Thermal power plant equipment health degree self-evaluation system based on deep learning

Abstract

The invention discloses a thermal power plant equipment health degree self-evaluation system based on deep learning, and the system comprises a data collection and preprocessing module which is used for collecting and preprocessing multi-source operation data of thermal power plant equipment; the feature extraction module is used for constructing a multi-dimensional equipment operation feature matrix; the deep learning modeling module is used for inputting the multi-dimensional equipment operation characteristic matrix into the improved NSRNet model for training, and generating a trained improved NSRNet model; the health degree calculation module is used for generating an equipment health degree index sequence; the health trend analysis module is used for forming an equipment health trend analysis result; the health assessment module is used for calculating a comprehensive health score value; the health early warning module is used for generating an equipment health risk report; and the visual interaction module is used for obtaining the updated improved NSRNet model, realizing intelligent, self-learning and real-time evaluation of the thermal power plant equipment health state, and improving the safety and stability of equipment operation.

Description

Technical Field

[0001] This invention relates to the field of health management of thermal power plant equipment, and in particular to a self-assessment system for the health of thermal power plant equipment based on deep learning. Background Technology

[0002] As a crucial component of my country's energy structure, thermal power plants rely heavily on the continuous and stable operation of large and complex units such as boilers, turbines, generators, cooling systems, and auxiliary equipment. With increasing unit capacity and more complex operating conditions, equipment operates under prolonged high-temperature, high-pressure, and high-load environments, leading to various equipment failures exhibiting multi-source, sudden, and cumulative characteristics. To ensure the safe and economical operation of these units, achieving intelligent monitoring and self-assessment of the operating status and health of thermal power plant equipment has become a key research direction for intelligent operation and maintenance in the power industry.

[0003] Currently, health assessment of thermal power plant equipment mainly relies on traditional condition diagnosis methods based on expert experience and health assessment methods based on statistical models. The former typically employs manual inspections, experience-based judgments, or threshold-based rule analysis to make single-point assessments of monitored parameters such as temperature, vibration, current, and pressure. These methods are limited by the subjectivity of expert experience and the complexity of equipment operating conditions, making it difficult to develop universally applicable health measurement models and accurately reflect the multidimensional state characteristics of equipment during dynamic operation. The latter, on the other hand, establishes statistical models such as linear regression, principal component analysis, and support vector machines based on historical monitoring data, achieving health status identification through mathematical feature extraction. However, this type of method has limited ability to model nonlinear features and cannot effectively capture the temporal dependencies and potential high-dimensional nonlinear features between multi-source data, resulting in insufficient accuracy and robustness of health assessment results. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based self-assessment system for the health of thermal power plant equipment. This invention fully utilizes advanced algorithmic technologies such as the improved NSRNet model, symbolic representation reasoning, industrial multi-source data fusion, and dynamic feature optimization to construct an intelligent equipment health assessment system with self-learning, self-adaptation, and interpretability. The system takes the operating data of key thermal power plant equipment as input, and through multi-dimensional feature extraction and deep network analysis, automatically establishes a nonlinear mapping relationship between equipment operating characteristics and health status, achieving quantitative assessment, trend prediction, and risk warning of equipment health.

[0005] According to an embodiment of the present invention, a deep learning-based self-assessment system for the health of thermal power plant equipment includes:

[0006] The data acquisition and preprocessing module is used to collect and preprocess multi-source operating data from thermal power plant equipment.

[0007] The feature extraction module is used to extract features from the preprocessed multi-source operating data and construct a multi-dimensional equipment operating feature matrix.

[0008] The deep learning modeling module is used to input the multi-dimensional device running feature matrix into the improved NSRNet model for training. The improved NSRNet model introduces a dynamic symbolic attention mechanism and a symbolic representation optimization mechanism to generate the trained improved NSRNet model.

[0009] The health calculation module is used to input real-time collected equipment operation data into the trained improved NSRNet model, and generate an equipment health index sequence based on the input encoding layer and feature transfer structure.

[0010] The health trend analysis module is used to calculate the rate of change of equipment health based on the equipment health index sequence, and generate equipment health trend analysis results.

[0011] The health assessment module is used to calculate a comprehensive health score based on the results of equipment health trend analysis.

[0012] The health warning module is used to generate a device health risk report when the comprehensive health score exceeds a preset threshold range;

[0013] The visualization and interaction module is used to display equipment health risk reports and health status change trends in a visual interface, and to conduct bidirectional data interaction with the distributed control system of the thermal power plant. The operating status feedback data is used to update and improve the weight parameters and learning rate of the NSRNet model, resulting in an updated and improved NSRNet model.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Collect and preprocess multi-source operational data from thermal power plant equipment;

[0016] Feature extraction is performed on the preprocessed multi-source operational data to construct a multi-dimensional equipment operation feature matrix;

[0017] The multi-dimensional device operation feature matrix is ​​input into the improved NSRNet model for training. A dynamic symbolic attention mechanism and a symbolic representation optimization mechanism are introduced to obtain the trained improved NSRNet model.

[0018] The real-time collected device operation data is input into the trained and improved NSRNet model to generate a sequence of device health indexes.

[0019] The rate of change of equipment health is calculated based on the equipment health index sequence to form the results of equipment health trend analysis.

[0020] Based on the results of equipment health trend analysis, the health status of the equipment is graded and assessed, and a comprehensive health score is calculated.

[0021] When the comprehensive health score exceeds the preset threshold range, the equipment health status monitoring stage is entered, and an equipment health risk report is generated.

[0022] The equipment health risk report is displayed through a visual interface and interacts with the distributed control system of the thermal power plant in two directions to update the data, resulting in an updated and improved NSRNet model.

[0023] Optionally, the multi-source operating data includes temperature, pressure, vibration, current, voltage, flow rate, and operating condition parameters, and the preprocessing includes noise filtering, outlier removal, missing value completion, data normalization, and time synchronization alignment.

[0024] Optionally, the construction of the multidimensional device operation feature matrix specifically includes:

[0025] The monitoring parameters of each moment of the preprocessed multi-source operation data are used to form a device operation feature vector, which is formed by arranging the parameter categories and sampling order in sequence.

[0026] The equipment operation feature vector is normalized to obtain a normalized equipment operation feature vector, which is obtained by performing a linear scaling transformation on the value of each equipment operation parameter.

[0027] The normalized equipment operation feature vector is subjected to feature dimensionality reduction and feature fusion processing. The principal component analysis method is used to extract the equipment operation features to form an equipment operation feature matrix. The extraction process performs eigenvalue decomposition on the covariance matrix of the normalized equipment operation feature vector and extracts the principal component vector corresponding to the largest eigenvalue as the main equipment operation features. The principal component score vector is calculated based on the linear combination relationship between each principal component vector and the normalized equipment operation feature vector, and the principal component score vector is summarized in chronological order.

[0028] The equipment operation feature matrix is ​​time-series encoded, and the equipment operation feature matrices at different time points are arranged in chronological order to form the equipment operation feature sequence.

[0029] The device operation feature sequence is mapped to a unified deep feature space using deep mapping processing to generate a deep device operation feature representation and construct a multi-dimensional device operation feature matrix. The construction process involves arranging the deep device operation feature representation according to time sequence and feature dimension, with rows representing the time sequence and columns representing the feature dimension, to form a multi-dimensional device operation feature matrix.

[0030] Optionally, obtaining the improved NSRNet model after training specifically includes:

[0031] The improved NSRNet model is trained by inputting a multi-dimensional device operation feature matrix into it. The improved NSRNet model includes an input encoding module, a deep symbolic feature learning module, a symbolic structure adaptive optimization module, and a symbolic expression generation and device health calculation module. The input encoding module performs non-linear encoding transformation on the device operation feature vector to form a deep device operation feature sequence. The deep symbolic feature learning module generates a candidate set of device health calculation functions by introducing a dynamic symbolic attention mechanism. The symbolic structure adaptive optimization module obtains the optimal mapping parameter set through a symbolic expression optimization mechanism. The symbolic expression generation and device health calculation module performs structural reconstruction and parameter loading to generate the trained improved NSRNet model.

[0032] In the input encoding module, the device operation feature vectors of the multi-dimensional device operation feature matrix are embedded and mapped to form a deep device operation feature sequence. The deep device operation feature sequence is formed by linearly combining the device operation feature vectors of each dimension with the corresponding weight parameters and applying a non-linear activation function, and outputting the mapping results in chronological order.

[0033] In the deep symbolic feature learning module, feature extraction is performed on the deep device operation feature sequence, a dynamic symbolic attention mechanism is introduced, and a candidate set of device health calculation functions is constructed. The construction process involves assigning dynamic weights based on the degree of correlation between symbol nodes and the importance of time series features, and performing weighted fusion on the deep device operation feature sequence to form a candidate set of device health.

[0034] In the symbol structure adaptive optimization module, based on the device health candidate set, a symbol expression optimization mechanism is introduced to jointly optimize the mapping parameters, symbol expression parameters, and symbol structure of the deep device operation feature sequence to obtain the objective function value.

[0035] In the joint optimization process, iterative training that minimizes the objective function value is used, and the weights of the mapping parameters, symbolic representation parameters, and symbolic structure are gradually adjusted through the adaptive update method of gradient descent. When the objective function converges to the preset accuracy threshold, the optimal set of mapping parameters is obtained.

[0036] In the symbol representation generation and device health calculation module, the parameter weights of each symbol node in the symbol structure and the optimal mapping parameter set are combined and hierarchically according to the network topology order. The improved NSRNet model is then restructured and loaded with parameters to generate the trained improved NSRNet model.

[0037] Optionally, the generation of the device health index sequence specifically includes:

[0038] Real-time operating data of thermal power plant equipment is collected, and noise filtering, data smoothing and normalization preprocessing are performed on the real-time operating data to obtain real-time operating characteristic data.

[0039] The real-time running feature data is input into the trained improved NSRNet model to calculate the device health index. The device health index is obtained by performing linear transformation and nonlinear activation processing on the real-time running feature data through the input encoding layer to obtain deep features. In the feature propagation structure, the deep features are propagated layer by layer through matrix operations of inter-layer weight parameters, and a normalized activation function is applied after weighted summation of each feature.

[0040] The device health indexes at each sampling time are combined in chronological order to form a device health index sequence.

[0041] Optionally, the formation of the device health trend analysis results specifically includes:

[0042] The equipment health index sequence is subjected to time-series difference calculation to obtain the equipment health change rate sequence. The equipment health change rate sequence is obtained by subtracting the equipment health index of the previous time point from the equipment health index of the current time point and dividing by the equipment health index of the previous time point.

[0043] The equipment health rate change series is smoothed by using a moving average method to weight the equipment health rate change series of adjacent time periods, resulting in a smoothed equipment health rate change series.

[0044] Trend analysis is performed on the smoothed equipment health rate change sequence to form equipment health trend analysis results. The equipment health trend analysis results are formed by calculating the directionality and amplitude of the smoothed equipment health rate change sequence in the time dimension, extracting the characteristics of rising, falling and stable changes, and forming them in chronological order.

[0045] Optionally, the calculation of the comprehensive health score specifically includes:

[0046] Based on the results of equipment health trend analysis, an equipment health status assessment sample set is constructed. The construction process involves extracting the equipment health change rate, health index and key feature parameters for each time period from the equipment health trend analysis results, and then aggregating and labeling them according to feature correlation to form the equipment health status assessment sample set.

[0047] The characteristic parameters of each parameter in the equipment health status assessment sample set are normalized to form a standardized feature matrix;

[0048] Based on the standardized feature matrix, the health status of each device sample is classified. The health status classification is divided according to the preset health level threshold. When the feature value in the standardized feature matrix is ​​greater than the health level threshold, it is determined to be a healthy level. When the feature value in the standardized feature matrix is ​​less than the health level threshold, it is determined to be an abnormal level. The classification result is obtained, and the device health level matrix is ​​formed.

[0049] The health score value corresponding to each health level is calculated based on the equipment health level matrix. The health score value is obtained by weighted summation of the weights of each feature parameter of the equipment health level matrix and the feature score.

[0050] The health scores of each health level are comprehensively calculated to obtain a comprehensive health score, which is a weighted average of all health scores and their corresponding weights.

[0051] Optionally, the generation of the device health risk report specifically includes:

[0052] A pre-set warning trigger threshold for the comprehensive health score is established. When the comprehensive health score exceeds the warning trigger threshold range, the device enters the health status monitoring phase.

[0053] During the equipment health status monitoring phase, the comprehensive health score value at each time series moment is calculated and analyzed to generate an equipment health status identification sequence. The generation of the equipment health status identification sequence is as follows: when the comprehensive health score value is less than the comprehensive health score value safety threshold, the current equipment operating status is marked as equipment health risk status; when the comprehensive health score value is between the comprehensive health score value safety threshold and the comprehensive health score value warning trigger threshold, the current equipment operating status is marked as equipment health concern status; when the comprehensive health score value is greater than the comprehensive health score value warning trigger threshold, the current equipment operating status is marked as equipment health normal status.

[0054] Based on the equipment health status identifier sequence, time segments in the equipment health risk state and equipment health concern state are filtered to obtain the change characteristics of the equipment health change rate and the comprehensive health score value in the corresponding time segments, and the equipment health risk level is calculated. The equipment health risk level includes the change range of the equipment health index, the fluctuation degree of the equipment health change rate, and the deviation degree of the comprehensive health score value.

[0055] An equipment health risk report is generated based on the equipment health risk level. The equipment health risk report includes the equipment health risk level, the trend of equipment health risk changes, key operating parameters, and equipment health early warning information.

[0056] Optionally, obtaining the updated and improved NSRNet model specifically includes:

[0057] Input the equipment health risk report into the equipment health visualization module, and graphically process the changing trends of equipment health risk level, equipment health index sequence and comprehensive health score to generate equipment health visualization data;

[0058] Based on the equipment health visualization data, an equipment health visualization interface is generated. The equipment health visualization interface includes a time series curve of equipment health, a distribution map of equipment health risk levels, and a trend map of comprehensive health score. The equipment health risk level is updated in real time through a dynamic refresh mechanism.

[0059] The data in the equipment health visualization interface is exchanged bidirectionally with the distributed control system of the thermal power plant, and the key operating parameters, equipment health risk level and equipment health early warning information in the equipment health risk report are synchronized to the distributed control system of the thermal power plant to form operating status feedback data.

[0060] Based on the operational status feedback data, the weight parameters and learning rate of the improved NSRNet model are adaptively updated to obtain the updated improved NSRNet model. The update process involves the system calculating the error between the prediction result of the improved NSRNet model and the actual equipment health based on the operational status feedback, adjusting the weight parameters of the improved NSRNet model through backpropagation, and dynamically correcting the learning rate according to the convergence rate of the error.

[0061] The beneficial effects of this invention are:

[0062] This invention proposes an intelligent health self-assessment system for thermal power plant equipment by combining deep learning and symbolic representation optimization techniques. It overcomes the limitations of traditional health assessments that rely on human experience, linear models, or fixed threshold judgments, realizing a transformation from static diagnosis to dynamic intelligent prediction in equipment condition assessment. The system uses an improved NSRNet model as its core, introducing a dynamic symbolic attention mechanism and a symbolic representation optimization mechanism into the model structure. This allows the model to dynamically allocate weights based on the relevance of equipment operating characteristics and capture nonlinear relationships, thereby significantly improving the accuracy and stability of equipment health assessment. Simultaneously, the symbolic representation optimization mechanism enhances the model's interpretability, enabling it not only to output results but also to reveal the inherent logical relationships of health changes, overcoming the black-box nature of traditional deep learning models.

[0063] In terms of data processing and feature construction, this invention establishes a unified multi-dimensional equipment operation feature matrix, fusing multi-source data such as temperature, pressure, current, and vibration under a unified dimension, and extracting key features using principal component analysis and deep mapping methods to ensure the representativeness and validity of the input data. The system obtains the equipment health index and rate of change through time series modeling, realizing a dynamic representation of the trend of equipment health status changes. This enables early identification of signs of health degradation, providing a basis for preventative equipment maintenance.

[0064] This invention further proposes a comprehensive health scoring mechanism that integrates health index, rate of change, and trend characteristics to achieve quantitative assessment and classification of equipment health status. When the comprehensive health score exceeds a set threshold, the system automatically triggers a health early warning module to generate a health risk report, thus shifting from passive detection to proactive early warning. Through a health visualization module, the system displays the health index, scoring trend, and risk level in real time and engages in bidirectional data interaction with the distributed control system of the thermal power plant. This allows the model parameters to be adaptively updated based on equipment operation feedback, forming a continuously optimizing self-learning mechanism.

[0065] Through the above methods, this invention achieves high-precision assessment and intelligent management of the health status of thermal power plant equipment. The system possesses advantages such as strong data fusion capabilities, high modeling accuracy, interpretable assessment results, fast response speed, and strong adaptability. It can operate stably for extended periods under complex operating conditions, effectively reducing equipment failure rates and the risk of unplanned downtime. This significantly improves the safety, reliability, and economy of thermal power plant equipment operation, providing efficient and reliable technical support for intelligent operation and maintenance of thermal power plants. Attached Figure Description

[0066] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0067] Figure 1 This is an overall flowchart of a deep learning-based self-assessment system for the health of thermal power plant equipment proposed in this invention.

[0068] Figure 2 This is a schematic diagram of the module structure of an improved NSRNet model for a deep learning-based self-assessment system for the health of thermal power plant equipment proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figure 1-2A deep learning-based self-assessment system for the health of thermal power plant equipment includes:

[0071] The data acquisition and preprocessing module is used to collect and preprocess multi-source operating data from thermal power plant equipment.

[0072] The feature extraction module is used to extract features from the preprocessed multi-source operating data and construct a multi-dimensional equipment operating feature matrix.

[0073] The deep learning modeling module is used to input the multi-dimensional device running feature matrix into the improved NSRNet model for training. The improved NSRNet model introduces a dynamic symbolic attention mechanism and a symbolic representation optimization mechanism to generate the trained improved NSRNet model.

[0074] The health calculation module is used to input real-time collected equipment operation data into the trained improved NSRNet model, and generate an equipment health index sequence based on the input encoding layer and feature transfer structure.

[0075] The health trend analysis module is used to calculate the rate of change of equipment health based on the equipment health index sequence, and generate equipment health trend analysis results.

[0076] The health assessment module is used to calculate a comprehensive health score based on the results of equipment health trend analysis.

[0077] The health warning module is used to generate a device health risk report when the comprehensive health score exceeds a preset threshold range;

[0078] The visualization and interaction module is used to display equipment health risk reports and health status change trends in a visual interface, and to conduct bidirectional data interaction with the distributed control system of the thermal power plant. The operating status feedback data is used to update and improve the weight parameters and learning rate of the NSRNet model, resulting in an updated and improved NSRNet model.

[0079] In this embodiment, the modules are interconnected using the following method:

[0080] Collect and preprocess multi-source operational data from thermal power plant equipment;

[0081] Feature extraction is performed on the preprocessed multi-source operational data to construct a multi-dimensional equipment operation feature matrix;

[0082] The multi-dimensional device operation feature matrix is ​​input into the improved NSRNet model for training. A dynamic symbolic attention mechanism and a symbolic representation optimization mechanism are introduced to obtain the trained improved NSRNet model.

[0083] The real-time collected device operation data is input into the trained improved NSRNet model to generate the current device health index sequence and calculate the device health change rate;

[0084] The health status of equipment is graded and assessed based on the equipment health index sequence and the equipment health change rate, and a comprehensive health score is calculated.

[0085] When the comprehensive health score exceeds the preset threshold range, the equipment health status monitoring stage is entered, and an equipment health risk report is generated.

[0086] The equipment health risk report is displayed through a visual interface and interacts with the distributed control system of the thermal power plant in two directions to update the data, resulting in an updated and improved NSRNet model.

[0087] In this embodiment, the multi-source operating data includes temperature, pressure, vibration, current, voltage, flow rate, and operating condition parameters. The preprocessing includes noise filtering, outlier removal, missing value completion, data normalization, and time synchronization alignment.

[0088] In this embodiment, the construction of the multi-dimensional device operation feature matrix specifically includes:

[0089] The monitoring parameters of each moment of the preprocessed multi-source operation data are used to form a device operation feature vector, which is formed by arranging the parameter categories and sampling order in sequence.

[0090] The equipment operation feature vector is normalized to obtain a normalized equipment operation feature vector, which is obtained by performing a linear scaling transformation on the value of each equipment operation parameter.

[0091] The normalized equipment operation feature vector is subjected to feature dimensionality reduction and feature fusion processing. The principal component analysis method is used to extract the equipment operation features to form an equipment operation feature matrix. The extraction process performs eigenvalue decomposition on the covariance matrix of the normalized equipment operation feature vector and extracts the principal component vector corresponding to the largest eigenvalue as the main equipment operation features. The principal component score vector is calculated based on the linear combination relationship between each principal component vector and the normalized equipment operation feature vector, and the principal component score vector is summarized in chronological order.

[0092] The equipment operation feature matrix is ​​time-series encoded, and the equipment operation feature matrices at different time points are arranged in chronological order to form the equipment operation feature sequence.

[0093] The device operation feature sequence is mapped to a unified deep feature space using deep mapping processing to generate a deep device operation feature representation and construct a multi-dimensional device operation feature matrix. The construction process involves arranging the deep device operation feature representation according to time sequence and feature dimension, with rows representing the time sequence and columns representing the feature dimension, to form a multi-dimensional device operation feature matrix.

[0094] In this embodiment, obtaining the improved NSRNet model after training specifically includes:

[0095] The improved NSRNet model is trained by inputting a multi-dimensional device operation feature matrix into it. The improved NSRNet model includes an input encoding module, a deep symbolic feature learning module, a symbolic structure adaptive optimization module, and a symbolic expression generation and device health calculation module. The input encoding module performs non-linear encoding transformation on the device operation feature vector to form a deep device operation feature sequence. The deep symbolic feature learning module generates a candidate set of device health calculation functions by introducing a dynamic symbolic attention mechanism. The symbolic structure adaptive optimization module obtains the optimal mapping parameter set through a symbolic expression optimization mechanism. The symbolic expression generation and device health calculation module performs structural reconstruction and parameter loading to generate the trained improved NSRNet model.

[0096] In the input encoding module, the device operation feature vectors of the multi-dimensional device operation feature matrix are embedded and mapped to form a deep device operation feature sequence. The deep device operation feature sequence is formed by linearly combining the device operation feature vectors of each dimension with the corresponding weight parameters and applying a non-linear activation function, and outputting the mapping results in chronological order.

[0097] In the deep symbolic feature learning module, feature extraction is performed on the deep device operation feature sequence, a dynamic symbolic attention mechanism is introduced, and a candidate set of device health calculation functions is constructed. The construction process involves assigning dynamic weights based on the degree of correlation between symbol nodes and the importance of time series features, and performing weighted fusion on the deep device operation feature sequence to form a candidate set of device health.

[0098] In the symbolic structure adaptive optimization module, based on the device health candidate set, a symbolic representation optimization mechanism is introduced to jointly optimize the mapping parameters, symbolic representation parameters, and symbolic structure of the deep device operation feature sequence, obtaining the objective function value:

[0099] ;

[0100] in, The objective function value, For time series moments The equipment health index monitoring value, For the device's operational feature vector, The device health index is used as a candidate set for device health. For depth mapping parameters, To represent parameters using symbols, It is a symbolic structure. and The regularization coefficient is . For symbolic structure complexity constraint functions, To minimize the computation, The total number of equipment operating samples. Let L be the L2 norm of the vector;

[0101] In the joint optimization process, iterative training that minimizes the objective function value is used, and the weights of the mapping parameters, symbolic representation parameters, and symbolic structure are gradually adjusted through the adaptive update method of gradient descent. When the objective function converges to the preset accuracy threshold, the optimal set of mapping parameters is obtained.

[0102] In the symbol representation generation and device health calculation module, the parameter weights of each symbol node in the symbol structure and the optimal mapping parameter set are combined and hierarchically according to the network topology order. The improved NSRNet model is then restructured and loaded with parameters to generate the trained improved NSRNet model.

[0103] In this embodiment, the generation of the device health index sequence specifically includes:

[0104] Real-time operating data of thermal power plant equipment is collected, and noise filtering, data smoothing and normalization preprocessing are performed on the real-time operating data to obtain real-time operating characteristic data.

[0105] The real-time running feature data is input into the trained improved NSRNet model to calculate the device health index. The device health index is obtained by performing linear transformation and nonlinear activation processing on the real-time running feature data through the input encoding layer to obtain deep features. In the feature propagation structure, the deep features are propagated layer by layer through matrix operations of inter-layer weight parameters, and a normalized activation function is applied after weighted summation of each feature.

[0106] The device health indexes at each sampling time are combined in chronological order to form a device health index sequence.

[0107] In this embodiment, the formation of the equipment health trend analysis results specifically includes:

[0108] The equipment health index sequence is subjected to time-series difference calculation to obtain the equipment health change rate sequence. The equipment health change rate sequence is obtained by subtracting the equipment health index of the previous time point from the equipment health index of the current time point and dividing by the equipment health index of the previous time point.

[0109] The equipment health rate change series is smoothed by using a moving average method to weight the equipment health rate change series of adjacent time periods, resulting in a smoothed equipment health rate change series.

[0110] Trend analysis is performed on the smoothed equipment health rate change sequence to form equipment health trend analysis results. The equipment health trend analysis results are formed by calculating the directionality and amplitude of the smoothed equipment health rate change sequence in the time dimension, extracting the characteristics of rising, falling and stable changes, and forming them in chronological order.

[0111] In this embodiment, the calculation of the comprehensive health score specifically includes:

[0112] Based on the results of equipment health trend analysis, an equipment health status assessment sample set is constructed. The construction process involves extracting the equipment health change rate, health index and key feature parameters for each time period from the equipment health trend analysis results, and then aggregating and labeling them according to feature correlation to form the equipment health status assessment sample set.

[0113] The characteristic parameters of each parameter in the equipment health status assessment sample set are normalized to form a standardized feature matrix;

[0114] Based on the standardized feature matrix, the health status of each device sample is classified. The health status classification is divided according to the preset health level threshold. When the feature value in the standardized feature matrix is ​​greater than the health level threshold, it is determined to be a healthy level. When the feature value in the standardized feature matrix is ​​less than the health level threshold, it is determined to be an abnormal level. The classification result is obtained, and the device health level matrix is ​​formed.

[0115] The health score value corresponding to each health level is calculated based on the equipment health level matrix. The health score value is obtained by weighted summation of the weights of each feature parameter of the equipment health level matrix and the feature score.

[0116] The health scores of each health level are comprehensively calculated to obtain a comprehensive health score, which is a weighted average of all health scores and their corresponding weights.

[0117] In this embodiment, the generation of the device health risk report specifically includes:

[0118] A pre-set warning trigger threshold for the comprehensive health score is established. When the comprehensive health score exceeds the warning trigger threshold range, the device enters the health status monitoring phase.

[0119] During the equipment health status monitoring phase, the comprehensive health score value at each time series moment is calculated and analyzed to generate an equipment health status identification sequence. The generation of the equipment health status identification sequence is as follows: when the comprehensive health score value is less than the comprehensive health score value safety threshold, the current equipment operating status is marked as equipment health risk status; when the comprehensive health score value is between the comprehensive health score value safety threshold and the comprehensive health score value warning trigger threshold, the current equipment operating status is marked as equipment health concern status; when the comprehensive health score value is greater than the comprehensive health score value warning trigger threshold, the current equipment operating status is marked as equipment health normal status.

[0120] Based on the equipment health status identifier sequence, time segments in the equipment health risk state and equipment health concern state are filtered to obtain the change characteristics of the equipment health change rate and the comprehensive health score value in the corresponding time segments, and the equipment health risk level is calculated. The equipment health risk level includes the change range of the equipment health index, the fluctuation degree of the equipment health change rate, and the deviation degree of the comprehensive health score value.

[0121] An equipment health risk report is generated based on the equipment health risk level. The equipment health risk report includes the equipment health risk level, the trend of equipment health risk changes, key operating parameters, and equipment health early warning information.

[0122] In this embodiment, obtaining the updated and improved NSRNet model specifically includes:

[0123] Input the equipment health risk report into the equipment health visualization module, and graphically process the changing trends of equipment health risk level, equipment health index sequence and comprehensive health score to generate equipment health visualization data;

[0124] Based on the equipment health visualization data, an equipment health visualization interface is generated. The equipment health visualization interface includes a time series curve of equipment health, a distribution map of equipment health risk levels, and a trend map of comprehensive health score. The equipment health risk level is updated in real time through a dynamic refresh mechanism.

[0125] The data in the equipment health visualization interface is exchanged bidirectionally with the distributed control system of the thermal power plant, and the key operating parameters, equipment health risk level and equipment health early warning information in the equipment health risk report are synchronized to the distributed control system of the thermal power plant to form operating status feedback data.

[0126] Based on the operational status feedback data, the weight parameters and learning rate of the improved NSRNet model are adaptively updated to obtain the updated improved NSRNet model. The update process involves the system calculating the error between the prediction result of the improved NSRNet model and the actual equipment health based on the operational status feedback, adjusting the weight parameters of the improved NSRNet model through backpropagation, and dynamically correcting the learning rate according to the convergence rate of the error.

[0127] Example 1:

[0128] This embodiment uses a 600MW supercritical coal-fired power unit as the research object to illustrate the practical application process and technical effects of the deep learning-based self-assessment system for the health of thermal power plant equipment of the present invention. The system is deployed in the central control room of the power plant and acquires multi-source operating data of key equipment such as boilers, turbines, and generators through a data acquisition and monitoring system and an equipment management system. The system collects multi-dimensional signals in real time, including main steam temperature, turbine vibration amplitude, generator current, voltage, main feedwater pressure, cooling water flow rate, and flue gas oxygen content, with a sampling period of 1 minute, and operates continuously for 30 days.

[0129] During the data acquisition phase, the system first performs outlier removal, missing value interpolation, and normalization on the acquired raw operational data to ensure data integrity and consistency. Subsequently, the feature extraction module transforms multi-source operational signals into a unified-dimensional equipment operational feature matrix. After principal component analysis and deep embedding mapping, the system automatically selects the features most relevant to the equipment's health status, achieving a data dimensionality compression rate of 74.5% and a feature information fidelity exceeding 95%.

[0130] During the model training phase, the multi-dimensional device running feature matrix is ​​input into the improved NSRNet model. The model adaptively adjusts the weight relationships between symbol nodes through a dynamic symbolic attention mechanism and jointly optimizes the symbol structure by combining a symbolic representation optimization mechanism, thus completing the functional mapping between features and the health index. After 500 rounds of training, the mean squared error of the health index prediction on the validation set is reduced to 0.0117, which is about 31.2% lower than that of the traditional LSTM model, and the prediction stability and convergence speed are significantly improved.

[0131] During system operation, real-time collected equipment operating characteristics are input into the trained improved NSRNet model, which then generates an equipment health index and its rate of change. The system calculates a comprehensive health score based on the health index's trend and classifies it into three states: healthy, concerning, and risky. When the comprehensive health score drops below a preset threshold, the system automatically triggers the equipment health early warning module and generates a risk report, providing decision-making reference for operators.

[0132] Table 1 Key Operational Monitoring Data of the Self-Assessment System for Equipment Health in Thermal Power Plants

[0133] Main steam temperature (°C) Steam turbine vibration (mm / s) Generator current (A) Main water supply pressure (MPa) Health Index Health status change rate (%) Overall health score health status System warning 542.5 1.26 985 17.3 0.962 -0.4 0.951 normal no 542.8 1.58 988 17.2 0.928 -1.1 0.907 normal no 542.4 1.93 991 17.0 0.884 -2.3 0.862 focus on no 541.9 2.27 990 16.8 0.836 -3.4 0.791 risk yes 541.6 2.64 988 16.6 0.802 -4.0 0.752 risk yes

[0134] As shown in Table 1, under healthy conditions, the main steam temperature remained stable at around 542℃, the turbine vibration amplitude was low, and the health index remained above 0.95, with no system warnings triggered. As operating time increased, the turbine vibration amplitude gradually rose, the main feedwater pressure slightly decreased, and the health index showed a slow downward trend, with the overall health score dropping from 0.95 to 0.86. The system automatically marked the equipment's operating status as "health concern." When the vibration amplitude further increased to above 2.3 mm / s and the overall health score dropped below 0.80, the system determined that the equipment had entered a health risk state and triggered the health warning module to generate a risk report.

[0135] On-site inspection and verification revealed an abnormally high temperature and low lubricating oil pressure in the third-stage bearing of the turbine's low-pressure cylinder, indicating potential early mechanical wear. The system identified the risk trend approximately 16 hours in advance, about 35% earlier than traditional vibration threshold-based alarm methods. This result demonstrates that the method of this invention can achieve highly sensitive early warning in the early stages of equipment health degradation, significantly improving the operational safety and maintenance efficiency of thermal power plant equipment.

[0136] Compared with traditional models, the improved NSRNet model of this invention reduces the overall health score prediction error by 28.9% and achieves a health status classification accuracy of 96.2%, while the traditional CNN model only achieves 89.5%. Under various operating conditions, the health trend output by the model remains highly consistent with the actual operating conditions of the equipment, indicating that the system has good dynamic adaptability and stability.

[0137] In summary, this embodiment fully verifies the technical feasibility and engineering applicability of the system of the present invention. It can realize real-time monitoring, intelligent assessment and early warning of the health status of thermal power plant equipment in complex operating environments, and provide reliable data support and technical guarantee for intelligent operation and maintenance of power plants.

Claims

1. A deep learning-based thermal power plant equipment health self-evaluation system, characterized in that, The application relates to a power plant equipment health monitoring method based on improved NSRNet model. The method comprises the following steps: a data acquisition and preprocessing module is used for acquiring and preprocessing multi-source operation data of power plant equipment; a feature extraction module is used for extracting features of the preprocessed multi-source operation data and constructing a multi-dimensional equipment operation feature matrix; a deep learning modeling module is used for inputting the multi-dimensional equipment operation feature matrix into an improved NSRNet model for training, the improved NSRNet model introduces a dynamic symbolic attention mechanism and a symbolic expression optimization mechanism, and a trained improved NSRNet model is generated; a health degree calculation module is used for inputting real-time collected equipment operation data into the trained improved NSRNet model, and generating an equipment health index sequence according to an input coding layer and a feature transmission structure; a health trend analysis module is used for calculating an equipment health change rate according to the equipment health index sequence, and forming an equipment health trend analysis result; a health assessment module is used for calculating a comprehensive health score value according to the equipment health trend analysis result; a health warning module is used for generating an equipment health risk report when the comprehensive health score value is greater than a preset threshold range; 2. The deep learning-based thermal power plant equipment health self-assessment system according to claim 1, characterized in that, a visual interactive module is used for displaying the equipment health risk report and the health degree change trend in a visual interface, and performing bidirectional data interaction with a power plant distributed control system, and feeding back operation state data to update weight parameters and learning rates of the improved NSRNet model, so that an updated improved NSRNet model is obtained. The modules are realized through the following methods: multi-source operation data of power plant equipment are acquired and preprocessed; features of the preprocessed multi-source operation data are extracted, and a multi-dimensional equipment operation feature matrix is constructed; the multi-dimensional equipment operation feature matrix is input into an improved NSRNet model for training, a dynamic symbolic attention mechanism and a symbolic expression optimization mechanism are introduced, and a trained improved NSRNet model is obtained; real-time collected equipment operation data are input into the trained improved NSRNet model, and an equipment health index sequence is generated; an equipment health change rate is calculated based on the equipment health index sequence, and an equipment health trend analysis result is formed; a comprehensive health score value is calculated according to a hierarchical evaluation of the equipment health trend analysis result; when the comprehensive health score value is greater than a preset threshold range, an equipment health state monitoring stage is entered, and an equipment health risk report is generated; 3. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, the equipment health risk report is displayed through a visual interface, and bidirectional data interaction and updating are performed with a power plant distributed control system, and an updated improved NSRNet model is obtained.

4. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The multi-source operation data comprise temperature, pressure, vibration, current, voltage, flow and operation condition parameters, and the preprocessing comprises noise filtering, abnormal value elimination, missing value completion, data normalization and time synchronization alignment. The construction of the multi-dimensional equipment operation feature matrix specifically comprises the following steps: each time point monitoring parameter of the preprocessed multi-source operation data is used to form an equipment operation feature vector, and the equipment operation feature vector is sequentially arranged according to parameter categories and sampling sequences. The device operation feature vector is normalized to obtain a normalized device operation feature vector, which is obtained by performing linear scaling conversion on the value of each device operation parameter; The normalized device operation feature vector is subjected to feature dimension reduction and feature fusion processing, and the principal component analysis method is used to extract the operation features of the device to form a device operation feature matrix. The process of extraction performs eigenvalue decomposition on the covariance matrix of the normalized device operation feature vector, extracts the principal component vector corresponding to the maximum eigenvalue as the main device operation feature, calculates the principal component score vector according to the linear combination relationship between each principal component vector and the normalized device operation feature vector, and collects the principal component score vector in time sequence; The device operation feature matrix is subjected to time series coding, and the device operation feature matrices at different time nodes are arranged in time sequence to form a device operation feature sequence; The device operation feature sequence is mapped to a unified deep feature space by using deep mapping processing to generate a deep device operation feature representation and construct a multi-dimensional device operation feature matrix. The process of construction arranges the deep device operation feature representation in time sequence and feature dimension to form a multi-dimensional device operation feature matrix, with rows representing time series and columns representing feature dimensions.

5. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The trained improved NSRNet model specifically includes: The multi-dimensional device operation feature matrix is input into the improved NSRNet model for training. The improved NSRNet model includes an input coding module, a deep symbolic feature learning module, a symbolic structure adaptive optimization module, and a symbolic expression generation and device health degree calculation module. The input coding module refers to nonlinear coding transformation of the device operation feature vector to form a deep device operation feature sequence. The deep symbolic feature learning module refers to the generation of a device health degree calculation function candidate set by introducing a dynamic symbolic attention mechanism. The symbolic structure adaptive optimization module refers to obtaining an optimal mapping parameter set through a symbolic expression optimization mechanism. The symbolic expression generation and device health degree calculation module refers to structure reconstruction and parameter loading to generate a trained improved NSRNet model; In the input coding module, the device operation feature vectors of the multi-dimensional device operation feature matrix are subjected to embedding mapping processing to form a deep device operation feature sequence. The deep device operation feature sequence is formed by linearly combining the device operation feature vectors of each dimension with the corresponding weight parameters and applying a nonlinear activation function to output the mapping results in time sequence; In the deep symbolic feature learning module, the deep device operation feature sequence is subjected to feature extraction, and a dynamic symbolic attention mechanism is introduced to construct a device health degree calculation function candidate set. The construction process assigns dynamic weights to the deep device operation feature sequence according to the correlation degree between the symbolic nodes and the importance of the time sequence features, and performs weighted fusion to form a device health degree candidate set. In the symbol structure adaptive optimization module, based on the device health degree candidate set, a symbol expression optimization mechanism is introduced, the mapping parameters, the symbol expression parameters and the symbol structure of the deep device running feature sequence are jointly optimized, and a target function value is obtained; In the joint optimization process, the mapping parameters, the symbol expression parameters and the symbol structure are gradually adjusted by using the iterative training of the minimum target function value and the adaptive update method of gradient descent, and when the target function converges to a preset precision threshold, an optimal mapping parameter set is obtained; In the symbol expression generation and device health degree calculation module, the parameter weights of the optimal mapping parameter set and each symbol node in the symbol structure are combined, and hierarchical combination is performed according to the network topology order, the structure reconstruction and parameter loading of the improved NSRNet model are performed, and the trained improved NSRNet model is generated.

6. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The generation of the device health degree index sequence specifically includes: Collecting real-time running data of the power plant equipment, and performing noise filtering, data smoothing and normalization preprocessing on the real-time running data to obtain real-time running feature data; The real-time running feature data is input into the trained improved NSRNet model, and the device health degree index is calculated, the device health degree index is obtained by inputting the encoding layer to perform linear transformation and nonlinear activation processing on the real-time running feature data to obtain deep features, and the deep features are propagated layer by layer through matrix operation of interlayer weight parameters in the feature transmission structure, and then normalized activation function is formed after weighted summation of each feature; The device health degree indexes of each sampling time are combined in time sequence to form a device health degree index sequence.

7. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The formation of the device health trend analysis result specifically includes: The device health degree index sequence is subjected to time series difference calculation to obtain a device health degree change rate sequence, the device health degree change rate sequence is obtained by subtracting the device health degree index of the previous time point from the device health degree index of the current time point and dividing by the device health degree index of the previous time point; The device health degree change rate sequence is smoothed, and the device health degree change rates of adjacent time periods are weighted and averaged by using the moving average method to obtain a smoothed device health degree change rate sequence; Based on the smoothed device health degree change rate sequence, trend analysis is performed to form a device health trend analysis result, the device health trend analysis result is obtained by calculating the direction and amplitude of the smoothed device health degree change rate sequence in the time dimension, extracting the rising, falling and stable change characteristics and forming them in time sequence.

8. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The calculation of the comprehensive health score value specifically includes: Based on the device health trend analysis result, a device health state evaluation sample set is constructed, the construction process is to extract the device health degree change rate, the health degree index and the key feature parameters of each time period in the device health trend analysis result, and to aggregate and label the device health state evaluation sample set according to the feature correlation; Each feature parameter in the device health state evaluation sample set is normalized to form a standardized feature matrix; The health state of each equipment sample is graded based on the standardized feature matrix, the health state grading is divided according to a preset health level threshold, when the feature value in the standardized feature matrix is greater than the health level threshold, it is determined as a health level, when the feature value in the standardized feature matrix is less than the health level threshold, it is determined as an abnormal level, a grading result is obtained, and an equipment health level matrix is formed; The health score value corresponding to each health level is calculated according to the equipment health level matrix, the health score value is obtained by weighted summation of the weight of each feature parameter of the equipment health level matrix and the feature score; The health score values of each health level are comprehensively calculated to obtain a comprehensive health score value, which is a weighted average result of all health score values and corresponding weights.

9. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The generation of the equipment health risk report specifically includes: A warning trigger threshold of the comprehensive health score value is preset, when the comprehensive health score value is greater than the warning trigger threshold of the comprehensive health score value, the equipment health state monitoring stage is entered; In the equipment health state monitoring stage, the comprehensive health score value of each time sequence moment is calculated and analyzed to generate an equipment health state identification sequence, when the comprehensive health score value is less than a safety threshold of the comprehensive health score value, the current equipment running state is marked as an equipment health risk state, when the comprehensive health score value is between the safety threshold of the comprehensive health score value and the warning trigger threshold of the comprehensive health score value, the current equipment running state is marked as an equipment health attention state, and when the comprehensive health score value is greater than the warning trigger threshold of the comprehensive health score value, the current equipment running state is marked as an equipment health normal state; Based on the equipment health state identification sequence, time segments in the equipment health risk state and the equipment health attention state are screened to obtain the equipment health degree change rate and the change characteristics of the comprehensive health score value corresponding to the time segments, and the equipment health risk level is calculated, the equipment health risk level includes the equipment health degree index change amplitude, the equipment health degree change rate fluctuation degree and the comprehensive health score value deviation degree; The equipment health risk report is generated according to the equipment health risk level, the equipment health risk report includes the equipment health risk level, the equipment health risk change trend, the key operation parameter and the equipment health warning information.

10. The deep learning-based thermal power plant equipment health self-assessment system according to claim 2, characterized in that, The updated improved NSRNet model is obtained specifically as follows: The equipment health risk report is input into the equipment health visualization module, the equipment health risk level, the equipment health degree index sequence and the change trend of the comprehensive health score value are graphically processed to generate equipment health visualization data; The equipment health visualization interface is generated according to the equipment health visualization data, the equipment health visualization interface includes an equipment health degree time sequence curve, an equipment health risk level distribution graph and a comprehensive health score value trend graph, and the equipment health risk level is updated in real time through a dynamic refreshing mechanism; The data in the device health visualization interface is bidirectionally interacted with the distributed control system of the thermal power plant, and key operation parameters, device health risk levels and device health warning information in the device health risk report are synchronized to the distributed control system of the thermal power plant, so as to form operation state feedback data. According to the operation state feedback data, the weight parameters and the learning rate of the improved NSRNet model are adaptively updated, and an updated improved NSRNet model is obtained. The updating process is that the system calculates the error between the prediction result of the improved NSRNet model and the actual device health degree according to the operation state feedback, adjusts the weight parameters of the improved NSRNet model through the back propagation method, and dynamically corrects the learning rate according to the convergence rate of the error.

Images