Temperature anomaly monitoring method and system for secondary reheating unit

Abstract

The invention discloses a temperature anomaly monitoring method and system for a secondary reheating unit, and relates to the technical field of temperature prediction. Determining the equipment points whose target correlation coefficients are greater than a preset correlation value according to the historical data set, and recording the equipment points as related equipment points; acquiring temperature data of a target equipment point and operation data of all related equipment points through a target time period to obtain first temperature data and an operation data set, performing data splitting on the first temperature data, and performing reconstruction to obtain a second sub-data set; and obtaining a predicted temperature of the target equipment point according to the operation data set and the second sub-data set, and if the predicted temperature is greater than a preset temperature threshold, giving an alarm. Through calculation of the target correlation coefficient, other parameters closely related to the target equipment point can be identified, so that a more comprehensive monitoring system is constructed, potential information between the first temperature data and the operation data set is mined through data splitting and reconstruction, the predicted temperature is finally determined, and the accuracy of temperature prediction is improved.

Description

Technical Field

[0001] This invention belongs to the field of temperature prediction technology, specifically relating to a method and system for monitoring temperature anomalies in reheat units. Background Technology

[0002] With the continued growth of global energy demand and increasingly stringent environmental requirements, the power generation industry is facing the dual challenges of improving energy efficiency and reducing pollution emissions. In thermal power systems, effectively improving thermal energy utilization, reducing fuel consumption, and minimizing pollution emissions have become key issues. Reheat cycle is a technology that effectively improves the thermal efficiency of steam turbine units. Initially, single reheat technology mainly involved returning steam after it had performed work in the high-pressure cylinder to the boiler for reheating, thus raising its temperature before it entered the intermediate or low-pressure cylinder. This avoids the rapid temperature drop during steam expansion, which would lead to increased humidity, thereby maintaining high cycle efficiency while extending the service life of the turbine blades. With technological advancements, the efficiency improvement of single reheat has gradually reached its limit. To further improve thermal energy utilization efficiency, researchers proposed double reheat technology, which involves heating the steam twice after it passes through the high-pressure and intermediate-pressure cylinders, ensuring that the steam maintains a high temperature when entering the turbine at different pressure stages, thereby achieving higher thermal efficiency.

[0003] However, current temperature prediction models often rely on physical or statistical models. Due to the complexity of double reheat systems, these models do not fully consider other influencing factors, resulting in insufficient accuracy of the prediction results. This leads to problems in various adjustments and decisions made based on these predicted temperatures in subsequent operations, affecting the stability of the system. Summary of the Invention

[0004] The purpose of this invention is to address the problem that insufficient accuracy of prediction results leads to issues in subsequent operations, such as adjustments and decisions based on these predicted temperatures, which in turn affect the stability of the system. Therefore, this invention proposes a method and system for monitoring temperature anomalies in secondary reheat units.

[0005] In a first aspect of this invention, a method for monitoring temperature anomalies in a double reheat unit is first proposed, the double reheat unit comprising multiple equipment points, the method comprising:

[0006] Historical data of all equipment points are obtained to obtain a historical dataset. Based on the historical dataset, the target correlation coefficient between the target equipment point and all equipment points in the first equipment point set is calculated. Equipment points with a target correlation coefficient greater than a preset correlation value are recorded as related equipment points. The first equipment point set is all equipment points in the secondary reheat unit except for the target equipment point.

[0007] The first temperature data and the operation data of all relevant equipment points are obtained by collecting temperature data of the target equipment point and operation data of all related equipment points within the target time period. The first temperature data is split into a first sub-data set, and the first sub-data set is reconstructed to obtain a second sub-data set.

[0008] Based on the complexity of the running dataset and the second subset dataset, the temperature prediction model is reconstructed to obtain the reconstructed temperature prediction model.

[0009] The predicted temperature of the target device is obtained based on the running dataset and the second sub-dataset. If the predicted temperature is not within the preset temperature range, an alarm is issued.

[0010] Optionally, the step of reconstructing the temperature prediction model based on the complexity of the running dataset and the second subset of datasets to obtain the reconstructed temperature prediction model includes:

[0011] The data complexity level is determined based on the characteristic correlation of the running dataset and the temporal fluctuation amplitude of the second subset, and the data complexity level includes high complexity, medium complexity and low complexity.

[0012] A hierarchical reconstruction strategy is adopted for different complexity levels: for high-complexity data, the feature extraction capability is enhanced by increasing the number of deformable convolutional layers in the residual network; for medium-complexity data, the number of attention heads and dynamic masking mechanism of the multi-head attention model are adjusted to strengthen feature association capture; for low-complexity data, the hidden layer dimension and regularization coefficient of the gated recurrent unit are optimized.

[0013] Based on the reconstructed sub-models of various complexity levels, the reconstructed temperature prediction model is generated by dynamic weight fusion, and the dynamic weights are updated in real time according to the prediction accuracy of each sub-model.

[0014] Optionally, the hierarchical reconstruction strategy further includes an adversarial perturbation verification step:

[0015] Constrained adversarial perturbations are generated for input data of various complexity levels. The adversarial perturbations are generated by the projective gradient descent algorithm and satisfy the condition that the perturbation amplitude does not exceed the device measurement accuracy threshold.

[0016] The perturbated input data is fed into the reconstructed sub-model to calculate the increase in prediction error. If the increase exceeds a preset threshold, the network depth or attention mechanism parameters of the corresponding sub-model are adjusted backtrackingly until the increase in error is lower than the threshold.

[0017] Optionally, the generation of the constrained adversarial perturbation specifically includes:

[0018] Initialize the perturbation vector to zero, calculate the gradient of the input data based on the loss function of the current sub-model, and iteratively update the perturbation vector along the gradient ascent direction;

[0019] After each iteration, the perturbation vector is projected onto a preset L∞ norm constraint space. The boundary values ​​of the L∞ norm constraint space are set according to the historical data fluctuation range of the device points, and the perturbation amplitude of a single feature does not exceed a preset proportion of its maximum value.

[0020] After a preset number of iterations, the final adversarial perturbation is obtained, which is used to verify the robustness of the reconstructed sub-model to minor disturbances.

[0021] Optionally, calculating the correlation coefficient between the target device point and all device points in the first device point set based on the historical dataset includes:

[0022] For each target device point and each device point in the first device point set, extract the target device point and the time series data recorded by the device point in the historical dataset to obtain the first target device point data and the first device point data;

[0023] If the data of the first device point is of the same data type as the data of the first target device point, then the first correlation coefficient between the data of the first target device point and the data of the first device point is calculated to obtain the target correlation coefficient between the device point and the target device point;

[0024] If the data types of the first device point data and the first target device point data are different, then the second correlation coefficient between the first target device point data and the first device point data is calculated to obtain the target correlation coefficient between the device point and the target device point.

[0025] Optionally, splitting the first temperature data to obtain a first subset includes:

[0026] The first temperature data is preprocessed to obtain the second temperature data, and preset white noise is added to the second temperature data to obtain the third temperature data.

[0027] Empirical mode decomposition is performed on the third temperature data to obtain an intrinsic mode function set. For each intrinsic mode function in the intrinsic mode function set, a clustering algorithm is used to obtain the intrinsic mode functions related to the second temperature data to obtain a similar mode function set.

[0028] For each similar intrinsic mode function in the set of similar mode functions, the correlation between the signal-to-noise ratio of the similar intrinsic mode function and the second temperature data is calculated to obtain the weight of the similar intrinsic mode function. The target sub-signal is obtained by multiplying the weight by the similar intrinsic mode function.

[0029] The reconstructed signal is obtained by superimposing all target sub-signals. The difference between the reconstructed signal and the second temperature data is calculated to obtain the target residual. If the target residual meets the preset condition, the similar mode function set is recorded as the first subset.

[0030] Optionally, reconstructing the first subset of data to obtain the second subset includes:

[0031] For the first subset of data, the complexity level of each subset is calculated using sample entropy and denoted as the target entropy value.

[0032] Sort all first sub-data according to the target entropy value to obtain the target entropy value sub-data set. Divide the target entropy value sub-data set according to the first preset entropy value and the second preset entropy value to obtain the first sub-data segment, the second sub-data segment and the third sub-data segment. The sum of the first sub-data segment, the second sub-data segment and the third sub-data segment is denoted as the second sub-data set.

[0033] Optionally, obtaining the predicted temperature of the target device point based on the running dataset and the second subset dataset includes:

[0034] For each type of data in the running dataset and the first sub-data segment, the data is converted to grayscale to obtain a one-dimensional array. The one-dimensional arrays of all types of data are concatenated to obtain the target image. The target image is then substituted into the first deep learning model to obtain the first preset temperature.

[0035] Substitute the second sub-data segment and the running dataset into the multi-head attention model to obtain the second preset temperature;

[0036] Substitute the third sub-data segment and the running dataset into the second deep learning model to obtain the third preset temperature;

[0037] The predicted temperature of the target device point is obtained by weighted fusion of the first preset temperature, the second preset temperature and the third preset temperature.

[0038] Optionally, the adversarial perturbation verification step further includes incremental model optimization of the perturbation feedback:

[0039] For sub-models whose error increase exceeds the threshold after verification by adversarial perturbation, the feature dimension with the largest prediction deviation in the perturbation sample is extracted. Based on the SHAP value weight of the feature dimension, the parameters of the layer corresponding to the low contribution feature in the model are frozen, and only the network layer associated with the high contribution feature is incrementally trained.

[0040] Incremental training employs an adversarial knowledge distillation loss function, with the original model serving as the teacher model and the sub-model to be optimized serving as the student model. By minimizing the KL divergence output by the teacher and student models and the classification loss of adversarial perturbation samples, the sub-model can enhance its resistance to perturbations of key features while retaining historical knowledge. The temperature coefficient of the KL divergence is dynamically adjusted according to the perturbation amplitude.

[0041] In a second aspect of the invention, a temperature anomaly monitoring system for a secondary reheat unit is provided, the secondary reheat unit comprising multiple equipment points, the system comprising:

[0042] The relevant equipment point confirmation module is used to obtain historical data of all equipment points to obtain a historical dataset, calculate the target correlation coefficient between the target equipment point and all equipment points in the first equipment point set based on the historical dataset, and record the equipment points whose target correlation coefficient is greater than a preset correlation value as relevant equipment points; the first equipment point set is all equipment points in the secondary reheat unit except for the target equipment point.

[0043] The data splitting module is used to collect temperature data of the target device point and operation data of all related device points through a target time period to obtain first temperature data and operation dataset, split the first temperature data to obtain a first subset dataset, and reconstruct the first subset dataset to obtain a second subset dataset.

[0044] The device point temperature prediction module is used to obtain the predicted temperature of the target device point based on the running dataset and the second sub-dataset, and to issue an alarm if the predicted temperature is not in the preset temperature range.

[0045] Optionally, the relevant equipment point confirmation module includes:

[0046] The device point data extraction module is used to extract the target device point and the time series data recorded by the device point in the historical dataset for each device point in the first device point set to obtain the first target device point data and the first device point data.

[0047] The first correlation coefficient calculation module is used to calculate the first correlation coefficient between the first target device point data and the first device point data if the data types of the first device point data and the first target device point data are the same, so as to obtain the target correlation coefficient between the device point and the target device point.

[0048] The second correlation coefficient calculation module is used to calculate the second correlation coefficient between the first target device point data and the first device point data if the data types of the first device point data and the first target device point data are different, thereby obtaining the target correlation coefficient between the device point and the target device point.

[0049] Optionally, the data splitting module includes:

[0050] The data preprocessing module is used to preprocess the first temperature data to obtain the second temperature data, and add preset white noise to the second temperature data to obtain the third temperature data.

[0051] The empirical mode decomposition module is used to perform empirical mode decomposition on the third temperature data to obtain an intrinsic mode function set. For each intrinsic mode function in the intrinsic mode function set, a clustering algorithm is used to obtain the intrinsic mode functions related to the second temperature data to obtain a similar mode function set.

[0052] The weight generation module is used to calculate the correlation between the signal-to-noise ratio of each similar intrinsic mode function in the set of similar intrinsic mode functions and the second temperature data to obtain the weight of the similar intrinsic mode function, and to obtain the target sub-signal by multiplying the weight by the similar intrinsic mode function.

[0053] The first subset data set confirmation module is used to superimpose all target sub-signals to obtain a reconstructed signal, calculate the difference between the reconstructed signal and the second temperature data to obtain the target residual, and if the target residual meets the preset conditions, the similar mode function set is recorded as the first subset data set.

[0054] Optionally, the data splitting module further includes:

[0055] The target entropy calculation module is used to calculate the complexity level of each first sub-data set based on the sample entropy, and record it as the target entropy value.

[0056] The dataset segmentation module is used to sort all first sub-data according to the target entropy value to obtain a target entropy value sub-dataset, and to segment the target entropy value sub-dataset according to a first preset entropy value and a second preset entropy value to obtain a first sub-data segment, a second sub-data segment, and a third sub-data segment; the sum of the first sub-data segment, the second sub-data segment, and the third sub-data segment is denoted as the second sub-dataset.

[0057] Optionally, the device point temperature prediction module includes:

[0058] The first preset temperature confirmation module is used to convert the grayscale value of each type of data in the running dataset and the first sub-data segment to obtain a one-dimensional array, to concatenate the one-dimensional arrays of all types of data to obtain a target image, and to substitute the target image into the first deep learning model to obtain the first preset temperature.

[0059] The second preset temperature confirmation module is used to substitute the second sub-data segment and the running dataset into the multi-head attention model to obtain the second preset temperature.

[0060] The third preset temperature confirmation module is used to substitute the third sub-data segment and the running dataset into the second deep learning model to obtain the third preset temperature.

[0061] The preset temperature fusion module is used to perform weighted fusion of the first preset temperature, the second preset temperature and the third preset temperature to obtain the predicted temperature of the target device point.

[0062] The beneficial effects of this invention are:

[0063] This invention proposes a method for monitoring temperature anomalies in double reheat units, through...

[0064] Historical data from all device points is acquired to obtain a historical dataset. Based on this dataset, the target correlation coefficient between the target device point and all device points in the first set of device points is calculated. Device points with a correlation coefficient greater than a preset value are identified as related device points. Temperature data from the target device point and operational data from all related device points are collected over a target time period to obtain first temperature data and an operational dataset. The first temperature data is split into a first subset, and this subset is then reconstructed to obtain a second subset. The predicted temperature of the target device point is determined based on the operational dataset and the second subset. If the predicted temperature exceeds a preset temperature threshold, an alarm is triggered. By calculating the target correlation coefficient, other parameters closely related to the target device point can be identified, thus building a more comprehensive monitoring system. Furthermore, by splitting and reconstructing the data, potential information between the first temperature data and the operational dataset is mined to ultimately determine the predicted temperature, improving the accuracy of temperature prediction. Attached Figure Description

[0065] The invention will now be further described with reference to the accompanying drawings.

[0066] Figure 1 A flowchart of a method for monitoring temperature anomalies in a reheat unit is provided as an embodiment of the present invention;

[0067] Figure 2 This invention provides a framework diagram for a temperature anomaly monitoring system for a secondary reheat unit. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0069] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] This invention provides a method for monitoring temperature anomalies in a double reheat unit. See also... Figure 1 , Figure 1 This is a flowchart illustrating a method for monitoring temperature anomalies in a double reheat unit, provided as an embodiment of the present invention. The double reheat unit includes multiple equipment points, and the method includes the following steps:

[0071] S101, Obtain historical data of all device points to obtain historical dataset, calculate target correlation coefficient between target device point and all device points in the first device point set based on historical dataset, and record device points with target correlation coefficient greater than preset correlation value as related device points;

[0072] S102, by collecting temperature data of the target device point and operation data of all related device points through the target time period, a first temperature data and an operation dataset are obtained. The first temperature data is split to obtain a first subset of data. The first subset of data is reconstructed to obtain a second subset of data.

[0073] S103, Based on the complexity of the running dataset and the second subset dataset, reconstruct the temperature prediction model to obtain the reconstructed temperature prediction model.

[0074] S104, based on the reconstructed temperature prediction model, the predicted temperature of the target device point is obtained, and if the predicted temperature is not in the preset temperature range, an alarm is issued.

[0075] The first set of equipment points consists of all equipment points in the secondary reheat unit except for the target equipment point.

[0076] Based on the temperature anomaly monitoring method for a reheat unit provided by the embodiments of the present invention, other parameters closely related to the target equipment point can be identified by calculating the target correlation coefficient, thereby constructing a more comprehensive monitoring system. Furthermore, by splitting and reconstructing data, potential information between the first temperature data and the operating dataset is mined, and the predicted temperature is finally determined, thus improving the accuracy of temperature prediction.

[0077] In one implementation, reconstructing the temperature prediction model based on the complexity of the running dataset and the second subset dataset to obtain the reconstructed temperature prediction model includes:

[0078] The data complexity level is determined based on the characteristic correlation of the running dataset and the temporal fluctuation amplitude of the second subset, and the data complexity level includes high complexity, medium complexity and low complexity.

[0079] A hierarchical reconstruction strategy is adopted for different complexity levels: for high-complexity data, the feature extraction capability is enhanced by increasing the number of deformable convolutional layers in the residual network; for medium-complexity data, the number of attention heads and dynamic masking mechanism of the multi-head attention model are adjusted to strengthen feature association capture; for low-complexity data, the hidden layer dimension and regularization coefficient of the gated recurrent unit are optimized.

[0080] Based on the reconstructed sub-models of various complexity levels, the reconstructed temperature prediction model is generated by dynamic weight fusion, and the dynamic weights are updated in real time according to the prediction accuracy of each sub-model.

[0081] Specifically, in this embodiment, the temperature prediction model is reconstructed based on the complexity of the running dataset and the second subset dataset to better adapt the model to data with different characteristics in the double reheat unit, thereby improving the accuracy of temperature prediction. The specific process is as follows:

[0082] First, determining the data complexity level requires considering two characteristics: the feature correlation of the operational dataset and the temporal fluctuation amplitude of the second subset. Feature correlation describes the strength of the correlation between operating parameters of different equipment points in the operational dataset, such as the correlation between the opening / closing degree of a valve and the turbine temperature. This indicator can be obtained by calculating the mutual information between features or the topological connection strength based on a graph neural network. Temporal fluctuation amplitude measures the drastic change of temperature data over time in the second subset, such as the magnitude and frequency of temperature rises and falls within a short period, which can be calculated using the standard deviation or coefficient of variation of the time series. Based on these two indicators, the data is divided into three levels: high complexity, medium complexity, and low complexity. High-complexity data typically exhibits complex feature correlations and drastic temporal fluctuations, such as multi-equipment collaborative data during unit startup. Medium-complexity data has intermediate levels of feature correlation and temporal fluctuations, such as routine data during stable unit operation. Low-complexity data has simple feature correlations and mild temporal fluctuations, such as single-equipment temperature data under stable load.

[0083] It's important to note that a layered reconstruction strategy is employed for different complexity levels because data with varying characteristics places different demands on the model. For high-complexity data, using a residual network and increasing the number of deformable convolutional layers enhances the model's ability to extract high-dimensional, non-linear features. Deformable convolutional layers can dynamically adjust the sampling position of the convolutional kernels by learning offsets, thus more accurately capturing subtle local features in the data and adapting to complex changes in operating conditions. For medium-complexity data, adjusting the number of attention heads and the dynamic masking mechanism in the multi-head attention model strengthens the model's ability to capture correlations between multiple features. Adjusting the number of attention heads adapts to the density of feature correlations, while the dynamic masking mechanism filters out irrelevant features and focuses on key correlations. For low-complexity data, optimizing the hidden layer dimension and regularization coefficient of the gated recurrent unit avoids overfitting while ensuring prediction accuracy. Adjusting the hidden layer dimension adapts to the memory requirements of time-series data, while the regularization coefficient controls the model's complexity.

[0084] When dynamically weighting and fusing reconstructed sub-models based on different complexity levels, the dynamic weights are updated according to the real-time prediction accuracy of each sub-model. For example, if a sub-model performs better in recent predictions, its weight will be increased accordingly. This fusion method can combine the advantages of each sub-model, enabling the reconstructed temperature prediction model to maintain high prediction performance under different operating conditions.

[0085] In one implementation, the hierarchical reconstruction strategy further includes an adversarial perturbation verification step:

[0086] Constrained adversarial perturbations are generated for input data of various complexity levels. The adversarial perturbations are generated by the projective gradient descent algorithm and satisfy the condition that the perturbation amplitude does not exceed the device measurement accuracy threshold.

[0087] The perturbated input data is fed into the reconstructed sub-model to calculate the increase in prediction error. If the increase exceeds a preset threshold, the network depth or attention mechanism parameters of the corresponding sub-model are adjusted backtrackingly until the increase in error is lower than the threshold.

[0088] In this embodiment, the adversarial perturbation verification step in the hierarchical reconstruction strategy aims to test the robustness of the reconstructed sub-model, that is, whether the model can maintain stable predictive performance when faced with minor disturbances. This is crucial for the safety monitoring of secondary reheat units. In actual industrial scenarios, minor disturbances such as sensor noise and data transmission errors are difficult to avoid, and the model needs to have the ability to resist such disturbances.

[0089] Specifically, when generating constrained adversarial perturbations, the projected gradient descent algorithm is used because it can efficiently find the optimal perturbation under constraints. The constraint is reflected in the fact that the perturbation amplitude does not exceed the device's measurement accuracy threshold. This threshold is determined by the device's hardware characteristics, such as the measurement error range of a temperature sensor, ensuring that the generated perturbation is a physically reasonable and acceptable small disturbance, rather than an extreme perturbation that is detached from reality.

[0090] It should be noted that feeding the perturbated input data into the reconstructed sub-model and calculating the prediction error increase is to quantify the model's sensitivity to perturbations. The error increase is the ratio of the difference between the prediction error before and after the perturbation to the original error. If the increase exceeds a preset threshold, it indicates that the model is too sensitive to minor perturbations and may misjudge due to noise or other interference in actual operation. In this case, it is necessary to backtrack and adjust the network depth or attention mechanism parameters of the corresponding sub-model—for example, increasing the network depth to enhance feature extraction capabilities, or adjusting the number of attention heads to strengthen the weights of key features—until the error increase is below the threshold, ensuring that the model can still work stably when perturbations are present.

[0091] For example, the adversarial perturbation verification step is not a one-time operation, but forms a closed loop with the hierarchical reconstruction strategy: after the model is reconstructed, the robustness is tested through perturbation verification. If it fails to meet the requirements, the model is readjusted until both accuracy and robustness requirements are met. This process can significantly improve the reliability of the model in complex industrial environments.

[0092] In one implementation, the generation of the constrained adversarial perturbation specifically includes:

[0093] Initialize the perturbation vector to zero, calculate the gradient of the input data based on the loss function of the current sub-model, and iteratively update the perturbation vector along the gradient ascent direction;

[0094] After each iteration, the perturbation vector is projected onto a preset L∞ norm constraint space. The boundary values ​​of the L∞ norm constraint space are set according to the historical data fluctuation range of the device points, and the perturbation amplitude of a single feature does not exceed a preset proportion of its maximum value.

[0095] After a preset number of iterations, the final adversarial perturbation is obtained, which is used to verify the robustness of the reconstructed sub-model to minor disturbances.

[0096] In this embodiment of the application, the process of generating constrained adversarial perturbations is to ensure the "minimum" and "effectiveness" of the perturbations through refined optimization steps—minimum means that the perturbation amplitude is within a physically reasonable range, and effective means that the perturbation can significantly expose the vulnerability of the model.

[0097] Initializing the perturbation vector to zero allows for iterative generation of the optimal perturbation from a undisturbed state. The gradient of the input data is calculated based on the loss function of the current sub-model, and the perturbation vector is iteratively updated along the gradient's ascending direction. The principle is that the loss function reflects the model's prediction error, and the gradient's ascending direction is the direction that increases the loss function—that is, the perturbation direction that maximizes the model's prediction error. By iterating along this direction, the smallest perturbation that most significantly affects the model's output can be found.

[0098] It should be noted that after each iteration, the perturbation vector is projected onto a preset L∞ norm constraint space. The L∞ norm constraint space is used to limit the perturbation amplitude in each feature dimension, ensuring that the perturbation of a single feature does not exceed a reasonable range. The boundary values ​​of this space are set according to the historical data fluctuation range of the device points. For example, if the historical temperature fluctuation range of a certain device is small, its corresponding boundary value will be tightened accordingly. The perturbation amplitude of a single feature does not exceed a preset percentage of its maximum value in order to further ensure the smallness of the perturbation and avoid the distortion of the verification results due to excessive perturbation of a single feature.

[0099] This application is not limited to this. The preset number of iterations is to strike a balance between computational efficiency and perturbation quality—too few iterations may result in inaccurate perturbations, while too many will increase computational costs. The resulting adversarial perturbation is used to verify the robustness of the reconstructed sub-model to minor perturbations. If the model can still maintain stable prediction performance after adding the perturbation, its robustness is satisfactory; otherwise, further model optimization is needed.

[0100] In one possible implementation, the adversarial perturbation verification step further includes incremental optimization of the model for perturbation feedback:

[0101] For sub-models whose error increase exceeds the threshold after verification by adversarial perturbation, the feature dimension with the largest prediction deviation in the perturbation sample is extracted. Based on the SHAP value weight of the feature dimension, the parameters of the layer corresponding to the low contribution feature in the model are frozen, and only the network layer associated with the high contribution feature is incrementally trained.

[0102] Incremental training employs an adversarial knowledge distillation loss function, with the original model serving as the teacher model and the sub-model to be optimized serving as the student model. By minimizing the KL divergence output by the teacher and student models and the classification loss of adversarial perturbation samples, the sub-model can enhance its resistance to perturbations of key features while retaining historical knowledge. The temperature coefficient of the KL divergence is dynamically adjusted according to the perturbation amplitude.

[0103] In this embodiment, the incremental optimization step of the model after adversarial perturbation verification is to make targeted improvements to the model vulnerabilities found in the verification, thereby improving the robustness and generalization ability of the model. This process forms a closed loop of "verification-optimization", enabling the model to continuously adapt to complex working conditions.

[0104] The feature dimension with the largest prediction bias in the perturbed samples is extracted because these dimensions are the model's weak points—the model is most prone to misclassification when these features are perturbed. Parameters are adjusted based on the SHAP weights of these feature dimensions, where the SHAP weights measure the contribution of a feature to the model output. These weights are calculated using the SHAP (SHapley Additive ex Planations) algorithm, which is based on Shapley values ​​in game theory and quantifies the impact of each feature on the prediction result. The parameters of the layers corresponding to low-contribution features in the model are frozen, and incremental training is performed only on the network layers associated with high-contribution features. The aim is to strengthen the model's ability to process key features while avoiding the destruction of existing model knowledge, thus reducing the waste of computational resources.

[0105] It should be noted that incremental training employs an adversarial knowledge distillation loss function, using the original model as the teacher model and the sub-model to be optimized as the student model. This is to allow the sub-model to retain the historical knowledge of the original model while learning new knowledge (the ability to resist perturbations). This loss function consists of two parts: the KL divergence of the outputs of the teacher and student models is used to measure the difference in their output distributions, ensuring that the sub-model inherits the prediction style of the teacher model; the classification loss for adversarial perturbation samples forces the sub-model to learn the correct prediction rules on perturbation samples, enhancing its resistance to perturbations.

[0106] For example, the temperature coefficient of the KL divergence is dynamically adjusted according to the perturbation amplitude. The principle is that the temperature coefficient controls the smoothness of the distribution matching. The larger the perturbation amplitude, the higher the sensitivity of the model to such perturbations. In this case, reducing the temperature coefficient allows the sub-model to focus more on the key output features of the teacher model, strengthening its resistance to perturbations. Conversely, increasing the temperature coefficient maintains the flexibility of the model. Through this optimization process, the sub-model can significantly improve its resistance to perturbations of key features while retaining its original predictive ability, ultimately improving the overall performance of the reconstructed temperature prediction model.

[0107] Specifically, the following sections provide a detailed explanation of the model reconstruction for different levels of complexity. It can be understood that this application employs different reconstructed sub-models for high-complexity, medium-complexity, and low-complexity data:

[0108] First, it should be noted that the unreconstructed prediction model in this embodiment is as follows:

[0109] First, we construct mutual information between the data to measure the strength of the association between two feature variables. The value range is [0, +∞), and the larger the value, the stronger the association.

[0110] For characteristic variables X (such as valve opening degree) and Y (such as turbine temperature), mutual information is defined as:

[0111]

[0112] Where p(x,y) is the joint probability density function of X and Y, and p(x) and p(y) are the marginal probability density functions of X and Y, respectively.

[0113] The sample data of features X and Y are discretized (e.g., binned with equal width) to obtain a discrete value sequence; the joint probability distribution p(x,y) and marginal probability distributions p(x) and p(y) are calculated; the mutual information value is calculated by substituting them into the above formula. The larger the value, the stronger the correlation between the two features. In this embodiment, the mutual information is based on the information entropy theory, which is more suitable for nonlinear correlation scenarios (such as the complex coupling relationship of equipment parameters in a double reheat unit) than the traditional correlation coefficient.

[0114] Then, by constructing a feature association graph, the overall association strength between features is quantified using the graph's topological structure. Key metrics include the Laplacian matrix spectral entropy. Specifically:

[0115] Association graph construction: using device point features as nodes, and edge weights w between nodes. ij The results are obtained by normalizing the mutual information of features i and j: (k,l represents all feature pairs).

[0116] Laplace matrix: L = DW, where D is the degree matrix (D ii =∑ j w ij W is the adjacency matrix (with elements w) ij ).

[0117] Spectral entropy: Eigenvalues ​​λ1, λ2, ..., λ3 are obtained by performing eigenvalue decomposition on L. n Spectral entropy is defined as:

[0118]

[0119] In the specific calculation, the mutual information of all feature pairs is calculated to construct the adjacency matrix W; then the Laplacian matrix L is derived and eigenvalue decomposition is performed; finally, the spectral entropy H is calculated. L The larger the value, the more complex the topological structure of the feature association (the more diverse the association).

[0120] The spectral entropy of this application embodiment characterizes the topological complexity of the graph through the eigenvalue distribution of the Laplace matrix, reflecting the overall diversity of feature associations, and is suitable for evaluating the feature association strength of multiple devices operating collaboratively in a double reheat unit.

[0121] The three reconstruction methods in this application will be described in detail below.

[0122] 1. Residual Networks (ResNet) and Deformable Convolutions

[0123] Residual network structure:

[0124] The core component is the residual block, which solves the gradient vanishing problem in deep networks through skip connections. The formula is:

[0125] y=F(x,{W i})+x;

[0126] Where x is the input F(x, {W i}), is the residual function (composed of convolutional layers, batch normalization layers, etc.), and y is the output.

[0127] Deformable convolution

[0128] By adding offset learning to the traditional convolution, the convolution kernel can adapt to the geometric deformation of features. The formula is as follows:

[0129]

[0130] Where p is the current position, and K is the kernel size (e.g., K = 9 for 3×3). k It is the original offset of the convolution kernel sampling points, Δp k It is the learned additional offset, w k These are convolution weights.

[0131] Specifically, for highly complex data (such as temperature data during the unit startup phase), a ResNet with 8 residual blocks is initialized; when the feature correlation spectral entropy H_L exceeds the threshold, the number of residual blocks is increased (e.g., increased to k), and the convolutional layer in each residual block is replaced with a deformable convolution.

[0132] Simultaneously optimize the convolution weights w through backpropagation. k and offset Δp k, enabling the model to adaptively capture local deformation features in the data. In the embodiments of this application, deformable convolution dynamically adjusts the sampling positions by learning offsets, enhancing the ability to extract non-linear deformation features (such as local patterns of sudden temperature rises and drops) in high-complexity data. The deep structure of the residual network ensures the layer-by-layer abstraction of high-dimensional features.

[0133] Multi-Head Attention Model

[0134] The multi-head attention model captures different association patterns between features through multiple parallel attention heads, and the dynamic masking mechanism is used to filter out irrelevant associations.

[0135] Core formula:

[0136] Scaled Dot-Product Attention:

[0137]

[0138] where Q (query), K (key), and V (value) are linear transformation matrices of features, and d k is the dimension of Q / K, and the scaling factor is used to avoid gradient vanishing.

[0139] Multi-Head Attention:

[0140] MultiHead(Q,K,V) = Concat(head1,...,head h )W O ;

[0141] where, h is the number of attention heads, is a learnable parameter.

[0142] Dynamic Masking: Apply a mask M (M ij = 0 indicates that features i and j are irrelevant) to the attention weight matrix. The formula is:

[0143]

[0144] Specifically, for medium-complexity data (such as multi-device collaboration data during stable operation), initialize an 8-head attention model; calculate the sparsity S (proportion of non-zero mutual information) of feature associations. If S < x1%, increase the number of heads (fine-grained capture of weak associations). If S > x2%, reduce the number of heads (coarse-grained aggregation of strong associations); generate a dynamic mask M based on the SHAP value (for example, set the mask to 0 for features with SHAP value < 0.1) to filter out irrelevant associations.

[0145] In this embodiment, multi-head attention learns different association patterns in parallel using multiple sets of parameters, dynamically adjusts the number of heads to adapt to the density of associations, and the masking mechanism reduces noise interference and improves the accuracy of capturing key associations in moderately complex data.

[0146] 3. Gated Cyclic Unit (GRU)

[0147] Gated recurrent units are used to process recurrent neural network variants of time-series data. They control the updating and forgetting of information through a gating mechanism and are suitable for low-complexity time-series data (such as temperature sequences under stable loads).

[0148] Core formula:

[0149] Update gate (controls the percentage of historical information retained):

[0150] z t =σ(W z ·[h t-1 ,x t ]);

[0151] Reset gate (controls the proportion of historical information forgotten):

[0152] r t =σ(W r ·[h t-1 ,x t ]);

[0153] Candidate hidden state:

[0154]

[0155] Final hidden state:

[0156]

[0157] Where, x t It is the input at time t, h t-1 Let σ be the hidden state at time t-1, and σ be the sigmoid activation function. It is the element-wise product, W z W r W h These are learnable parameters.

[0158] Specifically,

[0159] For low-complexity data (such as temperature series under stable load), initialize the GRU hidden layer dimension (e.g., 64);

[0160] Calculate the autocorrelation coefficient (ACF) of the time series data (to measure the correlation strength between adjacent time points). If ACF(1) > 0.7, increase the dimension of the hidden layer (e.g., 128) to enhance long memory. If ACF(1) < 0.3, decrease the dimension (e.g., 32) to avoid overfitting.

[0161] Add L2 regularization (add to the loss function) (Item), to suppress parameter overfitting.

[0162] In this embodiment, GRU dynamically adjusts the retention and forgetting of temporal information through update gates and reset gates to adapt to the smooth fluctuation characteristics of low-complexity data, while the adjustment of the hidden layer dimension balances the temporal memory capability and model complexity.

[0163] The following section provides a detailed explanation of the binding adversarial perturbation, SHAP value, adversarial knowledge distillation loss function, and weight fusion in this application.

[0164] Projected gradient descent (PGD) generates adversarial perturbations.

[0165] Definition: Generating the minimum perturbation under norm constraints to maximize the model's prediction error, the formula is:

[0166]

[0167] Where δ is the perturbation vector, ∈ is the L∞ norm constraint (controlling the maximum perturbation of a single feature), L is the model loss function (such as mean squared error), f is the prediction model, x is the input data, and y is the true label.

[0168] Specifically, assuming the initial perturbation δ0 = 0; iterative update (e.g., m rounds): a. Calculate the gradient of the loss function with respect to the input: b. Update the perturbation along the gradient ascent direction: (α is the step size); c. Project onto the constrained space: δ t+1 =clip(δ t+1 (ensure that the perturbation of each feature is within [-∈, ∈]); after the iteration, δ is obtained. * .

[0169] In this embodiment, the gradient ascent direction is the fastest direction that increases the loss function, and the optimal perturbation is approximated through iteration; the L∞ norm constraint ensures that the perturbation conforms to the sensor measurement error range, making the perturbation more reasonable in industrial scenarios.

[0170] SHAP value

[0171] In this embodiment, the contribution of each feature to the model output is quantified based on the Shapley value from game theory, using the following formula:

[0172]

[0173] Where F is the set of all features, S is a subset of F that does not contain feature i, and f(S) is the model output using only the features in subset S.

[0174] Specifically, for target feature i, enumerate all feature subsets S that do not contain i; calculate the model output difference between subset S and S∪i to measure the marginal contribution of i; and sum them by weighted Shapley value weights (factorial ratio of subset size) to obtain φ. i (Positive values ​​indicate that the feature promotes prediction, while negative values ​​indicate that it inhibits it).

[0175] In this embodiment, the SHAP value quantifies the importance of features by fairly allocating the contribution of features to the output (considering all subset combinations), providing a basis for incremental model optimization—features with high SHAP values ​​correspond to key perceptual dimensions of the model, and their anti-perturbation capabilities need to be strengthened.

[0176] Adversarial knowledge distillation

[0177] Through knowledge transfer between teacher and student models, the student model retains the capabilities of the teacher model while enhancing its robustness against disturbances. The loss function is:

[0178] L AKD =L CE (y,S(x))+αL KL (T(x),S(x))+βL Adv (D,S,T);

[0179] in:

[0180] L CE It is the classification cross-entropy loss of the student model S (with the true label y);

[0181] L KL It is the KL divergence output by the teacher-student model. T(x) is the output of the teacher model;

[0182] L Adv It is an adversarial loss (the loss that the discriminator D distinguishes between S(x) and T(x);

[0183] α and β are weighting coefficients.

[0184] Specifically, let the original model be teacher T, the model to be optimized be student S, and initialize the discriminator D;

[0185] Alternating training: a. Fix S and T, train D to distinguish between S(x+δ) and T(x+δ) (δ is the adversarial perturbation); b. Fix D and T, minimize L AKDUpdate S so that its output is close to y and T(x) and can fool D; dynamically adjust the temperature coefficient τ of KL divergence during training (e.g., decrease τ when the perturbation amplitude is large to enhance the learning of the key outputs of the teacher model).

[0186] In this embodiment, knowledge distillation transmits the teacher model's "dark knowledge" (output distribution) through KL divergence, and adversarial loss forces the student model to mimic the teacher's robustness on perturbation samples.

[0187] Dynamic weight fusion algorithm

[0188] By dynamically adjusting the weights to fuse the outputs of multiple models through real-time evaluation of sub-model performance, the formula is as follows:

[0189]

[0190] in, w is the predicted value of the k-th sub-model. k Its weights are obtained by normalizing the recent prediction accuracy of the sub-model (e.g., 1 / MAE):

[0191]

[0192] Specifically, the MAE (mean absolute error) of each sub-model on the validation set is calculated every n periods; the weights w are updated according to the formula above. k The weights of sub-models with higher accuracy are increased; the fused output As the final prediction result.

[0193] In this embodiment, dynamic weight fusion tracks the performance of sub-models in real time, adaptively allocates weights, and combines the advantages of sub-models with high, medium and low complexity to maintain the stability of overall prediction accuracy when the unit operating conditions change.

[0194] In one possible implementation, by monitoring the temperature of the target equipment and the operating data of related equipment in real time, potential anomalies can be detected in a timely manner. When the predicted temperature exceeds the preset temperature range, the system will issue an alarm to remind operators to avoid equipment failure or accidents, thereby ensuring the safe operation of the equipment. The preset temperature range is determined by technicians, and no operation will be performed if the predicted temperature is within the preset temperature range.

[0195] In one implementation, the double reheat unit includes multiple equipment points, which can be valves, generators, air inlets, cylinders, cylinder inlets, cylinder exhaust ports, reheaters, reheater inlets, reheater exhaust ports, preheater inlets, preheater exhaust ports, turbines, etc. The historical data for all equipment points can be datasets collected over N target time periods prior to the current time, where the specific number of N is determined by technical personnel. The collected datasets can include valve opening / closing degree, generator power, air inlet velocity, cylinder temperature, cylinder pressure, and cylinder inlet / outlet velocity. The historical datasets include: inlet temperature, cylinder exhaust temperature, cylinder inlet pressure, cylinder exhaust pressure, reheater temperature, reheater inlet temperature, reheater exhaust temperature, reheater inlet pressure, reheater exhaust pressure, preheater inlet temperature, preheater exhaust temperature, preheater inlet pressure, preheater exhaust pressure, turbine temperature, and turbine pressure. All historical datasets are aligned according to time series. Preset values ​​and target time periods are determined by technical personnel. The operational data for the relevant equipment points are the data collected as described above.

[0196] In one implementation, by analyzing historical data and calculating correlation coefficients, devices that are highly correlated with the target device can be identified, providing a scientific basis for fault diagnosis and decision-making. By collecting temperature and operational data in real time and combining data splitting and reconstruction techniques, the flexibility and accuracy of data processing can be improved, thereby helping managers better understand the mutual influence between devices and improve the accuracy of decision-making.

[0197] In one embodiment, calculating the correlation coefficient between the target device point and all device points in the first device point set based on the historical dataset includes:

[0198] For each device point in the target device point and the first device point set, extract the target device point and the time series data of that device point recorded in the historical dataset to obtain the first target device point data and the first device point data.

[0199] If the data of the first device point is of the same data type as the data of the first target device point, then the first correlation coefficient between the data of the first target device point and the data of the first device point is calculated to obtain the target correlation coefficient between the device point and the target device point.

[0200] If the data types of the first device point data and the first target device point data are different, then the second correlation coefficient between the first target device point data and the first device point data is calculated to obtain the target correlation coefficient between the device point and the target device point.

[0201] In one implementation, by extracting time-series data of the target device point and other device points, a precise correlation coefficient can be calculated, which can identify potential influencing factors of the target device point, thereby improving the accuracy of the analysis. By adopting different correlation coefficient calculation methods, the method can be adapted to the association analysis between different data types, improving its applicability and enabling effective association calculations even between dissimilar data.

[0202] In one implementation, the first target device point data and the first device point data are time-series aligned data. If the data types of the first device point data and the first target device point data are the same, it means that all device points corresponding to temperature-type data in the first target device point data are calculated using a first correlation coefficient, while device points corresponding to non-temperature-type data in the first target device point data are calculated using a second correlation coefficient. The first correlation coefficient can be Pearson correlation coefficient, Spearman rank correlation coefficient, Kendall correlation coefficient, distance correlation coefficient, etc.; the second correlation coefficient can be multi-column binary correlation coefficient, point binary correlation coefficient, multinomial logistic regression, etc. The target correlation coefficient between the device point and the target device point obtained through the first correlation coefficient is denoted as the first coefficient, and the target correlation coefficient between the device point and the target device point obtained through the second correlation coefficient is denoted as the second coefficient. The corresponding device points calculated using the first correlation coefficient are... A first correlation coefficient threshold is used to determine whether a device point is related to a target device point. A second correlation coefficient threshold is also used to determine whether a device point is related to a target device point. The first and second correlation coefficient thresholds can be the same or different, but the first threshold is usually higher. Data calculated using the first correlation coefficient represents two directly measurable sets of data, making these coefficients more sensitive and accurate in measuring correlation between data. Data calculated using the second correlation coefficient is used between different types of data. Since the statistical relationship between different types of data is often weaker, these coefficients generally have a more lenient response to the actual relationship, resulting in a lower second correlation coefficient threshold compared to the first. Both the first and second correlation coefficient thresholds are determined by technical personnel.

[0203] In one implementation, selecting equipment point data that is highly correlated with the target equipment point for prediction helps to make more accurate predictions. Highly correlated equipment point data is more likely to capture abnormal correlations between equipment points, providing reliable data support for fault prediction and early warning. Analyzing only equipment points with high correlation reduces the amount of computation and improves speed. The calculation of the correlation coefficient enables the system to identify the most likely associated equipment points. When the temperature prediction is abnormal, the system can quickly identify which equipment points may be affected, which helps to quickly diagnose and locate the problem.

[0204] In one embodiment, splitting the first temperature data to obtain a first subset includes:

[0205] The first temperature data is preprocessed to obtain the second temperature data, and preset white noise is added to the second temperature data to obtain the third temperature data.

[0206] Empirical mode decomposition is performed on the third temperature data to obtain an intrinsic mode function set. For each intrinsic mode function in the intrinsic mode function set, a clustering algorithm is used to obtain the intrinsic mode functions related to the second temperature data to obtain a similar mode function set.

[0207] For each similar intrinsic mode function in the set of similar mode functions, the correlation between the signal-to-noise ratio of the similar intrinsic mode function and the second temperature data is calculated to obtain the weight of the similar intrinsic mode function. The target sub-signal is obtained by multiplying the similar intrinsic mode function by the weight.

[0208] The reconstructed signal is obtained by superimposing all target sub-signals. The difference between the reconstructed signal and the second temperature data is calculated to obtain the target residual. If the target residual meets the preset conditions, the set of similar mode functions is recorded as the first subset.

[0209] In one implementation, white noise is added to the second temperature data. After empirical mode decomposition, meaningful signal components can be better separated, and noise interference in the original temperature data can be suppressed. The empirical mode decomposition process can effectively decompose intrinsic mode functions containing different frequency characteristics, which include target signals and noise components. Subsequent clustering algorithms are used to filter out intrinsic mode functions related to the original temperature data, thereby preserving key signals and reducing the impact of noise.

[0210] In one implementation, data preprocessing includes common techniques such as data cleaning, noise reduction, and data alignment. Specifically, to obtain the third temperature data by adding preset white noise to the second temperature data, the white noise sequence is added point-by-point to the second temperature data. For each intrinsic mode function (IMF), the mean, standard deviation, kurtosis, skewness, and frequency components are calculated as features. The extracted features are organized into a feature matrix, with each row corresponding to an IMF. The features of the second temperature data are organized into a vector. A clustering algorithm is used to calculate the correlation between the features of the feature matrix and the features of the second temperature data row by row. For each IMF feature vector in the feature matrix, a correlation value is obtained. The correlation value is compared to a correlation threshold. If the correlation value is greater than the threshold, the IMF is included in the similar mode function set. The correlation threshold is determined by technical personnel. The clustering algorithm can be K-means clustering, hierarchical clustering, etc.

[0211] In one implementation, the initial weights are obtained using the formula W = αA + βB, where W is the weight, α is the first constant coefficient, β is the second constant coefficient, A is the signal-to-noise ratio, B is the correlation value between the similar intrinsic mode function and the second temperature data, and the sum of α and β is 1. Typically, α is 0.5 and β is 0.5. After calculating the weights corresponding to all similar intrinsic mode functions, normalization is performed so that the sum of all weights is 1.

[0212] In one implementation, if the target residual does not meet the preset conditions, empirical mode decomposition is performed on the reconstructed signal, and then the above steps are repeated until the target residual meets the preset conditions, which are determined by technical personnel.

[0213] In one implementation, a clustering algorithm is used to obtain a set of similar mode functions related to the second temperature data. This allows for the extraction of core features of temperature changes at the device points, improving the identification of potential abnormal patterns in the system and enhancing the accuracy of temperature data interpretation. For each mode function in the set of similar mode functions, its signal-to-noise ratio is calculated and weighted, thereby strengthening the effective signal and improving the resolution and accuracy of the reconstructed signal. By superimposing all weighted target sub-signals, the resulting reconstructed signal more accurately represents the actual temperature change characteristics of the device.

[0214] In one embodiment, reconstructing the first subset of data to obtain the second subset includes:

[0215] For the first subset of data, the complexity level of each subset is calculated using sample entropy and denoted as the target entropy value.

[0216] Sort all first sub-data according to the target entropy value to obtain the target entropy value sub-data set. Divide the target entropy value sub-data set into first sub-data segment, second sub-data segment and third sub-data segment according to the first preset entropy value and the second preset entropy value. The sum of the first sub-data segment, the second sub-data segment and the third sub-data segment is denoted as the second sub-data set.

[0217] In one implementation, the complexity level is calculated using sample entropy, and the data is sorted and segmented according to the entropy value. This provides an effective tool for data analysis and lays the foundation for subsequent data feature fusion. The first preset entropy value and the second preset entropy value are determined by technical personnel, and the first preset entropy value is greater than the second preset entropy value. The first sub-data segment is the first sub-data with a target entropy value greater than or equal to the first preset entropy value. The second sub-data segment is the first sub-data with a target entropy value greater than the second preset entropy value and a target entropy value less than the first preset entropy value. The third sub-data segment is the first sub-data with a target entropy value less than or equal to the second preset entropy value.

[0218] In one embodiment, obtaining the predicted temperature of the target device point based on the running dataset and the second subset dataset includes:

[0219] For each type of data in the running dataset and the first sub-data segment, the data is converted to grayscale to obtain a one-dimensional array. The one-dimensional arrays of all types of data are concatenated to obtain the target image. The target image is then substituted into the first deep learning model to obtain the first preset temperature.

[0220] Substitute the second sub-data segment and the running dataset into the multi-head attention model to obtain the second preset temperature;

[0221] Substitute the third sub-data segment and the running dataset into the second deep learning model to obtain the third preset temperature;

[0222] The predicted temperature of the target device point is obtained by weighted fusion of the first preset temperature, the second preset temperature and the third preset temperature.

[0223] In one implementation, by combining multiple data sources and different types of models, the target device points can be analyzed from multiple perspectives, thereby improving the accuracy of prediction; grayscale conversion and image stitching can transform complex data into a processable form, enabling deep learning models to better capture the features in the data and further improve the accuracy of data prediction.

[0224] In one implementation, a first sub-data segment with a high target entropy value is combined with the running dataset. This dataset has a high level of complexity. Grayscale conversion and image stitching are performed, and the stitched image is then fed into a first deep learning model to obtain a first preset temperature. This method yields a more accurate first preset temperature. The first deep learning model can be a convolutional neural network, ResNet, etc. The second sub-data segment stores the first sub-data segment with a relatively high target entropy value. Because its complexity is low, a multi-head attention model is used to process the second sub-data segment and the running dataset to obtain a second preset temperature. This multi-head attention model can be a model using a multi-head attention mechanism, such as Transformer, XLNet, Longformer, etc. The third sub-data segment stores the first sub-data segment with a low target entropy value. Because its complexity is the lowest, a second deep learning model is used to process the third sub-data segment and the running dataset to obtain a third preset temperature. The second deep learning model can be a small model such as a feedforward neural network, GRU, TCN, etc.

[0225] Based on the same inventive concept, this invention also provides a temperature anomaly monitoring system for a double reheat unit. See also Figure 2 , Figure 2 This invention provides a framework diagram of a temperature anomaly monitoring system for a secondary reheat unit, comprising multiple equipment points. The system includes:

[0226] The relevant equipment point confirmation module is used to obtain historical data of all equipment points to obtain a historical dataset, calculate the target correlation coefficient between the target equipment point and all equipment points in the first equipment point set based on the historical dataset, and record the equipment points whose target correlation coefficient is greater than the preset correlation value as relevant equipment points; the first equipment point set is all equipment points in the secondary reheat unit except for the target equipment point.

[0227] The data splitting module is used to collect temperature data of the target equipment point and operation data of all related equipment points through the target time period to obtain the first temperature data and operation dataset. The first temperature data is split to obtain the first subset dataset, and the first subset dataset is reconstructed to obtain the second subset dataset.

[0228] The equipment point temperature prediction module is used to obtain the predicted temperature of the target equipment point based on the running dataset and the second subset dataset. If the predicted temperature is not within the preset temperature range, an alarm will be issued.

[0229] Based on the temperature anomaly monitoring system for a reheat unit provided by this invention, other parameters closely related to the target equipment point can be identified by calculating the target correlation coefficient, thereby constructing a more comprehensive monitoring system. Furthermore, by splitting and reconstructing data, potential information between the first temperature data and the operating dataset is mined, and the predicted temperature is finally determined, thus improving the accuracy of temperature prediction.

[0230] In one embodiment, the relevant device point confirmation module includes:

[0231] The equipment point data extraction module is used to extract the target equipment point and the time series data of the equipment point recorded in the historical dataset for each equipment point in the target equipment point and the first equipment point set to obtain the first target equipment point data and the first equipment point data.

[0232] The first correlation coefficient calculation module is used to calculate the first correlation coefficient between the first target device point data and the first device point data if the data types of the first device point data and the first target device point data are the same, so as to obtain the target correlation coefficient between the device point and the target device point.

[0233] The second correlation coefficient calculation module is used to calculate the second correlation coefficient between the first target device point data and the first device point data if the data types of the first device point data and the first target device point data are different, so as to obtain the target correlation coefficient between the device point and the target device point.

[0234] In one embodiment, the data splitting module includes:

[0235] The data preprocessing module is used to preprocess the first temperature data to obtain the second temperature data, and to add preset white noise to the second temperature data to obtain the third temperature data.

[0236] The Empirical Mode Decomposition (EMD) module is used to perform EMD on the third temperature data to obtain an intrinsic mode function (IMF) set. For each IMF in the IMF set, a clustering algorithm is used to obtain IMFs related to the second temperature data to obtain a similar mode function set.

[0237] The weight generation module is used to calculate the correlation between the signal-to-noise ratio of each similar intrinsic mode function in the similar mode function set and the second temperature data to obtain the weight of the similar intrinsic mode function, and then multiply the weight by the similar intrinsic mode function to obtain the target sub-signal.

[0238] The first subset confirmation module is used to superimpose all target sub-signals to obtain a reconstructed signal, calculate the difference between the reconstructed signal and the second temperature data to obtain the target residual, and if the target residual meets the preset conditions, the similar mode function set is recorded as the first subset.

[0239] In one embodiment, the data splitting module further includes:

[0240] The target entropy calculation module is used to calculate the complexity level of each first sub-data set based on the sample entropy, and denoted as the target entropy value.

[0241] The dataset segmentation module is used to sort all first sub-data according to the target entropy value to obtain the target entropy value sub-dataset, and to segment the target entropy value sub-dataset according to the first preset entropy value and the second preset entropy value to obtain the first sub-data segment, the second sub-data segment and the third sub-data segment; the sum of the first sub-data segment, the second sub-data segment and the third sub-data segment is denoted as the second sub-dataset.

[0242] In one embodiment, the device point temperature prediction module includes:

[0243] The first preset temperature confirmation module is used to convert the grayscale value of each type of data in the running dataset and the first sub-data segment to obtain a one-dimensional array, to concatenate the one-dimensional arrays of all types of data to obtain the target image, and to substitute the target image into the first deep learning model to obtain the first preset temperature.

[0244] The second preset temperature confirmation module is used to input the second sub-data segment and the running dataset into the multi-head attention model to obtain the second preset temperature.

[0245] The third preset temperature confirmation module is used to input the third sub-data segment and the running dataset into the second deep learning model to obtain the third preset temperature.

[0246] The preset temperature fusion module is used to perform weighted fusion of the first preset temperature, the second preset temperature and the third preset temperature to obtain the predicted temperature of the target device point.

[0247] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for monitoring temperature anomalies in a double reheat unit, the double reheat unit comprising multiple equipment points, characterized in that, The method includes: Historical data of all equipment points are obtained to obtain a historical dataset. Based on the historical dataset, the target correlation coefficient between the target equipment point and all equipment points in the first equipment point set is calculated. Equipment points with a target correlation coefficient greater than a preset correlation value are recorded as related equipment points. The first equipment point set is all equipment points in the secondary reheat unit except for the target equipment point. The first temperature data and the operation data of all relevant equipment points are obtained by collecting temperature data of the target equipment point and operation data of all related equipment points within the target time period. The first temperature data is split into a first sub-data set, and the first sub-data set is reconstructed to obtain a second sub-data set. Based on the complexity of the running dataset and the second subset dataset, the temperature prediction model is reconstructed to obtain the reconstructed temperature prediction model. The predicted temperature of the target device is obtained based on the reconstructed temperature prediction model. If the predicted temperature is not within the preset temperature range, an alarm is issued.

2. The method for monitoring temperature anomalies in a double reheat unit according to claim 1, characterized in that, The step of reconstructing the temperature prediction model based on the complexity of the running dataset and the second subset dataset, to obtain the reconstructed temperature prediction model, includes: The data complexity level is determined based on the characteristic correlation of the running dataset and the temporal fluctuation amplitude of the second subset, and the data complexity level includes high complexity, medium complexity and low complexity. A hierarchical reconstruction strategy is adopted for different complexity levels: for high-complexity data, the feature extraction capability is enhanced by increasing the number of deformable convolutional layers in the residual network; for medium-complexity data, the number of attention heads and dynamic masking mechanism of the multi-head attention model are adjusted to strengthen feature association capture; for low-complexity data, the hidden layer dimension and regularization coefficient of the gated recurrent unit are optimized. Based on the reconstructed sub-models of various complexity levels, the reconstructed temperature prediction model is generated by dynamic weight fusion, and the dynamic weights are updated in real time according to the prediction accuracy of each sub-model.

3. The method for monitoring temperature anomalies in a double reheat unit according to claim 2, characterized in that, The hierarchical reconstruction strategy also includes an adversarial perturbation verification step: Constrained adversarial perturbations are generated for input data of various complexity levels. The adversarial perturbations are generated by the projective gradient descent algorithm and satisfy the condition that the perturbation amplitude does not exceed the device measurement accuracy threshold. The perturbated input data is fed into the reconstructed sub-model to calculate the increase in prediction error. If the increase exceeds a preset threshold, the network depth or attention mechanism parameters of the corresponding sub-model are adjusted backtrackingly until the increase in error is lower than the threshold.

4. The method for monitoring temperature anomalies in a double reheat unit according to claim 3, characterized in that, The generation of the constraint-based adversarial perturbation specifically includes: Initialize the perturbation vector to zero, calculate the gradient of the input data based on the loss function of the current sub-model, and iteratively update the perturbation vector along the gradient ascent direction; After each iteration, the perturbation vector is projected onto a preset L∞ norm constraint space. The boundary values ​​of the L∞ norm constraint space are set according to the historical data fluctuation range of the device points, and the perturbation amplitude of a single feature does not exceed a preset proportion of its maximum value. After a preset number of iterations, the final adversarial perturbation is obtained, which is used to verify the robustness of the reconstructed sub-model to minor disturbances.

5. The method for monitoring temperature anomalies in a double reheat unit according to claim 1, characterized in that, The correlation coefficient between the target device point and all device points in the first device point set is calculated based on the historical dataset, including: For each target device point and each device point in the first device point set, extract the target device point and the time series data recorded by the device point in the historical dataset to obtain the first target device point data and the first device point data; If the data of the first device point is of the same data type as the data of the first target device point, then the first correlation coefficient between the data of the first target device point and the data of the first device point is calculated to obtain the target correlation coefficient between the device point and the target device point; If the data types of the first device point data and the first target device point data are different, then the second correlation coefficient between the first target device point data and the first device point data is calculated to obtain the target correlation coefficient between the device point and the target device point.

6. The method for monitoring temperature anomalies in a double reheat unit according to claim 1, characterized in that, The first subset of data obtained by splitting the first temperature data includes: The first temperature data is preprocessed to obtain the second temperature data, and preset white noise is added to the second temperature data to obtain the third temperature data. Empirical mode decomposition is performed on the third temperature data to obtain an intrinsic mode function set. For each intrinsic mode function in the intrinsic mode function set, a clustering algorithm is used to obtain the intrinsic mode functions related to the second temperature data to obtain a similar mode function set. For each similar intrinsic mode function in the set of similar mode functions, the correlation between the signal-to-noise ratio of the similar intrinsic mode function and the second temperature data is calculated to obtain the weight of the similar intrinsic mode function. The target sub-signal is obtained by multiplying the weight by the similar intrinsic mode function. The reconstructed signal is obtained by superimposing all target sub-signals. The difference between the reconstructed signal and the second temperature data is calculated to obtain the target residual. If the target residual meets the preset condition, the similar mode function set is recorded as the first subset.

7. The method for monitoring temperature anomalies in a double reheat unit according to claim 6, characterized in that, The second subset of data obtained by reconstructing the first subset includes: For the first subset of data, the complexity level of each subset is calculated using sample entropy and denoted as the target entropy value. Sort all first sub-data according to the target entropy value to obtain the target entropy value sub-data set. Divide the target entropy value sub-data set according to the first preset entropy value and the second preset entropy value to obtain the first sub-data segment, the second sub-data segment and the third sub-data segment. The sum of the first sub-data segment, the second sub-data segment and the third sub-data segment is denoted as the second sub-data set.

8. The method for monitoring temperature anomalies in a double reheat unit according to claim 7, characterized in that, The predicted temperature of the target device point is obtained based on the running dataset and the second subset dataset, including: For each type of data in the running dataset and the first sub-data segment, the data is converted to grayscale to obtain a one-dimensional array. The one-dimensional arrays of all types of data are concatenated to obtain the target image. The target image is then substituted into the first deep learning model to obtain the first preset temperature. Substitute the second sub-data segment and the running dataset into the multi-head attention model to obtain the second preset temperature; Substitute the third sub-data segment and the running dataset into the second deep learning model to obtain the third preset temperature; The predicted temperature of the target device point is obtained by weighted fusion of the first preset temperature, the second preset temperature and the third preset temperature.

9. The method for monitoring temperature anomalies in a double reheat unit according to claim 4, characterized in that, The adversarial perturbation verification step also includes incremental model optimization based on perturbation feedback: For sub-models whose error increase exceeds the threshold after verification by adversarial perturbation, the feature dimension with the largest prediction deviation in the perturbation sample is extracted. Based on the SHAP value weight of the feature dimension, the parameters of the layer corresponding to the low contribution feature in the model are frozen, and only the network layer associated with the high contribution feature is incrementally trained. Incremental training employs an adversarial knowledge distillation loss function, with the original model serving as the teacher model and the sub-model to be optimized serving as the student model. By minimizing the KL divergence output by the teacher and student models and the classification loss of adversarial perturbation samples, the sub-model can enhance its resistance to perturbations of key features while retaining historical knowledge. The temperature coefficient of the KL divergence is dynamically adjusted according to the perturbation amplitude.

10. A temperature anomaly monitoring system for a double reheat unit, the double reheat unit comprising multiple equipment points, characterized in that, The system includes: The relevant device point confirmation module is used to obtain historical data of all device points to obtain a historical dataset, calculate the target correlation coefficient between the target device point and all device points in the first device point set based on the historical dataset, and record the device points whose target correlation coefficient is greater than a preset correlation value as relevant device points. The first set of equipment points consists of all equipment points in the double reheat unit except for the target equipment point; The data splitting module is used to collect temperature data of the target device point and operation data of all related device points through a target time period to obtain first temperature data and operation dataset, split the first temperature data to obtain a first subset dataset, and reconstruct the first subset dataset to obtain a second subset dataset. The device point temperature prediction module is used to obtain the predicted temperature of the target device point based on the running dataset and the second sub-dataset, and to issue an alarm if the predicted temperature is not in the preset temperature range.

Images