Water-cooled wall three-dimensional thermal displacement real-time monitoring and health assessment method and system

Abstract

The present application relates to the field of boiler water wall monitoring, and provides a water wall three-dimensional thermal displacement real-time monitoring and health assessment method and system, the method comprising: acquiring multi-source monitoring data of the boiler water wall for space-time correlation analysis to obtain a multi-dimensional monitoring feature set; performing parameter threshold comparison and deviation analysis on the multi-dimensional monitoring feature set, constructing a thermal stress coupling model, analyzing parameter correlation, and obtaining a risk feature mapping network; performing thermal displacement trend prediction and stress concentration assessment on the risk feature mapping network based on a dynamic correlation prediction model to obtain an equipment health risk atlas; performing cracking risk assessment on the equipment health risk atlas based on a health state evaluation index system to obtain an equipment health state report, generating maintenance decision suggestions according to the report, and realizing real-time abnormality early warning and maintenance strategy suggestion pushing through edge computing. The present application can effectively early warn the cracking risk of the boiler water wall caused by uneven temperature and expansion, and improve the operation safety of the unit.

Description

Technical Field

[0001] This invention relates to the field of boiler water-cooled wall monitoring technology, and in particular to a method and system for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls. Background Technology

[0002] As the core equipment of a thermal power generation system, the safe and stable operation of the boiler directly affects the reliability and economy of the entire boiler unit. The water-cooled wall is the most important heating surface in the boiler, typically composed of tens of thousands of parallel tubes forming a four-sided enclosed structure around the furnace: the left side wall, right side wall, front wall, and rear wall, with a hollow furnace chamber in the center. The pulverized coal combustion temperature inside the furnace can reach over 1000℃. Water circulates inside the water-cooled wall tubes, absorbing heat from the high-temperature flame and flue gas to heat the water into high-temperature, high-pressure steam, completing the energy conversion. The upper part of the water-cooled wall connects to the upper header, and the lower part connects to the lower header. The entire structure is suspended from a large plate beam by supports and hangers, expanding vertically downwards when heated.

[0003] Under ideal operating conditions, the temperature of the four water-cooled walls at the same elevation should be consistent, with uniform expansion and balanced structural stress. However, due to factors such as uneven combustion organization within the furnace, unbalanced aerodynamic field distribution, and fuel distribution deviations, the water-cooled walls at different areas at the same elevation experience uneven heat absorption during actual operation, leading to localized overheating or significant temperature gradient differences. This non-uniform temperature distribution not only reduces the local strength of the material but also causes inconsistent thermal expansion in different parts, resulting in additional thermal stress in the tubes and connecting components. When local stress exceeds the allowable stress of the material, it easily triggers the initiation and propagation of cracks in the water-cooled wall tubes, ultimately leading to leakage accidents. More seriously, the water-cooled wall is composed of numerous slender tubes, with a dense structure and high pressure resistance. Once a single tube cracks, it will directly cause unplanned shutdowns of the boiler unit, resulting in huge economic losses. Statistics show that more than half of the unplanned boiler unit shutdowns originate from water-cooled wall leakage problems. Therefore, real-time monitoring of the water-cooled walls is necessary. However, traditional monitoring methods are mostly limited to single-point temperature or pressure measurements, which are difficult to fully reflect the thermal deformation state of the water-cooled wall in three-dimensional space and its dynamic coupling relationship with stress and load. They also lack the ability to perceive and quantify key risk factors such as uneven thermal displacement and excessive temperature deviation in real time. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and provides a method and system for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls.

[0005] One aspect of the present invention provides a method for real-time monitoring and health assessment of three-dimensional thermal displacement of a water-cooled wall, comprising: acquiring multi-source monitoring data of a boiler water-cooled wall; performing spatiotemporal correlation analysis on the multi-source monitoring data to obtain a multi-dimensional monitoring feature set; performing parameter threshold comparison and deviation analysis on the multi-dimensional monitoring feature set to construct a thermal stress coupling model, analyzing parameter correlation, and obtaining a risk feature mapping network; constructing a dynamic correlation prediction model; performing thermal displacement trend prediction and stress concentration assessment on the risk feature mapping network based on the dynamic correlation prediction model to obtain an equipment health risk map; establishing a health status evaluation index system; performing cracking risk assessment on the equipment health risk map based on the health status evaluation index system to obtain an equipment health status report; generating maintenance decision suggestions based on the equipment health status report, and realizing real-time anomaly warning and maintenance strategy suggestion push through edge computing.

[0006] Optionally, the multi-source monitoring data includes temperature signals, three-dimensional thermal displacement signals, stress signals, and unit load signals at different locations at the same altitude. Spatiotemporal correlation analysis is performed on the multi-source monitoring data to obtain a multi-dimensional monitoring feature set, including: extracting spatiotemporal distribution features from the temperature signals at different locations at the same altitude to obtain temperature field gradient features; performing three-dimensional coordinate transformation on the three-dimensional thermal displacement signals to obtain spatial deformation features; performing time-domain waveform analysis on the stress signals to obtain stress concentration features; extracting time-series trends from the unit load signals to obtain load fluctuation features; performing coupling analysis based on the temperature field gradient features and the load fluctuation features to obtain a thermal load correlation coefficient; correcting the spatial deformation features based on the thermal load correlation coefficient to obtain true thermal displacement features; and performing spatiotemporal registration based on the true thermal displacement features, the temperature field gradient features, the stress concentration features, and the load fluctuation features to construct a multi-dimensional monitoring feature set.

[0007] Optionally, the temperature signals at different locations at the same height are subjected to spatiotemporal distribution feature extraction to obtain temperature field gradient features. This includes: spatially discretizing the water-cooled wall monitoring area using a finite element mesh generation method to obtain grid node temperature data; calculating the deviation between the temperature of each grid node and the corresponding theoretical design value in the grid node temperature data to construct a temperature deviation matrix; performing spatial interpolation based on the temperature deviation matrix to generate a temperature field distribution cloud map; and calculating the temperature difference and relative deviation ratio at different locations at the same height based on the temperature field distribution cloud map to construct a temperature gradient feature vector and obtain the temperature field gradient features.

[0008] Optionally, the multidimensional monitoring feature set is subjected to parameter threshold comparison and deviation analysis to construct a thermal stress coupling model, analyze parameter correlation, and obtain a risk feature mapping network. This includes: comparing the temperature parameters, thermal displacement parameters, and stress parameters in the multidimensional monitoring feature set with their corresponding theoretical design values, and obtaining parameter exceedance indexes based on the comparison results; calculating the differences in temperature parameters, thermal displacement parameters, and stress parameters at different locations at the same height to construct a lateral deviation feature spectrum; performing dimensionless processing on the lateral deviation feature spectrum to obtain a relative deviation proportionality coefficient; constructing a thermal stress coupling model based on the parameter exceedance index and the relative deviation proportionality coefficient; performing multivariate correlation analysis on the temperature field gradient characteristics, the actual thermal displacement characteristics, and the unit load characteristics to obtain a cross-parameter correlation matrix; identifying key influencing factors based on the cross-parameter correlation matrix and constructing a parameter correlation network; and extracting risk features based on the parameter correlation network and the thermal stress coupling model to obtain the risk feature mapping network.

[0009] Optionally, the lateral deviation characteristic spectrum is dimensionless to obtain a relative deviation ratio coefficient, including: setting a deviation threshold range; normalizing the lateral deviation characteristic spectrum to obtain a standardized deviation value; using the analytic hierarchy process (AHP) to determine the weighting coefficients of the temperature parameter, the thermal displacement parameter, and the stress parameter; constructing a weighted deviation matrix based on the standardized deviation value and the weighting coefficients; and calculating the relative deviation ratio between each element in the weighted deviation matrix and the corresponding theoretical design value to obtain the relative deviation ratio coefficient.

[0010] Optionally, a dynamic correlation prediction model is constructed, and the risk feature mapping network is used to predict thermal displacement trends and assess stress concentration based on the dynamic correlation prediction model to obtain an equipment health risk map. This includes: constructing a long short-term memory network model and an attention mechanism model; fusing the long short-term memory network model and the attention mechanism model to construct an initial correlation prediction model; training the initial correlation prediction model using historical monitoring data of the boiler water-cooled wall to obtain a dynamic correlation prediction model; extracting temporal features from the risk feature mapping network using a sliding time window method, and inputting the extracted features into the dynamic correlation prediction model to obtain thermal displacement trend prediction results; constructing a stress development trend curve based on the thermal displacement trend prediction results and stress concentration assessment results; and comparing the stress development trend curve with a preset safety threshold to obtain an equipment health risk map.

[0011] Optionally, the health status evaluation index system includes parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators. Based on the health status evaluation index system, a crack risk assessment is performed on the equipment health risk map to obtain an equipment health status report. This includes: quantifying the equipment health risk map based on the health status evaluation index system to obtain equipment health indicators; standardizing the equipment health indicators to obtain corresponding normalized health indices; weighting and fusing the normalized health indices to obtain a comprehensive risk score; classifying the equipment crack risk level based on the comprehensive risk score; performing sensitivity analysis on the normalized health indices to obtain corresponding risk contribution indicators; and generating an equipment health status report based on the equipment crack risk level and the risk contribution indicators.

[0012] Optionally, the normalized health index is weighted and fused to obtain a comprehensive risk score, including: constructing a risk assessment matrix to map the normalized health index to the corresponding risk probability; using a fuzzy comprehensive evaluation method to calculate the membership degree of the risk probability to obtain the corresponding fuzzy risk vector; determining the index weight of the fuzzy risk vector based on the risk contribution index; using the index weight to perform a weighted summation of the fuzzy risk vector; mapping the weighted result to the [0,100] scoring interval to obtain a comprehensive risk score.

[0013] Optionally, maintenance decision recommendations are generated based on the equipment health status report, including: identifying high-risk areas and key influencing parameters based on the equipment health status report, matching them with a historical cracking case library, and inferring possible cracking modes; assessing the urgency of maintenance based on the equipment cracking risk level and thermal displacement trend prediction results, and generating maintenance priorities; formulating maintenance window recommendations based on the maintenance priorities and unit operation plans; generating targeted maintenance plans based on the inferred possible cracking modes and risk contribution indicators; and integrating the maintenance window recommendations and the targeted maintenance plans into the maintenance decision recommendations.

[0014] In another aspect, the present invention provides a real-time monitoring and health assessment system for three-dimensional thermal displacement of a water-cooled wall, comprising: at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores a computer program that, when executed by at least one of the processors, implements the real-time monitoring and health assessment method for three-dimensional thermal displacement of a water-cooled wall as described in any one of claims 1 to 9.

[0015] Compared to existing technologies, this invention overcomes the limitations of traditional monitoring methods that rely on single parameters and cannot comprehensively reflect the true thermal state of the structure by integrating multi-source monitoring data such as temperature, three-dimensional thermal displacement, stress, and unit load of the boiler water-cooled wall, and conducting spatiotemporal correlation analysis and multi-dimensional feature extraction. By constructing a thermal-stress coupling model and a risk feature mapping network, it achieves quantitative characterization and correlation analysis of key risk factors such as temperature unevenness, limited expansion, and stress concentration at the same elevation of the boiler water-cooled wall, improving the ability to identify abnormal states of the water-cooled wall under complex operating conditions. Furthermore, by introducing a dynamic correlation prediction model, combined with a long short-term memory network and attention mechanism, it can accurately predict the development trend of thermal displacement and the evolution law of stress, and capture potential cracking risks in advance. By establishing a health status evaluation index system and generating equipment health status reports accordingly, it achieves a systematic assessment from data to risk level. Finally, based on edge computing, it realizes localized real-time early warning and intelligent push of maintenance strategies, significantly improving response speed and the pertinence of operation and maintenance decisions. This invention enables all-time, multi-scale, and dynamic monitoring and health assessment of the three-dimensional thermal deformation behavior of water-cooled walls. It can effectively warn of cracking risks caused by uneven combustion, such as local overheating, expansion imbalance, and excessive thermal stress, significantly reducing the probability of unplanned unit shutdowns and improving the safety, stability, and economy of boiler operation. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 A flowchart of a method for real-time monitoring and health assessment of three-dimensional thermal displacement of a water-cooled wall provided in one embodiment of the present invention; Figure 2 This is a flowchart illustrating the process of comparing and analyzing parameter thresholds and deviations of a multidimensional monitoring feature set, constructing a thermal stress coupling model, analyzing parameter correlation, and obtaining a risk feature mapping network in a real-time monitoring and health assessment method for three-dimensional thermal displacement of a water-cooled wall, as provided in another embodiment of the present invention. Figure 3 This is a flowchart illustrating the process of constructing a dynamic correlation prediction model, using a risk feature mapping network to predict thermal displacement trends and assess stress concentration based on the dynamic correlation prediction model, in a method for real-time monitoring and health assessment of three-dimensional thermal displacement of a water-cooled wall, as provided in another embodiment of the present invention, to obtain a health risk map of the equipment. Detailed Implementation

[0018] The main objective of this invention is to provide a method and system for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls, in order to solve the technical problem that due to uneven combustion in the furnace, the temperature distribution at the same elevation of the water-cooled wall is uneven, causing differences in thermal expansion and concentration of thermal stress. Existing monitoring methods are unable to accurately perceive the risk of three-dimensional thermal displacement and multi-parameter coupling in real time, and cannot effectively warn of unplanned unit shutdowns caused by cracking of the water-cooled wall.

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0020] One embodiment of the present invention provides a method for real-time monitoring and health assessment of three-dimensional thermal displacement of a water-cooled wall, the process of which is as follows: Figure 1 As shown, it includes steps S100 to S500.

[0021] Step S100: Acquire multi-source monitoring data of the boiler water-cooled wall, perform spatiotemporal correlation analysis on the multi-source monitoring data, and obtain a multi-dimensional monitoring feature set. The boiler water-cooled wall is a heating surface structure composed of numerous parallel tubes arranged around the furnace inside the boiler, used to absorb heat from high-temperature flue gas and heat the working fluid water into steam. The boiler water-cooled wall is used to complete the conversion of combustion heat energy into working fluid heat energy and is a key component of boiler energy transfer.

[0022] Multi-source monitoring data can be a collection of real-time operational data from different physical quantities and sensor types. Specifically, it can include temperature data, three-dimensional thermal displacement data, stress data, and unit load data, which can be obtained by synchronously acquiring signals such as temperature, three-dimensional thermal displacement, stress, and unit load through a distributed sensor network. For example, temperature data can reflect the thermal state parameters of the working fluid on or inside the water-cooled wall tube at a specific location and time, used to identify local overheating areas and temperature distribution inhomogeneity. This data can be acquired through thermocouples, infrared thermometers, or fiber optic grating sensors. Three-dimensional thermal displacement data can describe the displacement of the water-cooled wall tube in three spatial directions after heating, used to reflect the spatial non-uniformity of thermal expansion behavior and reveal areas of restricted expansion. Three-dimensional thermal displacement data can be acquired through laser displacement sensors, visual measurement systems, or calculated using strain-displacement conversion models. Stress data can be the internal mechanical response values ​​of the water-cooled wall tube due to differences in thermal expansion and structural constraints, used to assess whether the material is approaching its allowable stress limit. This data can be obtained through indirect measurement using strain gauges or based on thermo-mechanical coupling inversion calculations. Unit load data can be operating parameters characterizing the current output power or steam flow of the boiler. It is used as an external disturbance variable to correlate and analyze the relationship between thermal response and operating status. Unit load data can be obtained by reading signals such as main steam flow and power generation from the distributed control system (DCS) connected to the boiler unit.

[0023] When acquiring multi-source monitoring data of boiler water-cooled walls, sensor signals such as temperature, three-dimensional thermal displacement, stress, and unit load can be collected simultaneously to obtain temperature data, three-dimensional thermal displacement data, stress data, and unit load data respectively. Furthermore, when acquiring multi-source monitoring data of boiler water-cooled walls, data from multiple sensor types can be uniformly acquired through an industrial IoT platform, or multi-source synchronization can be achieved using timestamp alignment and data interpolation processing, thus providing a comprehensive raw data foundation for subsequent analysis.

[0024] When performing spatiotemporal correlation analysis on multi-source monitoring data, correlation models between parameters can be established in both time series and spatial topology dimensions. Furthermore, this analysis can reveal the impact mechanism of uneven combustion conditions on the overall thermodynamic state of water-cooled walls by constructing spatiotemporal graph convolutional networks to extract neighborhood correlations or using sliding windows to calculate cross-regional cross-correlation coefficients. The multidimensional monitoring feature set can be obtained by extracting discriminative feature vectors from the spatiotemporal correlation analysis results. Further, obtaining the multidimensional monitoring feature set can be achieved through principal component analysis to reduce dimensionality and retain the main variation directions, or by using an autoencoder to learn low-dimensional embedding representations, thereby compressing data dimensionality and retaining key risk information.

[0025] Step S200 involves comparing parameter thresholds and performing deviation analysis on the multidimensional monitoring feature set, constructing a thermal stress coupling model, analyzing parameter correlations, and obtaining a risk feature mapping network. Specifically, parameter threshold comparison can be an operation that compares data in the multidimensional monitoring feature set with preset safety limits point-by-point or region-by-region, used to initially identify abnormal states exceeding the normal range. Deviation analysis can be a quantitative process of calculating the degree of parameter differences in different areas at the same elevation, used to identify structural risks such as uneven temperature and expansion imbalance.

[0026] When performing parameter threshold comparison and deviation analysis on a multidimensional monitoring feature set, the characteristic values ​​of various data in the multidimensional monitoring feature set can be compared with the corresponding safety thresholds, and the degree of difference between various data in the same elevation area can be calculated. Furthermore, when performing parameter threshold comparison and deviation analysis on a multidimensional monitoring feature set, a fixed engineering threshold can be set for hard decision-making, or a dynamic threshold such as the 3σ principle can be used to adapt to changes in working conditions, in order to identify two types of risks: explicit exceedances and implicit imbalances.

[0027] A thermo-stress coupling model is a mathematical model that describes the physical coupling relationship between temperature, displacement, and stress fields, used to quantify the distribution of additional thermal stress caused by temperature inhomogeneity. Thermo-stress coupling models can be constructed based on thermoelastic theory, combined with finite element methods or analytical methods. For example, when constructing a thermo-stress coupling model, a mapping relationship between temperature, displacement, and stress can be established based on thermoelastic theory. Furthermore, simplified analytical formulas can be used to quickly estimate thermal stress, or a pre-trained finite element surrogate model can be invoked for high-precision inversion, thereby achieving a physical deduction from observable temperature to displacement and stress.

[0028] Parameter correlation can refer to the statistical or physical dependencies between different variables in a multidimensional monitoring feature set, supporting the construction of risk feature mapping networks and revealing hidden failure modes. When analyzing parameter correlation, correlation coefficients, mutual information, or causal strength between multidimensional features can be calculated. For example, Granger causality tests can be used to analyze time-series causality, or graph attention networks can be used to learn the weights between features, thereby discovering non-intuitive multi-parameter coupled failure modes.

[0029] A risk feature mapping network can be a model that expresses the correlations between key risk factors in the form of a graph structure or neural network, used to achieve a joint representation of risks such as temperature unevenness, expansion restriction, and stress concentration. For example, a risk feature mapping network can be constructed through association rule mining, graph embedding, or attention weight learning. When obtaining the risk feature mapping network, the results of parameter correlation analysis can be organized into a network structure, where nodes represent risk factors and edges represent correlation strengths. Furthermore, a weighted undirected graph can be constructed to represent correlations, or a directed graph can be established to represent causal influence paths, thereby structurally expressing the corresponding characteristics of multiple risk factors.

[0030] Step S300: Construct a dynamic correlation prediction model. Based on this model, perform thermal displacement trend prediction and stress concentration assessment on the risk feature mapping network to obtain an equipment health risk map. Specifically, the dynamic correlation prediction model can be a prediction architecture that integrates time series modeling and an attention mechanism to capture the evolution patterns of thermal displacement and stress, and predict future thermal displacement trends and stress concentration areas in advance. For example, the dynamic correlation prediction model can be obtained by combining a Long Short-Term Memory (LSTM) network with an attention mechanism. The LSTM network can be a recurrent neural network with a gating mechanism to process time series data with long-term dependencies, specifically used to model the historical evolution paths of thermal displacement and stress. The attention mechanism can be a module that calculates the importance weights of inputs at different time steps or spatial locations to enhance the predictive sensitivity for key risk periods or regions.

[0031] Thermal displacement trend prediction can estimate the direction and magnitude of three-dimensional thermal displacement changes in a water-cooled wall over a future period, used to predict the development trend of expansion imbalance. Stress concentration assessment can quantitatively determine whether stress in a local area of ​​the water-cooled wall tends to concentrate or exceed limits, used to identify potential crack initiation locations. When performing thermal displacement trend prediction and stress concentration assessment on a risk feature mapping network based on a dynamic correlation prediction model, the current risk characteristics of the water-cooled wall can be input into the dynamic correlation prediction model to obtain the displacement and stress distribution for future periods output by the model. For example, when performing thermal displacement trend prediction and stress concentration assessment on a risk feature mapping network based on a dynamic correlation prediction model, the thermal displacement change over the next 15 minutes can be predicted in a rolling manner, or Monte Carlo simulation can be combined to assess prediction uncertainty, thereby achieving forward-looking risk identification.

[0032] The equipment health risk map is a spatial distribution map of risk that integrates prediction results with the current state of the water-cooled wall, allowing for the identification of high-risk areas. The equipment health risk map maps the output of a dynamic correlation prediction model onto the geometry of the water-cooled wall, visually presenting the spatial distribution and temporal evolution of cracking risk. After obtaining the equipment health risk map, the prediction results of the dynamic correlation prediction model can be overlaid with the current state of the water-cooled wall to generate a spatial distribution map with risk levels. Furthermore, when obtaining the equipment health risk map, risk thermography can be rendered on a 3D water-cooled wall model, or a 2D unfolded icon can be generated to annotate risk coordinates, thus intuitively displaying high-risk areas and their evolution trends.

[0033] Step S400: Establish a health status evaluation index system. Based on this system, conduct a crack risk assessment on the equipment health risk map to obtain an equipment health status report. Specifically, the health status evaluation index system can be an assessment framework composed of multiple dimensions of indicators used to quantify the health level of equipment. For example, the health status evaluation index system can design indicator weights and grading rules based on expert knowledge and data-driven methods to transform complex risk information into interpretable, graded health statuses. When establishing the health status evaluation index system, multi-dimensional evaluation indicators and their weights can be defined, and health level classification rules can be set. Specifically, the Analytic Hierarchy Process (AHP) can be used to determine the weights of each evaluation indicator, or fuzzy comprehensive evaluation can be used to handle uncertainty, thereby achieving a mapping from continuous risk values ​​to discrete health levels.

[0034] Cracking risk assessment is a comprehensive judgment process based on health risk mapping to determine the likelihood of leakage in water-cooled walls. It outputs a risk level to support decision-making. When assessing cracking risk based on equipment health risk mapping using a health status evaluation index system, the health risk mapping can be scored and graded according to the system. Equipment health status reports can be structured documents containing health levels, risk areas, trend predictions, and recommended measures, providing maintenance personnel with an actionable summary of the health status.

[0035] When obtaining equipment health status reports, structured reports can be generated by integrating crack risk assessment results, risk areas, trend predictions, and recommended measures. These reports can be in PDF format (text and images) or JSON format (documents) for system access, providing readable and traceable evidence for health diagnosis.

[0036] Step S500: Maintenance decision suggestions are generated based on the equipment health status report, and real-time anomaly warnings and maintenance strategy recommendations are pushed through edge computing. Specifically, maintenance decision suggestions can be targeted operation and maintenance intervention plans generated from the equipment health status report, used to guide maintenance priorities, adjust combustion strategies, or arrange downtime plans. When generating maintenance decision suggestions based on the equipment health status report, specific operation and maintenance actions can be output by matching a preset rule base or calling a strategy generation model. For example, when generating maintenance decision suggestions based on the equipment health status report, suggestions can be generated based on an if-then rule engine, or suggestions can be optimized using a reinforcement learning strategy network, thereby transforming the diagnostic results, i.e., the equipment health status report, into executable instructions. Edge computing can be a computing paradigm that performs data processing and intelligent inference on local devices close to the data source, used to reduce communication latency and achieve localized real-time warnings and strategy pushes. For example, edge computing can deploy lightweight models on industrial edge gateways or embedded controllers. Real-time anomaly warnings can be an alarm mechanism that is triggered immediately when a risk indicator is detected to exceed a threshold or when an increase in predicted risk is detected, to shorten the time from risk identification to alarm response. Maintenance strategy recommendation push can be achieved by sending the generated maintenance decision recommendations to the operation and maintenance terminal or control system through a communication interface to realize closed-loop intelligent operation and maintenance.

[0037] For example, in a scenario where combustion deviation leads to local overheating of the rear wall, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: During operation of a 600MW coal-fired unit, the flame center shifted to the rear wall area due to an imbalance in the primary air ratio. In this case, the temperature of the upper part of the rear wall was detected by a distributed temperature sensor array to be significantly higher than that of the front wall at the same elevation, and the three-dimensional displacement sensor simultaneously showed that the vertical expansion of this area was abnormally increased. After spatiotemporal correlation analysis of multi-source monitoring data, the temperature gradient and displacement covariance characteristics of the "rear wall-upper part" sub-region were extracted. Parameter threshold comparison revealed that the temperature exceeded the limit, and deviation analysis confirmed uneven expansion of the four walls. The thermal stress coupling model calculation showed that the additional thermal stress in this area had reached 90% of the material's allowable value. The risk feature mapping network correlated temperature, displacement, stress, and load, and the dynamic correlation prediction model, based on the LSTM-Attention architecture, predicted that the stress would continue to rise within the next 30 minutes. The equipment health risk map highlighted the upper part of the rear wall as a high-risk area, and the health status evaluation system assessed it as "medium risk," generating maintenance decision suggestions such as "adjusting the burnout air ratio and strengthening the rear wall ash blowing." The warning information was pushed to the central control room within 5 seconds through edge computing nodes, avoiding potential leakage accidents.

[0038] For example, multi-source monitoring data includes temperature signals, three-dimensional thermal displacement signals, stress signals, and unit load signals at different locations at the same altitude. Spatiotemporal correlation analysis is performed on the multi-source monitoring data to obtain a multi-dimensional monitoring feature set, including: extracting spatiotemporal distribution features from the temperature signals to obtain temperature field gradient features; performing three-dimensional coordinate transformation on the three-dimensional thermal displacement signals to obtain spatial deformation features; performing time-domain waveform analysis on the stress signals to obtain stress concentration features; extracting time-series trends from the unit load signals to obtain load fluctuation features; performing coupling analysis based on temperature field gradient features and load fluctuation features to obtain a thermal load correlation coefficient; correcting the spatial deformation features based on the thermal load correlation coefficient to obtain true thermal displacement features; and performing spatiotemporal registration based on true thermal displacement features, temperature field gradient features, stress concentration features, and load fluctuation features to construct a multi-dimensional monitoring feature set.

[0039] Specifically, temperature signals can be time-series data reflecting local thermal states collected by temperature sensors at specific locations on the boiler water-cooled wall. These signals are used to construct the temperature field distribution and identify localized overheating caused by uneven combustion. For example, temperature signals can be acquired in real-time using thermocouples, infrared sensors, or fiber optic grating sensors. Three-dimensional thermal displacement signals can be raw measurements describing the displacement changes of the water-cooled wall tubes in three spatial directions, used to reflect the spatial non-uniformity of thermal expansion behavior. For example, three-dimensional thermal displacement signals can be acquired using laser ranging, visual tracking, or eddy current sensors. Stress signals can be time-varying signals characterizing the internal mechanical response of the water-cooled wall tubes, used to identify whether the material is under high-risk stress. For example, stress signals can be obtained by indirectly measuring with strain gauges and then converting the measured stress, or by inversion using a thermo-mechanical model. Unit load signals can be time-series external operating parameters reflecting the current operating power level of the boiler, used as disturbance variables to analyze their modulation effect on the thermodynamic response. For example, unit load signals can be obtained by reading signals such as main steam flow, power generation, or fuel input rate from the DCS.

[0040] Different locations at the same height can be multiple monitoring areas distributed circumferentially along the furnace at the same elevation of the boiler water-cooled wall. These areas are used for lateral comparison of the thermal state differences of the wall in four directions: left, right, front, and rear. When acquiring multi-source monitoring data of the boiler water-cooled wall, temperature, displacement, stress, and global load signals of the wall in four directions can be collected simultaneously at the same elevation, thereby obtaining temperature signals, three-dimensional thermal displacement signals, stress signals, and unit load signals at different locations at the same height. Furthermore, this operation can achieve synchronous sampling at multiple points at the same height by deploying a ring sensor array, or ensure time alignment of multi-source data through a time synchronization protocol such as the Precision Time Protocol (PTP), thereby establishing a laterally comparable multi-source heterogeneous data foundation to support spatial non-uniformity analysis.

[0041] When extracting the spatiotemporal distribution features of temperature signals to obtain temperature field gradient features, the temperature difference between measuring points at different locations at the same elevation can be calculated, or the spatial gradient of the temperature field can be fitted. For example, this operation can use the finite difference method to calculate the circumferential gradient, or use radial basis function interpolation to construct a continuous temperature field and then take the derivative, thereby quantifying the heat distribution deviation caused by uneven combustion. The temperature field gradient feature can be a quantitative index reflecting the rate of change of spatial temperature, used to characterize the degree of uneven combustion and identify high-temperature gradient regions. For example, the temperature field gradient feature can be obtained by calculating the temperature difference between adjacent measuring points or fitting the spatial derivative of the temperature field.

[0042] When performing 3D coordinate transformation on a 3D thermal displacement signal to obtain spatial deformation features, the displacement vectors of each sensor in its local coordinate system can be transformed to the boiler's overall coordinate system. For example, this operation can be performed by aligning the coordinate systems using a rigid body transformation matrix, or by using a calibration plate to determine the sensor pose before performing the transformation, thereby achieving a consistent spatial representation of displacement data at different locations. Specifically, 3D coordinate transformation can be a mathematical transformation process that maps the original 3D thermal displacement signal from the sensor's local coordinate system to the boiler's unified global coordinate system, used to achieve a consistent spatial representation of displacement data at different locations. In this embodiment, 3D coordinate transformation is applied to the 3D thermal displacement signal, and the output is used to construct spatial deformation features. Spatial deformation features can be thermal displacement representations with a unified spatial reference system obtained after 3D coordinate transformation, used to support cross-regional deformation comparison and structural deformation pattern recognition. For example, spatial deformation features can be obtained by normalizing or decomposing the displacement vectors obtained from the 3D coordinate transformation.

[0043] When performing time-domain waveform analysis on stress signals to obtain stress concentration characteristics, the amplitude, duration, rate of change, and other time-domain properties of the stress signal can be analyzed. This operation can set thresholds to trigger event detection and statistically analyze the duration of exceeding limits, or extract higher-order statistics such as kurtosis and skewness of the stress waveform, thereby identifying abnormal stress states caused by limited expansion. Time-domain waveform analysis is the process of analyzing the waveform shape, amplitude, frequency, and other characteristics of stress signals in the time dimension, used to identify abnormal stress fluctuations or sustained high stress states. Stress concentration characteristics can be quantitative indicators extracted from time-domain waveform analysis that characterize abnormal increases in local stress, used to indicate the location of potential crack initiation. For example, stress concentration characteristics can be constructed based on parameters such as the peak value, duration, and rate of rise of the stress signal.

[0044] When extracting time-series trends from unit load signals to obtain load fluctuation characteristics, trend terms and fluctuation components can be separated from the original unit load signals. This operation can employ low-pass filters to extract slowly varying trends, or use Hilbert-Huang transforms to analyze non-stationary fluctuations, thereby characterizing the dynamic characteristics of external operating condition disturbances. Time-series trend extraction can be an operation that separates long-term trends, periodic fluctuations, or abrupt changes from unit load signals to characterize the dynamic impact of external operating conditions on the thermal system. Load fluctuation characteristics can be time-series features that characterize the amplitude, frequency, and direction of unit load changes, used to correlate and analyze the relationship between load disturbances and thermal response hysteresis. For example, load fluctuation characteristics can be obtained through differential analysis, spectral analysis, or rate of change calculations.

[0045] When performing coupled analysis based on temperature field gradient characteristics and load fluctuation characteristics to obtain the heat load correlation coefficient, the statistical dependence or physical response relationship between these two types of characteristics can be calculated. This operation can construct linear / nonlinear regression models to estimate the correlation strength, or calculate mutual information or Granger causality indices, thereby quantifying the modulating effect of load changes on heat distribution. The heat load correlation coefficient can be a dimensionless coefficient that quantifies the coupling strength between the temperature field gradient and load fluctuations, used to reveal the modulation mechanism of external load changes on internal heat distribution. The heat load correlation coefficient can be calculated through cross-correlation analysis, regression models, or information entropy methods.

[0046] When correcting spatial deformation characteristics based on heat load correlation coefficients to obtain true thermal displacement characteristics, the heat load correlation coefficients can be used to perform proportional correction or residual compensation on the spatial deformation characteristics. This operation can adjust the displacement amplitude by weighting according to the correlation coefficient, or subtract the non-thermal displacement component predicted by load fluctuations from the original displacement, thereby eliminating spurious displacements caused by non-thermal factors such as mechanical constraints and measurement drift. The true thermal displacement characteristics can be spatial deformation characteristics corrected by the heat load correlation coefficients, eliminating interference from non-thermal factors, and used to more accurately reflect the displacement caused by pure thermal expansion, improving the accuracy of subsequent modeling. The true thermal displacement characteristics are obtained by weighting or compensating the original spatial deformation characteristics using the heat load correlation coefficients. When constructing a multi-dimensional monitoring feature set based on spatiotemporal registration of true thermal displacement characteristics, temperature field gradient characteristics, stress concentration characteristics, and load fluctuation characteristics, these four types of features can be aligned and combined into a structured feature vector under a unified spatiotemporal grid. This operation can aggregate features and interpolate them to a unified spatial node within a fixed time window, or construct a graph structure with monitoring points as nodes and spatiotemporal adjacency as edges, thereby forming a high-dimensional, semantically rich, and physically consistent input dataset, i.e., a multidimensional monitoring feature set. Spatiotemporal registration can be the process of aligning multi-source heterogeneous features to a unified reference frame in the temporal and spatial dimensions, ensuring that multidimensional features can be fused and analyzed at the same spatiotemporal granularity.

[0047] Taking the example of false expansion signals caused by rapid load fluctuations, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: A 300MW coal-fired unit frequently scalds peak loads under Automatic Generation Control (AGC) mode, with the load dropping from 70% to 40% within 10 minutes. The original three-dimensional thermal displacement signal shows abnormal contraction in the middle of the front wall, suggesting obstructed expansion. However, the synchronously acquired temperature signal indicates that the temperature field gradient in this area is gentle, with no local cooling phenomenon. Under these circumstances, by extracting the time-series trend of the unit's load signal, it was found that the load decrease rate reached 8% / min. Coupled analysis of the temperature field gradient and load fluctuations yielded a thermal load correlation coefficient of -0.75, indicating that a rapid load decrease would trigger overall contraction. Based on this, the spatial deformation characteristics were corrected, and the global contraction component caused by the sudden load drop was eliminated. The obtained true thermal displacement characteristics show that the actual expansion behavior in the middle of the front wall is normal, and the stress concentration characteristics are not exceeded. The finally constructed multi-dimensional monitoring feature set eliminates the risk of false alarms and avoids unnecessary shutdown inspections.

[0048] For example, extracting the spatiotemporal distribution features of the temperature signal to obtain temperature field gradient features includes: spatially discretizing the water-cooled wall monitoring area using a finite element mesh generation method to obtain grid node temperature data; calculating the deviation between the temperature of each grid node in the grid node temperature data and the corresponding theoretical design value to construct a temperature deviation matrix; performing spatial interpolation based on the temperature deviation matrix to generate a temperature field distribution cloud map; and calculating the temperature difference and relative deviation ratio at different locations at the same height based on the temperature field distribution cloud map to construct a temperature gradient feature vector and obtain the temperature field gradient features.

[0049] The finite element method (FEM) mesh generation is a numerical modeling technique that divides a continuous water-cooled wall monitoring area into a finite number of discrete elements, i.e., a mesh. This provides a spatial topological framework for sparse temperature measurement points, supporting full-field temperature reconstruction and gradient calculation. For example, the FEM mesh generation method can use triangular or quadrilateral elements to structurally or unstructure the furnace walls based on geometry and sensor layout. The water-cooled wall monitoring area can be the physical area on the boiler furnace walls where temperature sensors are deployed, serving as the spatial range for temperature field modeling and analysis. Spatial discretization is a mathematical process that transforms a continuous spatial domain into a finite number of discrete nodes or elements, enabling the continuous temperature field to be numerically represented and calculated. Spatial discretization can be achieved through the FEM mesh generation method, outputting the temperature data of the mesh nodes.

[0050] Grid node temperature data can be a set of temperature values ​​obtained by interpolation or allocation on finite element grid nodes, serving as the basic data unit for constructing a structured temperature field. Grid node temperature data can map measured temperature signals to the nearest grid node or distribute them to multiple grid nodes using inverse distance weighting. When spatially discretizing the water-cooled wall monitoring area using the finite element mesh generation method to obtain grid node temperature data, a finite element mesh can be constructed based on the sensor layout and furnace geometry, and the measured temperature signals can be allocated to the corresponding grid nodes. This operation can be achieved by generating unstructured meshes and mapping temperatures using commercial finite element analysis software or open-source computational fluid dynamics software, or by performing nearest neighbor allocation based on regular quadrilateral meshes, thereby transforming sparse, unstructured measurement points into structured spatial data to support full-field modeling.

[0051] The theoretical design value corresponding to the temperature of each grid node can be the expected temperature value at each location of the water-cooled wall under ideal, uniform combustion and stable operating conditions. This value serves as a benchmark to evaluate actual temperature deviations and reflects the quality of combustion organization. This theoretical design value can be calibrated based on a boiler thermodynamic calculation model or historical steady-state operating data. The deviation between the temperature of each grid node and its corresponding theoretical design value can be the difference between these values. This difference is used to quantify the degree of local overheating or underheating and to indicate areas of uneven combustion. This deviation can be obtained by subtracting the corresponding design temperature from the temperature of each grid node point by point. The temperature deviation matrix can be a data structure that organizes the deviation values ​​corresponding to all grid nodes in matrix form, providing structured input for subsequent spatial interpolation and feature extraction. For example, the temperature deviation matrix can arrange the deviation values ​​according to the grid topology, forming a two-dimensional or three-dimensional array.

[0052] When constructing a temperature deviation matrix by calculating the deviation between the temperature of each grid node and its corresponding theoretical design value in the grid node temperature data, the measured temperature value can be subtracted from the theoretical design value for each grid node. The calculation result is then stored in the matrix according to the grid index to obtain the temperature deviation matrix. This operation can be achieved by dynamically updating the theoretical design value using a sliding window to adapt to varying loads, or by introducing a confidence interval to filter out spurious deviations caused by noise, thereby highlighting areas of abnormal combustion and separating the influence of operating conditions from equipment status.

[0053] Spatial interpolation is a mathematical method for reconstructing a continuous spatial field based on discrete node data, used to generate a smooth, fully covered temperature distribution image. A temperature field distribution cloud map can be an image that visualizes the spatial distribution of temperature on the surface of a water-cooled wall in a color-mapped format, used to visually display high / low temperature regions caused by uneven combustion. In this embodiment, the temperature field distribution cloud map can render the interpolated continuous temperature field as a pseudo-color image. When generating the temperature field distribution cloud map based on spatial interpolation using the temperature deviation matrix, an interpolation algorithm can be applied to fill the grid nodes with continuous temperature values ​​and then visualize the rendering. This operation can be achieved by using Gaussian process regression for probabilistic interpolation and outputting uncertainty, or by combining physical constraints such as energy conservation to correct the interpolation results, thereby achieving continuous and visualized global temperature distribution and revealing hidden uneven regions.

[0054] Temperature differences at different locations at the same height can be the absolute temperature difference between measuring points or areas at different locations within the same elevation, such as the left and right walls, directly reflecting the degree of circumferential thermal expansion incompatibility. Relative deviation ratios can be the ratio of the temperature difference to a reference temperature, such as the average temperature or theoretical design value, used to standardize temperature difference indicators and support comparisons across operating conditions. For example, the relative deviation ratio can be calculated as the temperature difference between two measuring points divided by the reference temperature to eliminate the influence of load levels. Temperature gradient feature vectors can be structured vectors composed of multiple temperature differences and relative deviation ratios, used as input features for machine learning models or risk assessment modules to characterize thermal nonuniformity. For example, temperature gradient feature vectors can organize the difference and ratio terms in a fixed order, such as left → right → front → back.

[0055] When calculating the temperature difference and relative deviation ratio at different locations at the same elevation to construct a temperature gradient feature vector and obtain the temperature field gradient characteristics, representative azimuth points can be selected at different orientations at the same elevation. The temperature deviation and relative deviation ratio between each pair of azimuth points can be calculated, and the calculation results can be combined into a temperature gradient feature vector to obtain the temperature gradient characteristics. This operation can be achieved by using principal component analysis to reduce dimensionality and retain the main gradient directions, or by introducing directional weights, such as higher weights for front and rear walls, to construct a weighted feature vector, thereby generating quantifiable, comparable, and input-into-model thermal nonuniformity characteristics.

[0056] Taking the deflection of the swirl burner causing localized overheating of the right wall as an example, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: In a subcritical boiler, due to the deflection of the primary air in the swirl burner, the flame adheres to the wall and scours the upper part of the right wall. Temperature sensors deployed in this area show a local temperature of 520°C, while the left wall at the same elevation is only 460°C. In this case, the finite element mesh generation method is used to divide the four walls into 20×30 meshes, and the measured temperature is mapped to the mesh nodes; the theoretical design value of the temperature signal is set to 480°C, and the temperature on the right wall is calculated. The deviation of some grid nodes reached +40℃, while the deviation of the upper grid nodes on the left wall was -20℃. Based on this, a temperature deviation matrix was constructed. A continuous temperature field cloud map was obtained by radial basis function interpolation, clearly showing the high-temperature patch on the right. Further calculation of the pairwise temperature difference between the four walls at the same elevation showed a maximum of 60℃, with a relative deviation ratio of 12.5%. Based on this, an 8-dimensional temperature gradient feature vector containing 4 absolute differences and 4 relative ratios was constructed and input into the thermal stress coupling model. This identified a significant risk of expansion restriction on the right wall, triggering a moderate warning. It was recommended to adjust the burner angle to avoid creep damage to the pipes.

[0057] For example, refer to Figure 2 The process involves comparing and analyzing parameter thresholds and deviations of the multidimensional monitoring feature set, constructing a thermal stress coupling model, analyzing parameter correlations, and obtaining a risk feature mapping network, including steps S201 to S207.

[0058] S201: The temperature, thermal displacement, and stress parameters in the multi-dimensional monitoring feature set are compared with their corresponding theoretical design values. The degree of parameter exceedance is then determined based on the comparison results. Specifically, the temperature parameter can be a numerical variable in the multi-dimensional monitoring feature set representing the local thermal state of the water-cooled wall, used to compare with the corresponding theoretical design value to determine if overheating has occurred. The thermal displacement parameter can be a feature variable in the multi-dimensional monitoring feature set describing the actual three-dimensional displacement of the water-cooled wall tubes after heating, used to reflect whether the expansion behavior deviates from the design expectations. The stress parameter can be a feature variable in the multi-dimensional monitoring feature set representing the intensity of the internal mechanical response of the water-cooled wall tubes, used to assess whether the structure is approaching the material strength limit. The theoretical design value can be the expected value of each physical parameter of the boiler water-cooled wall under ideal operating conditions, namely the temperature, thermal displacement, and stress parameters. The theoretical design value can be derived from boiler thermal calculation sheets, structural design specifications, or finite element simulation benchmarks, used as a reference benchmark to determine whether the measured parameters are abnormal.

[0059] When comparing temperature, thermal displacement, and stress parameters in a multi-dimensional monitoring feature set with their corresponding theoretical design values, the deviations between the measured characteristics and the corresponding theoretical design values ​​of these parameters can be calculated point by point to obtain the comparison results. This operation can be achieved by calculating the absolute difference to determine whether limits are exceeded or by calculating the relative error percentage to assess the degree of deviation, thereby identifying whether a single parameter exceeds the safety boundary. When obtaining the parameter exceedance index, the comparison results can be converted into a standardized exceedance score or index. This operation can be achieved by using 0-1 normalization to map the comparison results to the [0, 1] interval or by setting multi-level thresholds to generate discrete levels, thereby quantifying the severity of the anomaly and supporting subsequent model weighting.

[0060] S202: Calculate the differences in temperature parameters, thermal displacement parameters, and stress parameters at different locations at the same height to construct a lateral deviation characteristic spectrum. Different locations at the same height can be spatial positions of the boiler water-cooled wall at the same elevation, such as the left wall, right wall, front wall, and rear wall, serving as the spatial division basis for lateral deviation analysis. Temperature parameter differences can be pairwise differences between temperature parameters at different locations at the same height, used to quantify the non-uniformity of circumferential temperature distribution in the furnace. Thermal displacement parameter differences can be differences in thermal displacement parameters at different locations at the same height, used to reflect the spatial inconsistency of expansion behavior. Stress parameter differences can be differences in stress parameters at different locations at the same height, used to reveal localized stress concentrations caused by restricted expansion.

[0061] When calculating temperature, thermal displacement, and stress differences at different locations at the same height, pairwise or mean differences can be calculated for the corresponding parameters of the four walls (left, right, front, and rear walls) within a fixed elevation. This operation can be achieved by calculating the maximum-minimum difference as a non-uniformity index or by constructing a fully connected difference matrix to preserve the complete topology, thereby capturing circumferential non-uniformity and revealing combustion skew effects. When constructing the lateral deviation characteristic spectrum, the difference results corresponding to each elevation can be stacked by height to form a feature sequence. This operation can be achieved by storing the deviation at each elevation in vector form or by visualizing the height-circumferential deviation as an image, thus forming a deviation profile distributed along the furnace height.

[0062] S203: Dimensionless processing of the lateral deviation characteristic spectrum yields a relative deviation proportionality coefficient. Specifically, the dimensionless processing of the lateral deviation characteristic spectrum involves converting deviations with different physical units into a unified relative proportion. This operation can be achieved by normalization through division by their respective theoretical design values ​​or by Z-score standardization to eliminate dimensional differences and enable the fusion and comparison of multiple physical quantities. When obtaining the relative deviation proportionality coefficient, the dimensionless deviation value can be used as a new feature. The relative deviation proportionality coefficient can be the dimensionless lateral deviation characteristic value, representing the relative magnitude of the deviation relative to the design reference, used as input to the thermal stress coupling model, supporting multi-parameter fusion modeling.

[0063] S204: Construct a thermal stress coupling model based on the parameter exceedance index and the relative deviation ratio coefficient.

[0064] Specifically, when constructing a thermo-stress coupling model based on parameter exceedance indices and relative deviation ratios, these two indices can be used as input variables, combined with the thermoelastic equation to establish a stress prediction relationship. This operation can be achieved by constructing a linear weighted combination model or training a neural network surrogate model to fit a nonlinear relationship, thereby integrating both absolute exceedance and relative non-uniformity risk information.

[0065] S205: Perform multivariate correlation analysis on temperature field gradient characteristics, actual thermal displacement characteristics, and unit load characteristics to obtain a cross-parameter correlation matrix. The temperature field gradient characteristics can be the rate of change of temperature in space extracted from a multidimensional monitoring feature set, used to characterize the intensity of flame deflection or wall-attached combustion. The actual thermal displacement characteristics can be the feature representation obtained by correcting the spatial deformation characteristics corresponding to the actual three-dimensional thermal displacement signal based on the heat load correlation coefficient, used to distinguish it from theoretical expansion and reflect the actual structural response. The unit load characteristics can be the time-series or steady-state characteristics related to the thermodynamic response extracted from the unit load signal, used as external disturbance variables in the multivariate correlation analysis.

[0066] When performing multivariate correlation analysis on temperature field gradient characteristics, actual thermal displacement characteristics, and unit load characteristics, joint analysis of statistical dependencies among three or more variables can be conducted. This operation can be achieved through partial correlation analysis, canonical correlation analysis, or multi-element mutual information calculation, thereby revealing the coupling law between temperature gradient, displacement response, and load change. The cross-parameter correlation matrix can be the result of correlation coefficients between temperature field gradient characteristics, actual thermal displacement characteristics, and unit load characteristics stored in matrix form. The cross-parameter correlation matrix can be used to quantify the linear or nonlinear correlation strength between multi-source heterogeneous characteristics. For example, the cross-parameter correlation matrix may include, but is not limited to, linear correlation coefficient matrices, rank correlation matrices, and information entropy correlation matrices.

[0067] S206: Identify key influencing factors based on cross-parameter correlation matrices and construct parameter correlation networks.

[0068] Specifically, when identifying key influencing factors based on cross-parameter correlation matrices, features with correlation weights higher than a threshold or ranking high can be selected as key factors. This operation can be achieved by setting a fixed threshold for screening or by using principal component loadings to identify important variables, thereby reducing dimensionality and focusing on dominant risk variables. Key influencing factors can be feature variables with high correlation weights in the cross-parameter correlation matrix that play a dominant role in risk evolution, used to simplify model complexity and focus on core risk drivers. For example, key influencing factors may include, but are not limited to, the dominant temperature gradient direction, the maximum displacement response region, and load-sensitive frequency bands.

[0069] When constructing parameter correlation networks, directed or undirected graph models can be built using key influencing factors as nodes and their correlation strength as edges. This operation can be achieved by constructing an undirected weighted graph to represent correlation or a directed graph to represent causal direction, thereby visualizing and structurally expressing the causal or correlation paths between multiple parameters.

[0070] S207: Risk feature extraction is performed based on the parameter association network and the thermal stress coupling model to obtain a risk feature mapping network. Specifically, when extracting risk features based on the parameter association network and the thermal stress coupling model, highly discriminative risk representations can be extracted by fusing the output of the thermal stress coupling model with those from the parameter association network. This operation can be achieved by concatenating the model output with the network embedding vector or by aggregating coupling information through a graph neural network, thereby enhancing the physical interpretability and data adaptability of the risk features. When obtaining the risk feature mapping network, the extracted risk features can be organized into a node-edge structure to identify risk propagation paths, forming an interpretable and predictable risk knowledge graph.

[0071] For example, in a scenario where localized overheating occurs due to flames adhering to the front wall under low-load conditions, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: A 300MW coal-fired unit, operating at 40% load, experienced flame adhesion to the front wall due to improper secondary air distribution. The temperature parameter of the upper part of the front wall, extracted from the multi-dimensional monitoring feature set, was 620℃, exceeding the theoretical design value of 580℃ under this load, resulting in a temperature exceedance index of +6.9%. Simultaneously, the temperature difference between the four walls at the same elevation was calculated, with the front wall 85℃ higher than the rear wall. A lateral deviation feature spectrum was constructed, showing a significant bulge in the front wall area. After dimensionless processing, the relative deviation ratio coefficient was obtained as follows: 0.147; The thermal stress coupling model, combining the relative deviation proportionality coefficient and the over-limit index, outputs that the additional thermal stress on the upper part of the front wall reaches 185 MPa; On the other hand, multivariate correlation analysis found that the temperature field gradient characteristics of the front wall and the unit load characteristics are highly correlated with each other in the low-frequency range (r = 0.82). The cross-parameter correlation matrix identifies "low load - strong gradient of the front wall" as the key influencing factor, and the parameter correlation network reveals that there is a strong path between it and the displacement response of the front wall; Finally, by extracting risk features and integrating the above information, a risk feature mapping network is generated, which clearly marks the failure chain of "low load wall-attached combustion → front wall overheating → expansion restriction → stress concentration", supporting subsequent early warning and combustion adjustment.

[0072] For example, the lateral deviation characteristic spectrum is dimensionless to obtain the relative deviation proportion coefficient, including: setting a deviation threshold range, normalizing the lateral deviation characteristic spectrum to obtain standardized deviation values; using the analytic hierarchy process to determine the weighting coefficients of temperature parameters, thermal displacement parameters, and stress parameters respectively; constructing a weighted deviation matrix based on the standardized deviation values ​​and weighting coefficients; and calculating the relative deviation proportion of each element in the weighted deviation matrix with the corresponding theoretical design value to obtain the relative deviation proportion coefficient.

[0073] The deviation threshold interval can be a pre-defined normalized mapping range for different physical parameters, used to compress the original deviation values ​​into a unified dimensionless interval. For example, the deviation threshold interval can be set based on historical operating data statistics or design specifications to define the maximum acceptable deviation range for each parameter. Normalization can be a mathematical transformation that linearly or non-linearly maps the original differences in the transverse deviation characteristic spectrum to [0, 1] or other standard intervals according to the deviation threshold interval. The standardized deviation value can be a dimensionless deviation value obtained after normalization, reflecting the relative position of the actual deviation within the allowable range. The analytic hierarchy process (AHP) can be a multi-criteria decision-making method that transforms expert experience and qualitative judgment into quantitative weights. For example, the AHP can determine the weights by constructing a judgment matrix, calculating eigenvectors, and performing consistency checks. The AHP can also invite experts from boiler design, operation, and maintenance to jointly score, or combine the results of fault tree analysis (FTA) to assist in constructing a judgment matrix, thereby assigning weight coefficients that conform to the failure mechanism to the three types of parameters: temperature, thermal displacement, and stress.

[0074] The weighting coefficients for temperature parameters can be numerical values ​​assigned to the relative importance of temperature deviations in the overall risk assessment using the analytic hierarchy process (AHP), reflecting the leading role of local overheating as an early sign. Similarly, the weighting coefficients for thermal displacement parameters can be numerical values ​​assigned to the relative importance of thermal displacement deviations in the risk assessment using the AHP, reflecting the bridging role of expansion imbalance on structural constraints and stress accumulation. The weighting coefficients for stress parameters can be numerical values ​​assigned to the relative importance of stress deviations in the risk assessment using the AHP, emphasizing the ultimate impact of stress on whether the material cracks. The weighted deviation matrix can be a matrix constructed by multiplying the standardized deviation values ​​element-wise with the corresponding weighting coefficients for temperature, thermal displacement, and stress parameters. It represents the comprehensive deviation state with importance weights, used to integrate deviation magnitude and parameter importance, improving the accuracy of risk characterization. The theoretical design values ​​corresponding to each element in the weighted deviation matrix can be the expected baseline values ​​for each physical quantity (temperature, displacement, and stress) of the boiler water-cooled wall under ideal uniform heating conditions.

[0075] When setting deviation threshold ranges, maximum allowable deviation ranges can be specified for temperature difference, thermal displacement difference, and stress difference as upper and lower limits for normalization. For example, when setting deviation threshold ranges, stress deviation threshold ranges can be set according to ASME standards, or temperature deviation critical ranges can be deduced based on historical non-shutdown events, thereby ensuring comparability and engineering rationality of different physical quantities after normalization. When normalizing the lateral deviation characteristic spectrum, the differences of each parameter in the lateral deviation characteristic spectrum can be linearly scaled according to their corresponding deviation threshold ranges, and the output normalized numerical result is the standardized deviation value. The standardized deviation value serves as the input for subsequent weighted calculations, providing a multi-parameter deviation representation under a unified scale.

[0076] When using the analytic hierarchy process (AHP) to determine the weighting coefficients for temperature, thermal displacement, and stress parameters, experts can be organized to compare the relative importance of these three parameters in the water-cooled wall cracking process pairwise, construct a judgment matrix, and solve for the weights. For example, when determining the weighting coefficients for temperature, thermal displacement, and stress parameters using AHP, experts from boiler design, operation, and maintenance can be invited to jointly score the parameters, or the judgment matrix can be constructed using fault tree analysis results. This introduces engineering mechanism knowledge and avoids misjudgments caused by equal-weighted averaging. When constructing a weighted deviation matrix based on standardized deviation values ​​and weighting coefficients, the standardized deviation value of each parameter can be multiplied by its corresponding weighting coefficient, organized into a matrix according to spatial location. For example, the weighted deviation matrix can be constructed by multiplying elements-by-element to generate a weighted matrix of the same dimension, or by grouping by elevation to construct a block diagonal weighted structure, thus allowing the deviation of high-risk parameters such as stress parameters to have a greater weight in the comprehensive evaluation.

[0077] When calculating the relative deviation ratio of each element in the weighted deviation matrix to its corresponding theoretical design value, the percentage deviation of each element relative to its corresponding theoretical design value can be calculated. Specifically, when calculating the relative deviation ratio of each element in the weighted deviation matrix to its corresponding theoretical design value, if the corresponding theoretical design value is zero, the mean of nearby elevations is used as a substitute, or the logarithmic deviation is used to reduce the influence of extreme values, thereby generating a physically meaningful proportionality coefficient for easy comparison across working conditions. The final obtained relative deviation proportionality coefficient can be a dimensionless, weighted, and relativized set of deviation coefficients in the final output. The relative deviation proportionality coefficient can serve as a key input to the thermal stress coupling model to improve the model's sensitivity and accuracy to real risks.

[0078] Taking the compound deviation caused by the right-side wall burner tilt under high load as an example, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: When a 1000MW ultra-supercritical coal-fired unit is running at 90% load, the primary air velocity of the right-side wall burner is too low, causing the flame to adhere to the wall. Calculations show that the temperature difference between the four walls at the same elevation reaches 95℃, the thermal displacement difference is 12mm, and the stress difference is 45MPa. The deviation threshold ranges for temperature, thermal displacement, and stress parameters are set to [0, 120℃], [0, 15mm], and [0, 60MPa], respectively. After normalization, the standardized deviation values ​​for temperature, thermal displacement, and stress parameters are 0.79, 0.80, and 0.75, respectively. Through the analytic hierarchy process (AHP), experts determined that the stress parameter has the greatest direct impact on cracking and assigned it a weighting coefficient of [missing value]. The temperature parameter, due to its early warning value, is assigned a weighting coefficient of 0.3, and the thermal displacement parameter, as an intermediate variable, is assigned a weighting coefficient of 0.2. A weighted deviation matrix is ​​constructed by multiplying the standardized deviation values ​​of the temperature, thermal displacement, and stress parameters by their corresponding weighting coefficients, resulting in values ​​of 0.237, 0.160, and 0.375 for each element. Each element in this weighted deviation matrix is ​​then compared with its corresponding theoretical design values ​​(temperature difference 0, displacement difference 0, and stress difference 0) to calculate the relative deviation proportionality coefficient. Since the theoretical design value of the stress term is 0, the historical average of 5 MPa is used as the benchmark, resulting in a relative deviation proportionality coefficient of 7.5. This high coefficient, the relative deviation proportionality coefficient of 7.5, is identified as a high-risk signal by the thermal-stress coupling model, triggering an early warning.

[0079] For example, refer to Figure 3 A dynamic correlation prediction model is constructed, and the risk feature mapping network is used to predict thermal displacement trends and assess stress concentration based on the dynamic correlation prediction model to obtain the equipment health risk map, including steps S301 to S306.

[0080] S301: Constructing a Long Short-Term Memory (LSTM) network model and an attention mechanism model. The LSTM network model is a recurrent neural network with forget gates, input gates, and output gates, used to model time-series data with long-term dependencies. In this case, it's used to learn the long-term dynamic patterns in the evolution of thermal displacement and stress history of water-cooled walls. The attention mechanism model is a dynamic weighting mechanism that assigns different weights based on the importance of each time step of the input sequence to the current prediction task. Here, it's used to enhance the model's sensitivity to critical anomalous periods such as sudden load changes or local overheating, improving prediction accuracy. When constructing the LSTM network model and the attention mechanism model, LSTM network structures and attention weight calculation modules can be designed separately. This can be achieved by building LSTM layers using PyTorch and customizing the attention module, or by combining the built-in LSTM and attention layers in TensorFlow Keras, thus providing the basic components for subsequent fusion.

[0081] S302: Integrate the Long Short-Term Memory (LSTM) network model with the attention mechanism model to construct an initial association prediction model. This initial association prediction model is an untrained prediction architecture formed by the preliminary fusion of the LSTM network model and the attention mechanism model, serving as the trainable foundation for the dynamic association prediction model. For example, the initial association prediction model can be one or more of the following: a serial LSTM-Attention structure, a parallel LSTM-Attention fusion structure, or a multi-head attention-enhanced LSTM. The initial association prediction model can use the hidden states of the LSTM as input to the attention mechanism, calculating a context vector for the final prediction output. When fusing the LSTM network model with the attention mechanism model to construct the initial association prediction model, the hidden state sequence output by the LSTM can be input into the attention mechanism to generate a weighted context vector for the final prediction. For example, this operation can be achieved by appending a Bahdanau attention layer after the LSTM, or by using Transformer-style multi-head attention instead of traditional attention, thus forming a joint architecture with temporal modeling and key segment focusing capabilities.

[0082] S303: Train the initial correlation prediction model using historical monitoring data of the boiler water-cooled wall to obtain a dynamic correlation prediction model. The historical monitoring data consists of multi-source monitoring data of the boiler water-cooled wall collected over a past period, including samples from normal and abnormal operating conditions. This data is used to train the initial correlation prediction model, enabling it to have generalization prediction capabilities. For example, the historical monitoring data can be a cleaned and labeled time-series dataset extracted from a historical database or data lake of the boiler water-cooled wall system.

[0083] When training the initial correlation prediction model using historical monitoring data to obtain the dynamic correlation prediction model, the historical time-series features corresponding to the historical monitoring data can be used as input, and the corresponding actual thermal displacement and stress can be used as labels. The model parameters can be optimized through backpropagation. Furthermore, this operation can be achieved by using the mean squared error loss function for end-to-end training, or by introducing a physical constraint loss term to enhance the physical consistency of the model, thereby enabling the dynamic correlation prediction model to predict the thermo-mechanical coupling behavior under actual working conditions.

[0084] S304: A sliding time window method is used to extract temporal features from the risk feature mapping network, and the extracted features are input into a dynamic association prediction model to obtain the thermal displacement trend prediction results. The sliding time window method is a technique that divides continuous time-series data into fixed-length windows and slides them over time to generate sample sequences. Here, it is used to ensure that the data input to the dynamic association prediction model has contextual continuity and temporal structure. For example, the sliding time window method can include, but is not limited to, one or more of fixed-step sliding windows, overlapping sliding windows, and adaptive-length sliding windows. The temporal features can be a sequence of dynamic features extracted from the risk feature mapping network in chronological order, used as input to the dynamic association prediction model to characterize the evolution of the risk state. For example, temporal features can be obtained by sampling the output of the risk feature mapping network at different times using a sliding time window.

[0085] When using the sliding time window method to extract temporal features from a risk feature mapping network, continuous segments can be extracted from the network's temporal output with a fixed window length and step size. For example, this operation can be achieved by using a 5-minute window and a 1-minute step size to extract temporal features from the risk feature mapping network, or by dynamically adjusting the window length according to the rate of change of operating conditions, thereby generating structured temporal samples that meet the input requirements of a dynamic correlation prediction model. When inputting the extracted features into the dynamic correlation prediction model to obtain thermal displacement trend prediction results, the extracted temporal features can be fed into a trained dynamic correlation prediction model to obtain the future multi-step thermal displacement prediction values ​​output by the model. For example, this operation can be achieved by predicting the displacement over the next 10 time steps, or by outputting a probability prediction with a confidence interval, thus enabling a forward-looking estimate of the water-cooled wall expansion behavior. The thermal displacement trend prediction results can be three-dimensional thermal displacement prediction values ​​for several future time steps output by the dynamic correlation prediction model, reflecting the future trend of the water-cooled wall expansion behavior and used to deduce stress evolution.

[0086] S305: Construct a stress development trend curve based on the thermal displacement trend prediction results and stress concentration assessment results. The stress concentration assessment results can be quantitative indicators representing locally high-stress areas extracted from the current risk feature mapping network, used to combine with the thermal displacement trend to construct the stress development trend curve. The stress development trend curve is the future stress evolution trajectory synthesized based on the thermal displacement trend prediction results and stress concentration assessment results, used to visually demonstrate whether the stress in key areas tends to exceed limits. The stress development trend curve can be obtained by converting the predicted thermal displacement into stress using a thermal-stress coupling model, and then correcting it by superimposing the current stress concentration assessment results. When constructing the stress development trend curve based on the thermal displacement trend prediction results and stress concentration assessment results, the predicted thermal displacement can be substituted into the thermal-stress coupling model to calculate the stress, and the calculated stress can be corrected by integrating the current stress concentration assessment results. The stress development trend curve is then constructed based on the corrected stress. The stress development trend curve can be generated by calculating the stress value at each prediction time point by point, or by using interpolation smoothing to generate a continuous curve, thus generating a continuous curve reflecting future stress evolution.

[0087] S306: Compare and analyze the stress development trend curve with a preset safety threshold to obtain an equipment health risk map. The preset safety threshold can be the allowable stress of the material, the upper limit of stress or displacement set by the operating procedures, and is used as a baseline for judging the equipment health risk level. When comparing and analyzing the stress development trend curve with the preset safety threshold to obtain the equipment health risk map, it can be determined whether the stress development trend curve exceeds or approaches the preset safety threshold, and the risk level can be marked on the water-cooled wall geometric model. For example, if the stress in a certain area of ​​the water-cooled wall exceeds the limit within the next 30 minutes, that area can be marked as high-risk, or a heat map can be generated based on the degree of exceedance, thus transforming the abstract prediction result into a visualized and operable risk distribution map, i.e., an equipment health risk map.

[0088] For example, in scenarios where rapid load increases or decreases cause thermal stress fluctuations, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: During peak-shaving operation, the load of a coal-fired unit rapidly decreases from 80% to 40%. Through a risk feature mapping network, it is identified that the temperature in the middle region of the front wall drops sharply but the displacement lags behind. The time series features of the most recent 30 minutes are extracted by a sliding time window and input into a trained dynamic correlation prediction model. The LSTM part of the dynamic correlation prediction model captures the stress rebound phenomenon under similar load changes in the region in the past. The attention mechanism highlights the first 5 minutes of the load decrease as the critical window. The dynamic correlation prediction model outputs that the thermal displacement will continue to shrink in the next 20 minutes. Combined with the current stress concentration assessment results, the constructed stress development trend curve shows that the region will experience a tensile stress peak in 15 minutes, approaching the allowable limit of the material. The stress development trend curve is compared with the preset safety threshold, and the middle region of the front wall is marked as "high risk" on the equipment health risk map, triggering an early warning and suggesting that further load reduction be postponed to avoid the initiation of thermal fatigue cracks.

[0089] For example, the health status evaluation index system includes parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators. Based on the health status evaluation index system, a crack risk assessment is performed on the equipment health risk map to obtain an equipment health status report. This includes: quantifying the equipment health risk map based on the health status evaluation index system to obtain equipment health indicators; standardizing the equipment health indicators to obtain corresponding normalized health indices; weighting and fusing the normalized health indices to obtain a comprehensive risk score; classifying the equipment crack risk level based on the comprehensive risk score; performing sensitivity analysis on the normalized health indices to obtain corresponding risk contribution indicators; and generating an equipment health status report based on the equipment crack risk level and risk contribution indicators.

[0090] Specifically, the health status evaluation index system is an assessment framework for quantifying the health status of water-cooled walls, consisting of four dimensions: parameter exceedance, lateral deviation, trend change, and coupled risk. It includes corresponding parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators, used to structure the multidimensional risk information in the equipment health risk map into a calculable and comparable set of indicators. Parameter exceedance indicators are quantitative indicators reflecting whether monitored parameters exceed safety thresholds, used to identify explicit exceedance risks, such as local temperature or stress exceeding the allowable value of the material. Lateral deviation indicators characterize the degree of parameter difference between different areas at the same elevation, such as between four walls, used to quantify expansion imbalance or uneven temperature distribution caused by uneven combustion. Trend change indicators describe the dynamic characteristics of key parameters over time, used to capture early deterioration signals such as accelerated thermal displacement expansion or rapid stress accumulation. Coupled risk indicators reflect the degree of composite risk when multiple risk factors coexist, such as high temperature, high stress, and large displacement, used to identify high-risk conditions under nonlinear superposition effects.

[0091] When constructing a health status evaluation index system, the physical meaning, calculation logic, and applicable scope of four types of indicators—parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators—can be defined separately. For example, this operation can be achieved by setting indicator formulas based on expert experience, or by optimizing the indicator structure through inversion of historical fault data, thereby constructing a comprehensive evaluation dimension covering static exceedances, spatial unevenness, dynamic evolution, and multi-factor coupling. When quantifying equipment health risk maps based on the health status evaluation index system to obtain equipment health indicators, the spatial-temporal risk data in the equipment health risk map can be mapped to the calculation models of the four types of indicators—parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators—to obtain the corresponding equipment health indicators. For example, this operation can be achieved by independently calculating the four types of indicators—parameter exceedance indicators, lateral deviation indicators, trend change indicators, and coupled risk indicators—for each water-cooled wall region and then aggregating them, or by directly extracting statistical features as indicator values ​​at the global level, thereby realizing the conversion from graphical risk representation to numerical indicators.

[0092] Equipment health indicators are raw scoring vectors obtained by quantifying the equipment health risk map based on a health status evaluation index system. These vectors serve as intermediate assessment results before standardization. When standardizing equipment health indicators, mathematical transformations can be applied to map indicators with different dimensions to a unified interval, such as 0–1, to obtain the corresponding normalized health index. The normalized health index is a unified scale value obtained after dimensionless processing of the equipment health indicators, used to eliminate differences in dimensions and orders of magnitude between indicators and support subsequent fusion calculations. When weighted fusion of the normalized health indices to obtain a comprehensive risk score, weights can be assigned according to the importance of the equipment health indicators and combined linearly or non-linearly. This operation can be achieved by using fixed weights, such as expert scoring, for weighted summation. The comprehensive risk score can be a single risk metric generated by weighted fusion of the normalized health indices according to their corresponding weights, providing a quantitative representation of the overall health status and supporting risk level classification. When classifying equipment cracking risk levels based on the comprehensive risk score, continuous comprehensive risk scores can be mapped to preset discrete risk level intervals. The risk level of equipment cracking can be a discrete risk level based on a comprehensive risk score, which helps maintenance personnel quickly understand the current risk status of the equipment.

[0093] When conducting sensitivity analysis on the normalized health index, the marginal impact or contribution ratio of the normalized health index to the comprehensive risk score can be calculated, thereby obtaining the corresponding risk contribution index. Sensitivity analysis can be a quantitative method to assess the degree of influence of each normalized health index on the comprehensive risk score, used to identify dominant risk factors and support root cause localization. The risk contribution index can be a representation of the relative contribution ratio of each health dimension, such as lateral deviation and trend change, to the overall risk, used to reveal key risk sources and guide targeted intervention measures. When generating an equipment health status report based on the equipment cracking risk level and risk contribution index, the risk level and the risk contribution of each dimension can be integrated into structured text or visual charts to obtain the equipment health status report. For example, this operation can be achieved by generating a PDF report containing a risk map, indicator radar chart, and recommendation list. The equipment health status report can be a structured diagnostic document generated by integrating the equipment cracking risk level and risk contribution index, used to provide interpretable and traceable basis for operation and maintenance decisions.

[0094] For example, taking the high risk in the middle of the front wall caused by multi-factor coupling as an example, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: During the operation of a coal-fired unit, the equipment health risk map shows that there is a moderate temperature exceeding the limit in the middle of the front wall, a lateral expansion deviation that is significantly greater than that of the other three walls, and the thermal displacement shows an accelerating upward trend. The parameters in the health status evaluation index system are calculated to be: 0.65 for the parameter exceeding the limit, 0.82 for the lateral deviation, 0.78 for the trend change, and 0.90 for the coupling risk. After Min-Max standardization, the normalized health index is obtained. The health index, using expert weights (i.e., the weights of each indicator in the health status evaluation index system are 0.2, 0.3, 0.25, and 0.25 respectively), is weighted and fused with the normalized health index to obtain a comprehensive risk score of 0.79, which is classified as "high risk". Sensitivity analysis shows that the contribution of coupled risk indicators reaches 42%, indicating that the coexistence of high temperature, large displacement and high stress is the main cause. Based on this, an equipment health status report is generated, clearly marking "high risk in the middle of the front wall, mainly due to the coupling of multiple factors", and recommending "immediate adjustment of the burner angle, strengthening of soot blowing in this area, and arranging a shutdown inspection within 72 hours".

[0095] For example, a weighted fusion of normalized health indices to obtain a comprehensive risk score includes: constructing a risk assessment matrix to map the normalized health indices to the corresponding risk probabilities; using a fuzzy comprehensive evaluation method to calculate the membership degree of the risk probabilities to obtain the corresponding fuzzy risk vector; determining the index weights of the fuzzy risk vector based on the risk contribution index; using the index weights to perform a weighted summation of the fuzzy risk vector; and mapping the weighted result to the scoring interval [0, 100] to obtain the comprehensive risk score.

[0096] The risk assessment matrix can be a pre-defined transformation table or function that maps a normalized health index to its corresponding risk probability. This transforms the dimensionless health index into a probability value with engineering significance, facilitating subsequent fuzzy processing. For example, the risk assessment matrix can be constructed based on historical failure data statistics or expert experience, establishing a nonlinear mapping between the input normalized health index and the output risk probability. The risk probability can be a numerical value representing the likelihood of failure or anomaly corresponding to a certain health indicator, used as input for fuzzy comprehensive evaluation to reflect the potential risk level of each dimension. The risk probability can be obtained by looking up a table or calculating a function from the normalized health index using the risk assessment matrix. The fuzzy comprehensive evaluation method is a multi-index comprehensive evaluation method that utilizes fuzzy mathematics theory to handle uncertainties and boundary fuzzy problems. It overcomes the discontinuity of traditional hard threshold division in the state transition region and improves the robustness of the evaluation.

[0097] Membership degree calculation maps risk probabilities to multiple risk levels, such as low, medium, and high, based on a pre-defined membership function. This quantifies the fuzzy classification of indicators at different risk levels, resulting in a soft classification. A fuzzy risk vector, composed of membership degrees corresponding to each risk level, represents the distribution of the current state in the fuzzy risk space and serves as the basis for weighted fusion, preserving the uncertainty of risk information. For example, a fuzzy risk vector can be formed by concatenating the membership degrees calculated for the risk probabilities of each health dimension.

[0098] When constructing a risk assessment matrix and mapping normalized health indices to corresponding risk probabilities, a corresponding index-probability mapping relationship can be designed for each type of normalized health index, such as lateral deviation and trend changes. The risk probability corresponding to each type of normalized health index is then determined based on this index-probability mapping relationship. When using the fuzzy comprehensive evaluation method to calculate the membership degree of risk probabilities, for each risk probability, its membership degree at each level can be calculated based on a preset low / medium / high risk level membership function, and then these membership degrees are combined to form a fuzzy risk vector. The indicator weight is a coefficient used to weight and fuse the components of each fuzzy risk vector, reflecting the importance of different health dimensions to the overall risk, and is used to give high-sensitivity or high-contribution indicators a greater weight in the comprehensive score. This indicator weight can be dynamically determined based on the risk contribution indicators obtained in the preceding steps, rather than being fixedly allocated.

[0099] When determining the weights of a fuzzy risk vector based on its risk contribution index and performing a weighted summation, the risk contribution index can be normalized and used as the corresponding weight. This weight can then be used to weight and synthesize the components of each dimension in the fuzzy risk vector. For example, this operation can be achieved by directly using the Shapley value as the weight or by non-linearly amplifying the risk contribution index before normalizing and assigning weights, thereby achieving adaptive strengthening of risk-sensitive dimensions and improving the diagnostic value of the score. When mapping the weighted result to the [0, 100] scoring range, the fuzzy risk value after weighted fusion can be defuzzified, such as by using the centroid method and linearly scaling it to the 0–100 range, to obtain a comprehensive risk score. Furthermore, this operation can be achieved by using the maximum membership principle for defuzzification and mapping or by using a weighted average method for defuzzification followed by linear transformation, thereby generating an intuitive, comparable final score that supports threshold alarms.

[0100] For example, generating maintenance decision recommendations based on equipment health status reports includes: identifying high-risk areas and key influencing parameters based on the equipment health status report, matching them with a historical cracking case library, and inferring possible cracking modes; assessing maintenance urgency based on equipment cracking risk levels and thermal displacement trend prediction results, and generating maintenance priorities; formulating maintenance window recommendations based on maintenance priorities and unit operation plans; generating targeted maintenance plans based on inferred possible cracking modes and risk contribution indicators; and integrating maintenance window recommendations and targeted maintenance plans into maintenance decision recommendations.

[0101] High-risk areas can be localized locations in water-cooled walls marked in equipment health reports as having a high probability of cracking. These areas serve as target areas for maintenance resources, guiding targeted interventions. Key influencing parameters can be monitoring variables or characteristic indicators that play a dominant role in the current high-risk state, revealing the causes of risk and supporting the development of targeted measures. The historical cracking case library can be a structured database storing past water-cooled wall leakage events, including information such as failure location, operating conditions, crack morphology, and root cause. This provides a basis for analogical reasoning about the current risk state and assists in inferring potential failure modes. Furthermore, the historical cracking case library can be constructed through a fault archiving system, maintenance records, and expert annotations, and is continuously updated. For example, the historical cracking case library can include a set of material creep cracking cases, a set of thermal fatigue cracking cases, and a set of failure cases caused by the combined effects of corrosion and stress. Crack modes can be classification tags describing the failure mechanism and crack propagation characteristics of water-cooled wall tubes, used to guide the selection of maintenance schemes and distinguish situations requiring replacement, repair, or adjustment of operating strategies.

[0102] When identifying high-risk areas and key influencing parameters based on equipment health status reports, the risk map and indicator decomposition results in the reports can be analyzed to extract spatial location and dominant features. For example, this operation can extract the corresponding coordinates of the "high-risk" label in the equipment health status report through rule matching, or it can be achieved by parsing unstructured diagnostic text using text parsing technology, thereby locating the source of the problem and providing input for subsequent matching and decision-making. When matching against a historical cracking case library, the parameter feature vector of the current high-risk area can be compared with the failure features in the historical cracking case library to infer possible cracking patterns. This operation can be achieved by using a similar case retrieval algorithm to find the most similar historical cases, or by using a semantic similarity calculation model to calculate semantic similarity, thereby realizing a mapping from phenomenon to mechanism and improving the depth of diagnosis.

[0103] The equipment cracking risk level is a discrete risk severity indicator output from the health status evaluation index system, used as one of the inputs for maintenance urgency assessment. The thermal displacement trend prediction result can be a three-dimensional thermal deformation evolution sequence over a future period, output by a dynamic correlation prediction model, used to reflect whether the risk is worsening, thereby determining the intervention window. The thermal displacement trend prediction result can be used in conjunction with the equipment cracking risk level in the maintenance urgency assessment. The maintenance urgency assessment can be a quantitative judgment of whether immediate intervention is needed, based on a comprehensive risk level and trend prediction, used to avoid over-maintenance or delayed response, achieving on-demand maintenance. Maintenance priority can be a maintenance task ranking identifier generated based on the urgency assessment, used to guide resource allocation across multiple regions or units. For example, maintenance priorities can include emergency handling, planned handling, and monitoring observation.

[0104] When assessing maintenance urgency based on equipment cracking risk levels and thermal displacement trend predictions, the need for immediate intervention can be determined by integrating static risk levels with dynamic trend change rates. This can be achieved by setting combination rules, such as high risk + accelerated displacement → urgent intervention, or by using a trained urgency judgment model to output urgency labels, thereby enabling dynamic risk response grading and avoiding misjudgments based on static thresholds. When generating maintenance priorities, the maintenance urgency assessment results can be converted into sortable task levels. For example, this can be achieved by mapping the maintenance urgency assessment results to a 1-5 level numerical priority, or by generating color codes such as red / yellow / green indicators, thus supporting multi-task scheduling and optimized resource allocation.

[0105] Unit operation plans can be load scheduling and outage window information for a future period published by the power plant dispatch system, used to constrain the feasibility boundaries of maintenance window recommendations. For example, unit operation plans can be read from the Manufacturing Information System (MIS) or the dispatch platform interface. Maintenance window recommendations can be optimal maintenance time intervals recommended by combining maintenance priorities and unit operation plans, used to minimize power generation losses while ensuring safety. When formulating maintenance window recommendations based on maintenance priorities and unit operation plans, time windows that meet priority requirements can be selected during periods of low unit load or planned outages.

[0106] Risk contribution indicators can quantify the relative importance weight of each key influencing parameter to the current cracking risk, supporting the generation of targeted maintenance plans and identifying the main controlling factors. Furthermore, risk contribution indicators can be calculated through feature importance scoring, model attention weights, or parameter sensitivity analysis. Targeted maintenance plans can be combinations of specific operation and maintenance measures customized based on the inferred cracking mode and the root cause of the risk, used to improve the effectiveness of intervention and avoid resource waste caused by generalized treatment. For example, targeted maintenance plans may include burner tilt adjustment plans, local ash removal enhancement measures, and damaged pipe section replacement plans. When generating targeted maintenance plans based on cracking modes and risk contribution indicators, the inferred failure mechanisms and main controlling parameters can be mapped to a pre-set measure knowledge base, and the corresponding pre-set measures in the knowledge base can be combined to form targeted maintenance plans. For example, this operation can trigger corresponding plan templates based on if-then rules, or call a causal reasoning model to infer the causal chain and recommend measures, thereby ensuring that maintenance actions are aligned with the root cause and improving the repair success rate.

[0107] When integrating maintenance window recommendations and targeted maintenance plans into maintenance decision recommendations, the time recommendations and measures can be combined into structured operation and maintenance instructions. For example, this operation can generate work orders containing fields such as time, location, measures, and responsible persons, or output an API format for the equipment management system to call and implement, thereby forming a complete and executable closed-loop decision output.

[0108] Taking the differentiated handling of risks in multiple regions as an example, the application of the three-dimensional thermal displacement real-time monitoring and health assessment method for water-cooled walls provided in this embodiment includes: the equipment health status report of a coal-fired unit simultaneously marked two abnormal areas: the upper part of the rear wall was a high-risk area, and the middle part of the left wall was a medium-risk area; the key influencing parameters of the high-risk area, namely the upper part of the rear wall, were identified as "temperature gradient > 80℃ / m" and "vertical displacement deviation + 12mm". Matching the historical cracking case library, it was found that its similarity with the case of "thermal fatigue crack along the pipe axis" reached 92%; while the middle part of the left wall was mainly affected by "local ash accumulation leading to heat transfer deterioration", and matched as "perforation caused by local corrosion" mode; combined with the thermal displacement trend prediction, it was shown that the displacement growth rate of the rear wall area reached 0.5mm / min, and the maintenance priority of the rear wall area was assessed as "urgent", while the maintenance priority of the left wall was "planned". Based on the fact that the unit has no planned shutdown for the next 7 days but has a Class C maintenance window on the 8th day, the following maintenance window recommendations are made: immediately adjust the burnout air ratio and strengthen infrared monitoring in the rear wall area; if no improvement is made within 4 hours, apply for temporary load reduction; schedule partial cleaning of the left wall during the maintenance window on the 8th day; the final maintenance decision recommendation includes two differentiated solutions and corresponding window periods, and the maintenance decision recommendation is pushed to the operation and maintenance platform.

[0109] Another embodiment of the present invention provides a real-time monitoring and health assessment system for three-dimensional thermal displacement of a water-cooled wall, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, and when the computer program is executed by the at least one processor, it implements the real-time monitoring and health assessment method for three-dimensional thermal displacement of a water-cooled wall as described in the above embodiment.

[0110] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0111] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0112] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for real-time monitoring and health assessment of three-dimensional thermal displacement of a water-cooled wall, characterized in that, include: Multi-source monitoring data of the boiler water-cooled wall are acquired, and spatiotemporal correlation analysis is performed on the multi-source monitoring data to obtain a multi-dimensional monitoring feature set; The multidimensional monitoring feature set is subjected to parameter threshold comparison and deviation analysis, a thermal stress coupling model is constructed, parameter correlation is analyzed, and a risk feature mapping network is obtained. A dynamic correlation prediction model is constructed, and the risk feature mapping network is used to predict thermal displacement trends and assess stress concentration based on the dynamic correlation prediction model to obtain an equipment health risk map. Establish a health status evaluation index system, and conduct a crack risk assessment on the equipment health risk map based on the health status evaluation index system to obtain an equipment health status report; Maintenance decision recommendations are generated based on the equipment health status report, and real-time anomaly warnings and maintenance strategy recommendations are pushed through edge computing.

2. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 1, characterized in that, The multi-source monitoring data includes temperature signals, three-dimensional thermal displacement signals, stress signals, and unit load signals at different locations at the same altitude; Spatiotemporal correlation analysis is performed on the multi-source monitoring data to obtain a multi-dimensional monitoring feature set, including: The temperature signal is subjected to spatiotemporal distribution feature extraction to obtain temperature field gradient features; the three-dimensional thermal displacement signal is subjected to three-dimensional coordinate transformation to obtain spatial deformation features. The stress signal is subjected to time-domain waveform analysis to obtain stress concentration characteristics; the unit load signal is subjected to time-series trend extraction to obtain load fluctuation characteristics; and the temperature field gradient characteristics and the load fluctuation characteristics are coupled and analyzed to obtain the heat load correlation coefficient. Based on the heat load correlation coefficient, the spatial deformation characteristics are corrected to obtain the true thermal displacement characteristics; based on the true thermal displacement characteristics, the temperature field gradient characteristics, the stress concentration characteristics, and the load fluctuation characteristics, spatiotemporal registration is performed to construct a multidimensional monitoring feature set.

3. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 2, characterized in that, The temperature signal is subjected to spatiotemporal distribution feature extraction to obtain temperature field gradient features, including: The monitoring area of ​​the water-cooled wall was spatially discretized using the finite element mesh generation method to obtain the temperature data of the mesh nodes; the deviation between the temperature of each mesh node in the mesh node temperature data and the corresponding theoretical design value was calculated, and a temperature deviation matrix was constructed. Spatial interpolation is performed based on the temperature deviation matrix to generate a temperature field distribution cloud map. Based on the temperature field distribution cloud map, the temperature difference and relative deviation ratio of different parts at the same height are calculated, and a temperature gradient feature vector is constructed to obtain the temperature field gradient features.

4. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 2, characterized in that, The multidimensional monitoring feature set is subjected to parameter threshold comparison and deviation analysis. A thermal stress coupling model is constructed, parameter correlation is analyzed, and a risk feature mapping network is obtained, including: The temperature parameters, thermal displacement parameters, and stress parameters in the multidimensional monitoring feature set are compared with their corresponding theoretical design values, and the parameter exceedance index is obtained based on the comparison results. The differences in temperature parameters, thermal displacement parameters, and stress parameters at different locations at the same height are calculated to construct a lateral deviation feature spectrum. The lateral deviation feature spectrum is then dimensionless to obtain the relative deviation ratio coefficient. Based on the parameter exceedance index and the relative deviation ratio coefficient, a thermal stress coupling model is constructed; multivariate correlation analysis is performed on the temperature field gradient characteristics, the actual thermal displacement characteristics and the unit load characteristics to obtain a cross-parameter correlation matrix; Key influencing factors are identified based on the cross-parameter correlation matrix, and a parameter correlation network is constructed. Risk features are extracted based on the parameter correlation network and the thermal stress coupling model to obtain the risk feature mapping network.

5. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 4, characterized in that, The transverse deviation characteristic spectrum is dimensionless to obtain the relative deviation proportionality coefficient, including: A deviation threshold range is set, and the lateral deviation feature spectrum is normalized to obtain a standardized deviation value. The weighting coefficients of the temperature parameter, the thermal displacement parameter, and the stress parameter are determined using the analytic hierarchy process (AHP); a weighted deviation matrix is ​​constructed based on the standardized deviation value and the weighting coefficients. Calculate the relative deviation ratio between each element in the weighted deviation matrix and the corresponding theoretical design value to obtain the relative deviation ratio coefficient.

6. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 1, characterized in that, A dynamic correlation prediction model is constructed. Based on the dynamic correlation prediction model, the risk feature mapping network is used to predict thermal displacement trends and assess stress concentration to obtain an equipment health risk map, including: A long short-term memory network model and an attention mechanism model are constructed; the long short-term memory network model and the attention mechanism model are fused to construct an initial association prediction model; the initial association prediction model is trained using historical monitoring data of the boiler water-cooled wall to obtain a dynamic association prediction model. The risk feature mapping network is extracted using a sliding time window method, and the extracted features are input into the dynamic correlation prediction model to obtain thermal displacement trend prediction results. Based on the thermal displacement trend prediction results and stress concentration assessment results, a stress development trend curve is constructed. The stress development trend curve is compared and analyzed with a preset safety threshold to obtain an equipment health risk map.

7. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 1, characterized in that, The health status evaluation index system includes parameter exceedance index, horizontal deviation index, trend change index, and coupling risk index; Based on the aforementioned health status evaluation index system, a crack risk assessment is performed on the equipment health risk map to obtain an equipment health status report, including: Based on the health status evaluation index system, the equipment health risk map is quantified to obtain equipment health indicators; The equipment health indicators are standardized to obtain the corresponding normalized health index; the normalized health index is weighted and fused to obtain a comprehensive risk score; based on the comprehensive risk score, the equipment cracking risk level is classified; sensitivity analysis is performed on the normalized health index to obtain the corresponding risk contribution index; and an equipment health status report is generated based on the equipment cracking risk level and the risk contribution index.

8. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 7, characterized in that, The normalized health index is weighted and fused to obtain a comprehensive risk score, including: A risk assessment matrix is ​​constructed, and the normalized health index is mapped to the corresponding risk probability. The membership degree of the risk probability is calculated using the fuzzy comprehensive evaluation method to obtain the corresponding fuzzy risk vector. The index weight of the fuzzy risk vector is determined based on the risk contribution index. The fuzzy risk vector is weighted and summed using the index weight. The weighted result is mapped to the scoring interval [0,100] to obtain a comprehensive risk score.

9. The method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls according to claim 1, characterized in that, Based on the equipment health status report, maintenance decision recommendations are generated, including: Based on the equipment health status report, high-risk areas and key influencing parameters are identified, and historical cracking case databases are matched to infer possible cracking modes. Maintenance urgency is assessed based on the equipment cracking risk level and thermal displacement trend prediction results, and maintenance priorities are generated. Based on the maintenance priorities and unit operation plan, maintenance window recommendations are formulated; based on the inferred possible cracking modes and risk contribution indicators, targeted maintenance plans are generated; the maintenance window recommendations and the targeted maintenance plans are integrated into the maintenance decision recommendations.

10. A real-time monitoring and health assessment system for three-dimensional thermal displacement of a water-cooled wall, characterized in that, include: At least one processor; as well as, A memory communicatively connected to at least one of the processors; wherein, The memory stores a computer program that, when executed by at least one of the processors, implements the method for real-time monitoring and health assessment of three-dimensional thermal displacement of water-cooled walls as described in any one of claims 1 to 9.

Images