Anti-interference type coal quality multi-parameter on-line monitoring system for power plant

Abstract

The application discloses an anti-interference type thermal power coal quality multi-parameter online monitoring system and relates to the technical field of thermal power generation. The application solves the technical problem that the data processing method of the prior art often adopts simple filtering or mean processing, the anti-interference capability is weak, the feature extraction is not pure, and the accuracy of the monitoring result is low. The application obtains a normalized fusion input vector by processing multi-source heterogeneous original data, and performs anti-interference feature extraction on a coal quality deep feature vector. The application generates a monitoring result corresponding to a plurality of parameters based on the coal quality deep feature vector, dynamically identifies and suppresses an interference component introduced by a change in an operation state in original signals by using working condition data, thereby refining a pure and robust coal quality deep feature vector, replacing a passive processing mode such as fixed threshold filtering or sliding average, improving the purity of feature extraction and the robustness to a complex field environment, and further improving the accuracy and stability of the monitoring result.

Description

Technical Field

[0001] This application belongs to the field of thermal power generation technology, specifically an anti-interference type online monitoring system for multiple parameters of coal quality entering the furnace in thermal power plants. Background Technology

[0002] In modern thermal power generation systems, the real-time accuracy of coal quality input directly determines combustion efficiency and the economic efficiency and safety of unit operation. Traditional coal quality analysis relies heavily on offline sampling and testing, which has significant time lag and cannot meet the needs of real-time adjustment of combustion air distribution. Anti-interference online monitoring technology for multiple parameters of coal input in thermal power plants aims to overcome the impact of complex industrial environments on measurement accuracy. Through advanced sensing and signal processing algorithms, it achieves continuous, real-time, and synchronous monitoring of key parameters of the coal input. This is the core foundation for achieving deep peak shaving, energy conservation and emission reduction, and intelligent control of thermal power units, and has significant strategic importance for improving the economic benefits of power plants and reducing pollutant emissions.

[0003] In the harsh environment of thermal power plants, noise and outliers can severely contaminate sensor signals. Existing data processing methods often employ simple filtering or averaging, which have weak anti-interference capabilities, resulting in impure feature extraction and low accuracy of monitoring results. Therefore, further improvements are needed for online monitoring systems of multiple parameters of coal quality entering the furnace in thermal power plants. Summary of the Invention

[0004] This application aims to solve at least one of the technical problems existing in the prior art; to this end, this application proposes an anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants, which is used to solve the technical problem that the data processing methods of the prior art often use simple filtering or averaging, which has weak anti-interference ability, resulting in impure feature extraction and low accuracy of monitoring results.

[0005] To achieve the above objectives, the first aspect of this application provides an anti-interference online monitoring system for multiple parameters of coal quality fed into a thermal power plant, comprising: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together; The data acquisition module acquires multi-source heterogeneous raw data through data acquisition equipment; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data. The data analysis module includes a feature extraction unit and a result generation unit; The feature extraction unit: processes multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized and fused input vectors; and performs anti-interference feature extraction on the regularized and fused input vectors to extract coal quality depth feature vectors. The result generation unit obtains monitoring results corresponding to several parameters by performing parameter analysis with physical information constraints on the coal quality depth feature vector.

[0006] This application, through the above steps, utilizes operating condition data to dynamically identify and suppress interference components introduced by changes in operating status, thereby extracting a pure and robust coal quality depth feature vector. It breaks through the limitations of passive processing methods such as fixed threshold filtering or moving average, and instead implements parameter analysis based on the coal quality depth feature vector with physical constraints to output monitoring results. This achieves active and adaptive anti-interference from the signal level to the feature level, effectively ensuring the accuracy and stability of online monitoring results while improving the purity of feature extraction and robustness to complex field environments.

[0007] Furthermore, the step of processing the multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized fusion input vectors includes: Extract multi-source heterogeneous raw data; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data; Calculate the delay time; the delay time refers to the time it takes for coal to travel from the coal feeder to the monitoring point; the delay time is expressed as the ratio between the transmission distance of coal from the coal feeder to the monitoring point and the belt speed; The aforementioned delay time is applied to the auxiliary data of the operating conditions, and a synchronous continuous data stream is obtained by combining it with the original coal quality signal data; Data slicing and normalization are performed on a synchronous continuous data stream to obtain data sample units and their corresponding normalized fusion input vectors. The data slicing and normalization process refers to cutting the synchronous continuous data stream according to a fixed time window to obtain several data sample units, and performing feature operations on the data within the data sample units to normalize them into normalized fusion input vectors of fixed dimensions. The normalized fusion input vectors include original coal quality signal features and auxiliary features of operating conditions.

[0008] This application first calculates the delay time, clearly defining the transmission delay of coal flow from the feeder to the monitoring point based on the ratio of transmission distance to belt speed. This delay is then applied to the auxiliary data of the operating conditions to achieve strict synchronization with the original coal quality signal data on the time axis, ensuring that both represent the same coal flow segment. The synchronized continuous data stream is sliced ​​according to a fixed time window to generate discrete data sample units. Finally, feature operations are performed on the data within the unit, regularizing the original data of different dimensions and types into a regularized fusion input vector of fixed dimensions, which explicitly includes two types of features: coal quality signal and auxiliary operating conditions. This fundamentally solves the problem that multi-source data cannot be directly and effectively fused due to transmission lag and different formats. This lays a precise data alignment foundation and a unified processing format for the subsequent feature extraction unit to perform accurate interference correlation and filtering, thus overcoming the defects of simple filtering methods caused by asynchronous and irregular data, resulting in feature mixing and inaccurate extraction.

[0009] Furthermore, the step of extracting coal quality depth feature vectors from the regularized fused input vector to resist interference includes: Extract the normalized and fused input vector corresponding to several data sample units; The regularized fusion input vector is input into the anti-interference attention feature extraction network to obtain coal quality depth feature vectors corresponding to several data sample units; The anti-interference attention feature extraction network consists of an encoding module, a working condition interference attention module, a feature purification module, and a fusion dimensionality reduction module, which is used to extract a pure coal quality depth feature vector from the mixed regular fusion input vector. The encoding module consists of two parallel sub-encoders; the sub-encoders include a coal quality signal sub-encoder and an operating condition sub-encoder; The coal quality signal sub-encoder is used to capture the local patterns, spectral correlations and sequence dependencies of the original coal quality signal features in the regularized fusion input vector, and outputs the preliminary coal quality feature tensor. The operating condition sub-encoder is used to encode the operating condition auxiliary features in the regularized fusion input vector into an operating condition interference feature vector; the operating condition interference feature vector represents all external interference states that affect coal quality measurement.

[0010] Furthermore, the operating condition interference attention module is used to generate an anti-interference mask from the operating condition interference feature vector through an attention mechanism; The input data for the working condition disturbance attention module are the preliminary coal quality feature tensor and the working condition disturbance feature vector. The process of generating the anti-interference mask includes: Using the working condition disturbance feature vector as the query and several feature channels of the preliminary coal quality feature tensor as key-value pairs, the attention weight between the working condition disturbance feature vector and several feature channels of the preliminary coal quality feature tensor is calculated through an attention mechanism; the attention weight represents the severity of disturbance to each component of the preliminary coal quality feature tensor under the current working condition. A softmax normalization operation is performed on several attention weights to obtain several normalized weights; An anti-interference mask is obtained by mapping the normalized weights using a mapping function.

[0011] Furthermore, the feature purification module is used to perform feature optimization on the preliminary coal quality feature tensor to obtain a purified preliminary coal quality feature tensor. The input data for the feature purification module are the preliminary coal quality feature tensor and the anti-interference mask; The purified preliminary coal quality characteristic tensor is obtained by performing a masking operation on the preliminary coal quality characteristic tensor and the anti-interference mask; the masking operation satisfies: Where M represents the anti-interference mask, Represented as the preliminary characteristic tensor of coal quality, This is represented as element-wise multiplication; This is represented as the preliminary characteristic tensor of the purified coal.

[0012] Furthermore, the fusion and dimensionality reduction module is used to perform feature fusion and dimensionality reduction on the preliminary feature tensor of the purified coal to obtain the coal depth feature vector; The fusion dimensionality reduction module is constructed from several fully connected layers; Both feature fusion and feature dimensionality reduction are achieved through fully connected layers; The feature fusion refers to nonlinearly combining and fusing information from several features in the preliminary feature tensor of purified coal to obtain a high-dimensional feature tensor. Feature dimensionality reduction refers to reducing the feature dimension of a high-dimensional feature tensor to obtain a coal quality depth feature vector.

[0013] This application extracts a preliminary coal quality feature tensor reflecting intrinsic coal quality information and a working condition interference feature vector representing external interference states from a regularized fusion input vector using a parallel sub-encoder. Then, an anti-interference mask is dynamically generated through an attention mechanism. This mechanism uses the working condition feature as a query, calculates its correlation weight with each coal quality feature channel, and transforms it into a mask through normalization and mapping functions, thereby quantifying the degree of interference to each feature. This mask is used to perform element-wise weighted purification of the coal quality features, significantly suppressing severely interfered feature components. A fully connected layer is used to perform nonlinear fusion and dimensionality reduction on the purified features, resulting in a highly concentrated and pure coal quality depth feature vector. This achieves dynamic and adaptive interference filtering based on the attention mechanism, accurately separating intrinsic coal quality information from working condition interference noise at the feature level. This overcomes the inherent defects of traditional fixed filtering methods that cannot adapt to complex and variable working conditions, fundamentally improving the purity and robustness of feature representation, and providing a reliable guarantee for final high-precision monitoring.

[0014] Furthermore, the monitoring results corresponding to several parameters obtained through parameter analysis with physical information constraints on the coal quality depth feature vector include: Extracting coal depth feature vectors; The coal quality depth feature vector is input into the coal quality multi-parameter regression model to obtain the monitoring results corresponding to several parameters; The coal quality multi-parameter regression model has physical information constraints; The coal quality multi-parameter regression model consists of a shared coding layer, a task-specific decoding layer, and a physical constraint correction layer. The shared coding layer is used to extract abstract features from the coal quality depth feature vector, and is used to predict several parameters in the task-specific decoding layer.

[0015] Furthermore, the task-specific decoding layer is used to map the abstract features output by the shared coding layer to predicted values ​​corresponding to several parameters; The task-specific decoding layer consists of several task branches, each corresponding to a parameter; the task branches are in parallel structure. The task branch consists of a fully connected layer and uses the GELU activation function.

[0016] Furthermore, the physical constraint correction layer is used to introduce physical information constraints to correct the predicted values ​​corresponding to several parameters output by the task-specific decoding layer, thereby obtaining the monitoring results corresponding to several parameters. The physical constraint correction layer includes industrial analysis normalization constraints and empirical relational constraints for calorific value industrial analysis. The industrial analysis normalization constraint refers to performing a softmax normalization operation on the predicted values ​​of industrial parameters to obtain the monitoring results of industrial parameters. The empirical calorific value is obtained by substituting the monitoring results of several industrial parameters into the calorific value calculation formula; The monitoring results corresponding to the calorific value are obtained by weighted fusion of the predicted calorific value and the empirical calorific value.

[0017] Furthermore, the loss function of the coal quality multi-parameter regression model satisfies: ;in, Expressed as total loss, This is represented as data loss. Represented as physical constraint loss; and Represented as hyperparameters, used to balance the contributions of data loss and physical constraint loss; The data loss is expressed as a weighted mean square error loss, which is the weighted mean square error between the predicted and actual values ​​of several parameters. The physical constraint loss is expressed as the weighted sum of the mass balance constraint loss and the calorific value empirical formula constraint loss; The mass balance constraint loss satisfies: ; i represents the sample number, and N represents the total number of samples; This is represented as the mass balance constraint loss; , , and These are expressed as percentage values ​​of the predicted values ​​for moisture, ash, volatile matter, and fixed carbon in the industrial parameters, respectively. The constraint loss of the empirical formula for calorific value satisfies: ;in, Let represent the predicted calorific value corresponding to the i-th sample. This is represented as the empirical heat value corresponding to the i-th sample; This is expressed as the constraint loss in the empirical formula for calorific value.

[0018] Compared with the prior art, the beneficial effects of this application are: 1. This application obtains data sample units and their corresponding regularized and fused input vectors by processing multi-source heterogeneous raw data; extracts coal quality depth feature vectors from the regularized and fused input vectors to improve anti-interference features; obtains monitoring results corresponding to several parameters by performing parameter analysis on the coal quality depth feature vectors with physical information constraints; dynamically identifies and suppresses interference components introduced by changes in operating status in the original signal using operating condition data, thereby refining pure and robust coal quality depth feature vectors, replacing passive processing methods such as fixed threshold filtering or moving average; performs parameter analysis based on the coal quality depth feature vectors with physical constraints, and outputs monitoring results, realizing active and adaptive anti-interference from the signal level to the feature level, improving the purity of feature extraction and robustness to complex field environments, and thus improving the accuracy and stability of online monitoring results.

[0019] 2. This application extracts the preliminary coal quality feature tensor reflecting the intrinsic information of coal quality and the working condition interference feature vector representing the external interference state from the regularized fusion input vector through a parallel sub-encoder. Then, an anti-interference mask is dynamically generated through an attention mechanism. This mechanism uses the working condition feature as the query, calculates its correlation weight with each feature channel of coal quality, and transforms it into a mask through normalization and mapping functions, thereby quantifying the degree of interference of each feature. The coal quality features are then purified element by element using this mask, significantly suppressing the feature components that are severely interfered with. The purified features are then nonlinearly fused and dimensionality reduced through a fully connected layer to obtain a highly concentrated and pure coal quality depth feature vector. This achieves dynamic and adaptive interference filtering based on the attention mechanism, which can accurately separate the intrinsic information of coal quality from the working condition interference noise at the feature level. This overcomes the inherent defects of traditional fixed filtering methods that cannot adapt to complex and variable working conditions, fundamentally improving the purity and robustness of feature expression, and providing a reliable guarantee for the final high-precision monitoring.

[0020] 3. This application extracts general high-level features from input features through a shared coding layer; then, a task-specific decoding layer composed of multiple parallel task branches maps the high-level features to the initial predicted values ​​of specific parameters such as calorific value, moisture, ash content, and volatile matter; a physical constraint correction layer introduces two types of core domain knowledge to forcibly correct the initial predictions to obtain monitoring results; finally, the model is trained through a composite loss function that combines data fitting and physical constraints. This function not only minimizes the error between the predicted and true values, but also forces the model to comply with the empirical relationship between mass conservation and calorific value industrial analysis through a penalty term. The core physical and chemical laws in the field of coal quality analysis are deeply embedded into the machine learning model in a computable form, ensuring that the model's output results are not only statistically consistent with the data, but also physically consistent and reasonable. This fundamentally solves the problem that purely data-driven models may produce absurd predictions that violate basic common sense, greatly improving the reliability and interpretability of monitoring results. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Fig. 1 This is a schematic diagram of the anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to this application; Fig. 2 This is a flowchart of the anti-interference online monitoring method for multiple parameters of coal quality in thermal power plants according to this application. Detailed Implementation

[0023] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] Please see Figs. 1-2 The first aspect of this application provides an anti-interference online monitoring system for multiple parameters of coal quality fed into a thermal power plant, comprising: a data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected together. Data acquisition module: Acquires multi-source heterogeneous raw data through data acquisition equipment; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data; The data analysis module includes a feature extraction unit and a result generation unit; Feature extraction unit: Processes multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized and fused input vectors; performs anti-interference feature extraction on the regularized and fused input vectors to extract coal quality depth feature vectors; Result generation unit: By performing parametric analysis with physical information constraints on the coal quality depth feature vector, the monitoring results corresponding to several parameters are obtained.

[0025] In this embodiment, data processing of multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized and fused input vectors includes: Extract multi-source heterogeneous raw data; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data; in this embodiment, the raw coal quality signal data includes raw spectrum, microwave attenuation and X-ray count, etc.; the auxiliary operating condition data includes coal feeder speed, belt scale instantaneous flow rate, coal seam thickness profile sequence, belt speed, vibration intensity and dust concentration, etc. Calculate the delay time; the delay time refers to the time between the coal quality and the monitoring point; the delay time is expressed as the ratio between the transmission distance of the coal quality from the coal feeder to the monitoring point and the belt speed; in this embodiment, considering the transmission delay of the coal quality from the coal feeder to the monitoring point, the signal data of each channel are time-stamped and spatially aligned based on the belt speed to ensure that the data processed at the same time corresponds to the same segment of coal quality; A delay time is applied to the auxiliary data of the operating conditions, and a synchronous continuous data stream is obtained by combining it with the original coal quality signal data; Data slicing and normalization are performed on a synchronous continuous data stream to obtain data sample units and their corresponding normalized fusion input vectors. Data slicing and normalization refers to trimming the synchronous continuous data stream according to a fixed time window to obtain several data sample units, and performing feature operations on the data within the data sample units to normalize them into a fixed-dimensional normalized fusion input vector. The normalized fusion input vector includes original coal quality signal features and auxiliary operating condition features. In this embodiment, feature operations refer to converting the data into several feature values ​​corresponding to coal quality and operating conditions. The size of the fixed time window is set based on experience; in this embodiment, the size of the fixed time window is set to 5 seconds. In this embodiment, the original coal quality signal features include, but are not limited to, spectral intensity sequences and microwave parameters; the auxiliary operating condition features include, but are not limited to, thickness mean / variance, vibration intensity, and coal feed rate.

[0026] This embodiment first calculates the delay time precisely based on the ratio of transmission distance to belt speed to clarify the transmission lag of coal flow from the feeder to the monitoring point. Then, this delay compensation is applied to the auxiliary data of the operating conditions to achieve strict synchronization with the original coal quality signal on the time axis, ensuring that both accurately correspond to the same coal flow segment. Next, the synchronized continuous data stream is sliced ​​according to a fixed time window to generate discrete data sample units. Finally, feature operations are performed on the data within the unit, integrating the original data of different dimensions and types into a fixed-dimensional regularized fusion input vector. This vector explicitly includes two types of features: coal quality signal and auxiliary operating conditions. This fundamentally solves the problem that multi-source data cannot be directly and effectively fused due to transmission lag and format differences. It provides a precise data alignment basis and a unified processing format for the subsequent feature extraction unit to achieve accurate interference correlation and filtering, effectively avoiding the feature mixing and extraction deviation problems caused by data asynchrony and irregularity in simple filtering methods.

[0027] In this embodiment, the anti-interference feature extraction of coal quality depth feature vector from the regularized fused input vector includes: Extract the normalized and fused input vector corresponding to several data sample units; The regularized and fused input vector is fed into the anti-interference attention feature extraction network to obtain coal quality depth feature vectors corresponding to several data sample units; The anti-interference attention feature extraction network consists of an encoding module, an operating condition interference attention module, a feature purification module, and a fusion dimensionality reduction module. It is used to extract a pure coal quality depth feature vector from a mixed regular fusion input vector. The encoding module consists of two parallel sub-encoders; the sub-encoders include a coal quality signal sub-encoder and an operating condition sub-encoder; The coal quality signal sub-encoder is used to capture the local patterns, spectral correlations, and sequence dependencies of the original coal quality signal features in the regularized and fused input vector, and outputs a preliminary coal quality feature tensor; in this embodiment, the coal quality signal sub-encoder adopts a one-dimensional convolutional neural network; The operating condition sub-encoder is used to encode the operating condition auxiliary features in the regularized fusion input vector into an operating condition disturbance feature vector; the operating condition disturbance feature vector characterizes all external disturbance states that affect coal quality measurement.

[0028] In this embodiment, the working condition interference attention module is used to generate an anti-interference mask from the working condition interference feature vector through an attention mechanism. The input data for the working condition disturbance attention module are the preliminary coal quality feature tensor and the working condition disturbance feature vector; The process of generating an anti-interference mask includes: Using the working condition disturbance feature vector as the query and several feature channels of the preliminary coal quality feature tensor as key-value pairs, the attention weight between the working condition disturbance feature vector and several feature channels of the preliminary coal quality feature tensor is calculated through an attention mechanism; the attention weight represents the severity of disturbance to each component of the preliminary coal quality feature tensor under the current working condition. A softmax normalization operation is performed on several attention weights to obtain several normalized weights; An anti-interference mask is obtained by mapping the normalized weights through a mapping function. In this embodiment, the mapping function is a learnable nonlinear function. In this embodiment, each element value in the anti-interference mask is between 0 and 1. The closer the element value is to 0, the more severe the interference of the corresponding feature channel is under the current operating condition and it needs to be suppressed. The closer the element value is to 1, the more reliable the feature channel is and it should be retained.

[0029] In this embodiment, the feature purification module is used to perform feature optimization on the preliminary coal quality feature tensor to obtain the purified preliminary coal quality feature tensor. The input data for the feature purification module are the preliminary coal quality feature tensor and the anti-interference mask; The purified preliminary coal quality characteristic tensor is obtained by performing a masking operation on the preliminary coal quality characteristic tensor and the anti-interference mask; the masking operation satisfies: Where M represents the anti-interference mask, Represented as the preliminary characteristic tensor of coal quality, This is represented as element-wise multiplication; This is represented as the purified preliminary coal quality feature tensor. In this embodiment, after the preliminary coal quality feature tensor passes through the feature purification module, the features that are severely disturbed in the preliminary coal quality feature tensor are significantly attenuated, while the robust features that are strongly correlated with the properties of the coal itself are preserved and highlighted, thus obtaining the purified preliminary coal quality feature tensor.

[0030] In this embodiment, the fusion and dimensionality reduction module is used to perform feature fusion and dimensionality reduction on the preliminary feature tensor of the purified coal to obtain the coal depth feature vector. The fusion dimensionality reduction module is constructed from several fully connected layers; Both feature fusion and feature dimensionality reduction are achieved through fully connected layers; Feature fusion refers to the process of nonlinearly combining and fusing information from several features in the preliminary feature tensor of purified coal to obtain a high-dimensional feature tensor. Feature dimensionality reduction refers to reducing the feature dimension of a high-dimensional feature tensor to obtain a coal quality depth feature vector.

[0031] In this embodiment, the monitoring results corresponding to several parameters are obtained by performing parameter analysis with physical information constraints on the coal quality depth feature vector, including: Extracting coal depth feature vectors; The coal quality depth feature vector is input into the coal quality multi-parameter regression model to obtain the monitoring results corresponding to several parameters; The coal quality multi-parameter regression model has physical information constraints; The coal quality multi-parameter regression model consists of a shared coding layer, a task-specific decoding layer, and a physical constraint correction layer; The shared coding layer is used to extract abstract features from the coal quality depth feature vector, and is used to predict several parameters in the task-specific decoding layer. In this embodiment, the abstract features are higher-level features. The shared coding layer consists of several fully connected layers. In this embodiment, the shared coding layer consists of two fully connected layers, and the activation function is the GELU activation function.

[0032] In this embodiment, the task-specific decoding layer is used to map the abstract features output by the shared coding layer to predicted values ​​corresponding to several parameters; the parameters in this embodiment include calorific value, moisture content, ash content, volatile matter, and fixed carbon, etc. The task-specific decoding layer consists of several task branches, with each task branch corresponding to a parameter; these task branches are in parallel structure. The task branches consist of fully connected layers and use the GELU activation function.

[0033] In this embodiment, the physical constraint correction layer is used to introduce physical information constraints to correct the predicted values ​​corresponding to several parameters output by the task-specific decoding layer, thereby obtaining the monitoring results corresponding to several parameters. The physical constraint correction layer includes industrial analysis normalization constraints and empirical relational constraints for calorific value industrial analysis. Industrial analysis normalization constraint refers to applying a softmax normalization operation to the predicted values ​​of industrial parameters to obtain the monitoring results of the industrial parameters. In this embodiment, the industrial parameters include moisture, ash, volatile matter, and fixed carbon. The monitoring results after softmax normalization ensure that the sum of the monitoring results of the industrial parameters is 100%. The empirical calorific value is obtained by substituting the monitoring results of several industrial parameters into the calorific value calculation formula; in this embodiment, the calorific value calculation formula adopts Mendeleev's calculation formula. The monitoring result corresponding to the calorific value is obtained by weighted fusion of the predicted calorific value and the empirical calorific value; the weight coefficients of the predicted calorific value and the empirical calorific value are set according to experience. In this embodiment, the weight coefficients of the predicted calorific value and the empirical calorific value are set to 0.4 and 0.6, respectively.

[0034] The loss function of the coal quality multi-parameter regression model in this embodiment satisfies: ;in, Expressed as total loss, This is represented as data loss. Represented as physical constraint loss; and Represented as hyperparameters, used to balance the contributions of data loss and physical constraint loss; The data loss is represented by the weighted mean square error loss, which is the weighted mean square error between the predicted and actual values ​​of several parameters. The physical constraint loss is expressed as the weighted sum of the mass balance constraint loss and the calorific value empirical formula constraint loss; the weight coefficients corresponding to the mass balance constraint loss and the calorific value empirical formula constraint loss are set according to experience. In this embodiment, the mass balance constraint loss and the calorific value empirical formula constraint loss are set to 0.5 and 0.5, respectively. The mass balance constraint loss satisfies: ; i represents the sample number, and N represents the total number of samples; This is represented as the mass balance constraint loss; , , and These represent the percentage values ​​of the predicted values ​​for moisture, ash, volatile matter, and fixed carbon in the industrial parameters, respectively; the mass balance constraint loss is used to constrain the sum of the predicted values ​​of the industrial parameters to 100%. The constraint loss of the empirical formula for calorific value satisfies: ;in, Let represent the predicted calorific value corresponding to the i-th sample. This is represented as the empirical heat value corresponding to the i-th sample; This is expressed as the constraint loss in the empirical formula for calorific value.

[0035] This embodiment first extracts general high-level features from the input features through a shared coding layer. Then, a task-specific decoding layer, composed of multiple parallel task branches, maps the high-level features to initial predicted values ​​of specific parameters such as calorific value, moisture content, ash content, and volatile matter. A physical constraint correction layer introduces two types of core domain knowledge to forcibly correct the initial predictions: on the one hand, it forces the sum of industrial analysis parameters to be 100% through normalization; on the other hand, it uses the corrected industrial parameters combined with empirical formulas to calculate empirical calorific values, and then weights and fuses these with the model's initial calorific value prediction to obtain the final calorific value. Finally, the model is trained using a composite loss function combining data fitting and physical constraints. This function not only minimizes the error between the predicted and true values ​​but also forces the model to adhere to the empirical relationship between mass conservation and calorific value industrial analysis through a penalty term. This deeply embeds the core physical and chemical laws of coal quality analysis into the machine learning model in a computable form, ensuring that the model output not only statistically matches the data but is also physically consistent and reasonable. This fundamentally solves the problem of absurd predictions that violate basic common sense that may arise from purely data-driven models, greatly improving the reliability and interpretability of monitoring results.

[0036] Some of the data in the above formula are calculated by removing dimensions and taking their numerical values. The formula is the closest to the real situation obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0037] The working principle of this application is as follows: First, multi-source heterogeneous raw data is acquired. Second, data processing is performed on the multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized fusion input vectors. Third, anti-interference feature extraction of coal quality depth feature vectors is performed on the regularized fusion input vectors. Fourth, monitoring results corresponding to several parameters are obtained through parameter analysis with physical information constraints on the coal quality depth feature vectors. Fifth, interference components introduced by changes in operating state in the original signal are dynamically identified and suppressed using operating condition data, thereby refining pure and robust coal quality depth feature vectors, replacing passive processing methods such as fixed threshold filtering or moving averages. Fifth, physical constraint parameter analysis is performed based on the coal quality depth feature vectors to output monitoring results. This achieves active and adaptive anti-interference from the signal level to the feature level, improving the purity of feature extraction and robustness to complex field environments. This, in turn, improves the accuracy and stability of online monitoring results, avoiding the problem that existing data processing methods often use simple filtering or averaging, resulting in weak anti-interference capabilities, impure feature extraction, and low accuracy of monitoring results.

[0038] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. An anti-interference type online monitoring system for multiple parameters of coal quality fed into thermal power plants, characterized in that, include: A data acquisition module and a data analysis module; the data acquisition module and the data analysis module are connected to each other; The data acquisition module acquires multi-source heterogeneous raw data through data acquisition equipment; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data. The data analysis module includes a feature extraction unit and a result generation unit; The feature extraction unit: processes multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized and fused input vectors; and performs anti-interference feature extraction on the regularized and fused input vectors to extract coal quality depth feature vectors. The result generation unit obtains monitoring results corresponding to several parameters by performing parameter analysis with physical information constraints on the coal quality depth feature vector.

2. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 1, characterized in that, The process of processing multi-source heterogeneous raw data to obtain data sample units and their corresponding regularized and fused input vectors includes: Extract multi-source heterogeneous raw data; the multi-source heterogeneous raw data includes raw coal quality signal data and auxiliary operating condition data; Calculate the delay time; the delay time refers to the time it takes for coal to travel from the coal feeder to the monitoring point; the delay time is expressed as the ratio between the transmission distance of coal from the coal feeder to the monitoring point and the belt speed; The aforementioned delay time is applied to the auxiliary data of the operating conditions, and a synchronous continuous data stream is obtained by combining it with the original coal quality signal data; Data slicing and normalization are performed on a synchronous continuous data stream to obtain data sample units and their corresponding normalized fusion input vectors. The data slicing and normalization process refers to cutting the synchronous continuous data stream according to a fixed time window to obtain several data sample units, and performing feature operations on the data within the data sample units to normalize them into normalized fusion input vectors of fixed dimensions. The normalized fusion input vectors include original coal quality signal features and auxiliary features of operating conditions.

3. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 1, characterized in that, The process of extracting coal quality depth feature vectors from the regularized fused input vector to resist interference includes: Extract the normalized and fused input vector corresponding to several data sample units; The regularized fusion input vector is input into the anti-interference attention feature extraction network to obtain coal quality depth feature vectors corresponding to several data sample units; The anti-interference attention feature extraction network consists of an encoding module, a working condition interference attention module, a feature purification module, and a fusion and dimensionality reduction module; The encoding module consists of two parallel sub-encoders; the sub-encoders include a coal quality signal sub-encoder and an operating condition sub-encoder; The coal quality signal sub-encoder is used to capture the local patterns, spectral correlations and sequence dependencies of the original coal quality signal features in the regularized fusion input vector, and outputs the preliminary coal quality feature tensor. The operating condition sub-encoder is used to encode the operating condition auxiliary features in the regularized fusion input vector into an operating condition interference feature vector; the operating condition interference feature vector represents all external interference states that affect coal quality measurement.

4. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 3, characterized in that, The operating condition interference attention module is used to generate an anti-interference mask from the operating condition interference feature vector through an attention mechanism. The input data for the working condition disturbance attention module are the preliminary coal quality feature tensor and the working condition disturbance feature vector. The process of generating the anti-interference mask includes: Using the working condition disturbance feature vector as the query and several feature channels of the preliminary coal quality feature tensor as key-value pairs, the attention weight between the working condition disturbance feature vector and several feature channels of the preliminary coal quality feature tensor is calculated through an attention mechanism; the attention weight represents the severity of disturbance to each component of the preliminary coal quality feature tensor under the current working condition. A softmax normalization operation is performed on several attention weights to obtain several normalized weights; An anti-interference mask is obtained by mapping the normalized weights using a mapping function.

5. The anti-interference type online monitoring system for multiple parameters of coal quality fed into thermal power plants according to claim 3, characterized in that, The feature purification module is used to perform feature optimization on the preliminary coal quality feature tensor to obtain the purified preliminary coal quality feature tensor. The input data for the feature purification module are the preliminary coal quality feature tensor and the anti-interference mask; The purified preliminary coal quality characteristic tensor is obtained by performing a masking operation on the preliminary coal quality characteristic tensor and the anti-interference mask; the masking operation satisfies: Where M represents the anti-interference mask, Represented as the preliminary characteristic tensor of coal quality, This is represented as element-wise multiplication; This is represented as the preliminary characteristic tensor of the purified coal.

6. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 3, characterized in that, The fusion and dimensionality reduction module is used to perform feature fusion and dimensionality reduction on the preliminary feature tensor of the purified coal to obtain the coal quality depth feature vector. The fusion dimensionality reduction module is constructed from several fully connected layers; Both feature fusion and feature dimensionality reduction are achieved through fully connected layers; The feature fusion refers to nonlinearly combining and fusing information from several features in the preliminary feature tensor of purified coal to obtain a high-dimensional feature tensor. Feature dimensionality reduction refers to reducing the feature dimension of a high-dimensional feature tensor to obtain a coal quality depth feature vector.

7. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 1, characterized in that, The monitoring results obtained by performing parametric analysis on the coal quality depth feature vector with physical information constraints include: Extracting coal depth feature vectors; The coal quality depth feature vector is input into the coal quality multi-parameter regression model to obtain the monitoring results corresponding to several parameters; The coal quality multi-parameter regression model has physical information constraints; The coal quality multi-parameter regression model consists of a shared coding layer, a task-specific decoding layer, and a physical constraint correction layer. The shared coding layer is used to extract abstract features from the coal quality depth feature vector, and is used to predict several parameters in the task-specific decoding layer.

8. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 7, characterized in that, The task-specific decoding layer is used to map the abstract features output by the shared coding layer to the predicted values ​​corresponding to several parameters; The task-specific decoding layer consists of several task branches, each corresponding to a parameter; the task branches are in parallel structure. The task branch consists of a fully connected layer and uses the GELU activation function.

9. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 7, characterized in that, The physical constraint correction layer is used to introduce physical information constraints to correct the predicted values ​​of several parameters output by the task-specific decoding layer to obtain the monitoring results corresponding to several parameters. The physical constraint correction layer includes industrial analysis normalization constraints and empirical relational constraints for calorific value industrial analysis. The industrial analysis normalization constraint refers to performing a softmax normalization operation on the predicted values ​​of industrial parameters to obtain the monitoring results of industrial parameters. The empirical calorific value is obtained by substituting the monitoring results of several industrial parameters into the calorific value calculation formula; The monitoring results corresponding to the calorific value are obtained by weighted fusion of the predicted calorific value and the empirical calorific value.

10. The anti-interference type online monitoring system for multiple parameters of coal quality in thermal power plants according to claim 7, characterized in that, The loss function of the coal quality multi-parameter regression model satisfies: ;in, Expressed as total loss, This is represented as data loss. Represented as physical constraint loss; and Represented as hyperparameters, used to balance the contributions of data loss and physical constraint loss; The data loss is expressed as a weighted mean square error loss, which is the weighted mean square error between the predicted and actual values ​​of several parameters. The physical constraint loss is expressed as the weighted sum of the mass balance constraint loss and the calorific value empirical formula constraint loss; The mass balance constraint loss satisfies: ; i represents the sample number, and N represents the total number of samples; This is represented as the mass balance constraint loss; , , and These are expressed as percentage values ​​of the predicted values ​​for moisture, ash, volatile matter, and fixed carbon in the industrial parameters, respectively. The constraint loss of the empirical formula for calorific value satisfies: ;in, Let represent the predicted calorific value corresponding to the i-th sample. This is represented as the empirical heat value corresponding to the i-th sample; This is expressed as the constraint loss in the empirical formula for calorific value.

Images