Shannon divergence between predicted probability distribution 1 and predicted probability distribution 3, and the Jensen-Shannon divergence between predicted probability distribution 2 and predicted probability distribution 3 are calculated. The sum of the three Jensen-Shannon divergences is used as the consistency loss. The consistency loss and the classification cross-entropy loss are weighted and summed according to the weight coefficients of 0.3 and 0.7 to obtain the total loss function. The parameters of the anomaly pattern recognition network model are updated by minimizing the total loss function.

[0038] The technical effects of the consensus regularization framework based on multi-view learning are as follows: By constraining the prediction consistency of the same data from different perspectives, it strengthens the invariant feature learning of the anomaly pattern recognition network model to data transformations. This enables the anomaly pattern recognition network model to extract more stable discriminative features when facing similar anomaly patterns such as abnormal temperature distribution, abnormal pressure fluctuations, and abnormal heat flux density. It also reduces the sensitivity of the anomaly pattern recognition network model to noise from a single perspective, thereby improving the accuracy of anomaly pattern classification. Furthermore, the consensus regularization framework based on multi-view learning introduces an adversarial training mechanism, generating adversarial perturbation samples during training. By minimizing the prediction loss of adversarial samples, it enhances the robustness of the anomaly pattern recognition network model. This approach enables the anomaly pattern recognition network model to maintain stable recognition performance under different operating conditions and noise interference. It employs mixed-precision training technology, using 16-bit floating-point calculations for parallel forward propagation of samples from multiple viewpoints and 32-bit floating-point calculations for gradient updates. This accelerates the multi-view learning training process and reduces memory usage. In principle, the consistency regularization framework based on multi-view learning utilizes the inherent consistency of anomaly patterns in different representation spaces. Through multi-view consistency constraints, the anomaly pattern recognition network model is forced to learn essential anomaly features independent of viewpoint changes, rather than relying on surface features from a single viewpoint. This enhances the network model's ability to distinguish between different anomaly patterns with similar features.

[0039] The trained anomaly pattern recognition network model refers to an anomaly pattern recognition network model that has reached the convergence condition on the validation set after pre-training and fine-tuning training. The convergence condition is that the classification accuracy on the validation set changes by less than 0.1% over 10 consecutive training cycles.

[0040] The deep anomaly features refer to the high-level abstract features extracted from the fused feature vector by the anomaly pattern recognition network model through multiple convolutional layers and recurrent neural network layers. The deep anomaly features include comprehensive characterization information of temperature distribution anomalies, pressure fluctuation anomalies, and heat flux density anomalies.

[0041] The abnormal mode classification results include temperature distribution abnormality classification, pressure fluctuation abnormality classification, heat flux density abnormality classification, and comprehensive abnormality classification. The temperature distribution abnormality classification includes local overheating mode, overall temperature rise mode, and temperature gradient abnormality mode. The pressure fluctuation abnormality classification includes pressure drop mode, pressure oscillation mode, and pressure rise mode. The heat flux density abnormality classification includes heat flux concentration mode, heat flux dispersion mode, and heat flux reversal mode. The comprehensive abnormality classification includes heat storage medium leakage mode, heat exchanger scaling mode, circulating pump efficiency reduction mode, and insulation layer aging mode.

[0042] The anomaly confidence score refers to the probability value output by the anomaly pattern recognition network model that indicates the current system operating state is abnormal. The anomaly confidence score ranges from [0, 1], and the closer the anomaly confidence score is to 1, the higher the degree of anomaly.

[0043] The preset anomaly threshold is determined based on the safety operation requirements of the thermal energy storage system and the statistics of historical anomalies. When the thermal energy storage system has high safety requirements, the preset anomaly threshold is set to be in the range [0.5, 0.7). When the thermal energy storage system allows a certain false alarm rate, the preset anomaly threshold is set to be in the range [0.7, 0.9].

[0044] The abnormality type identifiers include temperature distribution abnormality identifiers, pressure fluctuation abnormality identifiers, heat flux density abnormality identifiers, heat storage medium leakage identifiers, heat exchanger scaling identifiers, circulating pump efficiency reduction identifiers, and insulation layer aging identifiers. Each identifier corresponds to a different abnormality mode category, which is used to guide the selection of subsequent system control strategies.

[0045] The anomaly response strategy library stores system control schemes corresponding to different anomaly type identifiers, including upper and lower limit adjustment values ​​of the operating temperature range, increase / decrease ratio of the circulation flow parameter, and adjustment range of the heat exchange power parameter. After obtaining the corresponding system control parameter set according to the anomaly type identifier, the system control system executes parameter adjustment operations.

[0046] In addition, the present invention also provides a method for forming an abnormal pattern recognition system for thermal energy storage system by means of a computer, wherein the computer is provided with a storage medium, the storage medium stores program instructions, and the program instructions execute the above-mentioned abnormal pattern recognition method for thermal energy storage system when running in the computer.

[0047] The specific implementation methods of the above steps are described in detail below.

[0048] The specific implementation of step S01 involves deploying temperature sensor arrays, pressure sensors, and heat flux density sensors at key locations in the thermal energy storage system. The sampling frequency of the sensors is set to 1–10 Hz to ensure the capture of dynamic changes during system operation. The collected temperature distribution data is first preprocessed to remove outliers and noise generated during the acquisition process. Then, multi-scale wavelet decomposition technology is used to analyze the temperature distribution data. Daubechies wavelet or Morlet wavelet is selected as the wavelet basis function, and wavelet transform is performed on the temperature distribution data at 3–5 different scale levels. Through wavelet transform, the temperature signal can be decomposed into inverse... The components reflect different frequency characteristics. High-frequency components correspond to rapid temperature fluctuations and abrupt changes, while low-frequency components correspond to slow temperature changes and overall distribution patterns. Wavelet coefficients at each scale level, including detail coefficients and approximation coefficients, are extracted and organized into vectors according to time order and spatial location to construct a temperature feature vector set. This feature vector set can comprehensively reflect the temperature distribution characteristics of the thermal energy storage system at different time and spatial scales. The role of multi-scale wavelet decomposition is to overcome the limitations of traditional single-scale analysis methods, and it can simultaneously characterize the local detail features and global trend features of temperature distribution, providing rich feature information for subsequent anomaly pattern recognition.

[0049] The specific implementation of step S02 involves first performing time-domain analysis on the collected pressure fluctuation data to calculate the statistical characteristic parameters of the pressure fluctuations, including mean, variance, peak-to-peak value, and fluctuation amplitude. Then, a fast Fourier transform algorithm is used to transform the pressure fluctuation data from the time domain to the frequency domain, obtaining the spectral distribution of the pressure fluctuations. The frequency component with the largest amplitude in the spectrum is identified as the peak frequency. The peak frequency reflects the main oscillation characteristics of the pressure fluctuations and has an important indicative role in identifying potential resonance phenomena or periodic disturbances in the pressure system. The amplitude attenuation coefficient is calculated by identifying continuous pressure fluctuation peaks in the time domain. The ratio of the amplitudes of adjacent peaks is calculated, and the logarithm of the ratio sequence is taken for linear fitting. The slope of the fitted line is the amplitude attenuation coefficient, which reflects the attenuation rate of pressure fluctuations over time and can characterize the damping characteristics and energy dissipation mechanism of the system. The extracted peak frequencies, amplitude attenuation coefficients, and time-domain statistical feature parameters are organized into a pressure feature vector set. This feature vector set comprehensively describes the dynamic characteristics of pressure fluctuations from both time and frequency domains. The combination of time-domain and frequency-domain analysis can complementarily extract different dimensions of pressure fluctuation features, providing more comprehensive pressure system state information for anomaly pattern recognition.

[0050] The specific implementation of step S03 involves constructing a two-dimensional or three-dimensional heat flux density field from the collected heat flux density data. The mesh division of the heat flux density field is determined based on the spatial structure of the thermal energy storage system and the arrangement of measuring points. The gradient of the heat flux density field is calculated using the finite difference method or the finite element method to obtain the rate of change of heat flux density in each direction in space, forming a heat flux density gradient field. This gradient field reflects the direction and intensity distribution of heat transfer. Topological analysis of the gradient field is performed by calculating the Hessian matrix at each point in the gradient field. Based on the eigenvalues ​​of the Hessian matrix, it is determined whether a point is a critical point, including saddle points, maxima, and minima. Singularity points and saddle points correspond to transition regions in the heat flux density gradient field. Maximum points correspond to heat flux convergence regions, and minimum points correspond to heat flux divergence regions. By statistically analyzing the spatial coordinates, number, and topological connections of these singularity points, a singularity distribution feature is formed. This feature is then encoded into a feature vector, creating a heat flux feature vector set. Topological analysis of the heat flux density gradient field can reveal the heat transfer patterns and heat flux distribution structure within a thermal energy storage system. This is crucial for identifying abnormal patterns such as heat exchanger blockage and insulation layer damage. Compared to simple numerical statistics, topological analysis provides richer geometric information and spatial correlation features.

[0051] The specific implementation of step S04 involves aligning the temperature feature vector set, pressure feature vector set, and heat flux feature vector set according to the acquisition timestamps to ensure that feature data from different sensors correspond to the system state at the same moment. After alignment, the three feature vector sets are concatenated to form a fused feature vector containing multi-physics information of temperature, pressure, and heat flux. The dimension of the fused feature vector is equal to the sum of the dimensions of the three feature vector sets. To eliminate differences in dimensions and numerical ranges between different physical quantities, a maximum-minimum normalization method is used to normalize the fused feature vector. The normalization formula is to subtract the minimum value of each feature element from the minimum value of that feature in the entire dataset, and then divide by the feature's minimum value. The difference between the maximum and minimum values ​​maps the normalized feature values ​​to the range of 0 to 1. The purpose of normalization is to prevent features with large numerical ranges from dominating the model training process and to ensure that different physical field features have the same weight in subsequent analysis. The fused feature vectors after normalization are organized and stored in chronological order to form a system operating state feature library. This feature library records the multi-physical field coupling state of the thermal energy storage system at different times, providing input data for the anomaly pattern recognition network model. The feature fusion method can comprehensively utilize the complementary information of three physical quantities: temperature, pressure, and heat flow. Compared with the analysis of a single physical quantity, it can more comprehensively reflect the operating state and anomaly characteristics of the system.

[0052] The specific implementation of step S05 involves extracting historical anomaly pattern samples from the historical operation database of other thermal energy storage systems that are similar to the current thermal energy storage system in terms of structural type, capacity level, and operating conditions. These historical anomaly pattern samples contain labeled anomaly types and corresponding feature data. Since different thermal energy storage systems have differences in design parameters, operating environments, and measurement methods, directly using these historical samples may lead to data distribution shifts. Therefore, domain feature alignment is required. Domain feature alignment uses the maximum mean difference criterion as a measurement method. This criterion measures the distribution difference by calculating the distance between the mean vectors of two data distributions in the regenerative kernel Hilbert space. A feature transformation matrix is ​​constructed, and the source is minimized... The maximum mean difference between the system data distribution and the target system data distribution is used to align the feature space. The optimization process of the feature transformation matrix is ​​solved iteratively using gradient descent or quasi-Newton methods. The distribution of the aligned historical anomaly pattern samples in the feature space is closer to the data distribution of the current thermal energy storage system, reducing the domain offset between different systems. The aligned historical anomaly pattern sample set is used as the transfer learning source dataset. The role of this dataset is to use the anomaly pattern recognition knowledge already learned by the source system to accelerate the training process of the target system model through transfer learning, and improve the model's generalization ability on the target system. Especially when there are insufficient labeled samples in the target system, the transfer learning source dataset can effectively alleviate the data scarcity problem.

[0053] The specific implementation of step S06 is to construct a mathematical simulation model of the thermal energy storage system based on the first law of thermodynamics and the fundamental equations of heat transfer. This simulation model includes the energy conservation equation, momentum conservation equation, and heat transfer equation. The governing equations are discretized using the finite volume method or finite element method. Geometric parameters, material thermophysical property parameters, and boundary condition parameters of the simulation model are set to ensure consistency with the actual thermal energy storage system. Abnormal disturbance parameters are introduced into the simulation model to simulate different types of abnormal operating conditions. The leakage rate parameter of the heat storage medium is set to a range of 0 to 10% per hour, the scale thickness parameter of the heat exchanger is set to a range of 0 to 5 mm, the efficiency reduction rate parameter of the circulating pump is set to a range of 0 to 30%, and the aging coefficient parameter of the insulation layer is set to a range of 1.0 to 2.5. By randomly sampling or systematically setting different combinations of abnormal disturbance parameters within the set range, the simulation model is run to obtain temperature field distribution, pressure field distribution, and heat flow field distribution data corresponding to different abnormal modes. Feature vectors are extracted from the simulation results to form a simulation abnormal sample set. To increase the diversity and coverage of the training data... The scope of this study involves data augmentation of the simulated abnormal sample set, including three augmentation methods: random noise injection, time series slice reconstruction, and feature dimension shuffling. Random noise injection involves superimposing normally distributed random noise with a mean of 0 and a standard deviation of 5% to 15% of the standard deviation of the eigenvalues ​​onto the eigenvalues ​​of the simulated samples. Time series slice reconstruction involves dividing the time series data into windows of 10 to 30 time steps and then randomly recombining them to form new time series samples. Feature dimension shuffling involves randomly shuffling the dimensional order of the feature vectors while maintaining the integrity of the time series within each dimension. After data augmentation, an expanded training sample set is obtained, which is typically 3 to 5 times the size of the original simulation sample set. The role of the mathematical simulation model is to supplement abnormal samples that are difficult to obtain in the actual system, especially some rare or safety-risk abnormal operating conditions. Through simulation, a large number of labeled samples can be generated without affecting the operation of the actual system. Data augmentation further improves the diversity of training data and enhances the model's adaptability and robustness to data changes.

[0054] The specific implementation of step S07 involves merging the transfer learning source dataset and the extended training sample set according to the number of samples to form a hybrid training dataset. During the merging process, it is necessary to ensure that the feature dimensions of the two datasets are consistent. If the dimensions are inconsistent, feature alignment or padding is required. The hybrid training dataset is randomly shuffled to avoid the influence of sample order on model training. Then, the hybrid training dataset is divided into a training set and a validation set according to a ratio of 70% to 30%. The training set is used for learning and updating model parameters, and the validation set is used to monitor the model training process and evaluate the model's generalization performance. The partitioning process adopts a hierarchical sampling method to ensure that the proportion of various abnormal pattern samples in the training set and validation set is consistent with the proportion in the hybrid training dataset, avoiding the impact of class imbalance on model performance. The construction of the hybrid training dataset integrates historical data from similar systems and data generated based on simulation models, which can provide rich abnormal pattern information for model training and improve the model's ability to identify different abnormal patterns.

[0055] The specific implementation of step S08 involves constructing an abnormal pattern recognition network model. This model employs a deep learning architecture, including an input layer, a first convolutional layer, a second convolutional layer, a multi-head attention mechanism layer, a recurrent neural network layer, and a fully connected layer. The input layer receives a fused feature vector from a system runtime feature library. The first convolutional layer uses 32 to 64 convolutional kernels, with kernel sizes of 3×3 or 5×5, and employs the ReLU activation function to extract local feature patterns from the fused feature vector. The second convolutional layer uses 64 to 128 convolutional kernels, also employing the ReLU activation function, to extract higher-level global features. The number of attention heads in the multi-head attention mechanism layer is set according to the dimension of the fused feature vector. When the number of dimensions is less than 64, 4 attention heads are used; when the number of dimensions is between 64 and 128, 8 attention heads are used; and when the number of dimensions is greater than or equal to 128, 16 attention heads are used. The multi-head attention mechanism can weighted aggregate different feature channels, highlighting features more important for anomaly pattern recognition. The recurrent neural network layer uses a long short-term memory network or gated recurrent unit structure, with the number of hidden layer units set to 64 to 256 to capture long-term dependencies in time-series data. The fully connected layer includes two output branches: one branch outputs the anomaly pattern classification result using the Softmax activation function, and the other branch outputs the anomaly confidence score using the Sigmoid activation function. During the pre-training phase... Using a consistency regularization framework based on multi-view learning, three different data augmentation methods are applied to each fused feature vector in the training set to generate samples from three perspectives. Feature perturbation involves superimposing random perturbations of 10% to 20% of the feature standard deviation onto the original feature values. Time window shifting involves shifting the time series forward or backward by 1 to 5 time steps. Scaling transformation involves scaling the feature values ​​by 0.8 to 1.2 times. The samples from the three perspectives are then input into the anomaly pattern recognition network model to obtain three predicted probability distributions. The Jensen-Shannon divergence among the three predicted probability distributions is calculated, which measures the difference between the different predicted distributions. The sum of the Jensen-Shannon divergences of the three pairs of predicted probability distributions is used as the consistency loss, which is used to reduce the consistency of the data. The bundle model maintains consistent predictions for samples from different perspectives. The total loss function is obtained by weighting and summing the classification cross-entropy loss with weight coefficients of 0.3 and 0.7. Model parameters are updated using backpropagation and the Adam optimizer, with a learning rate set to 0.001 to 0.01. The pre-training phase lasts 50 to 100 training epochs. In the fine-tuning phase, the pre-trained model is further trained using the training set, with the learning rate reduced to 0.1 to 0.3 times that of the pre-training phase (typically 0.0001 to 0.001). Fine-tuning lasts 20 to 50 training epochs. During training, a validation set is used to monitor model performance. The model is considered successful when the classification accuracy on the validation set changes by less than 0 for 10 consecutive training epochs.At 1%, the model is considered to have reached convergence, and training is stopped, resulting in a completed anomaly pattern recognition network model. The consistency regularization framework based on multi-view learning strengthens the model's invariant feature learning to data transformations by constraining the consistency of predictions from different perspectives on the same data. This enables the model to extract more stable discriminative features, reduces its sensitivity to noise from a single perspective, and improves the accuracy and robustness of anomaly pattern classification.

[0056] The specific implementation of step S09 involves inputting the fused feature vector from the system operating state feature library into the trained anomaly pattern recognition network model. Through the model's forward propagation process, the fused feature vector is processed sequentially through convolutional layers, multi-head attention mechanism layers, and recurrent neural network layers. In these layers, the model extracts deep anomaly features. These deep anomaly features are high-level abstract feature representations obtained after multiple nonlinear transformations, which can comprehensively reflect the coupling relationship and intrinsic connection between temperature distribution anomalies, pressure fluctuation anomalies, and heat flux density anomalies. The model's output layer generates anomaly pattern classification results and anomaly confidence scores. The anomaly pattern classification results include specific anomaly types, such as local overheating patterns, pressure drop patterns, heat flux concentration patterns, and heat storage medium leakage patterns. The anomaly confidence score is a value between 0 and 1, representing the probability that the current system operating state is an anomaly. The closer the score is to 1, the greater the possibility of system anomaly. The classification results and confidence scores output by the model provide quantitative basis for subsequent anomaly judgment and response.

[0057] The specific implementation of step S10 involves comparing the anomaly confidence score with a preset anomaly threshold and calculating the difference between the two. The preset anomaly threshold is determined based on the safety operation requirements of the thermal energy storage system. For systems with high safety requirements, the threshold is set to 0.5 to 0.7. For systems that allow for a certain false alarm rate, the threshold is set to 0.7 to 0.9. When the difference is greater than 0, it indicates that the anomaly confidence score exceeds the preset threshold, and the current system operating state is determined to be an abnormal state. At this time, the anomaly type identifier corresponding to the anomaly mode classification result is extracted. The anomaly type identifier includes temperature distribution anomaly identifier, pressure fluctuation anomaly identifier, heat flux density anomaly identifier, thermal storage medium leakage identifier, heat exchanger scaling identifier, circulating pump efficiency reduction identifier, and insulation layer aging identifier. Each identifier corresponds to a specific anomaly mode, providing guidance for the subsequent selection of system control strategies. By setting a preset anomaly threshold, it is possible to identify and warn of system anomalies in the early stages, avoiding the expansion of anomalies that could lead to system failures or safety accidents.

[0058] The specific implementation of step S11 involves querying the anomaly response strategy library based on the extracted anomaly type identifier. This strategy library stores system control schemes corresponding to different anomaly types, including adjustment values ​​for the operating temperature range, adjustment ratios for circulation flow parameters, and adjustment magnitudes for heat exchange power parameters. For abnormal temperature distribution, it is usually necessary to adjust the upper or lower limit of the operating temperature range, with an adjustment magnitude of 5°C to 20°C. For abnormal pressure fluctuations, it is usually necessary to adjust the circulation flow parameter, with an adjustment ratio of 10% to 30%. For abnormal heat flux density, it is usually necessary to adjust the heat exchange power parameter, with an adjustment magnitude of 5% to 25%. For comprehensive anomalies such as leakage of the heat storage medium or... Scale buildup in heat exchangers requires the simultaneous adjustment of multiple parameters. Based on the obtained set of control parameters, the control system sends control commands to the thermal energy storage system to execute parameter adjustment operations. After adjustment, the change in the anomaly confidence score is continuously monitored. When the anomaly confidence score drops below the preset anomaly threshold, it indicates that the system anomaly has been effectively suppressed and the system has returned to normal operation. If the anomaly confidence score is still higher than the threshold after adjustment, further control efforts or other emergency measures need to be taken. The establishment of the anomaly response strategy library is based on the experience of handling historical anomalies and the analysis of system dynamic characteristics, which can provide decision support for rapid response and effective control under abnormal operating conditions.

[0059] Specifically, the principle of this invention is as follows: This invention can solve the core technical problem of insufficient accuracy in anomaly pattern recognition of thermal energy storage systems. Its fundamental principle lies in establishing a complete representation space for anomaly patterns through multi-physics feature fusion, and effectively extracting essential anomaly features using a deep neural network combined with a consistency regularization framework based on multi-view learning. From an information theory perspective, the accuracy of anomaly pattern recognition in thermal energy storage systems depends on the feature space's ability to distinguish anomaly patterns and the model's learning ability to discriminate features. The fundamental reason for the insufficient accuracy of traditional methods is that monitoring a single physical parameter leads to incomplete feature information, and shallow models have limited ability to express complex nonlinear anomaly features. This invention extracts temperature distribution data at different times through multi-scale wavelet decomposition. The scale-based variation features capture both the rapid temperature fluctuations reflected by high-frequency components and the slow temperature evolution trend reflected by low-frequency components. The peak frequency of pressure fluctuations is extracted through time-domain and frequency-domain analysis to reflect oscillation characteristics, and the amplitude attenuation coefficient reflects the energy dissipation law. The saddle points, maxima, and minima in the heat flux density gradient field and their spatial connections are identified through topological analysis. These multi-scale, multi-domain, and multi-structure feature extraction methods comprehensively characterize the features of anomalous patterns from different physical perspectives. The high-dimensional feature vector formed after feature fusion contains the coupling information of the temperature field, pressure field, and heat flux field, enabling anomalous patterns that were originally difficult to distinguish in a single feature space to form clearer inter-class boundaries in the fused feature space, thereby improving the discriminative ability of the feature space.

[0060] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0061] The specific implementation of step S01 involves deploying a temperature sensor array, a pressure sensor, and a heat flux density sensor at key locations in the thermal energy storage system. The sampling frequency of the sensors is set to 1–10 Hz. After preprocessing the collected temperature distribution data, multi-scale wavelet decomposition technology is used to analyze the temperature distribution data. The formula for wavelet decomposition is as follows: ; In the formula, For the first Scale No. The wavelet coefficients of the position are dimensionless. This is a scale parameter, with values ​​ranging from 1 to... ; For position parameters; for Temperature distribution data at any given time, in °C; For reference temperature, the empirical value is 50℃; Total sampling time, in seconds; For reference time, the default value is 1 second; Here are wavelet basis functions, defined as follows: ,in For normalized time , The mother wavelet function is the Daubechies wavelet or Morlet wavelet, which has localization properties in the normalized time domain.

[0062] The formula for constructing the temperature feature vector set is as follows: ; In the formula, It is a set of temperature feature vectors; The number of wavelet coefficients at the first scale; For the first The number of wavelet coefficients at the scale; The maximum decomposition scale is typically set to 3 to 5; for arbitrary scales... Number of wavelet coefficients It depends on the signal length and the characteristics of the wavelet basis function.

[0063] The parameter acquisition method is as follows: The temperature data is acquired in real time using temperature sensors, including the following steps: Step 1: Arrange 5 to 10 temperature sensors along the height of the wall of the thermal energy storage tank; Step 2: Set the sensor acquisition frequency to 1 to 10 Hz and continuously acquire temperature data; Step 3: Filter the acquired data to remove high-frequency noise. The value is determined based on the sampling frequency and the desired frequency range for analysis, and is typically set to 3 to 5.

[0064] The specific implementation of step S02 involves performing time-domain and frequency-domain analysis on the collected pressure fluctuation data. The peak frequency is obtained through Fast Fourier Transform, and the formula for calculating the spectral amplitude is as follows: ; In the formula, For frequency The corresponding normalized spectral amplitude is dimensionless. for Pressure fluctuation data at any given time, in Pa; For reference pressure, the empirical value is... Pa; Angular frequency, in rad / s; The imaginary unit satisfies ; Modular operation is used to calculate the modulus of a complex number.

[0065] The formula for determining the peak frequency is expressed as follows: ; In the formula, This represents the angular frequency corresponding to the peak frequency, expressed in rad / s.

[0066] The formula for calculating the amplitude attenuation coefficient is as follows: ; In the formula, This is the amplitude attenuation coefficient, in units of... ; The number of pressure peaks identified; The peak number; This serves as a reference time interval between adjacent peaks, typically ranging from 0.1 to 1.0 seconds. For the first The amplitude of each peak, in Pa, is obtained by peak detection of the pressure time series.

[0067] The formula for constructing the pressure feature vector set is as follows: ; In the formula, This is a set of pressure feature vectors, where each element is a dimensionless quantity. This is the average of the pressure fluctuation data, in Pa. This represents the standard deviation of the pressure fluctuation data, in Pa. This is the reference angular frequency, which defaults to 1 rad / s.

[0068] The parameter acquisition method is as follows: The pressure sensor is used to acquire data in real time, including step 1: installing pressure sensors at key locations in the circulation pipeline of the thermal energy storage system; step 2: setting the sensor acquisition frequency to 10 to 50 Hz; and step 3: performing bandpass filtering on the acquired data. The peak value is obtained through a peak detection algorithm, and the detection threshold is set to 1.5 to 2.5 times the standard deviation of pressure fluctuation. and Obtained through statistical analysis and calculation. , ,in This represents the total number of pressure data sampling points. The sampling point number, For the first Each sampling time.

[0069] The specific implementation of step S03 involves constructing a heat flux density field from the collected heat flux density data, and then using the finite difference method to calculate the gradient of the heat flux density field. The formula for calculating the gradient field is as follows: ; In the formula, For position The normalized heat flux density gradient at point is dimensionless. For position Heat flux density at , in units of ; and These are spatial coordinates, in meters (m). For reference heat flux density, the empirical value is ; This is a reference length, with a default value of 1m.

[0070] In topological analysis, singularities are identified by calculating the eigenvalues ​​of the Hessian matrix. The formula for calculating the Hessian matrix is ​​as follows: ; In the formula, For position The normalized Hessian matrix at point , where all matrix elements are dimensionless. The eigenvalues ​​of the Hessian matrix. and By solving the characteristic equation Obtain, among which For eigenvalues, It is the identity matrix. It is the determinant of a matrix. and All are dimensionless quantities used to determine the type of singularity.

[0071] The formula for constructing the heat flux feature vector set is as follows: ; In the formula, This is a set of heat flow feature vectors, where each element is a dimensionless quantity. The number of saddle points; This represents the number of maximum points. This represents the number of local minima. This represents the average distance between singular points, expressed in meters (m).

[0072] The parameter acquisition method is as follows: The heat flux density is measured using heat flux density sensors, including the following steps: Step 1: Arranging a heat flux density sensor array on the surface of the heat exchanger in the thermal energy storage system; Step 2: Setting the sensor measurement frequency to 0.1 to 1 Hz; Step 3: Constructing a continuous heat flux density field using interpolation methods. Singularities are determined through calculation. eigenvalues and Achieve, when When it is determined to be a saddle point, and When it is determined to be a local minimum point, and This is determined to be a maximum point. Average distance ,in The total number of singularities. and The singularity number, and The first The and the first The spatial coordinates of the singular points are given in meters.

[0073] The specific implementation of step S04 involves aligning the temperature feature vector set, pressure feature vector set, and heat flow feature vector set according to timestamps and then concatenating the vectors. The calculation formula for feature fusion is expressed as follows: ; In the formula, To fuse feature vectors, since , and All elements have been normalized to dimensionless quantities, therefore All elements are also dimensionless.

[0074] The normalization process uses the maximum-minimum normalization method, and the normalization formula is expressed as follows: ; In the formula, For the fusion feature vector of the th The normalized value of each element is dimensionless. The element index ranges from 1 to the total dimension of the fused feature vector. For the fusion feature vector of the th The original values ​​of each element; For the first The minimum value of each feature in the entire dataset; For the first The maximum value of each feature in the entire dataset; To prevent small constants from being divided by zero, the default value is 0. .

[0075] The formula for constructing the system operation status feature library is expressed as follows: ; In the formula, This is a system operation status feature library; for Time-normalized fused feature vector; For time indexing; This represents the total number of sampling times.

[0076] The specific implementation of step S05 involves extracting historical anomaly pattern samples from similar thermal energy storage system data sources, and aligning domain features using the maximum mean difference criterion. The formula for calculating the maximum mean difference is as follows: ; In the formula, The maximum mean difference is dimensionless. This represents the number of samples in the source domain. The number of samples in the target domain; and For source domain sample indexing; and Index the target domain samples; For the source domain Feature vectors of each sample; For the source domain Feature vectors of each sample; For the target domain Feature vectors of each sample; For the target domain Feature vectors of each sample; For the kernel function, a radial basis function is used. ,in and The normalized eigenvectors are dimensionless. For vectors and The Euclidean distance between them This is a kernel bandwidth parameter, typically ranging from 0.5 to 2.0; The feature scaling parameter is empirically defined as the standard deviation of the feature vector.

[0077] The objective function for optimizing the feature transformation matrix is ​​expressed as follows: ; In the formula, The feature transformation matrix has dimension 1. ; The source domain feature matrix has dimension 1. ; Let be the feature matrix of the target domain, with dimension . ; This is the regularization coefficient, typically ranging from 0.001 to 0.01; The Frobenius norm is defined as follows: ,in and For the row and column indices of the matrix elements, For matrix No. Line number Column elements; The dimension of the feature vector.

[0078] The parameter acquisition method is as follows: and Each feature vector is extracted from the historical database of the source domain and the current running data of the target domain, and each feature vector has been normalized to a dimensionless quantity. The solution is obtained by iterative optimization using the gradient descent method, with the number of iterations set to 100 to 500 and the learning rate set to 0.001 to 0.01.

[0079] The specific implementation of step S06 is to construct a mathematical simulation model of the thermal energy storage system based on the thermodynamic equation and the heat transfer equation. The energy conservation equation is expressed as follows: ; In the formula, Density of the thermal storage medium, in units of ; Specific heat capacity, unit: ; Temperature, in Kelvin (K). Time, in seconds; These are spatial coordinates, in meters (m). Thermal conductivity, in units of ; The heat source term per unit volume, in units of ; For reference density, the empirical value is 1000. ; For reference specific heat capacity, the empirical value is 4200. .

[0080] The injection formula for abnormal disturbance parameters is expressed as follows: ; ; ; ; In the formula, To account for the density of the heat storage medium after leakage, the unit is... ; The initial density is given in units of... ; The leakage rate is 0 to 10% per hour. The normalized scale thickness is dimensionless. This represents the initial scale thickness, in mm. The scaling growth rate is expressed in mm / h. For reference thickness, the default value is 1mm; The efficiency of the circulating pump is dimensionless. The initial efficiency ranges from 0.7 to 0.9 and is dimensionless. The efficiency decline rate ranges from 0% to 30%. The thermal conductivity of the insulation layer is expressed in units of 1000 ppm. ; This is the initial thermal conductivity, in units of... ; The aging factor ranges from 1.0 to 2.5.

[0081] The formula for random noise injection in data augmentation is as follows: ; In the formula, The enhanced sample after noise injection is dimensionless. This is a simulation sample, dimensionless; Let be a random variable that follows a standard normal distribution, with a mean of 0, a standard deviation of 1, and is dimensionless. The standard deviation of the simulated samples is dimensionless. To enhance the scaling parameter, an empirical value is the mean of the simulation samples, used to normalize the noise amplitude to a reasonable range; it is dimensionless.

[0082] The data augmentation formula for feature perturbation is expressed as follows: ; In the formula, The enhanced sample after feature perturbation is dimensionless; The perturbation amplitude coefficient follows a normal distribution with a mean of 0 and a standard deviation of 0.1 to 0.2. It is dimensionless and is used to generate view 1 samples in multi-view learning.

[0083] The data augmentation formula for time window shifting is expressed as follows: ; In the formula, The enhanced sample after time window shifting is dimensionless. The time shift is 1 to 5 time steps, expressed in seconds. This formula is used to generate viewpoint 2 samples in multi-view learning.

[0084] The formula for data augmentation through scaling is expressed as follows: ; In the formula, The augmented sample after scaling is dimensionless. is the scaling factor, ranging from 0.8 to 1.2, dimensionless, and this formula is used to generate viewpoint 3 samples in multi-view learning.

[0085] The specific implementation method of step S07 is the same as described above, and will not be repeated in detail here.

[0086] The specific implementation of step S08 involves constructing an abnormal pattern recognition network model. The number of attention heads in the multi-head attention mechanism layer is determined based on the dimension of the fused feature vector. Four attention heads are used when the dimension of the fused feature vector is between 1 and 64; eight attention heads are used when the dimension is between 64 and 128; and sixteen attention heads are used when the dimension is greater than or equal to 128. In the consistency regularization framework based on multi-view learning, the consistency loss uses the Jensen-Shannon divergence metric. The formula for calculating the Jensen-Shannon divergence is as follows: ; In the formula, To predict the probability distribution and The Jensen-Shannon divergence between them is dimensionless; The Kullback-Leibler divergence; and Given two probability distribution vectors, each probability distribution vector contains There are elements, among which This represents the total number of abnormal pattern categories. and Each element takes a value between 0 and 1, and the sum of all elements is 1. It is the arithmetic mean of two probability distributions.

[0087] The formula for calculating the Kullback-Leibler divergence is as follows: ; In the formula, and It is a probability distribution vector; This is a category index, with values ​​ranging from 1 to... ; and Categories The corresponding probability value ranges from 0 to 1 and is dimensionless. This represents the total number of abnormal pattern categories. To prevent small constants from having zero values ​​in logarithmic operations, the default value is 0. .

[0088] The formulas for calculating the consistency loss from the three perspectives are as follows: ; In the formula, This is a consistency loss, dimensionless; The predicted probability distribution vector is obtained by inputting the View 1 sample generated through feature perturbation into the abnormal pattern recognition network model. The predicted probability distribution vector is obtained by inputting the generated View 2 samples into the abnormal pattern recognition network model after the time window translation enhancement. The predicted probability distribution vectors are obtained after the viewpoint 3 samples are enhanced through scaling and input to the abnormal pattern recognition network model. The dimension of each probability distribution vector is . .

[0089] The formula for calculating the total loss function is as follows: ; In the formula, The total loss function is dimensionless. This is the consistency loss weighting coefficient, with a value of 0.3, and is dimensionless. The cross-entropy loss weight coefficient is 0.7, which is dimensionless. The cross-entropy loss is dimensionless.

[0090] The formula for calculating the classification cross-entropy loss is as follows: ; In the formula, For category The true label is generated using one-hot encoding when the sample belongs to the category. hour ,otherwise Dimensionless; The category predicted by the anomaly pattern recognition network model The probability value, ranging from 0 to 1, is dimensionless; To prevent small constants from having zero values ​​in logarithmic operations, the default value is 0. .

[0091] The specific implementation method of step S09 is the same as described above, and will not be repeated in detail here.

[0092] The specific implementation of step S10 is to calculate the difference between the anomaly confidence score and the preset anomaly threshold. The formula for calculating the difference is as follows: ; In the formula, The difference is dimensionless. The score represents the anomaly confidence level, with a value ranging from 0 to 1, and is dimensionless. The preset abnormal threshold is set to 0.5 to 0.9 according to safety requirements, and is dimensionless.

[0093] The specific implementation of step S11 involves querying the anomaly response strategy library based on the anomaly type identifier, obtaining the system control parameter set, and performing parameter adjustments. The formula for adjusting the operating temperature is expressed as follows: ; In the formula, The adjusted operating temperature is expressed in °C. The operating temperature before adjustment is expressed in °C. This is a temperature adjustment coefficient, ranging from 0.5 to 1.5, and is dimensionless. The temperature range is adjusted based on a reference value, typically ranging from 5 to 20°C. It is a symbolic function that returns 1 when the independent variable is greater than 0, -1 when it is less than 0, and 0 when it is equal to 0. It is dimensionless.

[0094] The formula for adjusting the circulation flow rate is expressed as follows: ; In the formula, The adjusted circulation flow rate, in units of ; The original circulating flow rate is expressed in units of [unit missing]. ; This is the flow rate adjustment factor, ranging from 0.1 to 0.3, and is dimensionless.

[0095] The formula for adjusting heat exchange power is expressed as follows: ; In the formula, The adjusted heat exchange power is expressed in kW. The heat exchange power before adjustment is expressed in kW. This is the power adjustment factor, ranging from 0.05 to 0.25, and is dimensionless.

[0096] The principles and effects of each formula are explained below: Spectrum Amplitude Formula And amplitude attenuation coefficient formula Pressure fluctuation characteristics were analyzed using a joint time-frequency domain analysis, with peak frequency... This method can identify resonant frequencies and periodic disturbance sources in a system. The amplitude attenuation coefficient, calculated by time averaging the logarithm of the ratio of continuous peak amplitudes, reflects the system's damping characteristics and energy dissipation mechanism. The logarithmic operation transforms the exponential decay of amplitude into a linear relationship, facilitating the acquisition of the attenuation coefficient through linear fitting. The combination of these two methods effectively distinguishes different anomaly modes such as sudden pressure drops, pressure oscillations, and gradual pressure increases. Compared to simple time-domain statistical analysis, which only provides limited information such as mean and variance, this method provides richer dynamic characteristics of the pressure system and offers earlier warnings for gradual anomalies such as decreased circulating pump efficiency. Frequency domain analysis identifies anomalous frequency components, while time-domain analysis quantifies the evolution rate of anomalies. The complementary nature of these two methods improves the accuracy and timeliness of anomaly identification. (Heat flux density gradient field formula) And Hessian matrix formula The spatial distribution characteristics of heat transfer were revealed by calculating the first and second derivatives of the heat flux density field, and the eigenvalue determination conditions of the Hessian matrix were also established. Identify saddle points. and Identify local minimum points. and Identifying maxima and singularities reveals the heat transfer paths and heat exchange efficiency distribution within a thermal energy storage system. Saddle points correspond to regions where heat flow direction changes, maxima correspond to heat flow convergence centers, and minima correspond to heat flow divergence centers. Compared to simple numerical statistics of heat flux density, which only reflect average levels, topological analysis methods provide deeper geometric information and spatial correlation characteristics. It has a stronger ability to discriminate abnormal patterns of heat flow concentration or dispersion caused by heat exchanger fouling and can achieve knowledge transfer for cross-system anomaly pattern recognition. (Maximum mean difference formula) And the optimization formula of the feature transformation matrix Domain adaptation is achieved by measuring the distance between the source and target domain data distributions in the reproducing kernel Hilbert space. This formula is based on the kernel embedding theorem, which states that the maximum difference between the means of two probability distributions is equal to the distance between their mean embeddings in the reproducing kernel Hilbert space. The radial basis function... Mapping eigenvectors to a high-dimensional space makes a non-linearly separable distribution linearly separable in that space; the eigenvector transformation matrix... The optimization achieves feature space alignment by minimizing the maximum mean difference between the transformed source and target domain data. The regularization term... This method prevents overfitting. Compared to directly using source domain data for transfer learning, which may lead to performance degradation due to domain shift, this method significantly reduces the negative impact of domain shift on model performance and improves the model's recognition accuracy on the target system. Especially when labeled samples are scarce in the target system, transfer learning can fully utilize the abnormal pattern recognition knowledge of the source domain, accelerating model training convergence and improving generalization ability. The advantage of this method is that it does not require specific assumptions about the data distribution of the source and target domains, making it widely applicable. Energy conservation equation. and abnormal disturbance parameter injection formula , , , The system response under different abnormal operating conditions was simulated by numerically solving the thermodynamic governing equations. The energy conservation equation is based on the first law of thermodynamics and Fourier's law of heat transfer. The left side represents the rate of change of energy per unit volume of the heat storage medium with time, and the first term on the right side represents the change of energy flux caused by heat conduction, while the second term represents the energy generated by the internal heat source. The leakage rate parameter... The mass loss process of the thermal storage medium and the scaling thickness parameters were simulated. The simulation demonstrated the process by which fouling accumulation on the heat exchanger surface leads to increased heat transfer resistance, and the pump efficiency reduction rate. The simulation investigated the process of reduced delivery capacity due to wear or failure of the circulating pump, and the aging coefficient of the insulation layer. This study simulates the process of increased heat loss due to the degradation of thermal insulation material performance. By systematically setting different combinations of abnormal disturbance parameters within a defined range and running the simulation model, simulation samples covering various abnormal modes can be generated. Compared to the difficulty and high cost of obtaining abnormal samples solely based on actual measurement data, the simulation method can obtain a large number of labeled samples without affecting the safe operation of the system. Especially for rare or safety-risk abnormal conditions such as severe leakage or complete pump failure, simulation data can effectively supplement the lack of training samples. Data augmentation formula , , and The diversity of samples was further enhanced by injecting random noise, feature perturbations, time window shifting, and scaling, thus improving the model's robustness to measurement noise, time shifts, and amplitude variations. These data augmentation methods enable the model to learn stable features that are insensitive to common data transformations, avoiding overfitting of the model to specific patterns in the simulation data. (Janssen-Shannon divergence formula) and the Kullback-Leibler divergence formula Measuring the differences between predicted probability distributions from different perspectives constrains the model's predictive consistency for different augmentation transformations of the same data. The Jensen-Shannon divergence is a symmetric divergence measure with values ​​ranging from 0 to infinity. It is 0 when two distributions are identical, and the greater the difference between the distributions, the larger the divergence value. This divergence is obtained by calculating the arithmetic mean of the Kullback-Leibler divergences of the two distributions and their respective mean distributions. The Kullback-Leibler divergence measures the distribution... Relative to distribution Information loss, when and When they are completely identical, the value is 0; consistency loss formula. The sum of three pairs of divergences among the predicted distributions of the three viewpoint samples was calculated. By minimizing this loss function, the model was constrained to produce consistent prediction results for different viewpoint samples generated by feature perturbations, time window shifts, and scale transformations. The total loss function formula is as follows: Combining consistency loss with classification cross-entropy loss A weighted combination of 0.3 and 0.7 is used. The classification cross-entropy loss is used to supervise the model's learning of accurate anomaly pattern classification, while the consistency loss is used to regularize the model's learning of stable features insensitive to data changes. This combination enables the model to accurately classify anomaly patterns and extract robust discriminative features. Compared to traditional single-supervised learning methods that easily overfit to the surface features of training data by simply minimizing the classification cross-entropy loss, the multi-view learning framework significantly improves the model's robustness to noise interference and changing operating conditions by strengthening the model's invariance to feature perturbations, time window shifts, and scale transformations. It reduces the model's over-reliance on the surface features of the training data, allowing the model to learn more fundamental anomaly pattern discriminative features. The theoretical basis of this method is the invariance learning principle, which states that true anomaly pattern discriminative features should remain unchanged under reasonable data changes. The consistency constraint forces the model to ignore surface changes introduced by data augmentation and focus on learning the intrinsic features of anomaly patterns. (System regulation parameter adjustment formula) , and By establishing a quantitative relationship between anomaly confidence scores and changes in control parameters, adaptive parameter adjustment based on the degree of anomaly is achieved. The sign function in the operating temperature adjustment formula determines the adjustment direction based on the sign of the difference between the anomaly confidence score and the threshold. The adjustment coefficient... and benchmark adjustment range The three formulas jointly determine the adjustment intensity. The circulating flow rate adjustment formula makes the flow rate change proportional to the degree of abnormal deviation. The more the abnormality confidence score exceeds the threshold, the greater the flow rate increase to strengthen the system circulation. The heat exchange power adjustment formula makes the power reduction proportional to the degree of abnormal deviation. The more the abnormality confidence score exceeds the threshold, the greater the power reduction to reduce the system load. The core of these three formulas is to transform discrete abnormality judgment results into continuous control actions. Compared with the fixed threshold on / off control strategy, which only takes a fixed-amplitude adjustment action when an abnormality is detected, which is prone to parameter mutations and system oscillations, this method can achieve a smoother parameter adjustment process and avoid the impact of frequent adjustments on system stability. , and The settings are optimized according to different anomaly types and system safety requirements, so that the control response can effectively suppress anomalies without excessively interfering with normal operation. At the same time, the closed-loop feedback mechanism continuously monitors the anomaly confidence score and dynamically adjusts parameters according to the control effect, ensuring that the control measures can effectively suppress the abnormal state and enable the system to quickly return to the normal operating range. This adaptive control strategy is more timely and accurate than manual experience-based adjustment, significantly improving the safety and reliability of thermal energy storage systems and reducing the risk of abnormal events evolving into serious failures.

[0097] To better understand and implement this invention, the following is a specific application scenario example 2: A molten salt thermal energy storage system uses a mixed molten salt of sodium nitrate and potassium nitrate as the heat storage medium, with an operating temperature range of 290–565°C, and is equipped with 12 circulating pumps and 8 sets of plate heat exchangers. In the initial stage of system operation, problems such as uneven temperature distribution and abnormal pressure fluctuations frequently occurred. Traditional monitoring methods were unable to accurately locate the root cause of the abnormalities, leading to decreased system efficiency and increased equipment wear.

[0098] The technical team first deployed 146 temperature sensors, 78 pressure sensors, and 52 heat flux density sensors at key locations in the thermal energy storage system, with a sampling frequency of 5Hz. During 72 hours of continuous system operation, the size of the temperature distribution data matrix collected was [data missing]. The size of the pressure fluctuation data matrix is The heat flux density data matrix is ​​of size The temperature distribution data was decomposed into five levels using the Daubechies 4 wavelet basis function to extract high-frequency and low-frequency components at each level. The first level of high-frequency components reflects rapid temperature changes in the 0.5–2.5 Hz range, while the fifth level of low-frequency components reflects a slower temperature trend in the 0–0.16 Hz range. The wavelet coefficients from the five levels were combined to form a 730-dimensional temperature feature vector set.

[0099] A 4096-point Fast Fourier Transform was performed on the pressure fluctuation data to obtain the spectral distribution. Analysis revealed that the peak frequencies were concentrated in three frequency bands: 0.87Hz, 1.34Hz, and 2.16Hz, corresponding to the impeller passage frequency of the circulating pump, the natural frequency of the pipe, and the vibration frequency of the heat exchanger, respectively. The amplitude attenuation coefficient of the pressure fluctuation data within a 300-second time window was calculated. By extracting the amplitude values ​​of 20 consecutive peaks and calculating the natural logarithm of their ratio, the amplitude attenuation coefficient was found to be 0.023. By combining peak frequency and amplitude attenuation coefficient with time-domain statistical characteristics, a 186-dimensional pressure feature vector set is established.

[0100] Gradient calculations were performed on the heat flux density data using a spatial grid, and the central difference method was employed to obtain the results. direction and The gradient components in the direction are determined, and the gradient magnitude and gradient direction angle are calculated. Topological analysis of the heat flux density gradient field identifies 23 saddle points, 17 maxima, and 15 minima. The spatial coordinates of these singular points are extracted, and the connections between them are recorded to form a topological skeleton, establishing a 295-dimensional heat flux feature vector set.

[0101] The temperature, pressure, and heat flow feature vector sets were aligned according to timestamps and then concatenated to form a 1211-dimensional fused feature vector. A maximum-minimum normalization method was used to process each dimension of the fused feature vector, mapping the values ​​to intervals. After normalization, a system operating state feature library containing 51,840 time-point samples was established.

[0102] The technical team extracted historical anomaly pattern samples from data sources of three similar molten salt thermal energy storage systems. These similar systems have capacities ranging from 400 to 600 cubic meters per second. Within the specified range, the operating temperature range is similar to that of the target system. A total of 8760 historical anomaly pattern samples were collected, including 1240 samples of heat storage medium leakage, 1680 samples of heat exchanger scaling, 1520 samples of decreased circulating pump efficiency, 980 samples of insulation aging, and 3340 samples of normal operation. These samples underwent domain feature alignment processing, and the maximum mean difference between the source system and the target system in the feature space was calculated to be 0.147. By mapping the source system samples to the target system feature space using a feature transformation matrix, the maximum mean difference was reduced to 0.038, thus establishing the transfer learning source dataset.

[0103] A mathematical simulation model of a thermal energy storage system based on the finite element method was constructed. The model includes fluid dynamics equations, energy conservation equations, and heat transfer equations. In the simulation model, the leakage rate of the thermal storage medium was set to a range of 0–8% per hour, and the fouling thickness of the heat exchanger was set to a range of 0–5 mm. The efficiency reduction rate of the circulating pump ranges from 0% to 35%, and the aging coefficient of the insulation layer ranges from 0% to 0.6. By combining different abnormal disturbance parameter values, 12,400 simulated abnormal samples were generated. Data augmentation processing was performed on the simulated abnormal sample set, including injecting normally distributed random noise with a standard deviation of 0.05, recombining sequences according to a 60-second time window, and randomly shuffling the feature dimensions, ultimately resulting in an expanded training sample set of 37,200 samples.

[0104] The 8760 samples from the transfer learning source dataset were merged with the 37200 samples from the extended training dataset to form a mixed training dataset of 45960 samples. The mixed training dataset was then divided into a training set of 32172 samples and a validation set of 13788 samples in a 7:3 ratio.

[0105] An anomaly pattern recognition network model is constructed. The input layer receives a 1211-dimensional fused feature vector. The first convolutional layer uses 64 convolutional kernels with a kernel size of 3 and a stride of 1 to extract local features, resulting in a 64-channel feature map. The second convolutional layer uses 128 convolutional kernels with a kernel size of 3 and a stride of 1 to extract global features, resulting in a 128-channel feature map. Since the 1211-dimensional fused feature vector falls within the range, a multi-head attention mechanism layer is set with 16 attention heads, each with a dimension of 64, to perform weighted aggregation on the 128-channel feature map. The recurrent neural network layer adopts a bidirectional long short-term memory network structure with 256 hidden units to capture temporal dependencies. The fully connected layer contains a 3-layer structure with 512, 256, and 15 neurons respectively. The output layer uses the Softmax activation function to output the classification results of 15 anomaly patterns and anomaly confidence scores.

[0106] The anomaly pattern recognition network model was pre-trained using the training set, employing a consistency regularization framework based on multi-view learning. For each fused feature vector in the training set, a feature perturbation with a standard deviation of 0.08 was applied to generate viewpoint 1 samples, a time window shift was applied to generate viewpoint 2 samples, and a scaling factor of 0.85–1.15 was applied to generate viewpoint 3 samples. The three viewpoint samples were then input into the anomaly pattern recognition network model to obtain the predicted probability distribution. The Jensen-Shannon divergence between viewpoint 1 and viewpoint 2 was calculated to be 0.124, between viewpoint 1 and viewpoint 3 to be 0.139, and between viewpoint 2 and viewpoint 3 to be 0.117. The consistency loss was obtained by summing the three Jensen-Shannon divergences, resulting in a consistency loss of 0.380. The classification cross-entropy loss was calculated to be 1.863, and the total loss function value was obtained by weighting the losses using coefficients of 0.3 and 0.7, resulting in a total loss function value of 1.418. The initial learning rate was set to 0.001, the batch size to 128, and the training duration to 50 epochs.

[0107] Fine-tuning was performed on the pre-trained model, with the learning rate reduced to 0.0002, and training continued for 30 epochs. The training process was monitored using a validation set, and convergence was defined as the classification accuracy on the validation set changing by less than 0.1% over 10 consecutive training epochs. The final trained anomaly pattern recognition network model achieved a classification accuracy of 94.7% on the validation set.

[0108] The technical team applied the trained anomaly pattern recognition network model during actual system operation. At a certain moment, the fused feature vector from the system's operational state feature library was input into the model, as shown below. Figure 2 As shown, the extracted deep anomaly features exhibit a clear clustering structure in 128-dimensional space. The model output anomaly pattern classification result is heat exchanger fouling pattern, with an anomaly confidence score of 0.827. Figure 3 As shown, the heat flux density gradient field distribution at each measuring point of the system at this moment indicates that there are multiple singular points clustered in the heat exchanger region, which verifies the accuracy of the model identification results.

[0109] The preset anomaly threshold is set to 0.65 based on system safety operation requirements. The difference between the anomaly confidence score of 0.827 and the preset anomaly threshold of 0.65 is calculated to be 0.177. Since the difference is greater than 0, the current system operating state is determined to be abnormal. The anomaly type identifier is extracted as heat exchanger scaling.

[0110] The system retrieves the corresponding set of control parameters by querying the abnormal response strategy database based on the heat exchanger fouling indicators. The control scheme includes lowering the upper limit of the operating temperature from 565℃ to 540℃, increasing the circulation flow rate from 85% to 95% of the rated flow rate, and increasing the heat exchange power from 120... Adjusted to 105 After the execution parameters were adjusted, the system ran for 30 minutes. Then, the running data was collected again and input into the model. The anomaly confidence score dropped to 0.58, which was lower than the preset anomaly threshold. The system then returned to normal operation.

[0111] The technical team inspected and verified the system, and found that there was indeed a 2.8 ohmstone inside heat exchanger No. 8. The thickness of the scale layer perfectly matched the model identification results. During the six months of continuous system operation, the anomaly pattern recognition method identified 3 heat storage medium leakage events, 17 heat exchanger scaling events, 9 circulating pump efficiency reduction events, and 5 insulation layer aging events, achieving an accuracy rate of 93.8%, effectively preventing major system failures. As shown in Table 1, the identification performance indicators of various anomaly patterns demonstrate that different types, such as abnormal temperature distribution, abnormal pressure fluctuations, and abnormal heat flux density, can be accurately identified.

[0112] Table 1 Performance Indicators for Various Anomaly Pattern Recognition

[0113] The technological advancement of this invention compared to traditional methods lies in the synergistic mechanism of multi-physics feature fusion and deep learning. Traditional methods mainly rely on threshold monitoring of a single physical quantity, failing to capture the coupled correlation features between temperature, pressure, and heat flux density, resulting in low accuracy and high false alarm rates in anomaly pattern recognition. This invention extracts multi-frequency features of temperature through multi-scale wavelet decomposition, captures the dynamic characteristics of pressure fluctuations through frequency domain analysis, and identifies spatial structural anomalies in heat flux density through topological analysis. These three types of features are then fused in the time domain to form a multi-dimensional feature vector, achieving a comprehensive representation of the system's operating state. A consistency regularization framework based on multi-view learning strengthens the model's invariant feature learning to data transformations, enabling the model to extract essential features of anomaly patterns from different perspectives rather than superficial features. The hybrid training strategy of using transfer learning source datasets and simulated anomaly sample sets utilizes historical experience from similar systems and covers a wider range of anomaly scenarios through mathematical simulation models, solving the problem of insufficient training samples for the target system. The multi-head attention mechanism layer adaptively adjusts the number of attention heads according to the dimension of the fused feature vector, allowing the model to flexibly focus on important information from different feature channels. Recurrent neural network layers capture temporal dependencies, enabling the model to identify the evolutionary trends of gradual anomaly patterns. The synergistic effect of these techniques transforms anomaly pattern recognition from passive threshold alarms to proactive pattern recognition, from single-point monitoring to system-level comprehensive analysis, and from experience-driven to data-driven approaches, significantly improving the accuracy, robustness, and generalization ability of anomaly pattern recognition in thermal energy storage systems.

[0114] It should be noted that the variables involved in this invention are explained in detail in Tables 2, 3, and 4 below.

[0115] Table 2. Variable Explanation Table (Part 1)

[0116] Table 3. Variable Explanation Table (Part Two)

[0117] Table 4. Variable Explanation Table (Part 3)

[0118] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal patterns in a thermal energy storage system, characterized in that, Includes the following steps: The system collects temperature distribution data, pressure fluctuation data, and heat flux density data. Multi-scale wavelet decomposition is performed on the temperature distribution data to establish a temperature feature vector set. Pressure fluctuation data is analyzed to establish a pressure feature vector set. Heat flux density data is processed to establish a heat flux feature vector set. The temperature, pressure, and heat flux feature vector sets are fused to establish a system operating status feature library. Historical anomaly pattern samples are extracted from similar thermal energy storage system data sources, and domain feature alignment is performed to establish a transfer learning source dataset. A mathematical simulation model of a thermal energy storage system is constructed, abnormal disturbance parameters are injected, a simulation abnormal sample set is generated, and data augmentation processing is performed to obtain an extended training sample set. The source dataset for transfer learning is mixed with the extended training sample set to create a hybrid training dataset; an anomaly pattern recognition network model is constructed and trained using the hybrid training dataset. Input the system operation status feature library into the trained anomaly pattern recognition network model, and output the anomaly pattern classification results and anomaly confidence scores; When the anomaly confidence score is greater than the preset anomaly threshold, the anomaly type identifier is extracted; the system control parameter set is obtained based on the anomaly type identifier, and the operating parameters of the thermal energy storage system are adjusted. Its core lies in using a consistency regularization framework based on multi-view learning for training.

2. The method according to claim 1, characterized in that, In the step of establishing the temperature feature vector set, high-frequency and low-frequency components of the temperature distribution data are extracted. The high-frequency components reflect the rapid change characteristics of temperature, while the low-frequency components reflect the slow change trend of temperature.

3. The method according to claim 2, characterized in that, In the step of establishing the pressure feature vector set, time-domain analysis and frequency-domain analysis are performed on the pressure fluctuation data to extract the peak frequency and amplitude attenuation coefficient of the pressure fluctuation data. The peak frequency refers to the frequency value corresponding to the maximum amplitude in the spectrum after the pressure fluctuation data is Fourier transformed, and the amplitude attenuation coefficient refers to the attenuation rate of the amplitude of the pressure fluctuation data with time in the time domain.

4. The method according to claim 3, characterized in that, In the step of establishing the heat flux feature vector set, gradient calculation is performed on the heat flux density data to obtain the heat flux density gradient field, and topological analysis is performed on the heat flux density gradient field to extract the singular point distribution characteristics of the heat flux density gradient field.

5. The method according to claim 4, characterized in that, In the step of establishing the system operation status feature library, the temperature feature vector set, pressure feature vector set, and heat flow feature vector set are fused to obtain a fused feature vector, and the fused feature vector is normalized. The feature fusion refers to aligning the temperature feature vector set, pressure feature vector set, and heat flow feature vector set according to the timestamp and then concatenating the vectors to form a fused feature vector containing multi-physics information.

6. The method according to claim 5, characterized in that, The similar thermal energy storage system data source refers to the historical operating database of other thermal energy storage systems that are similar to the current thermal energy storage system in terms of structural type, capacity level, and operating conditions.

7. The method according to claim 6, characterized in that, In the step of establishing the transfer learning source dataset, the historical abnormal pattern samples are aligned with the domain features. The domain feature alignment refers to calculating the distribution difference between the historical abnormal pattern samples and the current thermal energy storage system operation data in the feature space by similar thermal energy storage system data sources, using the maximum mean difference criterion to measure the distance between the two data distributions, and mapping the feature distribution of the historical abnormal pattern samples to a feature space similar to that of the current thermal energy storage system through feature transformation.

8. The method according to claim 7, characterized in that, In the step of generating the simulation anomaly sample set, abnormal disturbance parameters are injected into the mathematical simulation model of the thermal energy storage system. The mathematical simulation model of the thermal energy storage system refers to the mathematical description model of the thermal energy storage system established based on the thermodynamic equation and the heat transfer equation. The mathematical simulation model of the thermal energy storage system simulates the temperature field distribution, pressure field distribution and heat flow field distribution of the thermal energy storage system under different operating conditions through numerical solution methods. The abnormal disturbance parameters include the leakage rate parameter of the thermal storage medium, the scaling thickness parameter of the heat exchanger, the efficiency reduction rate parameter of the circulating pump, and the aging coefficient parameter of the insulation layer.

9. The method according to claim 8, characterized in that, The data augmentation process includes injecting random noise into the simulated abnormal sample set, reconstructing time series slices, and shuffling feature dimensions. Random noise injection refers to superimposing random noise that follows a normal distribution onto the feature values ​​of the simulated abnormal samples.

10. The method according to claim 9, characterized in that, The structure of the anomaly pattern recognition network model is as follows: the input layer receives the fused feature vector, the first convolutional layer extracts local features, the second convolutional layer extracts global features, the multi-head attention mechanism layer performs weighted aggregation of features, the recurrent neural network layer captures temporal dependencies, and the fully connected layer outputs the anomaly pattern classification result and anomaly confidence score.

Images