Neural network-based building insulation material deliquescence rate prediction system and method

Abstract

The application provides a neural network-based building thermal insulation material deliquescence rate prediction system and method, relates to the field of computer systems based on biological models, and fundamentally solves many deficiencies of the prior art in the field of building thermal insulation material deliquescence deterioration prevention and treatment by constructing a three-layer progressive intelligent system; the space-time graph network capable of capturing dynamic nonlinear causal relationship is used as a prediction core, and the forward-looking prediction of the deterioration trend is realized. On this basis, by introducing a risk potential field and a high-level decision-making agent, abstract prediction is converted into specific and executable tactical task planning, and the problem of disconnection between prediction and intervention is solved; finally, by introducing a low-level execution agent, micro-real-time self-adaptive control under the premise of following macro-strategy is realized, and the rigidity and blindness of traditional control methods are overcome.

Description

Technical Field

[0001] This invention relates to the field of computer systems based on biological models, specifically to a system and method for predicting the deliquescence rate of building insulation materials based on neural networks. Background Technology

[0002] The external composite insulation layer of large liquefied natural gas (LNG) storage tanks or nuclear power plant containment structures in coastal areas. The insulation layers of these facilities are exposed to extreme environments such as high salt spray, high humidity, and temperature fluctuations for extended periods, making them highly susceptible to deliquescence and degradation; traditional methods assume that the insulation material is static.

[0003] The existing technology, with publication number CN104299032A, entitled "A Method for Predicting Soil Corrosion Rate in Substation Grounding Grids," is fast, effective, and accurate, providing a reliable basis for predicting soil corrosion in substation grounding grids.

[0004] Existing monitoring and prediction technologies for the deliquescence and degradation of building insulation materials have the following three progressively worsening core technical defects when facing the complex service environments of critical facilities such as large coastal liquefied natural gas (LNG) storage tanks:

[0005] Limitations of Static Models: Traditional prediction models generally treat insulation materials as static media with constant physical properties, and their prediction logic is based on fixed heat and mass transfer equations. However, the actual deliquescence process is a dynamic, self-accelerating positive feedback loop: initial deliquescence alters the material's microstructure (e.g., increased porosity, enhanced connectivity), and this structural degradation, in turn, significantly accelerates subsequent deliquescence rates. Existing technologies cannot capture this dynamic nonlinear coupling relationship where "current degradation state affects future degradation rate," resulting in severely insufficient long-term prediction accuracy, often only allowing for lagging assessments after damage has already occurred.

[0006] The disconnect between prediction and intervention: Most existing technologies remain at the level of passively predicting degradation rates. They typically provide single risk indicators or degradation trend charts, but cannot translate these macroscopic predictions into specific, actionable intervention strategies. The system cannot autonomously identify the most threatening root causes of degradation, let alone generate precise control commands for proactive repair or prevention. This "predict-only, no-decision" approach significantly diminishes the engineering application value of the prediction results and fails to create an effective closed-loop maintenance system.

[0007] Lack of coordinated and adaptive control capabilities: Even in the few systems with proactive intervention capabilities, their control logic is often based on simple triggering mechanisms with preset thresholds, lacking adaptability to global situations and real-time changes. The system cannot perform optimal resource allocation under multiple objectives and constraints based on the overall energy budget, the risk level of different areas, and real-time feedback from microscopic damage. For example, it cannot intelligently decide whether to prioritize repairing a point with the highest current stress or prioritizing blocking an area that will deteriorate the fastest in the future. Nor can it dynamically fine-tune control commands based on real-time repair effects (such as changes in acoustic emission signals from microcapsule rupture) during repair tasks. This results in intervention behaviors that are often blind, inefficient, and may even lead to energy waste.

[0008] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a neural network-based system and method for predicting the deliquescence rate of building insulation materials, in order to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A neural network-based system for predicting the deliquescence rate of building insulation materials is applied to the insulation layer of buildings. The system is configured to perform the following operations:

[0012] Data acquisition module: used to collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer;

[0013] Deliquescence rate prediction module: Based on the macroscopic state data, a spatiotemporal graph network model is used to predict and generate a spatial distribution map of the deliquescence rate of the insulation layer at future times;

[0014] Risk potential field generation module: used to generate a repair unit stress potential field characterizing the risk of microscopic damage inside the insulation layer based on the spatial distribution map of the deliquescence rate at the future time; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit.

[0015] Hierarchical decision-making and collaborative execution control module: used to generate tactical mission objectives to guide repair or blocking operations based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate using a hierarchical reinforcement learning model;

[0016] Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

[0017] A method for predicting the deliquescence rate of building insulation materials based on neural networks, the method being used to execute the neural network-based building insulation material deliquescence rate prediction system, specifically including the following steps:

[0018] Step S1: Collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer;

[0019] Step S2: Based on the macroscopic state data, use a spatiotemporal graph network model to predict and generate a spatial distribution map representing the deliquescence rate of the insulation layer at future times;

[0020] Step S3: Based on the spatial distribution map of the deliquescence rate at the future time, generate a stress potential field of the repair unit that characterizes the risk of microscopic damage inside the insulation layer; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit.

[0021] Step S4: Based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate, a hierarchical reinforcement learning model is used to generate tactical mission objectives to guide repair or blocking operations;

[0022] Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. To address the static limitations of existing models, this network introduces a causal inference mechanism, enabling it to learn the true causal relationship between the "current degradation state" and the "future degradation rate" from historical data. This allows for the accurate capture and quantification of the dynamic nonlinear positive feedback effect of the deliquescence process. It can generate a high-precision, future-oriented spatial distribution map of the deliquescence rate, achieving a fundamental shift from "lagging assessment" to "forward-looking prediction."

[0025] 2. To address the disconnect between prediction and intervention, a high-level decision-making agent was designed within the "Risk Potential Field Generation Module" and the "Hierarchical Decision-Making and Collaborative Execution Control Module." The former creatively transforms the rate prediction map into a "Repair Unit Stress Potential Field" characterizing microscopic damage risk through a physical proxy model or explicit weighted model, mapping abstract rate data into quantifiable risk indicators directly related to physical damage. The latter further combines this potential field with the rate map and the system's total available energy, generating a "Tactical Mission Objective" containing clear repair priorities and intervention modes through reinforcement learning or explicit computation.

[0026] 3. To overcome the shortcomings of lacking coordinated and adaptive control, this invention introduces a low-level executive agent in the "hierarchical decision-making and coordinated execution control module." This agent receives the tactical mission objectives from the upper layer as "strategic instructions," while simultaneously sensing the acoustic emission signals of microcapsule rupture due to microscopic damage in real time. Following the strategic instructions, it uses algorithms such as Proximal Policy Optimization (PPO) to dynamically and precisely adjust the continuous control instructions for the path blocking unit and the self-repair trigger unit, seeking the optimal physical repair effect within the constraints of the mission. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall system flow of the present invention;

[0028] Figure 2 This is a logic block diagram of the data acquisition module of the present invention;

[0029] Figure 3 This is a logic block diagram of the deliquescence rate prediction module of the present invention;

[0030] Figure 4 This is a logic block diagram of the hierarchical decision-making and collaborative execution control module of the present invention. Detailed Implementation

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0032] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0033] Example 1:

[0034] Please see Figures 1 to 4 The present invention provides a technical solution:

[0035] A neural network-based system for predicting the deliquescence rate of building insulation materials is applied to the insulation layer of buildings. The system is configured to perform the following operations:

[0036] Data acquisition module: used to collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer;

[0037] Further explanation of the data acquisition module: The configuration of the data acquisition module is as follows: At multiple preset acquisition points on the insulation layer, macroscopic state data characterizing the macroscopic, slow chemical degradation process of the insulation layer, and microscopic damage data characterizing the microscopic, sudden physical damage events of the insulation layer are acquired; wherein, chemical bond vibration frequency shift and microcapsule rupture acoustic emission signal are extracted from the macroscopic state data and microcapsule rupture data, respectively; in this embodiment, the "chemical bond vibration frequency shift" is denoted as... The "acoustic emission signal of microcapsule rupture" is denoted as... ;

[0038] Furthermore, the Raman spectrometer includes multiple fiber Raman probes deployed inside the insulation layer in a preset grid layout; the acoustic emission sensor includes multiple piezoelectric acoustic emission sensors deployed in a grid layout corresponding to the fiber Raman probes, so as to realize the point-to-point acquisition of data at different locations in the insulation layer.

[0039] The chemical bond vibration frequency offset The shift in the vibrational frequency of the chemical bonds was measured by Raman spectroscopy. By periodically acquiring the Raman spectra of specific chemical bonds in the material through an array of fiber optic Raman probes deployed within the insulation layer, the difference between the characteristic peak position and the reference peak position in the initial state of the material is calculated. This difference is quantified into a scalar, reflecting the degree of change in the molecular structure of the material due to chemical degradation. In this embodiment, the "specific chemical bonds in the material" are characterized as C-C bonds in the polymer backbone or Si-O-Si bonds in inorganic fibers.

[0040] The acoustic emission signal of the microcapsule rupture Measured by an acoustic emission sensor. The acoustic emission signal from the microcapsule rupture. By deploying a piezoelectric acoustic emission sensor array within the insulation layer, the elastic waves released when the pre-embedded self-healing microcapsules rupture due to the propagation of microcracks inside the material are monitored in real time. The cumulative energy value of the elastic wave signal within a preset frequency band is extracted and quantified into a scalar to characterize the occurrence and severity of micro-damage events.

[0041] The data acquisition module is further configured to measure the chemical bond vibration frequency offset at each acquisition location. Acoustic emission signal from microcapsule rupture A weighted combined analysis is performed to calculate the comprehensive degradation risk index; in this embodiment, the comprehensive degradation risk index is denoted as... This embodiment implements a comprehensive degradation risk index. The calculation process is as follows:

[0042] 1.1) Shifting the original chemical bond vibration frequencies Mapping to dimensionless values ​​in the 0-1 interval yields the normalized frequency offset. This is used to eliminate the influence of dimensions and facilitate subsequent weighted calculations. The calculation is performed using a minimum-maximum normalization method. The calculation method involves using the currently collected chemical bond vibration frequency offset. Subtract a preset minimum offset threshold Then divide the resulting difference by the maximum offset threshold. With minimum offset threshold The difference. Among them, and The frequency offset range was predetermined by conducting accelerated aging tests on the insulation material of this model and recording the range of frequency offsets throughout the entire process from intact to complete failure.

[0043] 1.2) The original microcapsule rupture acoustic emission signal Mapping to dimensionless values ​​in the 0-1 interval to obtain normalized acoustic emission energy. Specifically, the acoustic emission signal from the microcapsule rupture during the current acquisition period is used. Subtract a background noise energy threshold Then, the difference is divided by the maximum expected energy of a single rupture event. Compared with background noise energy threshold The difference. Among them, and It was determined in advance by conducting tensile failure tests on material samples containing microcapsules under standard laboratory conditions and calibrating the energy range of their acoustic emission signals.

[0044] 1.3) Determine the contribution weight of macroscopic degradation Weight of contribution to micro-damage These two dimensionless parameters represent the relative importance of macroscopic chemical degradation and microscopic physical damage to the overall failure of the insulation layer, respectively, and their sum is 1. Specifically, the Analytic Hierarchy Process (AHP) is used to determine these parameters, as follows:

[0045] A second-order judgment matrix is ​​constructed to compare the relative importance of "long-term performance degradation due to chemical degradation" and "sudden failure due to microcracks" to the application scenario. The comparison results are obtained by taking the geometric mean of scores given by at least three technical experts in the field based on historical failure data and material mechanism analysis. Subsequently, the maximum eigenvalue of the judgment matrix and its corresponding normalized eigenvector are calculated, and the two components of this eigenvector are used as weights for macroscopic degradation contributions. Weight of contribution to micro-damage This embodiment requires calculating and verifying that the consistency ratio of the judgment matrix is ​​less than 0.1 to ensure the consistency and logic of expert judgments;

[0046] 1.4) Comprehensive Deterioration Risk Index It is a dimensionless risk index ranging from 0 to 1, and its value is directly proportional to the overall deterioration risk level of the insulation layer; the overall deterioration risk index Its value is equal to the normalized frequency offset. Weight of contribution to macroeconomic degradation The product of the two, plus the normalized acoustic emission energy. Weight of contribution to micro-damage The product. In this calculation, the input parameters are... and All are dimensionless values ​​between 0 and 1, and the weighting parameters are... and It is also a dimensionless value, therefore the output comprehensive degradation risk index Both are dimensionless values, but the physical dimensions remain consistent.

[0047] 1.5) The beneficial effect of the data acquisition module is that, through the above-described calculation process of the comprehensive deterioration risk index, a comprehensive, quantitative, and physically meaningful characterization of the deterioration state of the insulation layer is achieved. Its logical rationality and technical effectiveness are demonstrated through the following deductive process:

[0048] The comprehensive degradation risk index The calculation formula originates from the linear weighted summation method in multi-attribute decision theory; the process of parameter substitution, data calculation, and result derivation at the application level is as follows:

[0049] Step a1) Parameter acquisition and normalization: at time... At the collection location point Collected Simultaneously collected The preset normalization parameters are calculated as follows: , ; , Calculated according to the normalization formula: . .

[0050] Step a2) Weight determination and combinatorial optimization:

[0051] Scenario 1, focusing on long-term stability: Experts, using the analytic hierarchy process (AHP), determined that "long-term performance degradation due to chemical degradation" is significantly more important than "sudden failure due to microcracks." The weights were calculated and determined as follows: , .

[0052] Scenario 2: External insulation for high-rise buildings emphasizes impact resistance and toughness. Experts determined that "sudden failure caused by microcracks" is more critical than "chemical degradation." The weights were calculated and determined as follows: , .

[0053] This weighting mechanism allows the system to be customized according to different application needs and priorities by adjusting the weighting of macro-degradation contribution. and the contribution weight of micro-damage By combining these technologies and optimizing their application, risk assessments can be made more targeted.

[0054] Step a3) Result Calculation: In Scenario 1: In scenario two: The calculation process clearly demonstrates that even with the same original data, different weight configurations will yield different risk assessment results, reflecting the algorithm's operability and adaptability to specific engineering needs.

[0055] 1.6) Parameter positional relationship: In the comprehensive deterioration risk index In the calculation formula, the normalized frequency offset and normalized acoustic emission energy As a fundamental input variable, it serves as the starting point for the computational logic. Macroscopic degradation contribution weight. and the contribution weight of micro-damage As a regulating parameter, it is directly multiplied by the basic input variable, occupying the middle link in the logic chain, and plays a role in adjusting the strength of their respective influences. This parallel and weighted structure ensures that the two degradation mechanisms are evaluated independently before being integrated according to their importance, resulting in a clear structure that meets the technical objectives.

[0056] Overall Deterioration Risk Index With normalized frequency offset It exhibits a direct linear relationship, with its slope being the contribution weight of macroscopic degradation. This means that, all other things being equal, the more severe the chemical degradation, i.e., the greater the normalized frequency shift. Increase, overall deterioration risk index It increases linearly.

[0057] Overall Deterioration Risk Index With normalized acoustic emission energy It exhibits a direct linear relationship, with its slope representing the contribution weight of microscopic damage. This means that, all other things being equal, the more severe the microscopic physical damage (i.e., the normalized acoustic emission energy), the greater the impact. Increase), comprehensive deterioration risk index It increases linearly.

[0058] This embodiment is designed to ensure that the computational logic aligns with physical reality: any form of degradation will increase the overall risk. This is due to the normalized frequency offset. and normalized acoustic emission energy The value range is [0,1], and the macroscopic degradation contribution weight is... and the contribution weight of micro-damage The range of values ​​is [0,1] and satisfies Therefore, the comprehensive degradation risk index The output range is limited to the interval [0,1]; when the comprehensive degradation risk index... The closer the output is to 0, the closer the overall health status of the insulation layer is to perfect; at this point, there are two additive terms. and Given that all values ​​approach zero, and since the weights are non-zero, this necessitates normalizing the frequency offset. and normalized acoustic emission energy All approach 0. Normalized frequency offset Approaching 0 means a shift in the vibrational frequency of chemical bonds. Approaching its minimum value indicates that the material's chemical structure is stable and there is no significant degradation. Normalized acoustic emission energy Approaching 0 means the acoustic emission signal from microcapsule rupture. The levels are close to background noise, indicating that there is no significant microcrack generation and propagation within the material. Both degradation mechanisms are at extremely low levels, which is consistent with the technical goal of keeping the insulation layer in good condition.

[0059] When the comprehensive deterioration risk index The closer the output is to 1, the closer the insulation layer is to the critical state of complete failure. This indicates... The closer it gets to 1. Because... This mathematically requires normalizing the frequency offset. and / or normalized acoustic emission energy The closer it is to 1, the better. Normalized frequency offset Approaching 1 means a shift in the frequency of chemical bond vibrations. Approaching its maximum threshold indicates that the material's chemical structure has severely deteriorated. Normalized acoustic emission energy. A value approaching 1 indicates the acoustic emission signal upon microcapsule rupture. Reaching the maximum expected energy level of a single fracture event indicates that severe and widespread physical damage is occurring within the material. It should be noted that this embodiment will incorporate a comprehensive degradation risk index. The output values ​​are divided into the following three intervals:

[0060] Approaching complete failure range: Comprehensive degradation risk index Values ​​above R1; and R1 values ​​within the range of (0.85, 0.98); approaching the intact range: Comprehensive Deterioration Risk Index Values ​​below R²; and R² values ​​within the range of (0.01, 0.23); falling within the intermediate degradation range: Comprehensive Degradation Risk Index The values ​​are within the range (R2, R1). This embodiment provides the following data table and analysis:

[0061] Table 1. Example study of the comprehensive degradation risk index:

[0062] Monitoring point status Chemical bond vibration frequency shift (cm⁻¹) Acoustic emission signal of microcapsule rupture (mV·μs) Normalized frequency offset Normalized acoustic emission energy Macroeconomic degradation contribution weight Micro-damage contribution weight Overall Deterioration Risk Index Initially intact 0.2 55 0.02 0.005 0.7 0.3 0.0155 Slight chemical degradation 3 60 0.3 0.01 0.7 0.3 0.213 Microcracks appeared 1.5 550 0.15 0.5 0.7 0.3 0.255 Moderate composite degradation 5.5 450 0.55 0.4 0.7 0.3 0.505 Severe chemical degradation 8 80 0.8 0.03 0.7 0.3 0.569 Nearly expired 9.5 950 0.95 0.9 0.7 0.3 0.935

[0063] The table uses Scenario 1 (with a weight of ) as an example. , Taking the example of this invention, the innovative effect of the data acquisition module is clearly demonstrated.

[0064] Distinguishing between different degradation modes: Comparing the lines "slight chemical degradation" and "microcracks appearing," although the latter has a higher overall degradation risk index... The value (0.2550) is slightly higher than the former (0.2130), but their compositions are completely different. The risk of the former mainly comes from the normalized frequency offset. The latter mainly comes from normalized acoustic emission energy. This indicates that the overall degradation risk index... It not only provides the risk level but also implies the source of the risk, enabling downstream modules to take targeted measures and providing data support; targeted measures include, but are not limited to, considering corrosion prevention if the main cause is chemical degradation, and prioritizing repair if the main cause is physical damage.

[0065] Quantifying the combined degradation effect: Under the "moderate combined degradation" state, both degradation mechanisms contribute significantly, resulting in a final comprehensive degradation risk index. The value (0.5050) is the result of the weighted effect of the two, which accurately reflects the high-risk state under the combined effect.

[0066] Full lifecycle coverage: from "initially intact" with a low overall degradation risk index The value (0.0155) indicates a high overall degradation risk index, suggesting the index is on the verge of failure. The value (0.9350) effectively covers and quantifies the entire life cycle of thermal insulation materials, from brand new to failure, verifying the effectiveness of its dynamic range.

[0067] Deliquescence rate prediction module: Based on the macroscopic state data, a spatiotemporal graph network model is used to predict and generate a spatial distribution map of the deliquescence rate of the insulation layer at future times;

[0068] To further explain, the deliquescence rate prediction module is configured to: based on a comprehensive degradation risk index obtained from the data acquisition module. Using the macroscopic state data time series, including the thermal insulation layer, a causal inference spatiotemporal graph convolutional network is employed to predict and generate a spatial distribution map of the deliquescence rate at future times; in this embodiment, the spatial distribution map of the deliquescence rate is denoted as... ;

[0069] Furthermore, the causal inference spatiotemporal graph convolutional network includes graph convolutional layers, gated recurrent unit layers, and attention mechanism layers; the graph convolutional layers are used to capture the spatial dependencies between sensor nodes at each time step; the gated recurrent unit layers are used to learn the comprehensive degradation risk index of each node. The temporal evolution pattern; the attention mechanism layer is trained to calculate the Granger causality score between different nodes, and the information transfer in the graph convolution process is weighted according to the Granger causality score; the spatial distribution map of the deliquescence rate. The generation process, specifically the calculation flow, is broken down as follows:

[0070] The Granger causality score from node j to node i is labeled as The maximum theoretical deliquescence rate is identified as... The normalized node-predicted deliquescence rate is identified as... .

[0071] In this embodiment, the "sampling location point" is abstracted as a logical "node"; therefore, the set of all sensors on the insulation layer constitutes a set of nodes in a graph; that is, the "sampling location point" is represented as a "node".

[0072] 2.1) Calculate the nodal fundamental deliquescence rate of the current node i. : is the comprehensive degradation risk index based on the current node i An initial deliquescence rate value was calculated, in millimeters per year (mm / a). It reflects the instantaneous deliquescence rate corresponding to the current deterioration state; specifically, it is calculated using a preset empirical transformation function. This empirical transformation function is obtained through nonlinear regression analysis fitting based on extensive salt spray accelerated aging experimental data of the current insulation material. The node-based deliquescence rate... Its value is equal to a material deliquescence rate reference coefficient. With comprehensive degradation risk index The product of these factors, plus an environmental correction factor. Multiply by the exponent of each factor. Wherein is the base coefficient for the material deliquescence rate. This is a constant characterizing the inherent deliquescence resistance of the insulation material, determined by the standard ASTM-B117 salt spray test; environmental correction factor. It is a correction factor obtained by referring to the prefabrication table based on the average temperature and humidity of the macro environment in which the insulation layer is located;

[0073] 2.2) Calculate the causal influence rate increment of the current node i. The algorithm considers the accelerating effect of upstream neighboring nodes with strong causal relationships on the future deliquescence rate of the current node i, expressed in millimeters per year (mm / a). This characteristic is the core innovation of this deliquescence rate prediction module, reflecting the dynamic process of degradation propagation.

[0074] Increment of causal effect rate The causal relationship matrix output by the causal inference spatiotemporal graph convolutional network is obtained by weighted summation. The calculation process is as follows: Granger causal relationship scores between the current node i and all other nodes j are extracted from the causal inference spatiotemporal graph convolutional network and denoted as... Secondly, samples with scores higher than the causal association threshold were selected. The strong causal upstream nodes; then, the node-based deliquescence rate of each strong causal upstream node. Its corresponding Granger causality score Multiply them; finally, sum all these products to obtain the causal effect rate increment. Causal correlation threshold It is a statistical quantile that can effectively distinguish significant causal relationships from random fluctuations based on historical data analysis;

[0075] 2.3) Calculate the predicted deliquescence rate of the current node i. : This is the final output, a spatial distribution map of the deliquescence rate. The value corresponding to the predicted deliquescence rate of that node is expressed in millimeters per year (mm / a). The predicted deliquescence rate of the node... Its value is equal to the basic deliquescence rate of the node. With the rate increment of causal influence The sum. In this calculation, the input parameter is the basic deliquescence rate. With the rate increment of causal influence The units are all millimeters per year (mm / a), therefore the output deliquescence rate spatial distribution map The unit is also millimeters per year (mm / a), maintaining consistent physical dimensions. Repeating the above calculations for all nodes in the insulation layer will construct a complete spatial distribution map of the deliquescence rate. .

[0076] Furthermore, the beneficial effect of the deliquescence rate prediction module lies in its ability to achieve dynamic, forward-looking, and physically interpretable prediction of the deliquescence rate of the insulation layer through the calculation process of the node-predicted deliquescence rate. Its logical rationality and technical effectiveness are demonstrated through the following deductive process: the node-predicted deliquescence rate... The calculation formula is derived from the principle of effect superposition in physics, and Granger causality in time series analysis is used as the quantitative basis for effect transmission.

[0077] Monitor target node i in the insulation layer, which is potentially affected by an upstream neighbor node j.

[0078] Step a3) Calculate the basic deliquescence rate of node i. At time t, the comprehensive degradation risk index is obtained at node i. The default parameters are: , Calculated based on the empirical transformation function: .

[0079] Step a4) Calculate the causal effect rate increment The "weight" here refers to the causal correlation threshold. Setting the causal correlation threshold. It determines which causal paths are considered valid, thus affecting the final rate.

[0080] High-precision mode: To capture all potential degradation paths, the causal correlation threshold is adjusted. The value is set to a low value; in this embodiment, the 0.90 quantile is used.

[0081] High-confidence mode: To avoid noise interference, only the most significant degradation path is considered, and the causal correlation threshold is set. The value is set to a relatively high value; in this embodiment, the 0.99 quantile is used.

[0082] The system adopts a high-precision mode and sets... From the causal inference spatiotemporal graph convolutional network model, the data from node j to node i is obtained. .because This causal path is activated. The basic deliquescence rate of node j. Calculate the rate increment of causal effects: .

[0083] Step a5) Synthesize and normalize the final rate: Calculate the node-predicted deliquescence rate: Set the maximum theoretical deliquescence rate of the current insulation material. for Calculate the normalized node predicted deliquescence rate. .

[0084] Node-predicted deliquescence rate The comprehensive degradation risk index of node i They are in a direct proportional relationship. The increase leads to an increase in the basic deliquescence rate of node i. The linear increase leads to a higher predicted deliquescence rate at the nodes. Increase.

[0085] Node-predicted deliquescence rate Score of causality with Granger They are directly proportional. When At that time, the node predicts the deliquescence rate. The larger the value, The larger the value, the more accurate the node prediction of the deliquescence rate. Increase. This quantifies the logic that "the stronger the causal relationship, the greater the impact."

[0086] Node-predicted deliquescence rate With upstream nodes There is a direct proportional relationship. The faster the upstream node deteriorates, the stronger its "contagion" effect on the downstream. The larger the predicted deliquescence rate output of a node, the higher the risk of deliquescence faced by the node corresponding to that sampling location in the future, and the more severe the deterioration trend.

[0087] Predicting deliquescence rate using normalized nodes The output range is limited to the interval [0,1]. When the normalized node predicts the deliquescence rate... The closer the output is to 0, the lower the risk of deliquescence at future time points and the more stable the state of the node; normalized node prediction of deliquescence rate. When the value approaches 0, the following situations exist: The nodal prediction of deliquescence rate at molecular positions in the calculation formula It approaches 0. This requires its two additive terms to be close to 0. and All are close to 0. Approaching 0, requiring A value close to 0 indicates that the node itself is in a healthy state. Approaching 0 means the Granger causality score of all upstream nodes. All below or itself A low value indicates that the node has not been significantly affected by external degradation. This aligns with the technical objective of keeping the node in a safe and isolated state.

[0088] When normalized nodes predict deliquescence rate The closer the output is to 1, the higher the risk of deliquescence and the more severe the degradation trend of the node in the future. Normalized node prediction of deliquescence rate. When the value approaches 1, the following situations exist: The position of molecules in the calculation formula Approaching the maximum theoretical deliquescence rate This can be achieved through two main paths: 1) the node itself becomes extremely degraded. Approaching 1, leading to Large numerical value; 2) The node's own state is generally average, but it deteriorates very quickly due to one or more factors. (A high value greater than 0.8) and a very strong causal relationship ( The influence of upstream nodes (with high values) leads to The values ​​are large. The following is a data table and analysis of an example of the deliquescence rate prediction module:

[0089] Table 2. Example study of the deliquescence rate prediction module:

[0090] Monitoring Node Scenario Self-comprehensive deterioration risk index upstream node basic deliquescence rate Granger causality score Causal correlation threshold Increment of causal effect rate (mm / a) Node-predicted deliquescence rate Normalized node prediction of deliquescence rate Healthy isolated nodes 0.05 0.22 0.85 0.9 0 0.11 0.022 Self-moderately degraded node 0.5 0.22 0.82 0.9 0 1.1 0.22 Affected healthy nodes 0.1 3.3 0.98 0.9 3.23 3.45 0.69 Degraded collaborative acceleration node 0.6 3.3 0.97 0.9 3.2 4.52 0.904 Critical causal influence node 0.2 2.2 0.91 0.9 2 2.44 0.488 Weak causal relationship nodes 0.7 3.3 0.89 0.9 0 1.54 0.308

[0091] This form (setting) , , This clearly demonstrates the innovative effect of the deliquescence rate prediction module of this invention:

[0092] Differentiate the source of degradation: Compare "moderately degraded nodes" with "affected healthy nodes." The former... (0.220) is entirely due to its own higher [value / quality]. Contribution; while the latter's (0.690) Although one's own health ( Although the risk is only 0.10, its predictive risk is much higher due to the "contagion" from strong causal upstream nodes. This invention can clearly quantify and distinguish these two distinct sources of risk, which is impossible for traditional prediction models.

[0093] Quantifying synergistic effects: The "deteriorating synergistic acceleration node" illustrates the worst-case scenario. Its own state is poor ( ) and powerful external influences ( The superposition of these factors resulted in the highest prediction rate. This accurately simulates the accelerated spread of degradation in reality.

[0094] Effectiveness of causal gating: Comparing "weakly causally related nodes" with "affected healthy nodes." Although the upstream nodes of the former are more severely degraded, because... Not exceeding the causal association threshold ,That When the value is 0, the prediction rate is determined solely by itself. This demonstrates that the present invention, by using causal inference as a "gating switch," can effectively filter out indirect noise influences, making the prediction more robust.

[0095] In a specific embodiment of the present invention, in order to adapt to the limited precision of the digital processing system and improve computational efficiency, the system further includes a processing step: calculating the normalized node predicted deliquescence rate. With a preset rate determination lower threshold Compare; if the normalized node predicts the deliquescence rate Less than the lower limit threshold for rate determination In this embodiment, the lower limit threshold for rate determination is set. The initial value is 0.01; then, in the subsequent risk potential field generation module, this normalized node is used to predict the deliquescence rate. The value of is forcibly set to 0. This processing method is a common technique used by those skilled in the art when implementing numerical algorithms. It aims to effectively process parameters that are theoretically infinitely close to zero in engineering, thereby ignoring the minimum values ​​that have no practical impact on the final result.

[0096] Risk potential field generation module: used to generate a repair unit stress potential field characterizing the risk of microscopic damage inside the insulation layer based on the spatial distribution map of the deliquescence rate at the future time; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit.

[0097] To further explain, the risk potential field generation module is configured to: generate a spatial distribution map of deliquescence rate obtained from the deliquescence rate prediction module. Using a physical proxy model, a stress potential field of the repair unit is generated to characterize the risk of microscopic damage inside the insulation layer; in this embodiment, the stress potential field of the repair unit is denoted as... .

[0098] Furthermore, the physical proxy model is a pre-trained deep neural network whose input is the spatial distribution map of the deliquescence rate. The output is the stress potential field of the repair unit. The training process of this neural network is completed by supervised learning of deliquescence rate and microcapsule stress data generated by a large number of finite element simulations.

[0099] The construction and application of the physical proxy model described in 3.0) is specifically implemented in four stages: data generation, model construction, model training, and model deployment and inference.

[0100] The first stage is training dataset generation: the goal is to create a large number of high-fidelity (input-output) data pairs for supervised learning.

[0101] Defining the input space: The spatial distribution map of deliquescence rate is parametrically generated in a computer to produce tens of thousands of samples with different characteristics. These samples cover a wide range of operating conditions, including: uniform low / high rate distribution, linear gradient distribution, radial gradient distribution, and complex non-uniform distribution containing multiple random shapes and intensities of "hot spots".

[0102] Perform finite element simulation: Use commercial finite element analysis software (e.g., COMSOL-Multiphysics or ABAQUS) to build a two-dimensional or three-dimensional geometric model that is consistent with the actual insulation layer structure, and define the material properties and geometric dimensions of the microcapsules in the model.

[0103] Obtaining the true output values: Apply each deliquescence rate spatial distribution map sample generated in step 1 as a load condition to the finite element model, run the simulation, and calculate the von Mises equivalent stress borne by each microcapsule unit under the corresponding deliquescence rate distribution. Normalize the stress field obtained from the simulation to form the true stress potential field of the repair unit corresponding one-to-one with the input "deliquescence rate spatial distribution map," and label it as... Connect each "spatial distribution map of deliquescence rate" with its corresponding... Combined into a training data pair This forms the final training dataset.

[0104] The second stage, model construction, employs a U-Net network architecture suitable for image-to-image translation tasks. This architecture consists of symmetric encoder and decoder paths and skip connections between them.

[0105] 1. Encoder Path: Composed of a series of convolutional layers and max-pooling layers. Its function is to receive the "spatial distribution map of deliquescence rate" from the input, progressively extracting its hierarchical features from low-level texture to high-level abstraction, while reducing the spatial dimensionality. This is equivalent to the network "understanding" the pattern of deliquescence rate distribution.

[0106] 2. Decoder Path: Composed of a series of up-convolutional layers and standard convolutional layers. Its function is to progressively restore the abstract feature maps extracted by the encoder to the resolution of the original input image, ultimately generating the predicted stress potential field. This is equivalent to the network "drawing" its understanding results.

[0107] 3. Skip Connections: These connections directly concatenate the feature maps from each stage of the encoder path to the corresponding input stage of the decoder path. This is the core of the U-Net architecture. It allows the decoder to directly utilize uncompressed low-level features from the encoder, containing precise spatial location information, during image reconstruction. This ensures that the true value of the stress potential field of the output repair unit is precisely aligned with the input "hygroscopic rate spatial distribution map" in detail, avoiding spatial information loss caused by the encoding and decoding process.

[0108] The third stage, model training: The goal is to train the neural network on a training dataset so that its weight parameters converge, enabling it to accurately fit the spatial distribution of deliquescence rates. Complex nonlinear mapping relationships.

[0109] This embodiment uses mean-squared error (MSE) as the loss function to measure the model's predicted output. With finite element simulation true value The loss function aims to minimize pixel-level differences between the predicted values ​​at each node. It employs the Adam optimizer, which efficiently and stably adjusts network weights during training to minimize the loss function.

[0110] The training dataset is fed into the U-Net network in mini-batches. For each batch, a forward propagation is performed once to obtain the predicted values. The MSE loss is calculated, and then the gradient of the loss function with respect to the weights of each layer of the network is calculated using the backpropagation algorithm. Finally, the Adam optimizer updates the weights based on the gradient. This process is repeated for several epochs until the loss function converges to a sufficiently low level and performs well on independent validation sets.

[0111] The fourth stage, model deployment and inference: After training, the optimal network weight parameters are saved. In the actual running system, this physical proxy model is invoked as an independent computational module. When the upstream "deliquescence rate prediction module" generates a real-time "deliquescence rate spatial distribution map," this map is directly input into the U-Net model that has been loaded with pre-trained weights. The model only needs to perform one forward inference calculation to output the predicted stress potential field of the repair unit in a very short time and pass it to the downstream "hierarchical decision module."

[0112] 3.1) In another specific embodiment, the conversion process from the "spatial distribution map of deliquescence rate" to the "stress potential field of the repair unit" is achieved through an explicit weighted model. This explicit weighted model directly constructs a clear mathematical relationship based on physical insights. The specific calculation process of the explicit weighted model is broken down as follows:

[0113] The stress potential field of the repair unit The generation process, specifically the calculation flow, is as follows: Calculate the rate gradient scalar and denot it as... ; Rate gradient scalar Characterizing the node-predicted deliquescence rate The degree of spatial variation within its neighborhood. The more drastic the change in deliquescence rate, the greater the humidity and salt concentration gradient within the insulation material, resulting in greater internal stress. This is calculated using numerical differentiation methods; specifically, based on the spatial distribution map of the deliquescence rate. By applying a Sobel operator or a similar gradient operator, the velocity gradient vector at position i of each node is calculated. Then, the magnitude of this gradient vector is taken to obtain the velocity gradient scalar for that node. .

[0114] 3.2) Determine the contribution weight of the rate amplitude. With rate gradient contribution weight These two dimensionless parameters represent the relative importance of the absolute magnitude and spatial variation rate of the deliquescence rate to the stress inducing microscopic damage, respectively, and their sum is 1. Specifically, they are determined using a principal component analysis (PCA) method. The first step involves multiple simulations using finite element simulation software (such as ANSYS or ABAQUS), systematically changing the amplitude and gradient of the deliquescence rate and recording the corresponding von Mises equivalent stress on the microcapsule shell. Then, principal component analysis is performed on the simulation dataset containing the three variables: rate amplitude, rate gradient, and equivalent stress. Finally, the contribution weight of the rate amplitude is determined based on the absolute value ratio of the load coefficients corresponding to the rate amplitude and rate gradient in the first principal component. With rate gradient contribution weight .

[0115] 3.3) Calculate the stress potential field of the repair element at node i, and denot it as... Repair unit stress potential field It is a final dimensionless risk index between 0 and 1, whose value is directly proportional to the stress level and rupture probability of the repair unit embedded in the insulation layer; in this embodiment, the repair unit is characterized as a microcapsule; the stress potential field of the repair unit at node i. The value equals the contribution weight of the rate amplitude. The normalized node-predicted deliquescence rate The product of the ... With the normalized rate gradient scalar The product of . Normalization uses the min-max method to map each physical quantity to the [0,1] interval. In this calculation, all input parameters are normalized to dimensionless values, and the weights are also dimensionless; therefore, the output stress potential field of the repair unit is... Similarly, since these are dimensionless values, repeating the above calculations for all nodes in the insulation layer will construct the complete stress potential field of the repair unit. .

[0116] Furthermore, the beneficial effect of the risk potential field generation module lies in its ability to transform the abstract deliquescence rate prediction into a concrete, guideable microscopic risk distribution map through the calculation process of the stress potential field of the repair unit. Its logical rationality and technical effectiveness are demonstrated through the following derivation process: The normalized rate gradient scalar is identified as... The process of parameter input, data calculation, and result deriving at the application level of the risk potential field generation module is as follows:

[0117] The scenario is set to assess the microscopic damage risk of nodes at different locations in the insulation layer: the normalized node-predicted deliquescence rate of node i is obtained. By analyzing the spatial distribution of the entire deliquescence rate... Perform gradient calculations to obtain the normalized rate gradient scalar for that node. .based on Rate amplitude contribution weight With rate gradient contribution weight The combination of factors determines the system's sensitivity to different types of risks; for brittle insulation materials: these materials are more sensitive to stress gradients, and microcracks easily propagate at stress concentration points. Offline PCA analysis shows that the gradient contributes more, and the weight is set as ( , Prioritize the degradation boundary region. For tough insulation materials: these materials can better withstand stress concentration, but will fail when overall degradation reaches a certain level. The weight is set as ( , This embodiment focuses more on areas with high overall degradation. It is specifically designed for brittle materials and sets (…). , At node i, the measurement and calculation are obtained. , Calculate the stress potential field of the repair unit: .

[0118] In this embodiment, Predicted deliquescence rate with normalized nodes They are directly proportional. Under the conditions, Increase This also increases accordingly. This aligns with the physical law that "the faster the degradation rate, the higher the risk of microscopic damage."

[0119] With normalized rate gradient scalar They are directly proportional. Under the conditions, Increase This also increases accordingly. This quantifies the material mechanics principle that "the more drastic the change in degradation rate (i.e., the boundary between the degraded and healthy regions), the greater the internal stress gradient of the material, and the higher the risk of microscopic damage." This is because the input normalized nodes predict the deliquescence rate. and normalized rate gradient scalar The range of values ​​is [0,1], and Therefore, the stress potential field of the output repair unit is... The range of values ​​is also limited to the interval [0,1]; when the stress potential field of the repair unit... The closer the output is to 0, the lower the risk of microscopic damage to node i, and the safer the pre-embedded repair unit, without the need for immediate triggering; when the stress potential field of the repair unit... If the sum approaches 0, there exists a case where both addition terms approach 0. This means... and All values ​​approach 0. This describes an ideal healthy state: not only is the deliquescence rate of the node itself low, but the deliquescence rate of its surrounding area is also similarly low, with no drastic changes in rate. This aligns with the technical goal of a safe state where the material is homogeneous, stable, and free from stress concentration.

[0120] When the stress potential field of the repair unit The closer the output is to 1, the higher the risk of microscopic damage to the node, the greater the stress on the pre-embedded repair unit, and the higher the probability of breakage. It should be the target for priority repair or blocking. Approaching 1 indicates that the weighted sum approaches its maximum value. This can be achieved through two typical paths: 1) and All values ​​approaching 1 indicate that the node is on the boundary of a rapidly deteriorating region, representing the highest risk area. 2) In Even in brittle material scenarios, The value fluctuates around 0.6, but To approach 1, its It will also be high. This accurately identifies "crack tip" type risk points, meaning that although the overall degradation is not severe, the local stress concentration has reached a dangerous level, which is highly consistent with the technical goal of this invention to achieve precise tactical intervention. The following is a data table and analysis of an embodiment of the risk potential field generation module:

[0121] Table 3: Example Study of the Risk Potential Field Generation Module

[0122] Node Scenario Types Normalized node prediction of deliquescence rate Normalized rate gradient scalar Rate Amplitude Contribution Weight Rate gradient contribution weight Repair unit stress potential field Health Regional Center 0.05 0.02 0.3 0.7 0.029 Deterioration area center 0.9 0.1 0.3 0.7 0.34 Deterioration front boundary 0.45 0.95 0.3 0.7 0.8 Isolated high-speed degradation points 0.95 0.85 0.3 0.7 0.88 Slow and uniform deterioration region 0.3 0.05 0.3 0.7 0.125 Gradient mutation but moderate rate region 0.5 0.7 0.3 0.7 0.64

[0123] This table (with weights set to brittle material mode) , This clearly demonstrates the innovative effect of the risk potential field generation module of the present invention:

[0124] Accurately identify the highest risk point: compare the "deterioration region center" and the "deterioration front boundary". Although the former's normalized node predicted deliquescence rate (0.90) is much higher than the latter's (0.45), its final risk value is... (0.340) is much lower than the latter (0.800). This is because the risk potential generation module gives higher weights to the gradient ( This study successfully identified that the intense stress concentration at the "deterioration front boundary" was the primary cause of microscopic damage. This provides counterintuitive but physically accurate guidance for repair decisions.

[0125] Quantifying composite risks: The "isolated high-speed degradation point" combines high rate (0.95) and high gradient (0.85) to obtain the highest repair unit stress potential field (0.880), accurately identifying the most urgent repair target.

[0126] Distinguishing Risk Types: The gradient values ​​of "deterioration region centers" and "slowly uniformly deteriorating regions" are very low, and their risk values ​​are mainly contributed by the rate amplitude, indicating that the risk in these regions is "static" and sporadic. In contrast, the risk values ​​of "deterioration front boundaries" and "gradient abrupt change regions" are mainly contributed by the gradient, indicating that the risk in these regions is "dynamic" and linear. This invention can clearly distinguish and quantify these two different types of risk.

[0127] Hierarchical decision-making and collaborative execution control module: used to generate tactical mission objectives to guide repair or blocking operations based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate using a hierarchical reinforcement learning model;

[0128] Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

[0129] Further explanation: The hierarchical decision-making and collaborative execution control module is further configured to: based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate, generate a tactical mission objective containing repair priorities and intervention modes through a high-level decision-making agent; the tactical mission objective is defined as a structured data object used to transmit strategic intent to lower-level execution units and set the constraint framework for their actions. This ensures that the system can perform local operations under the guidance of global optimization, avoiding suboptimal decisions caused by focusing only on local states.

[0130] Further explanation: The high-level decision-making agent is further configured to take the current total available energy of the system as part of its state input, and its action space is defined as a set of discrete, structured tactical mission objectives; the generation process of the tactical mission objectives is guided by a preset reward function, the design principle of which is to give positive incentives to actions that can cause a preset decrease in the stress potential field of the repair unit, while applying negative penalties to actions that cause excessive consumption of system energy.

[0131] 4.0) The specific implementation process of the high-level decision-making agent generating the tactical mission objective is broken down as follows: The description of the "tactical mission objective generation process guided by reward function" is as follows:

[0132] Through extensive exploration and trial and error in a simulated environment, the high-level decision-making agent learns decision-making strategies that maximize long-term cumulative rewards, thereby ensuring that each tactical mission objective it generates is the optimal choice after comprehensively weighing short-term repair benefits against long-term resource sustainability. As a specific implementation, the high-level decision-making agent is trained using a deep Q-network algorithm. The reward function provides supervision signals for the network's training process, guiding the network weights to converge in a direction that accurately predicts the long-term cumulative rewards that different tactical mission objectives can bring. The specific implementation process of the single-step reward value used by the high-level decision-making agent during training and decision-making is broken down as follows:

[0133] The first step is to define the core structure of the single-step reward function: Let the single-step reward value be denoted as... This is a dimensionless scalar value, whose numerical value comprehensively reflects the positive benefits and negative costs of performing a tactical mission; the single-step reward value... It equals the stress improvement benefit minus the energy consumption cost.

[0134] The second step is the quantification of stress improvement benefits: This step aims to quantify the positive contribution to structural health brought about by repair or interruption operations; the stress improvement benefit is a positive number, the value of which is directly proportional to the decrease in the stress potential field of the repair unit caused by the intervention; the stress improvement benefit is equal to the product of the benefit conversion coefficient and the decrease in the stress potential field. The decrease in the stress potential field is denoted as... This represents the reduction in the total stress potential field of the entire insulation layer repair unit after executing the selected tactical mission objective, relative to the amount before execution. It is obtained by: using the aforementioned physical proxy model of this invention, performing a rapid forward inference on the expected state after the action is executed to obtain a predicted stress potential field, and then calculating the difference between this predicted and current total stress potential field; the benefit conversion coefficient is denoted as... This is a dimensionless positive constant used to convert the physical dimensions of stress reduction into a scale that matches the reward value. During the system design phase, based on the assessment of the importance of structural safety and to balance the magnitudes of benefits and costs in the reward function, it is determined either through expert calibration or through hyperparameter tuning during simulation training. .

[0135] The third step is the quantitative calculation of energy consumption costs: This step aims to quantify the resource costs required to execute a tactical mission. Energy consumption cost is a positive number, directly proportional to the system energy consumed in executing the mission. The energy consumption cost equals the product of the cost conversion factor and the mission execution energy consumption. Mission execution energy consumption is denoted as... This represents the estimated system energy consumption for executing the currently selected tactical mission objective. It is obtained by pre-storing a mission energy consumption lookup table within the system. This table is built based on offline calibration of the power characteristics of repair units (such as microwave resonators) and blocking units (such as cooling equipment). The corresponding predicted energy consumption value is directly retrieved based on the "mode" field (such as powerful, conventional, energy-saving) in the tactical mission objective. The cost conversion factor is denoted as... This is a dimensionless positive constant used to convert the physical dimensions of energy consumption into a scale that matches the reward value. Its value is determined similarly to the reward conversion coefficient, through expert calibration or hyperparameter tuning. Its core function is to adjust the system's preference between "repair effect" and "energy saving" during decision-making. Through the above process, the system can calculate a specific single-step reward value for each "state-action" pair. During the training of a deep Q-network, this reward value serves as the core feedback signal, used to iteratively update the Q-value function through the Bellman-equation. Ultimately, this enables the high-level decision-making agent to learn to select the tactical mission objective with the highest long-term cumulative reward expectation under any state.

[0136] 4.1) The system first analyzes the stress potential field of the input repair unit to identify potential risk areas that need attention. In the stress potential field map of the repair unit, the connected regions whose corresponding values ​​of the stress potential field of the repair unit exceed a preset risk activation threshold are represented as "candidate intervention areas". The candidate intervention area represents the local area in the insulation layer that needs to be evaluated for intervention. The method of obtaining it is to apply an image segmentation algorithm to the stress potential field of the repair unit. In this example, it is based on the connected component analysis of the threshold. The risk activation threshold used is calibrated by an offline Weibull distribution reliability model based on the brittleness of the insulation material, the design life and historical failure data, to determine a critical stress level that can balance false alarms and false alarms.

[0137] 4.2) For each identified candidate intervention area, the system calculates a comprehensive index to assess the urgency of its intervention; in this embodiment, the "comprehensive index" is characterized by a regional comprehensive risk level and denoted as . This is a dimensionless quantitative indicator, and its value is directly proportional to the current degree of damage and future deterioration trend of the area. The overall risk level of the area... The formula is: 4.3) Regional average stress potential field value: This represents the arithmetic mean of the stress potential fields of all repair units within the candidate intervention area, reflecting the current static risk level of the area. It is obtained by integrating or summing the stress potential field values ​​within the candidate intervention area. 4.4) Regional average deliquescence rate value: This represents the arithmetic mean of the deliquescence rate values ​​of all nodes within the candidate intervention area, reflecting the future dynamic deterioration trend of the area. It is obtained by integrating or summing the deliquescence rate values ​​within the candidate intervention area.

[0138] 4.5) Stress Contribution Weights and Rate Contribution Weights: These are dimensionless parameters, summing to 1, used to determine the relative importance of static risk and dynamic trend in the comprehensive assessment. Their numerical determination is as follows: A large dataset of "historical events" containing stress, rate, and whether structural failure ultimately occurred is collected. Principal Component Analysis (PCA) is applied to this dataset. Based on the absolute values ​​of the load coefficients corresponding to the stress and rate variables in the first principal component obtained from the analysis, these two weights are proportionally allocated. This method ensures that the weight allocation is based on the inherent physical correlation revealed by the data, rather than subjective settings.

[0139] 4.6) The system determines the intensity of action to be taken based on its own energy status: This embodiment sets intervention modes including but not limited to powerful mode, normal mode, and energy-saving mode; discrete labels are used to describe the level of resource consumption for the action; these labels are obtained by comparing the system's current total available energy with a set of preset energy thresholds. In this embodiment, powerful mode energy thresholds and normal mode energy thresholds are set. When the total available energy is higher than the powerful mode threshold, "powerful mode" is selected; when the total available energy is between the two, "normal mode" is selected; when it is lower than the normal mode threshold, "energy-saving mode" is selected. These thresholds are preset based on the energy consumption characteristics of the repair unit and the system's endurance requirements through offline energy balance calculations, and will not be elaborated further; the system's current total available energy in this embodiment is a real-time scalar value provided by the energy management unit, denoted as... ;

[0140] 4.7) The system integrates the above calculation results to form a structured instruction, which serves as the tactical mission objective, denoted as... This is a data object containing three key fields: location, priority, and pattern; the system selects the region with the highest overall risk level. The candidate intervention area is taken as the target location; the highest The value itself serves as the repair priority; and the intervention mode selected in step 4.6) is used as the mode field. The final generated tactical mission objective... The data is output to downstream execution units, thus completing a full computational process from situational awareness to strategic decision-making.

[0141] Furthermore, by identifying the regional average stress potential field value of the candidate intervention area (the regional average stress potential field value is...) ) and the regional average deliquescence rate (the regional average deliquescence rate is denoted as This involves weighting and combining data to transform this core idea into a directly calculable regional comprehensive risk level. This enables precise and dynamic prioritization of interventions. The parameter relationships and computational logic design of this mechanism are closely aligned with the technical objective of "identifying and prioritizing the areas with the most potential threats." The computational process first uses a risk activation threshold (identified as...) to determine the risk activation threshold. Initial screening was conducted to ensure that computing resources were concentrated in high-risk areas;

[0142] Subsequently, the core regional comprehensive risk level The calculation formula (whose initial source is the linear weighted model in multi-attribute decision theory) will and As input. These two parameters are in parallel in the formula, categorized by their respective stress contribution weights (the stress contribution weights are labeled as...). ) and rate contribution weight (identifying the rate contribution weight as Adjustments were made to ensure a strict proportional relationship between input and output: or Any increase will quantitatively lead to a decrease in the overall regional risk level. The increase perfectly aligns with the technical objective of "higher risk, higher performance targets." Ultimately, the system's current total available energy serves as a constraint on the final decision, determining the tactical mission objectives. The intervention model ensures a balance between the effectiveness of actions and the sustainability of resources. The overall regional risk level... The output value range is limited to the interval [0,1]; this design not only provides a standardized basis for subsequent priority sorting, but also gives the output value a clear technical meaning:

[0143] When the overall risk level of the region The closer the calculated output is to 1, the higher the overall risk level of the candidate intervention area, and the greater the likelihood that it will be selected as the highest priority intervention target. This occurs because the area... and If at least one or both of these values ​​are at a high level, it indicates that the area either has severe stress concentration, is undergoing severe deliquescence and deterioration, or both, posing the greatest threat to the overall structure of the insulation layer.

[0144] When the overall risk level of the region The closer the calculated output is to 0, the lower the overall risk level of the candidate intervention area, and the lower its priority for intervention. This occurs because the area... and All are at low levels, indicating that the region is currently stable and the deterioration trend is slow, and does not pose a major threat in the short term.

[0145] In a specific embodiment of the present invention, in order to adapt to the limited precision of the digital processing system and improve computational efficiency, the system further includes a processing step: calculating the comprehensive risk level of the region. The risk level of the region is compared with a preset risk ignoring threshold; if the overall risk level of the region is... If the risk level is less than the risk ignoring threshold, then in the subsequent priority ranking steps, the overall risk level of the region will be determined. The value is forcibly set to 0. This processing method is a common technique used by those skilled in the art when implementing numerical algorithms. It aims to effectively process parameters that theoretically approach the target value infinitely in engineering practice, thereby ignoring the minimum values ​​that have no practical impact on the final result. In this embodiment, the risk ignoring threshold is set to 0.01; to further illustrate the optimization effect of the weight allocation rule on the technical effect, an example is given below. When the system is running under the "preventive maintenance" strategy, the operator sets... , At this point, the system will prioritize processing... Compared to In areas with high activity, conversely, when the system is operating under an "emergency repair" strategy, the settings are... , At this point, the system will concentrate resources for processing. The highest-level area is targeted to eliminate the most serious structural hazards as quickly as possible. To verify the effectiveness of the above technology, a specific implementation example is provided, including parameter input, data calculation, and result generation at the application level, as recorded in the table below:

[0146] Table 4. Example Study of Hierarchical Decision-Making and Collaborative Execution Control Module:

[0147] Parameter name Corner A Window sill area B Pipe penetration point in area C South wall, central section D Roof joint area E North wall shadow area F Regional average stress potential field value 0.65 0.8 0.85 0.4 0.55 0.1 Regional average deliquescence rate 0.9 0.25 0.6 0.7 0.5 0.15 Stress contribution weight) 0.4 0.4 0.4 0.4 0.4 0.4 Rate contribution weight 0.6 0.6 0.6 0.6 0.6 0.6 Regional comprehensive risk level 0.79 0.47 0.7 0.58 0.53 0.13 Total available energy of the system [kJ] 850 850 850 850 850 850 Final intervention model Powerful Mode Normal mode Powerful Mode Normal mode Normal mode Non-intervention Final priority 1 5 2 3 4 6

[0148] The table above shows the weights set to favor future trends. Under these conditions, the system's decision-making process for six different candidate intervention zones. Although "window sill zone B" and "pipe penetration zone C" have the highest current stress values ​​( The values ​​are 0.80 and 0.85 respectively, but the system does not prioritize them. Instead, the "corner A area" (where the current stress is lower but the rate of deterioration is the fastest) is prioritized. It achieved the highest overall risk level. It is assigned the highest priority "1". This fully demonstrates that the present invention can go beyond simple static threshold judgment and achieve truly forward-looking risk prediction.

[0149] The priority of all regions is based on the overall regional risk level. The numerical value is uniquely and definitively determined, eliminating the uncertainty of human intervention and achieving automated and standardized decision-making. This example sets the energy threshold for the powerful mode to 750 kJ and the energy threshold for the normal mode to 300 kJ. Due to current energy... The unit is kJ. The system selected "powerful mode" for the highest priority areas A and C, and "normal mode" for the lower priority areas. Area F was not intervened in due to its low risk, which reflects the intelligence and rationality of resource allocation.

[0150] Risk activation threshold The determination of this threshold follows a dual-standard principle. First, based on statistics, it is set as the 95th percentile of the stress potential field of the repair unit in a large number of healthy insulation layer samples to filter out normal background noise. Second, based on materials mechanics, its value must not exceed 70% of the characteristic value of the initiation stress of micro-damage in the insulation material to ensure sufficient safety margin. The smaller of the two values ​​is taken as the final threshold.

[0151] Energy threshold ( , The threshold is determined based on operational requirements and equipment energy consumption. Energy threshold for normal mode. It is set to the total energy required to support the system completing three full "normal mode" repair cycles, plus a 20% safety reserve. Powerful mode energy threshold. Set as This includes the energy required for a complete "Power Mode" repair cycle. This rule ensures that the system has sufficient energy reserves to handle multiple routine tasks that may occur afterward, before selecting a high-energy-consumption mode.

[0152] 5.1) Further explanation of the hierarchical decision-making and collaborative execution control module: The hierarchical decision-making and collaborative execution control module further includes a low-level execution agent; the low-level execution agent is configured to receive the tactical mission objective generated by the high-level decision-making agent as the carrier of the top-level strategic intent. And combined with the acoustic emission signals of microcapsule rupture that characterize the dynamics of microscopic damage, acquired in real time from the sensor network. This generates and outputs continuously adjustable blocking control commands for the path blocking unit and continuously adjustable repair control commands for the self-repair triggering unit. In this embodiment, the blocking control commands are identified as... The repair control command is identified as This decision-making model, which combines strategic commands with real-time tactical awareness, is one of the core innovations of this invention. It empowers the system to perform dynamic and refined optimal control while strictly adhering to global planning.

[0153] The low-level executive agent is trained using the Proximal Policy Optimization (PPO) algorithm. The state space of the low-level executive agent includes the tactical mission objective. and the acoustic emission signal of the microcapsule rupture Its action space is determined by the blocking control command. and repair control commands The resulting continuous control vector; the behavioral policy of this low-level executive agent is shaped by a pre-defined reward function, which is designed such that the output action conforms to the tactical mission objective in terms of both spatial location and resource consumption. The constraint can effectively lead to the rupture of the microcapsule and the acoustic emission signal. A high positive reward is obtained when the signal weakens; conversely, if the action violates the acoustic emission signal of the microcapsule rupture... If an instruction causes energy consumption to exceed its mode limits, a significant negative reward will be obtained.

[0154] The specific implementation process of the second single-step reward value used by the low-level executive agent during training and decision-making is broken down as follows:

[0155] The second single-step reward value of the lower-level executing agent is denoted as This is a dimensionless scalar value, the algebraic sum of three components, comprehensively reflecting the strategic compliance, physical effectiveness, and legality of a single-step action. The second single-step reward value. It equals the sum of the task compliance benefit and the damage repair efficiency benefit, minus the out-of-bounds penalty.

[0156] The task compliance benefit is denoted as This is a non-negative dimensionless numerical value, the value of which is related to the action's impact on the tactical mission objective. The degree of instruction matching is directly proportional to the task's compliance. The compliance reward is equal to the product of the basic reward, the spatial matching factor, and the pattern matching factor. The basic reward is set to a preset positive constant; in this example, it is set to 1.0, serving as the base score for compliance. The spatial matching factor is represented as a value between [0,1]. It is obtained by calculating the position of the actual action performed by the lower-level agent relative to the tactical mission objective. The Euclidean distance to a specified location is mapped using a Gaussian function. The factor is 1 when the distance is zero, and the factor approaches 0 as the distance increases.

[0157] The pattern matching factor is represented as a value between [0,1]. It is obtained by taking the actual output control power (blocking control command) as the value. and repair control commands (the corresponding total power) and tactical mission objectives The power limit corresponding to the specified mode (such as high power, normal, energy saving) is compared. If the limit is not exceeded, the mode matching factor is 1; if the limit is exceeded, the mode matching factor is 0.

[0158] The benefit of damage repair efficiency is denoted as This is a dimensionless value, and its value is related to the acoustic emission signal caused by the microcapsule rupture triggered by the control action. The decrease is directly proportional to the rate of decline. The aforementioned damage repair efficacy benefit... This equals the product of the performance conversion factor and the relative decline rate of the acoustic emission signal. The performance conversion factor is a preset positive constant; in this example, it is set to 2.0, used to adjust the weight of the repair effect in the total reward. The relative decline rate of the acoustic emission signal is a value between (-∞, 1], calculated as follows (before the action) -After the action Before the action The data is obtained by using an acoustic emission sensor to collect data before and after the action is performed. The integral energy value is calculated. The out-of-bounds penalty term is denoted as... This is a non-negative, infinite GINI penalty value, activated only when an explicit violation occurs. (The out-of-bounds penalty term...) The penalty is equal to the base penalty for violation multiplied by the violation flag. The base penalty is a preset, relatively large positive integer, 5.0 in this example, to ensure the penalty is strong enough to deter violations. The violation flag is a Boolean value (0 or 1). It is obtained by checking if the pattern matching factor is 0; if it is 0 (i.e., energy exceeding limits), the violation flag is set to 1; otherwise, it is 0. This embodiment includes the limitation of the low-level executive agent. Its technical effect is that it constructs a complete "strategic planning-tactical execution" closed loop, ensuring the final accurate implementation of the top-level strategic intent. The logical rationale of this mechanism is that, through the low-level executive agent as the tactical executor, its decision-making logic (whose initial source is reinforcement learning theory in operations research) is entirely determined by the second single-step reward value. Guided by. The design is key to achieving the core technological goal of "autonomous and optimized execution under strategic guidance".

[0159] The structure and parameter position relationship of this reward function ensure that the learning direction of the low-level executive agent is highly consistent with the technical objectives of this invention. The calculation formula will include task compliance benefits. and benefits of damage repair efficacy As a positive incentive, the penalty for exceeding the boundary will be... As a negative penalty item. This ensures that all actions of the lower-level execution agents must be within the tactical mission objective. Within the established framework, the strategic unity of the operation was maintained; the benefits of damage repair effectiveness were also ensured. This drives the lower-level execution agent to utilize real-time [processing / processing] while following instructions. Feedback is used to continuously optimize the blocking control commands. and repair control commands To maximize the physical effect; while the out-of-bounds penalty item This sets a red line for their behavior, ensuring the safe and compliant operation of the system.

[0160] When the second single-step reward value When the calculated output is a large positive value (in this example, "large positive value" is set to 1.0), the action of the lower-level agent is a successful, high-quality execution. This occurs because the action's... and All are positive, and A value of zero indicates that the lower-level agent not only accurately executed the instructions from the higher level, but also achieved significant repair or blocking effects at the physical level.

[0161] When the second single-step reward value When the calculated output is a large negative value (in this example, "large negative value" is set to -1.0), the action of the lower-level executing agent is a failed and penalized violation. This occurs because of the out-of-bounds penalty term. Once activated, its massive negative value dominates the entire reward signal, even though the action unexpectedly brings some damage repair efficiency benefits. The system will also determine that it is an error that must be avoided in future strategies. To verify the above technical effect, a specific implementation example of parameter substitution, data calculation and result derivation at the application level is provided and recorded in the following table:

[0162] Table 5: Example studies of low-level executive agents:

[0163]

[0164] The table above clearly illustrates the behavior evaluation mechanism of low-level executive agents in different scenarios.

[0165] Comparing Scene 1 and Scene 3, although the actions in Scene 3 also resulted in a physical repair effect ( However, due to an incorrect execution location (space matching factor of 0), it caused... The reward is 0, resulting in a total reward far lower than that of scenario 1, where perfect execution was achieved. This demonstrates that the reward mechanism in this invention effectively ensures that the low-level execution agent "does the right thing." In scenario 4, the low-level execution agent exceeded the power limit of the energy-saving mode in pursuit of a possible effect, immediately triggering a huge over-limit penalty. This results in a final reward of -5.00. This strong negative feedback, through the PPO algorithm's training process, effectively prevents similar future behavior from the low-level agent. Scenario 6 illustrates a special case where the low-level agent perfectly follows the instructions, but due to unknown physical reasons, its actions actually exacerbate the damage. (rising), leading to The reward function of this invention can capture this situation and provide a relatively low reward, guiding the low-level executive agent to explore more effective control methods in the future. This demonstrates the system's adaptability to the complexity of the real world. In scenario 5, the low-level executive agent simultaneously activates the repair and blocking units, achieving excellent repair results. The system received a reward second only to perfect execution, demonstrating its ability to learn and encourage complex and efficient strategies for multi-unit collaborative work.

[0166] Example 2: A method for predicting the deliquescence rate of building insulation materials based on neural networks, wherein the method is used to execute the neural network-based building insulation material deliquescence rate prediction system, and the specific steps include:

[0167] Step S1: Collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer;

[0168] Step S2: Based on the macroscopic state data, use a spatiotemporal graph network model to predict and generate a spatial distribution map representing the deliquescence rate of the insulation layer at future times;

[0169] Step S3: Based on the spatial distribution map of the deliquescence rate at the future time, generate a stress potential field of the repair unit that characterizes the risk of microscopic damage inside the insulation layer; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit.

[0170] Step S4: Based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate, a hierarchical reinforcement learning model is used to generate tactical mission objectives to guide repair or blocking operations;

[0171] Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

[0172] It should be noted that all calculation formulas in this application employ regression analysis, including but not limited to machine learning algorithms, to deeply analyze the collected parameters and identify their natural trends and interrelationships. Specialized software, such as Python's Scikit-learn library or the R language, is used to automatically generate mathematical models that match the data. Then, cross-validation and other methods are used to objectively evaluate the model performance, and continuous feedback and optimization are combined to ensure that the created formulas truly reflect the inherent laws of the data, thereby guaranteeing their effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula undergo dimensionless processing within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless processing techniques include, but are not limited to, min-max-normalization and Z-score standardization.

[0173] The technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of this invention.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A neural network-based system for predicting the deliquescence rate of building insulation materials, applied to the insulation layer of buildings, characterized in that, The system is configured to perform the following operations: Data acquisition module: used to collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer; Deliquescence rate prediction module: Based on the macroscopic state data, a spatiotemporal graph network model is used to predict and generate a spatial distribution map of the deliquescence rate of the insulation layer at future times; Risk potential field generation module: used to generate a repair unit stress potential field characterizing the risk of microscopic damage inside the insulation layer based on the spatial distribution map of the deliquescence rate at the future time; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit. Hierarchical decision-making and collaborative execution control module: used to generate tactical mission objectives to guide repair or blocking operations based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate using a hierarchical reinforcement learning model; Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

2. The neural network-based building insulation material deliquescence rate prediction system according to claim 1, characterized in that: The data acquisition module is configured as follows: at multiple preset acquisition points on the insulation layer, macroscopic state data characterizing the macroscopic and slow chemical degradation process of the insulation layer, and microscopic damage data characterizing the insulation layer under microscopic and sudden physical damage events are acquired; wherein, chemical bond vibration frequency shift and microcapsule rupture acoustic emission signal are extracted from the macroscopic state data and microscopic damage data, respectively. The data acquisition module is further configured to perform weighted analysis on the chemical bond vibration frequency offset and the acoustic emission signal of microcapsule rupture at each acquisition location to calculate the comprehensive degradation risk index. The higher the output of the comprehensive deterioration risk index, the closer the insulation layer is to the critical state of complete failure.

3. The neural network-based building insulation material deliquescence rate prediction system according to claim 2, characterized in that: The deliquescence rate prediction module is configured as follows: Based on the time series of macroscopic state data, including the comprehensive deterioration risk index, obtained from the data acquisition module, a causal inference spatiotemporal graph convolutional network is used to predict and generate a spatial distribution map of the deliquescence rate of the insulation layer at future times. The causal inference spatiotemporal graph convolutional network includes graph convolutional layers, gated recurrent unit layers, and attention mechanism layers; The spatial distribution map of deliquescence rate includes the node-predicted deliquescence rate represented by each collection point. The higher the predicted deliquescence rate output of a node, the higher the risk of deliquescence and the more severe the degradation trend will be for the node corresponding to that sampling location in the future.

4. The neural network-based building insulation material deliquescence rate prediction system according to claim 3, characterized in that: The risk potential field generation module is configured as follows: Based on the spatial distribution map of deliquescence rate obtained from the deliquescence rate prediction module, a physical proxy model is used to generate a repair unit stress potential field that characterizes the risk of microscopic damage inside the insulation layer. The physical proxy model is a pre-trained deep neural network whose input is the spatial distribution map of the deliquescence rate and whose output is the stress potential field of the repair unit.

5. The neural network-based building insulation material deliquescence rate prediction system according to claim 3, characterized in that: The conversion process from "spatial distribution map of deliquescence rate" to "stress potential field of repair unit" is achieved through an explicit weighted model.

6. The neural network-based building insulation material deliquescence rate prediction system according to claim 5, characterized in that: The greater the stress potential field of the repair unit, the higher the risk of microscopic damage to the node, the greater the stress borne by the pre-embedded repair unit, and the more likely it is to be prioritized for repair or blocking.

7. The neural network-based building insulation material deliquescence rate prediction system according to claim 6, characterized in that: The hierarchical decision-making and collaborative execution control module is further configured as follows: Based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate, a tactical mission objective containing repair priorities and intervention modes is generated through a high-level decision-making agent. The tactical mission objective is defined as a structured data object used to convey strategic intent to lower-level execution units and set the framework for their actions.

8. The neural network-based building insulation material deliquescence rate prediction system according to claim 7, characterized in that: The high-level decision-making agent is further configured to take the system’s current total available energy as part of its state input, and its action space is defined as a set of discrete, structured tactical mission objectives. The generation process of the tactical mission objectives is guided by a preset reward function. The design principle of the reward function is to give positive incentives to actions that can cause a preset decrease in the stress potential field of the repair unit, while applying negative penalties to actions that cause excessive consumption of system energy. In the stress potential field diagram of the repair unit, the connected regions whose corresponding values ​​of the stress potential field of the repair unit exceed a preset risk activation threshold are characterized as "candidate intervention areas". For each identified candidate intervention area, a comprehensive regional risk score is calculated to assess the urgency of intervention. The higher the calculated output of the regional comprehensive risk level, the higher the comprehensive risk level of the candidate intervention area, and the more likely it is to be selected as the highest priority intervention target.

9. The neural network-based building insulation material deliquescence rate prediction system according to claim 8, characterized in that: The hierarchical decision-making and collaborative execution control module further includes a low-level execution agent; The low-level executive agent is configured to receive the tactical mission objective generated by the high-level decision-making agent as a carrier of the top-level strategic intent, and combine it with the microcapsule rupture acoustic emission signal characterizing the microscopic damage dynamics acquired in real time from the sensor network, in order to generate and output continuously adjustable blocking control commands for the path blocking unit and continuously adjustable repair control commands for the self-repair trigger unit. The state space of the low-level executive agent includes the tactical mission objective and the acoustic emission signal of the microcapsule rupture, while its action space is a continuous control vector composed of the blocking control command and the repair control command.

10. A method for predicting the deliquescence rate of building insulation materials based on neural networks, characterized in that: The method is used to execute the neural network-based building insulation material deliquescence rate prediction system according to any one of claims 1-9, and the specific steps include: Step S1: Collect macroscopic state data and microscopic damage data related to the deliquescence process of the insulation layer; Step S2: Based on the macroscopic state data, use a spatiotemporal graph network model to predict and generate a spatial distribution map representing the deliquescence rate of the insulation layer at future times; Step S3: Based on the spatial distribution map of the deliquescence rate at the future time, generate a stress potential field of the repair unit that characterizes the risk of microscopic damage inside the insulation layer; the region with a higher deliquescence rate in the spatial distribution map corresponds to the region with a higher potential field value in the stress potential field of the repair unit. Step S4: Based on the stress potential field of the repair unit and the spatial distribution map of the deliquescence rate, a hierarchical reinforcement learning model is used to generate tactical mission objectives to guide repair or blocking operations; Based on the tactical mission objective and the microscopic damage data, generate and output coordinated control commands for the path blocking unit and self-repair trigger unit preset in the insulation layer.

Images