A Dynamic Sensing and Early Warning Method for Boiler Corrosion Based on Multi-Source Information Fusion

Abstract

This invention discloses a dynamic sensing and early warning method for boiler corrosion based on multi-source information fusion, belonging to the field of industrial equipment health monitoring. The specific steps of this early warning method are as follows: collecting and preprocessing multi-dimensional data of each boiler's operation, and integrating the processed multi-dimensional data to generate corresponding state feature vectors. This invention can achieve a refined expression of corrosion risk, effectively improving the ability to identify corrosion hazards in advance and the level of operational safety, and enhancing the foresight and accuracy of early warning. At the same time, this invention can realize real-time monitoring and early warning of boiler corrosion, providing a guarantee for safe operation, ensuring that the early warning effect remains optimal with changes in time and operating conditions, and significantly improving the accuracy of corrosion rate prediction.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment health monitoring, and in particular to a method for dynamic sensing and early warning of boiler corrosion based on multi-source information fusion. Background Technology

[0002] During long-term operation, boilers are prone to corrosion, wear, and structural deterioration of metal components due to factors such as high temperature, high pressure, and media corrosion, which affects the safety and service life of the boiler. Traditional boiler corrosion monitoring methods mostly rely on periodic manual inspections or data from a single sensor, which have problems such as discontinuous data acquisition, limited monitoring dimensions, low prediction accuracy, and delayed early warning. At the same time, the boiler operating environment is complex and variable, and data from a single source cannot fully reflect the corrosion evolution process, making it difficult to identify corrosion risks in a timely manner and failing to effectively guide operation and maintenance and life management.

[0003] Existing methods for dynamic sensing and early warning of boiler corrosion are insufficient for achieving precise and visual representation of corrosion risks, and the identification of potential hazards is often delayed. Furthermore, existing methods typically rely on single or a few types of monitoring data, lacking in-depth fusion and real-time analysis of multi-source heterogeneous information, making it difficult to adaptively adjust to changes in operating conditions, and resulting in significant deviations in corrosion rate prediction. To address these issues, we propose a method for dynamic sensing and early warning of boiler corrosion based on multi-source information fusion. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a dynamic sensing and early warning method for boiler corrosion based on multi-source information fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A dynamic sensing and early warning method for boiler corrosion based on multi-source information fusion, the specific steps of which are as follows:

[0007] Ⅰ: Collect and preprocess multi-dimensional data of each boiler operation, and integrate the processed multi-dimensional data to generate corresponding state feature vectors;

[0008] II: Map each state feature vector to the corresponding risk node, analyze the causal relationship between each risk node, and construct the corresponding corrosion risk propagation chain;

[0009] III: Based on the feature vectors of each state and the corresponding corrosion risk propagation chain, predict the corrosion rate of the boiler under the current operating conditions and calculate the corresponding remaining service life assessment value.

[0010] IV: Based on the historical corrosion data of each boiler, dynamically optimize the corresponding early warning threshold boundaries, and adjust the corresponding early warning trigger conditions in combination with the current operating status of each boiler.

[0011] V: Real-time monitoring of the operational status of each corrosion risk propagation chain, and assessment and classification of risk levels based on the optimized early warning threshold boundaries.

[0012] As a further aspect of the present invention, the specific steps for generating the corresponding state feature vector in step I are as follows:

[0013] S1.1: Collect multi-dimensional data of each boiler operation, and use smoothing filter to filter noise in each dimension of the data. Then, use linear interpolation to supplement the missing values ​​of each dimension of the data. After that, map each dimension of the data to the interval [0, 1], classify each dimension of the data according to the boiler corrosion influencing factors, and form the corresponding influencing factor set.

[0014] S1.2: Real-time acquisition and processing of quality indicators of data from various dimensions, and simultaneous acquisition of dynamic change data of the corresponding boiler operating environment. Through real-time data sampling and statistical analysis, the reliability change trend of data from various dimensions and the dynamic characteristics of the external environment are continuously captured. At the same time, based on the known domain knowledge and historical data of boiler corrosion monitoring, basic rules for corrosion monitoring are established to determine the weight influence coefficients corresponding to different data quality indicators and the degree of influence of environmental changes on the correlation of data from various dimensions. Then, combined with the feedback of historical fusion effect, the initial threshold and gradient range of weight adjustment are set.

[0015] S1.3: Based on the real-time monitoring of quality indicators and environmental changes, and in accordance with the initial weight adjustment rules for corrosion monitoring, calculate the fusion weight of each dimension of data in real time, adjust the weight ratio of each dimension of data according to the weight adjustment rules, and then overwrite the corresponding historical weight values ​​with the current weight adjustment results of each dimension of data in real time to form an updated weight set.

[0016] S1.4: Select an appropriate weighted fusion algorithm for each dimension of data for integration processing. If it is numerical data or sensor data, use a weighted summation method for fusion calculation and allocate the calculation weights according to the real-time weights of each dimension of data. If it is structured data or environmental feature data, use a feature-level weighted fusion method to integrate the weights into the feature extraction process, strengthen the feature contribution of data whose weights exceed the preset threshold, and then integrate the data of each dimension into the corresponding state feature vector.

[0017] As a further aspect of the present invention, the specific steps for constructing the corresponding corrosion risk propagation chain in step II are as follows:

[0018] S2.1: Based on the influencing factors and state feature vectors of each boiler corrosion, the classification criteria for risk nodes are determined, and temperature anomalies, excessive medium corrosion components, metal loss, and excessive humidity are selected as key indicators. Each indicator is then defined as an independent corrosion risk node type. After that, according to the classification criteria and types of risk nodes, each state feature vector is mapped to the corresponding risk node, and a one-to-one or many-to-many relationship between the state feature vector and the risk node is established. The continuously changing feature data is transformed into quantitative indicators of node attributes, and the specific attribute values ​​of each node are obtained.

[0019] S2.2: Based on each state feature vector, perform attribute quantification processing on each mapped risk node, determine the attribute dimension of each risk node, and transform the qualitative information of each state feature vector into quantitative values ​​through statistical analysis and numerical conversion methods, transform the continuously changing state feature vector into quantitative indicators of risk node attributes, and finally obtain the attributes of each risk node.

[0020] S2.3: Based on the physicochemical principles of boiler corrosion and historical operating data, we sort out the potential causal relationships between each risk node, and then analyze the interaction mechanism of the indicators corresponding to different risk nodes in the corrosion process. If the corrosive components of the medium exceed the standard, the risk node will cause the value of the metal loss node to rise. If the temperature exceeds the preset range, the risk node will accelerate the correlation process. Then, by calculating the correlation coefficient and time correlation degree of the attribute values ​​of each risk node, we initially screen out the causal correlation node pairs that meet the preset requirements.

[0021] S2.4: Construct and train the correlation analysis model, and use the trained correlation analysis model to calculate the correlation strength value between each risk node, and sort them from high to low. Then, starting from the risk node whose correlation strength value exceeds the preset threshold, connect the upstream and downstream risk nodes in sequence according to the correlation strength value to form multiple initial corrosion risk propagation paths. Then integrate all corrosion risk propagation paths to construct the corresponding corrosion risk propagation chain.

[0022] As a further aspect of the present invention, the specific training steps of the correlation analysis model described in S2.4 are as follows:

[0023] P1.1: Collect risk node data from historical corrosion cases of different boilers, organize the risk node set corresponding to each corrosion case into a node feature matrix, construct an adjacency matrix based on the real causal relationship between each risk node, and quantify the correlation strength between risk node pairs with numerical values ​​to form a graph structure data sample. Then, divide the data into training set, validation set and test set according to a preset ratio.

[0024] P1.2: Construct an association analysis model and input the graph structure data in the training set into the association analysis model in batches. The association analysis model calculates the predicted value of the association strength of each node pair through forward propagation. Then, it calculates the error between the predicted value and the true value based on the cross-entropy loss function. The error signal is then propagated backward along the network layer of the association analysis model through the backpropagation algorithm. The weights and biases of each layer of the association analysis model are updated using the gradient descent algorithm.

[0025] P1.3: Set the number of training iterations and batch size. Record the loss value after each batch of training is completed. At the same time, evaluate the performance of the association analysis model with the validation set after each round of iteration. Monitor the trend of training set loss and validation set loss during training. If the validation set loss increases for several consecutive rounds or the training reaches the preset number of training rounds, stop training and complete the training of the association analysis model.

[0026] As a further aspect of the present invention, the specific steps for predicting the corrosion rate of the boiler under the current operating conditions in step III are as follows:

[0027] S3.1: Collect historical operating data and corrosion data of different types of boilers as source domain data, and simultaneously collect current operating data of each boiler as target domain data. Construct source domain datasets and target domain datasets respectively. Then, extract corrosion rate features from the source domain datasets, and construct corrosion prediction models based on the extracted features. At the same time, select regression algorithms as the core architecture of corrosion prediction models, use each feature as input and corrosion rate as output to train the corrosion prediction models, and adjust the model parameters through iterative optimization.

[0028] S3.2: Use statistical analysis methods to compare the differences in the characteristic distributions of source domain data and target domain data, calculate the characteristic mean and variance of the two types of data, analyze the reasons for the differences, determine the differences between the current boiler and historical boilers in various aspects, and set an adaptive loss function to dynamically adjust the parameters of the corrosion prediction model using the target domain data.

[0029] S3.3: Divide the target domain data into training set and validation set according to the proportion. Input the training set into the corrosion prediction model for secondary training. Update the model parameters using the minimum batch iteration method. Focus on optimizing the parameter weights that are strongly correlated with the current boiler corrosion characteristics. After each round of fine-tuning, use the validation set to evaluate the corrosion rate prediction accuracy of the model. If the prediction accuracy does not meet the preset standard, continue to adjust the learning rate and the number of iterations until the corrosion prediction model's prediction performance reaches the preset expectation.

[0030] S3.4: The state feature vectors and the corresponding corrosion risk propagation chains are used as inputs and fed into the optimized corrosion prediction model. The corrosion prediction model calculates and outputs the corrosion rate prediction values ​​of each boiler under the current operating conditions in real time, and continuously receives the updated state feature vectors and corrosion risk propagation chains, and the corrosion rate prediction results are updated in real time.

[0031] S3.5: Collect the corrosion data corresponding to the design life parameters and the operating time of each boiler, and combine the predicted corrosion rate results to calculate the total corrosion of each boiler through the cumulative corrosion calculation method. Then, combine the maximum allowable corrosion corresponding to each design life parameter to obtain the corresponding remaining allowable corrosion. After that, based on the predicted corrosion rate results, estimate the service time corresponding to each remaining allowable corrosion, and finally output the remaining service life assessment value of each boiler.

[0032] As a further aspect of the present invention, the specific steps for dynamically optimizing the corresponding early warning threshold boundary in step IV are as follows:

[0033] S4.1: Collect real-time operation status data and historical corrosion data of each boiler during long-term operation, and classify and organize them according to data attributes, dividing them into operation status feature set, corrosion status feature set and early warning result dataset. At the same time, divide each type of dataset into training sample set, verification sample set and test sample set according to preset ratio, and build an early warning threshold optimization sample library.

[0034] S4.2: Using the operation state feature set, corrosion state feature set and early warning result dataset as environmental state variables, set the corresponding environmental state space, quantify the current operation scenario and early warning status of each boiler, set the action space of the agent, determine the adjustment range and adjustment step size of the early warning threshold boundary, then construct the reward function, select the reinforcement learning algorithm that is suitable for the threshold optimization task, and construct the threshold adjustment model.

[0035] S4.3: Input the training sample set in the early warning threshold optimization sample library into the threshold adjustment model in batches. The agent selects the threshold adjustment action based on the current environmental state. At the same time, it calculates the loss value using the policy gradient method, updates the model parameters through the backpropagation algorithm, and optimizes the action selection strategy. After each iteration, the model performance is evaluated using the validation sample set until training is complete. Then, the operation status data and corrosion status data of each boiler are received in real time and input into the threshold adjustment model. The threshold adjustment model outputs the current optimal early warning threshold boundary adjustment scheme.

[0036] S4.4: Dynamically adjust the corresponding early warning thresholds according to the early warning threshold boundary adjustment scheme, while continuously collecting new operating data and early warning feedback results from each boiler, and supplementing the early warning threshold optimization sample library with new data. Optimize model parameters using the newly added samples, then establish a multi-dimensional evaluation system, set different evaluation indicators, and then use the test sample set and actual operating data of each boiler to comprehensively verify the optimized early warning thresholds. Analyze the degree of improvement in early warning effect after the optimization of each early warning threshold. If the degree of improvement meets the preset standard, maintain the current optimization strategy; otherwise, retrain the threshold adjustment model until the preset standard is reached.

[0037] As a further aspect of the present invention, the specific calculation formula for the fusion weights of the data in each dimension in real-time calculation as described in S1.3 is as follows:

[0038] ;

[0039] Indicates the first Data in each dimension at time Real-time fusion weights; Indicates the first Initial weights for each dimension of data; The weighting coefficients representing the influence of data quality indicators; Indicates the first Data in each dimension at time Quality rating; The weighting adjustment coefficient representing the environmental impact; Indicates the first Data in each dimension at time The quality score is affected by environmental changes.

[0040] The specific calculation formula for the weighted summation method described in S1.4 is as follows:

[0041] ;

[0042] In the formula, This represents the fusion result of numerical data; Indicates the number of numerical data dimensions involved in the fusion; Indicates the first Data in each dimension at time The standardized value;

[0043] The specific calculation formula for the feature-level weighted fusion method described in S1.4 is as follows:

[0044] ;

[0045] In the formula, Indicates the first One fusion feature; Indicates the relationship with the first The number of dimensions related to each feature; Indicates the first Data in each dimension at time For the The contribution value of each feature.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] 1. This invention collects multi-dimensional data on the operation of various boilers, performs smoothing filtering and noise reduction, linear interpolation to fill gaps in each dimension of the data, and then uniformly maps it to the [0, 1] interval. The data is then classified according to the factors influencing boiler corrosion, forming corresponding sets of influencing factors. Simultaneously, the quality indicators of each dimension of the data and the dynamic information of the boiler's operating environment are acquired in real time. Combined with domain knowledge of boiler corrosion monitoring and historical data, basic rules for corrosion monitoring are established, and the fusion weights of each dimension of data are dynamically calculated and adjusted. Furthermore, weighted summation or feature-level weighting is used to fuse the data according to the data type of each dimension, forming corresponding state feature vectors. Finally, based on the influencing factors and state feature vectors of each boiler's corrosion, the classification of risk nodes is determined. The standard uses abnormal temperature, excessive corrosive components in the medium, excessive metal loss, and excessive humidity as key indicators. It maps the characteristic vectors of each state to the quantitative attributes of each risk node. Based on the physicochemical mechanism of boiler corrosion and historical data analysis, it analyzes the causal relationship between each risk node. It screens related node pairs through correlation coefficient and time-series correlation analysis, constructs and trains a correlation analysis model, calculates the correlation strength value between each risk node, forms multiple corrosion risk propagation paths, and finally constructs the corresponding corrosion risk propagation chain. This enables dynamic identification and propagation analysis of boiler corrosion risks, achieves refined expression of corrosion risks, effectively improves the ability to identify corrosion hazards in advance and the level of operational safety, and enhances the foresight and accuracy of early warning.

[0048] This invention collects historical operation and corrosion data of different boilers as source domain data and simultaneously collects current boiler operation data as target domain data to construct a corrosion prediction model. It extracts corrosion rate features, trains an initial model using a regression algorithm, and compares the feature distribution differences between the source and target domain data to determine the differences between the current and historical boilers. An adaptive loss function is set, and the corrosion prediction model parameters are dynamically adjusted using target domain data. After secondary training and validation optimization, the corrosion prediction model's predictive performance reaches the preset expectation. Then, the feature vectors of each state and the corresponding corrosion risk propagation chain are input into the optimized model, which outputs the current corrosion rate of each boiler in real time and updates it dynamically. Simultaneously, it combines the corrosion amount data corresponding to the boiler's design life and operating time to accumulate... The remaining allowable corrosion amount is calculated based on the corrosion rate. The corresponding remaining service life assessment value is calculated and output based on the predicted corrosion rate results. Long-term real-time operation status and historical corrosion data of each boiler are collected, classified and organized into three feature sets and sample sets. An optimized sample library for early warning thresholds is constructed. An adaptive adjustment model for early warning thresholds is established based on reinforcement learning. The threshold boundary is optimized through reward function and strategy iteration. The model dynamically outputs the optimal early warning threshold under real-time data drive, and continuously updates the sample library and model parameters by combining the early warning threshold feedback. Finally, the threshold optimization effect is verified through a multi-dimensional evaluation system. It can realize real-time monitoring and early warning of boiler corrosion, provide a guarantee for safe operation, ensure that the early warning effect remains optimal with the change of time and operating conditions, and significantly improve the accuracy of corrosion rate prediction. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0050] Figure 1 This is a flowchart of a boiler corrosion dynamic sensing and early warning method based on multi-source information fusion proposed in this invention. Detailed Implementation

[0051] Reference Figure 1 A dynamic sensing and early warning method for boiler corrosion based on multi-source information fusion is proposed. The specific steps of this early warning method are as follows:

[0052] Collect and preprocess multi-dimensional data of each boiler operation, and integrate the processed multi-dimensional data to generate corresponding state feature vectors.

[0053] Specifically, multi-dimensional data on the operation of each boiler is collected, and noise in each dimension is filtered out using smoothing filters. Missing values ​​in each dimension are then supplemented using linear interpolation. The data is then mapped to the [0, 1] interval. Each dimension is categorized according to boiler corrosion influencing factors to form a corresponding influencing factor set. The quality indicators of the processed data are collected in real time, and dynamic changes in the boiler operating environment are simultaneously collected. Through real-time data sampling and statistical analysis, the reliability trends of each dimension and the dynamic characteristics of the external environment are continuously captured. Based on known domain knowledge and historical data of boiler corrosion monitoring, basic rules for corrosion monitoring are established to determine the weighting influence coefficients corresponding to different data quality indicators and the degree of influence of environmental changes on the correlation of each dimension's data. Finally, combined with historical fusion effect feedback, [the following steps are taken] to set... The initial threshold and gradient range for weight adjustment are determined based on real-time monitoring of quality indicators and environmental changes, and in accordance with the initial weight adjustment rules for corrosion monitoring. The fusion weights of each dimension of data are calculated in real-time, and the weight proportions of each dimension are adjusted according to the weight adjustment rules. The current weight adjustment results for each dimension are then overwritten with the corresponding historical weight values ​​in real-time, forming an updated weight set. A suitable weighted fusion algorithm is selected for each dimension of data for integration processing. For numerical data or sensor data, a weighted summation method is used for fusion calculation, and the weights are allocated according to the real-time weights of each dimension. For structured data or environmental feature data, a feature-level weighted fusion method is used, integrating the weights into the feature extraction process to enhance the feature contribution of data whose weights exceed the preset threshold. Finally, the data from each dimension are integrated into corresponding state feature vectors.

[0054] It should be further explained that the specific calculation formula for the fusion weights of data from each dimension in real time is as follows:

[0055] ;

[0056] Indicates the first Data in each dimension at time Real-time fusion weights; Indicates the first Initial weights for each dimension of data; The weighting coefficients representing the influence of data quality indicators; Indicates the first Data in each dimension at time Quality rating; The weighting adjustment coefficient representing the environmental impact; Indicates the first Data in each dimension at time The quality score is affected by environmental changes.

[0057] The specific calculation formula for the fusion calculation using the weighted summation method is as follows:

[0058] ;

[0059] In the formula, This represents the fusion result of numerical data; Indicates the number of numerical data dimensions involved in the fusion; Indicates the first Data in each dimension at time The standardized value.

[0060] The specific calculation formula for the feature-level weighted fusion method is as follows:

[0061] ;

[0062] In the formula, Indicates the first One fusion feature; Indicates the relationship with the first The number of dimensions related to each feature; Indicates the first Data in each dimension at time For the The contribution value of each feature.

[0063] Each state feature vector is mapped to a corresponding risk node, the causal relationship between each risk node is analyzed, and the corresponding corrosion risk propagation chain is constructed.

[0064] Specifically, based on the influencing factors and state feature vectors of each boiler corrosion, classification criteria for risk nodes are determined, and temperature anomalies, excessive levels of corrosive media components, metal loss, and excessive humidity are selected as key indicators. Each indicator is then defined as an independent corrosion risk node type. Following the classification criteria and types of risk nodes, each state feature vector is mapped to its corresponding risk node, establishing a one-to-one or many-to-many relationship between state feature vectors and risk nodes. Continuously changing feature data is transformed into quantitative indicators of node attributes, obtaining the specific attribute values ​​for each node. Based on each state feature vector, attribute quantification processing is performed on each mapped risk node to determine the attribute dimensions of each risk node. Through statistical analysis and numerical conversion methods, the qualitative information of each state feature vector is transformed into quantitative values, and the continuously changing state feature vectors are transformed into quantitative indicators of risk node attributes, ultimately yielding the final risk node value. Based on the physicochemical principles of boiler corrosion and historical operating data, the potential causal relationships between various risk nodes are identified. The interaction mechanisms of indicators corresponding to different risk nodes during the corrosion process are then analyzed. If the corrosive components of the medium exceed the standard, the risk node will cause the metal loss node value to increase. If the temperature exceeds the preset range, the risk node will accelerate the correlation process. Then, by calculating the correlation coefficient and temporal correlation degree of the attribute values ​​of each risk node, causal correlation node pairs that meet the preset requirements are initially screened. A correlation analysis model is constructed and trained. Using the trained correlation analysis model, the correlation strength value between each risk node is calculated and sorted from high to low. Then, starting from the risk node whose correlation strength value exceeds the preset threshold, upstream and downstream risk nodes are connected sequentially according to the ranking of correlation strength values ​​to form multiple initial corrosion risk propagation paths. Finally, all corrosion risk propagation paths are integrated to construct the corresponding corrosion risk propagation chain.

[0065] It should be further explained that the specific training steps for the association analysis model are as follows:

[0066] Risk node data from historical corrosion cases of different boilers were collected, and the risk node set corresponding to each corrosion case was organized into a node feature matrix. Then, an adjacency matrix was constructed based on the true causal relationship between each risk node. At the same time, the correlation strength between risk node pairs was quantified numerically to form a graph structure data sample. Then, the training set, validation set, and test set were divided according to a preset ratio to construct a correlation analysis model. The graph structure data in the training set was input into the correlation analysis model in batches. The correlation analysis model calculated the predicted value of the correlation strength of each node pair through forward propagation. Then, the error between the predicted value and the true value was calculated based on the cross-entropy loss function. Then, the error signal was propagated backward along the network layer of the correlation analysis model through the backpropagation algorithm. The weights and bias terms of each layer of the correlation analysis model were updated using the gradient descent algorithm. The number of training iterations and batch size were set. The loss value was recorded after each batch of training was completed. At the same time, the performance of the correlation analysis model was evaluated using the validation set after each round of iteration. During the training process, the change trend of the training set loss and the validation set loss was monitored. If the validation set loss increased continuously for several rounds or the training reached the preset training round, the training was stopped, and the training of the correlation analysis model was completed.

[0067] Reference Figure 1 A dynamic sensing and early warning method for boiler corrosion based on multi-source information fusion is proposed. The specific steps of this early warning method are as follows:

[0068] Based on the feature vectors of each state and the corresponding corrosion risk propagation chain, the corrosion rate of the boiler under the current operating conditions is predicted, and the corresponding remaining service life assessment value is calculated.

[0069] Specifically, historical operating data and corrosion data of different types of boilers are collected as source domain data, and current operating data of each boiler are simultaneously collected as target domain data. Source domain datasets and target domain datasets are constructed separately. Then, corrosion rate features are extracted from the source domain datasets, and a corrosion prediction model is built based on these features. A regression algorithm is selected as the core architecture of the corrosion prediction model. The model is trained using each feature as input and the corrosion rate as output. Model parameters are adjusted through iterative optimization. Statistical analysis methods are used to compare the feature distribution differences between the source domain data and the target domain data, and the mean and variance of the features of the two types of data are calculated to analyze the reasons for the differences and determine the differences between the current boiler and historical boilers in various aspects. An adaptive loss function is set, and the parameters of the corrosion prediction model are dynamically adjusted using the target domain data. The target domain data is divided into training and validation sets proportionally. The training set is input into the corrosion prediction model for secondary training, and the model parameters are updated using a minimum batch iteration method, focusing on optimizing the corrosion rate of the current boiler. The parameters with strong correlation to corrosion characteristics are weighted, and the corrosion rate prediction accuracy of the model is evaluated using a validation set after each round of fine-tuning. If the prediction accuracy does not meet the preset standard, the learning rate and number of iterations are adjusted until the corrosion prediction model's prediction performance reaches the preset expectation. Each state feature vector and the corresponding corrosion risk propagation chain are used as inputs and fed into the optimized corrosion prediction model. The corrosion prediction model calculates and outputs the corrosion rate prediction value of each boiler under the current operating conditions in real time, and continuously receives updated state feature vectors and corrosion risk propagation chains. The corrosion rate prediction results are updated in real time, and the design life parameters and corrosion amount data corresponding to the operating time of each boiler are collected. Combined with the predicted corrosion rate results, the total amount of corrosion that has occurred in each boiler is calculated using the cumulative corrosion amount calculation method. Then, combined with the maximum allowable corrosion amount corresponding to each design life parameter, the corresponding remaining allowable corrosion amount is obtained. Finally, based on the corrosion rate prediction results, the service time corresponding to each remaining allowable corrosion amount is estimated, and the remaining service life assessment value of each boiler is output.

[0070] Based on historical corrosion data of each boiler, the corresponding early warning threshold boundaries are dynamically optimized, and the corresponding early warning triggering conditions are adjusted in combination with the current operating status of each boiler.

[0071] Specifically, real-time operational status data and historical corrosion data from the long-term operation of each boiler are collected and categorized according to data attributes, into operational status feature sets, corrosion status feature sets, and early warning result datasets. Simultaneously, each dataset is further divided into training sample sets, validation sample sets, and test sample sets according to a preset ratio to construct an early warning threshold optimization sample library. Using the operational status feature sets, corrosion status feature sets, and early warning result datasets as environmental state variables, corresponding environmental state spaces are defined to quantify the current operating scenarios and early warning states of each boiler. The action space of the intelligent agent is also defined, determining the adjustment range and step size of the early warning threshold boundary. A reward function is then constructed, and a reinforcement learning algorithm suitable for the threshold optimization task is selected to build a threshold adjustment model. The training sample sets from the early warning threshold optimization sample library are input into the threshold adjustment model in batches. The intelligent agent selects threshold adjustment actions based on the current environmental state, and the loss value is calculated using the policy gradient method. The model parameters are updated using the backpropagation algorithm to optimize the action selection strategy. After each iteration, the model performance is evaluated using a validation sample set until training is complete. Then, the operational status data and corrosion status data of each boiler are received in real time and fed into the threshold adjustment model. The threshold adjustment model outputs the current optimal warning threshold boundary adjustment scheme. The corresponding warning threshold is dynamically adjusted according to the warning threshold boundary adjustment scheme. At the same time, new operating data and warning feedback results of each boiler are continuously collected and the new data are added to the warning threshold optimization sample library. The model parameters are optimized using the new samples. Then, a multi-dimensional evaluation system is established, and different evaluation indicators are set. The optimized warning thresholds are then fully validated using a test sample set and actual operating data of each boiler. The degree of improvement of the warning effect after the optimization of each warning threshold is analyzed. If the degree of improvement meets the preset standard, the current optimization strategy is maintained; otherwise, the threshold adjustment model is retrained until the preset standard is reached.

[0072] The system monitors the operational status of each corrosion risk propagation chain in real time and assesses and classifies risk levels based on the optimized early warning threshold boundaries.

Claims

1. A boiler corrosion dynamic sensing and early warning method based on multi-source information fusion, characterized in that, The early warning method comprises the following specific steps: I: Collect and pretreat multi-dimensional data of each boiler operation, integrate the processed multi-dimensional data, and generate corresponding state feature vectors; II: Map each state feature vector to a corresponding risk node, analyze the causal relationship between each risk node, and construct a corresponding corrosion risk propagation chain; the specific steps are as follows: S2.1: Based on the influencing factors and state feature vectors of each boiler corrosion, determine the classification standard of the risk node, and select temperature anomaly, medium corrosion component exceeding standard, metal loss and humidity exceeding standard as key indicators, and then define each indicator as an independent corrosion risk node type, and then map each state feature vector to the corresponding risk node according to the classification standard and type of the risk node, establish a one-to-one or many-to-many association relationship between the state feature vector and the risk node, convert the continuously changing feature data into quantified indicators of node attributes, and obtain the specific attribute value of each node; S2.2: Based on each state feature vector, attribute quantization processing is performed on each mapped risk node, the attribute dimension of each risk node is determined, and through statistical analysis and numerical conversion method, the qualitative information of each state feature vector is converted into quantitative value, and the continuously changing state feature vector is converted into quantified indicators of risk node attributes, and finally the attributes of each risk node are obtained; S2.3: Based on the physical and chemical principles of boiler corrosion and historical operation data, the potential causal relationship between each risk node is analyzed, and the interaction mechanism of the indicators corresponding to different risk nodes in the corrosion process is analyzed, if the medium corrosion component exceeds the standard, the risk node will cause the metal loss node value to rise, if the temperature exceeds the preset range, the risk node will accelerate the associated process, then the correlation coefficient and time sequence correlation degree of each risk node attribute value are calculated, and the causal associated node pairs meeting the preset requirements are preliminarily screened out; S2.4: Build and train the correlation analysis model, calculate the correlation strength value between each risk node using the trained correlation analysis model, and sort them from high to low, then take the risk node with correlation strength value exceeding the preset threshold as the starting point, connect the upstream and downstream risk nodes in turn according to the sorting of the correlation strength value, form multiple initial corrosion risk propagation paths, integrate all corrosion risk propagation paths, and construct a corresponding corrosion risk propagation chain; The specific training steps of the correlation analysis model are as follows: P1.1: Collect risk node data in different boiler historical corrosion cases, organize the risk node set corresponding to each corrosion case into a node feature matrix, construct an adjacency matrix according to the real causal relationship between each risk node, and quantify the correlation strength between the risk node pairs to form a graph structure data sample, and then divide the training set, validation set and test set according to the preset proportion; P1.2: Construct a correlation analysis model, and input the graph structure data in the training set into the correlation analysis model in batches. The correlation analysis model calculates the correlation strength prediction value of each node pair through forward propagation. Then, the error between the prediction value and the true value is calculated based on the cross-entropy loss function. Then, the error signal is transmitted in the reverse direction along the network layers of the correlation analysis model through the back propagation algorithm, and the weights and bias terms of each layer of the correlation analysis model are updated by using the gradient descent algorithm; P1.3: Set the number of training iterations and the batch size. Record the loss value after completing a batch of training. At the same time, evaluate the performance of the correlation analysis model using the validation set after completing one round of iteration. Monitor the training set loss and validation set loss trend during the training process. If the validation set loss increases continuously for multiple rounds or the training reaches the preset training round, stop the training, and complete the training of the correlation analysis model; III: According to the state feature vector and the corresponding corrosion risk propagation chain, the corrosion rate of the boiler under the current working condition is predicted, and the corresponding remaining service life evaluation value is calculated. The specific steps are as follows: S3.1: Collect the historical operation data and corrosion data of different types of boilers as source domain data, and synchronously collect the operation data of the current boilers as target domain data. Construct the source domain data set and the target domain data set respectively. Then, extract the corrosion rate features from the source domain data set. Based on the extracted features, a corrosion prediction model is constructed. A regression algorithm is selected as the core architecture of the corrosion prediction model. The features are input, and the corrosion rate is output. The corrosion prediction model is trained by iteratively optimizing and adjusting the model parameters; S3.2: Use statistical analysis method to compare the feature distribution difference of source domain data and target domain data, and calculate the feature mean and variance of the two types of data. Analyze the reasons for the difference, determine the differences between the current boiler and the historical boiler in various aspects, and set an adaptive loss function. Dynamically adjust the parameters of the corrosion prediction model using the target domain data; S3.3: Divide the target domain data into training set and validation set in proportion. Input the training set into the corrosion prediction model for secondary training. Update the model parameters using the minimum batch iteration method. Focus on optimizing the parameter weights that are strongly related to the corrosion characteristics of the current boiler. After each fine-tuning, evaluate the corrosion rate prediction accuracy of the model using the validation set. If the prediction accuracy does not meet the preset standard, continue to adjust the learning rate and iteration times until the prediction performance of the corrosion prediction model reaches the preset expectation; S3.4: Input the state feature vector and the corresponding corrosion risk propagation chain into the optimized corrosion prediction model. The corrosion prediction model outputs the corrosion rate prediction value of the current boiler under the existing operating condition through real-time calculation. The corrosion prediction model continuously receives the updated state feature vector and corrosion risk propagation chain, and updates the corrosion rate prediction result in real time. S3.5: Collect the design life parameters of each boiler and the corrosion amount data corresponding to the running time, and combine the predicted corrosion rate prediction results to calculate the total corrosion amount of each boiler that has occurred by the cumulative corrosion amount accounting method. Then, combine the allowed maximum corrosion amount corresponding to each design life parameter to obtain the corresponding remaining allowed corrosion amount. Then, according to the corrosion rate prediction results of each boiler, calculate the use time corresponding to each remaining allowed corrosion amount, and finally output the remaining service life evaluation value of each boiler; IV: Based on the historical corrosion data of each boiler, dynamically optimize the corresponding early warning threshold boundary, and at the same time, adjust the corresponding early warning trigger condition combined with the current running state of each boiler; V: Real-time monitor the running state of each corrosion risk propagation chain, and evaluate and divide the risk level combined with the optimized early warning threshold boundary.

2. The method according to claim 1, wherein, The specific steps of generating the corresponding state feature vector in step I are as follows: S1.1: Collect multi-dimensional data of each boiler running, filter noise in each dimension data by using smoothing filter, then supplement missing values of each dimension data by using linear interpolation method, then map each dimension data to [0, 1] interval, classify each dimension data according to boiler corrosion influencing factors, and form corresponding influencing factor set; S1.2: Real-time collect quality indicators of processed multi-dimensional data, and synchronously collect dynamic change data of corresponding boiler running environment, then continuously capture reliability change trend of each dimension data and dynamic characteristics of external environment through real-time data sampling and statistical analysis, and based on known field knowledge and historical data of boiler corrosion monitoring, establish corrosion monitoring basic rules to determine weight influence coefficient corresponding to different data quality indicators and influence degree of environmental change on correlation of each dimension data, then set initial threshold and gradient range of weight adjustment combined with historical fusion effect feedback; S1.3: According to real-time monitored quality indicators and environmental change, and according to initialized weight adjustment corrosion monitoring basic rules, real-time calculate fusion weight of each dimension data, and adjust weight proportion of each dimension data according to weight adjustment rules, then real-time cover historical weight value of each dimension data with current weight adjustment result to form updated weight set; S1.4: Select appropriate weighted fusion algorithm for each dimension data for integration processing, if it is numerical data or sensor data, use weighted summation method for fusion calculation, and calculate weight according to real-time weight of each dimension data; If it is structured data or environmental feature data, use feature level weighted fusion method, integrate weight into feature extraction process, strengthen feature contribution of weight exceeding preset threshold, then integrate each dimension data into corresponding state feature vector.

3. The method according to claim 1, characterized in that, The specific steps of dynamically optimizing the corresponding early warning threshold boundary in step IV are as follows: S4.1: Collect real-time operation state data and historical corrosion data of each boiler during long-term operation, classify and organize them according to data attributes, divide them into operation state feature set, corrosion state feature set and early warning result data set, and divide each data set into training sample set, validation sample set and test sample set according to a preset proportion, and construct early warning threshold optimization sample library; S4.2: Taking the operation state feature set, the corrosion state feature set and the early warning result data set as the environmental state variables, setting the corresponding environmental state space, quantifying the running scene and the early warning state of each boiler, setting the action space of the agent, determining the adjustment range and adjustment step of the early warning threshold boundary, then constructing the reward function, selecting the reinforcement learning algorithm suitable for threshold optimization task, and constructing the threshold adjustment model; S4.3: Input the training sample set in the early warning threshold optimization sample library into the threshold adjustment model in batches, the agent selects the threshold adjustment action based on the current environmental state, calculates the loss value by using the policy gradient method, updates the model parameters by using the back propagation algorithm, optimizes the action selection strategy, evaluates the model performance using the validation sample set after each iteration is completed, until the training is completed, then receives the operation state data and the corrosion state data of each boiler in real time, and inputs them into the threshold adjustment model, the threshold adjustment model outputs the current optimal early warning threshold boundary adjustment scheme; S4.4: According to the early warning threshold boundary adjustment scheme, dynamically adjust the corresponding early warning threshold, continuously collect new operation data and early warning feedback results of each boiler, and supplement new data to the early warning threshold optimization sample library, optimize the model parameters using the new samples, then establish a multi-dimensional evaluation system, set different evaluation indexes, and use the test sample set and the actual operation data of each boiler to verify the optimized early warning threshold, analyze the improvement degree of the early warning effect after optimizing each early warning threshold, if the improvement degree meets the preset standard, maintain the current optimization strategy; Otherwise, retrain the threshold adjustment model until the preset standard is met.

4. The method according to claim 2, characterized in that, The specific calculation formula of the real-time calculation of the fusion weight of each dimension data in S1.3 is as follows: ； Indicates the first Data in each dimension at time Real-time fusion weights; Indicates the first Initial weights for each dimension of data; The weighting coefficients representing the influence of data quality indicators; Indicates the first Data in each dimension at time Quality rating; The weighting adjustment coefficient representing the environmental impact; Indicates the first Data in each dimension at time The quality score is affected by environmental changes. The specific calculation formula of the fusion calculation by using the weighted summation method in S1.4 is as follows: ； In the formula, This represents the fusion result of numerical data; Indicates the number of numerical data dimensions involved in the fusion; Indicates the first Data in each dimension at time The standardized value; The specific calculation formula of the feature-level weighted fusion method in S1.4 is as follows: ； In the formula, represents the first fusion feature; represents the number of dimension data related to the first feature; represents the first dimension data at the moment of time; contribution value of the first feature.

Images