Cooling rate control method based on hot stress prediction during grinding wheel sintering cooling stage

Abstract

This invention provides a cooling rate control method based on thermal stress prediction during the sintering and cooling stage of grinding wheels. It generates standardized multimodal data through spatiotemporal registration and noise filtering, identifies key nodes in thermal stress evolution using finite element simulation, constructs a thermal stress response fingerprint spectrum specific to the material structure, and encodes it as a structured feature vector. This enables real-time characterization of the stress state without relying on meshes or material parameter libraries. A historical knowledge base is used to quickly map the feature vector to stress trends, enabling prediction and dynamic calibration of abnormal disturbances. Closed-loop error compensation generates high-confidence stress trend fragment flow, driving low-latency automatic adjustment of the cooling system. This invention improves the accuracy of thermal stress prediction and the efficiency of cooling control response, ensuring the performance and safety of the grinding wheel.

Description

Technical Field

[0001] This invention relates to the field of thermal stress prediction and control technology in the sintering and cooling process of precision ceramic grinding wheels, and particularly to a cooling rate control method based on thermal stress prediction during the sintering and cooling stage of grinding wheels. Background Technology

[0002] Currently, in the field of thermal stress prediction and cooling control technology for the sintering and cooling process of precision ceramic grinding wheels, mainstream technologies generally employ high-fidelity finite element numerical simulation, full-scale AI training models based on multimodal sensor data, or integrated end-to-end process control closed loops in the context of digital twins. These methods typically rely on accurate modeling of the grinding wheel material properties and a comprehensive digital description of the cooling environment, aiming for continuous reconstruction of the thermodynamic field and real-time extrapolation of the global stress distribution. In recent years, hybrid prediction schemes combining physics-driven and data-driven approaches have gradually become a trend, attempting to improve the accuracy of theoretical predictions while also considering the real-time response during field deployment. However, the existing technological system still faces the following prominent problems; On the one hand, thermal stress prediction based on traditional finite element models generally suffers from high computational resource consumption and limited simulation step size in high-dimensional geometry and complex working conditions, making it difficult to support sub-second or sub-millisecond online control requirements. Whenever environmental disturbances or sudden changes in working conditions occur during the cooling process (such as airflow disturbances or accidental opening of the furnace door), the simulation boundaries and parameters need to be re-initialized, which can easily lead to a serious lag in the output of the prediction model, ultimately limiting the timeliness and convergence of thermal stress prediction and cooling strategy control. On the other hand, while end-to-end "black box" AI models introduced in recent years have improved their empirical fitting capabilities through large-scale training on historical samples, they often suffer from complex model tuning, large generalization hyperparameter spaces, high dependence on model transparency and physical interpretability, and insufficient transferability to variations in different grinding wheel specifications or materials. Especially under dynamic operating conditions, "black box" models require full retraining or high-frequency weight updates, further lengthening the response chain and failing to meet the needs of rapid stress trend prediction and timely decision-making in the process field. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a cooling rate control method based on thermal stress prediction during the sintering and cooling stage of grinding wheel.

[0004] The technical solution of this invention is implemented as follows: a cooling rate control method based on thermal stress prediction during the sintering and cooling stage of a grinding wheel, comprising: S1: Acquire multiple batches of high spatiotemporal resolution temperature field scanning data and surface strain synchronous acquisition data of precision ceramic grinding wheels under typical cooling conditions, and use the temperature field scanning data and surface strain synchronous acquisition data as the original input primitives for constructing thermal stress response fingerprint spectrum; S2: Based on the original input primitives and the finite element simulation inversion results, identify five key response nodes: the initial thermal relaxation inflection point, the peak time of the circumferential gradient, the radial stress polarity reversal point, the starting position of the surface compressive stress saturation segment, and the pre-peak interval of the core tensile stress. Generate a thermal stress response fingerprint spectrum that characterizes the topological features of the stress evolution path under the coupling effect of heat conduction and thermal expansion. S3: Encode the thermal stress response fingerprint spectrum into a one-dimensional structured vector containing the relative position offset of each key response node on the standardized time axis, the local slope change rate, and the neighborhood fluctuation entropy value, and generate a spectrum feature vector for real-time decoding without the need for mesh generation or material parameter library dependence. S4: Based on historical real cooling process data, construct a map-stress trend mapping knowledge base consisting of map feature vectors and subsequent key region stress envelope annotations, and input the currently measured map feature vectors into the map-stress trend mapping knowledge base to perform nearest neighbor search and local weighted interpolation operations to generate initial stress trend inference results; S5: Monitor whether there are abnormal disturbances caused by sudden changes in ambient airflow or accidental opening of the furnace door in the actual cooling curve. If an abnormal disturbance is detected, the graph residual analysis module is triggered to calculate the structural similarity deviation between the current graph feature vector and the knowledge base sample, and generate an incremental fusion correction instruction for the neighboring sample with the largest deviation. S6: According to the incremental fusion correction instruction, the stress trends corresponding to the three neighboring samples with the largest deviation in the initial stress trend extrapolation result are locally corrected to generate corrected stress trend output data with dynamic calibration capability. S7: Encapsulate the corrected stress trend output data into a trend fragment stream containing only the stress change slope segment and confidence interval within the next two seconds, and generate a lightweight control command input source adapted to sub-millisecond decision cycles. S8: Receive the trend fragment flow and analyze the stress change slope segment and confidence interval in it. Based on the analysis results, adjust the cooling rate execution parameters of the cooling medium in real time to complete the low-latency closed-loop control action between thermal stress prediction and control strategy.

[0005] The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel provided by this invention has the following beneficial effects: (1) This invention proposes a new paradigm that transforms the thermal stress evolution process from “continuous physical field numerical reconstruction” to “key temporal pattern recognition”, which significantly improves the prediction response speed and engineering practicality. By extracting five key response nodes with strong temporal stability and structural specificity on the cooling curve—including the initial thermal relaxation inflection point, the peak moment of the circumferential gradient, and the radial stress polarity reversal point—a “thermal stress response fingerprint spectrum” that does not depend on specific physical quantity values ​​but can characterize the topological features of the stress evolution path is constructed, realizing a highly abstract expression of complex thermodynamic behavior. The spectrum encodes the relative position offset, local slope change rate, and neighborhood fluctuation entropy of each node on the standardized time axis in the form of a 15-dimensional structured vector. It can be quickly decoded and generated by only the measured surface temperature curve, without calling the material library or performing spatial discretization processing, which greatly reduces the data preprocessing and feature extraction overhead, and compresses the delay of the front-end perceived feature output to the order of tens of milliseconds. (2) This invention designs a closed-loop prediction architecture that coordinates a lightweight map matching engine with a dynamic calibration mechanism, effectively ensuring the robustness and adaptability of the system under unsteady conditions. The engine abandons the complex physical equation solving process and instead relies on a map-stress trend mapping knowledge base constructed from 200 sets of real cooling processes. It uses nearest neighbor retrieval combined with local weighted interpolation to realize the rapid deduction of stress trends, avoiding the large amount of online computing resources required by deep neural networks or reinforcement learning, and ensuring that a single match can be completed within 10ms. When there are abnormal disturbances such as sudden changes in environmental airflow or opening of the furnace door, the system automatically starts the map residual analysis module to evaluate the structural similarity deviation between the current map and historical samples, and performs incremental fusion correction on the stress trends corresponding to the three nearest samples with the largest deviation. It only updates the local model output without triggering global retraining, which maintains the continuity of prediction and enhances the adaptability to external disturbances. More importantly, the system transmits a "trend fragment stream" to the downstream controller—that is, it only pushes a segment of the stress change slope and its confidence interval within the next 2 seconds each time, rather than a complete prediction curve. This allows the control strategy module to complete the rate adjustment decision in sub-100ms time, which fully meets the dual stringent requirements of real-time performance and stability in the cooling process of precision ceramic grinding wheels. Attached Figure Description

[0006] Figure 1 The flowchart shows the cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to the present invention. Figure 2 This is a sub-flowchart of the cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to the present invention. Figure 3 This is another sub-flowchart of the cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. like Figure 1 As shown, this invention provides a cooling rate control method based on thermal stress prediction during the sintering cooling stage of a grinding wheel, specifically including: S1: Acquire multiple batches of high spatiotemporal resolution temperature field scanning data and surface strain synchronous acquisition data of precision ceramic grinding wheels under typical cooling conditions, and use the temperature field scanning data and surface strain synchronous acquisition data as the original input primitives for constructing thermal stress response fingerprint spectrum; S2: Based on the original input primitives and the finite element simulation inversion results, identify five key response nodes: the initial thermal relaxation inflection point, the peak time of the circumferential gradient, the radial stress polarity reversal point, the starting position of the surface compressive stress saturation segment, and the pre-peak interval of the core tensile stress. Generate a thermal stress response fingerprint spectrum that characterizes the topological features of the stress evolution path under the coupling effect of heat conduction and thermal expansion. S3: Encode the thermal stress response fingerprint spectrum into a one-dimensional structured vector containing the relative position offset of each key response node on the standardized time axis, the local slope change rate, and the neighborhood fluctuation entropy value, and generate a spectrum feature vector for real-time decoding without the need for mesh generation or material parameter library dependence. S4: Based on historical real cooling process data, construct a map-stress trend mapping knowledge base consisting of map feature vectors and subsequent key region stress envelope annotations, and input the currently measured map feature vectors into the map-stress trend mapping knowledge base to perform nearest neighbor search and local weighted interpolation operations to generate initial stress trend inference results; S5: Monitor whether there are abnormal disturbances caused by sudden changes in ambient airflow or accidental opening of the furnace door in the actual cooling curve. If an abnormal disturbance is detected, the graph residual analysis module is triggered to calculate the structural similarity deviation between the current graph feature vector and the knowledge base sample, and generate an incremental fusion correction instruction for the neighboring sample with the largest deviation. S6: According to the incremental fusion correction instruction, the stress trends corresponding to the three neighboring samples with the largest deviation in the initial stress trend extrapolation result are locally corrected to generate corrected stress trend output data with dynamic calibration capability. S7: Encapsulate the corrected stress trend output data into a trend fragment stream containing only the stress change slope segment and confidence interval within the next two seconds, and generate a lightweight control command input source adapted to sub-millisecond decision cycles. S8: Receive the trend fragment flow and analyze the stress change slope segment and confidence interval in it. Based on the analysis results, adjust the cooling rate execution parameters of the cooling medium in real time to complete the low-latency closed-loop control action between thermal stress prediction and control strategy.

[0009] Step S1: Acquire multiple batches of high spatiotemporal resolution temperature field scanning data and surface strain synchronous acquisition data of precision ceramic grinding wheels under typical cooling conditions, and use the temperature field scanning data and surface strain synchronous acquisition data as the original input primitives for constructing a thermal stress response fingerprint spectrum. Specifically, this includes: S1.1: Perform time reference calibration on the infrared thermal imaging array sensor and fiber optic strain sensing network deployed in the precision ceramic grinding wheel sintering furnace to eliminate clock drift error between multi-source heterogeneous sensors and generate a synchronous trigger signal with a unified timestamp. The time reference of the infrared thermal imaging array sensor and the fiber optic strain sensing network deployed in the precision ceramic grinding wheel sintering furnace is uniformly calibrated during the cooling stage. Independent signal pulse capture operation is performed to extract the original period parameters of the internal clock signals of the two types of sensors. The period parameters are input into the high-precision clock drift estimation module to generate drift data. The relative offset value is calculated by subtracting the drift data from the current time reference of each sensor. The sampling trigger time of the infrared thermal imaging array sensor is adjusted by compensating for the offset value, thereby aligning the acquisition window of the fiber optic strain sensing network in physical time and achieving dynamic synchronization. The adjusted trigger time is input into the unified timestamp generation module, which integrates the preset global time base signal in the furnace and outputs a synchronous trigger signal with a global unified timestamp. The synchronous trigger signal is bound to the start conditions of subsequent temperature field scanning and strain capture to ensure that the data acquisition of multi-source heterogeneous sensors remains consistent within the millisecond error range; By using the above processing method, the original sensor clock signal from the previous step is converted into a synchronous trigger signal with a unified timestamp, thereby improving the time base accuracy and optimizing the sampling synchronization of multi-source data acquisition. For example, an infrared thermal imaging array with a resolution of 640×480 pixels and a fiber optic strain sensing network with a sampling frequency of 50Hz are deployed inside a precision ceramic grinding wheel sintering furnace. The period parameters of their internal clock signals are extracted, where the clock period of the infrared array is 20ms and the clock period of the strain network is 20.0008ms. The drift is calculated to be 0.0008ms by a high-precision drift estimation module. The relative offset value is calculated and compensated using the following formula: in, For the infrared array clock cycle, The clock period of the fiber Bragg grating strain network. This is the relative offset value. The calculated offset value of 0.0008ms is applied to compensate for the sampling trigger time of the infrared array. After global synchronization time base fusion, a unified timestamp synchronization trigger signal is generated. This signal is used for the subsequent synchronous start of temperature field scanning and strain capture. It has been verified that the acquisition time difference between the two types of sensors is reduced to no more than 0.001ms, which significantly improves the spatiotemporal alignment accuracy of multi-source data sources. S1.2: Based on the synchronous trigger signal, perform high-frequency sampling processing on the radiation energy distribution of the grinding wheel surface during the cooling process to obtain a discrete temperature matrix sequence containing spatial coordinate information, and generate a high spatiotemporal resolution temperature field scanning raw data stream; S1.3: The synchronous trigger signal is used to drive the fiber optic grating strain sensing network to perform parallel capture processing of micro-deformation on the surface of the grinding wheel, so as to extract the wavelength offset sequence reflecting the thermal expansion behavior of the material and generate the original data stream of synchronous acquisition of surface strain. S1.4: Perform multi-dimensional spatiotemporal registration algorithm processing on the original data stream of the high spatiotemporal resolution temperature field scanning and the original data stream of the surface strain synchronous acquisition to establish a precise mapping relationship between the temperature gradient field and the strain evolution field at the pixel level and the measurement point level, and generate a spatiotemporally aligned multimodal fusion dataset. S1.5: Perform outlier removal and noise smoothing filtering based on the spatiotemporally aligned multimodal fusion dataset to remove spurious signal components introduced by environmental airflow disturbances or equipment vibrations, and generate standardized original input primitives for constructing thermal stress response fingerprints.

[0010] Step S2: Based on the original input primitives and the finite element simulation inversion results, identify five key response nodes: the initial thermal relaxation inflection point, the peak time of the circumferential gradient, the radial stress polarity reversal point, the starting position of the surface compressive stress saturation segment, and the precursor interval before the core tensile stress peaks. Generate a thermal stress response fingerprint spectrum characterizing the topological features of the stress evolution path under the coupling effect of heat conduction and thermal expansion. Specifically, this includes: S2.1: Perform spatiotemporal alignment processing on the high spatiotemporal resolution temperature field scanning data and surface strain synchronous acquisition data in the original input primitive to eliminate sensor sampling clock deviation and spatial coordinate mapping error, and generate a synchronous multi-source monitoring dataset with a unified spatiotemporal reference. S2.2: Perform transient thermo-structural coupled finite element simulation inversion operation based on the synchronous multi-source monitoring dataset to reconstruct the three-dimensional temperature gradient distribution and equivalent stress tensor field of the non-measured area inside the grinding wheel, and generate simulation inversion result data containing global thermodynamic state information; Based on the temperature field scanning matrix sequence and surface strain wavelength offset sequence in the synchronous multi-source monitoring dataset, a set of boundary conditions and initial condition parameters required for transient thermal-structural coupled finite element simulation is constructed. A three-dimensional finite element model is established based on the actual measured geometric dimensions of the grinding wheel, the material anisotropy coefficient, and the thermophysical parameters. A unit mesh with sufficient density is divided in the model to ensure the analytical accuracy of the internal non-measured regions. Temperature field data is mapped to mesh surface nodes and set as transient thermal input. At the same time, strain data is converted into equivalent thermal expansion coefficient constraints and applied to the corresponding structural nodes to achieve thermal-structural coupling. The transient operation, controlled by time step, is performed using a finite element solver to calculate the temperature gradient distribution and corresponding stress tensor field of the nodes inside the grinding wheel within each time step. The heat conduction equation and the thermal stress balance equation are used in a coordinated iterative process to converge the global thermodynamic state. The temperature gradient field is calculated using the following heat conduction equation: in, Where is the thermal diffusivity, and T is the node temperature. The rate of change of temperature with respect to time, For the temperature field Laplace operator; Using the thermal stress calculation formula: in, Where E is thermal stress and E is the elastic modulus. Use the initial reference temperature; Equivalent stress tensor field data are generated using the above formula; the three-dimensional temperature gradient distribution and corresponding stress tensor field at each time step are output as global thermodynamic state information for use by the subsequent automatic positioning algorithm of key response nodes. Through transient thermo-structural coupled finite element simulation inversion, the synchronous monitoring data from the previous step is transformed into a three-dimensional temperature gradient distribution and equivalent stress tensor covering the non-measured area inside the grinding wheel, realizing accurate reconstruction of thermal stress in physically invisible areas, and providing a complete thermodynamic data foundation for subsequent node feature extraction. For example, under typical cooling conditions of a precision ceramic grinding wheel with a diameter of 300 mm and a thickness of 20 mm, the finite element model is divided into 120,000 four-node thermal-structural coupled elements, and the material thermal diffusivity is... Set to 1.2×10 -5 m² / s, elastic modulus E set to 95 GPa, reference initial temperature The temperature was set at 800 K. In the simulation, the time step was 0.05 s. The measured surface temperature matrix was interpolated and mapped to the model surface nodes. The input strain data was converted into local thermal expansion coefficient boundaries and applied to the corresponding nodes. The temperature gradient distribution at each time step was calculated according to the heat conduction formula, and a global stress tensor field was generated by combining it with the thermal stress formula. At the end of the simulation, the stress in the peak region of the internal radial temperature gradient was 210 MPa, and the stress in the peak region of the circumferential gradient was 185 MPa. This result significantly improved the accuracy of internal stress prediction and achieved high-resolution virtual reconstruction of thermal stress in non-measured regions. S2.3: Using the time series derivative characteristics and extreme value detection algorithm in the simulation inversion results data, the abrupt change point of thermal relaxation rate, the peak moment of circumferential stress gradient, the instant of radial stress polarity reversal, the starting position of the saturation segment of surface compressive stress, and the precursor interval of core tensile stress peak are automatically located to generate a set of time series coordinates for five key response nodes. S2.4: Extract the relative position offset, local slope change rate and neighborhood fluctuation entropy of each node on the standardized time axis based on the time-series coordinate set of the five key response nodes, so as to quantify the stress evolution path topology under the coupling effect of heat conduction and thermal expansion, and generate a thermal stress response fingerprint spectrum with material-structure specificity. Based on the time-series coordinate set of the five key response nodes output by the preceding sub-step S2.3, a calculation model for the relative position offset of the corresponding nodes on the standardized time axis is established to eliminate the time-scale influence caused by the difference in cooling cycles of different batches. In this model, the total cooling time under typical cooling conditions is used as a normalization factor. The difference between the time sequence coordinates of each node and the timestamp of the cooling start reference point is proportionally converted to form a position offset data sequence under a unified time reference. For the aforementioned relative position offset data sequence, combined with the temperature field scan curves of the corresponding node neighborhood in the multimodal fusion dataset, a sliding window differentiation process is performed to calculate the instantaneous rate of change of the temperature curve at each node. This rate of change aims to measure the intensity of heat conduction rate at stress evolution inflection points, and is expressed by the formula: in, Represents the temperature value of the node's neighborhood. This is the current timestamp. The sliding window step size, This represents the rate of change of the local slope. The calculated set of local slope change rates is used to estimate Shannon entropy values ​​in the neighborhood of each node to form a fluctuation entropy index that reflects the non-uniformity of the material's microstructure and the degree of random perturbation in the stress evolution path. The Shannon entropy calculation formula is as follows: in, For Shannon entropy, Let be the probability distribution of the local slope values ​​in the neighborhood of the i-th node; The above-mentioned standardized time axis relative position offset, local slope change rate and neighborhood fluctuation entropy value are spliced ​​together according to the order of each node and encapsulated in a structured manner to form a thermal stress response fingerprint spectrum that can quantify the topological morphology of stress evolution path under the coupling effect of heat conduction and thermal expansion. This processing method transforms the temporal location results from the previous step into multidimensional feature data containing time location, conduction rate, and perturbation complexity, thereby enabling an accurate description of the material-structure specific stress evolution mode. For example, in a batch of precision ceramic grinding wheel cooling conditions with a cooling cycle of 120 seconds, the time coordinates of the five key nodes are 15 seconds, 40 seconds, 65 seconds, 90 seconds, and 110 seconds, respectively. Dividing the time difference of each node by the total cooling duration yields standardized offsets of 0.125, 0.333, 0.542, 0.75, and 0.917, respectively. The sampling step size is [not specified in the original text]. If the temperature drops from 850°C to 848°C in 0.5 seconds, then the rate of change of the local slope is... Based on the frequency distribution of the slope values ​​in the neighborhood, the Shannon entropy H is calculated to be 1.75 bits. The normalized offsets, corresponding local slope change rates, and fluctuation entropy values ​​of all five nodes are encapsulated into a 15-dimensional feature vector in node order, forming the thermal stress response fingerprint spectrum for this batch. This spectrum is used for millisecond-level stress trend extrapolation in subsequent retrieval and matching, effectively reducing the response delay between thermal stress prediction and control strategies, and significantly improving the real-time performance and robustness of cooling rate decisions. S2.5: Perform cross-batch working condition consistency verification and outlier removal on the thermal stress response fingerprint spectrum to ensure the stability of the spectrum features under different cooling rates and environmental disturbances, and generate a standard thermal stress response fingerprint spectrum that can be finally encoded into a structured vector.

[0011] like Figure 2As shown, step S3 involves encoding the thermal stress response fingerprint into a one-dimensional structured vector containing the relative position offsets of each key response node on the standardized time axis, the rate of change of local slope, and the entropy of neighborhood fluctuations. This generates a feature vector for real-time decoding that does not require mesh generation or material parameter library dependency. Specifically, this includes: S3.1: The timestamp data of five key response nodes in the thermal stress response fingerprint spectrum, namely the inflection point of initial thermal relaxation, the peak time of circumferential gradient, the point of radial stress polarity reversal, the starting position of the saturation segment of surface compressive stress, and the pre-peak interval of core tensile stress, are normalized and mapped based on the total cooling time under typical cooling conditions to eliminate the time scale influence caused by the difference in cooling cycles of different batches of grinding wheels, and generate a standardized time axis relative position sequence with a unified time reference. S3.2: Based on the standardized time axis relative position sequence, the sliding window differential algorithm is used to perform local slope calculation on the temperature field scanning data in the neighborhood of each key response node, so as to quantify the instantaneous change intensity of heat conduction rate at the stress evolution inflection point and generate a set of local slope change rates characterizing the severity of thermal shock. S3.3: For the fluctuation data of each node neighborhood in the set of local slope change rates, the Shannon entropy estimation algorithm is applied to perform neighborhood fluctuation entropy value calculation and processing to measure the degree of random disturbance and nonlinear complexity of the thermal stress evolution path at key turning points, and generate neighborhood fluctuation entropy value index that reflects the non-uniformity of the material microstructure. The input conditions include the set of local slope change rates output by S3.2 and the normalized time axis relative position sequence of key response nodes. The temperature field scan data of each node's neighborhood is used as the original numerical matrix for the fluctuation analysis. For the fluctuation data matrix in the neighborhood of each key response node, high-frequency component separation processing is performed to eliminate low-frequency trends and retain random perturbation components that reflect the non-uniformity of the material's microstructure. Based on the separated fluctuation components, a probability distribution function of the node neighborhood is constructed, and probability estimation is achieved by accumulating the frequency sequence through the sampling window, which prepares the input for subsequent entropy calculation. The Shannon entropy estimation algorithm is used to calculate the information content of the neighborhood probability distribution function of each node, and the entropy value of each node is normalized and mapped to adjust the entropy value distribution between different nodes to a uniform range so that the numerical scale in the feature vector is consistent. The normalized entropy index, together with the standardized time axis relative position sequence and the set of local slope change rates, forms a complete combination of topological features; By calculating the entropy value of neighborhood fluctuations, the set of local slope change rates is transformed into a quantitative index that reflects the nonlinear complexity of the thermal stress evolution path, thus realizing the measurement of material microstructure nonuniformity without relying on mesh generation or material parameter library. For example, during the cooling process of a precision ceramic grinding wheel, the temperature field scanning data in the neighborhood of the radial stress polarity reversal point is sampled with a window length of 50ms. After separating the fluctuation components, the data is divided into 20 discrete intervals. The frequency is accumulated to obtain the probability distribution of each interval. For example, the probability of interval 1 is 0.08, the probability of interval 2 is 0.12, and so on. The probabilities of each interval are then calculated. These probabilities are then applied to the Shannon entropy formula. For example, for interval 1, the contribution to the entropy value is calculated as follows: The cumulative contribution from all intervals yielded an entropy value of 4.15 bits. This entropy value was normalized to the [0,1] interval across the entire node range and incorporated into the 15-dimensional graph feature vector as a node neighborhood fluctuation entropy index. Verification showed that this index significantly improved under abrupt changes in cooling rate, indicating its ability to sensitively reflect the degree of random disturbance in the thermal stress evolution path at key turning points, effectively enhancing the rapid response performance of the stress prediction model. S3.4: The standardized time axis relative position sequence, local slope change rate set and neighborhood fluctuation entropy index are serially spliced ​​and vectorized according to the preset feature dimension sorting rules to construct a compact data structure containing fifteen-dimensional feature information and generate a one-dimensional structured vector for characterizing the topological features of the thermal stress evolution path. The standardized time axis relative position sequence, local slope change rate set, and neighborhood fluctuation entropy index obtained by processing S3.1 to S3.3 are used as combined inputs. The index position is calculated according to the preset feature dimension sorting rules to determine the specific arrangement order of various features in the final vector. The standardized timeline relative position sequence is indexed and mapped according to the order of five key response nodes, and placed in the first to fifth dimensions of the feature vector to ensure that the time reference information occupies the position first, so that the subsequent retrieval engine can quickly locate it. The set of local slope change rates is arranged sequentially according to the node correspondence order and placed in the sixth to tenth dimensions of the feature vector. The instantaneous change intensity feature of heat conduction rate is kept in a one-to-one correspondence with the time position feature, so as to realize the temporal proximity storage of features of the same node. The neighborhood fluctuation entropy index is arranged sequentially according to the corresponding node and placed in the eleventh to fifteenth dimensions of the feature vector to supplement the characterization of the random perturbation characteristics of the stress evolution path at key turning points; The three feature groups are sequentially written into a one-dimensional vector buffer using a serial splicing process. A vectorization encapsulation algorithm is used to assign fixed feature index labels to ensure that the position of each dimension in the feature space is stable and predictable. During the encapsulation process, a unified conversion of numerical types is performed, encoding all feature values ​​into the same floating-point precision to reduce the type conversion overhead of the computing engine during matching and retrieval. By sorting and serially splicing the feature dimensions, the result of the previous step is transformed into a one-dimensional structured vector containing fifteen feature information and a compact structure, thus realizing a complete representation of the topological features of the thermal stress evolution path in a low-dimensional space. For example, in a typical cooling condition of a precision ceramic grinding wheel, the standardized time axis relative position sequences of the five key response nodes are 0.12, 0.37, 0.55, 0.78, and 0.91, respectively; the local slope change rate sets are -2.35, 4.12, -1.87, 3.54, and -0.98, respectively; and the neighborhood fluctuation entropy indices are 0.45, 0.62, 0.51, 0.73, and 0.49, respectively. The three feature groups are arranged sequentially according to node order and given floating-point precision format, and then serially concatenated to obtain the vector buffer [0.12, 0.37, 0.55, 0.78, 0.91, -2.35, 4.12, -1.87, 3.54, -0.98, 0.45, 0.62, 0.51, 0.73, 0.49]. In the encapsulation algorithm, the vector is labeled with feature dimensions 0 to 14 according to a preset indexing rule, and the corresponding node feature index labels remain unchanged. When this vector is used as input to the graph matching engine, the retrieval latency is significantly reduced and the matching accuracy is greatly improved. The verification results show that millisecond-level stress trend inference can be achieved in the digital twin simulation environment, and the stress state can be independently characterized without relying on mesh generation or material parameter library. S3.5: Perform mesh-free dependency verification and material parameter decoupling verification on the generated one-dimensional structured vector to ensure that the vector can independently characterize the stress state by relying only on the measured surface temperature curve, and finally output a spectrum feature vector that is adapted to the real-time decoding requirements of the spectrum matching engine. A grid-free verification model is established for the generated one-dimensional structured vector. The feature values ​​of each dimension of the input vector are used as the verification input. The preset spatial mapping evaluator is called to calculate the reconstruction error of the vector under different grid partitioning assumptions and verify its stability under arbitrary grid generation strategies. A threshold comparison process is performed on the above reconstructed error sequence to remove vector dimensions whose errors exceed a preset upper limit under certain grid assumptions, ensuring that the remaining dimensions can stably characterize stress evolution without depending on a specific spatial discretization scheme. A material parameter decoupling verification process was established, pairing the vector with the cooling surface temperature curves of different materials in the material parameter library, and using the correlation coefficient calculation formula: in The correlation coefficient between the vector eigenvalues ​​and the eigenvalues ​​of the material temperature curve. For the set of vector eigenvalues, This is the set of characteristic values ​​of the material's temperature curve. These are the mean values ​​of the corresponding feature sets, and the dependence of the vector on the material parameters is determined by comparing the ρ values. Perform maximum value statistics on the ρ value of all material combinations and compare it with the preset material dependence threshold. When the maximum correlation coefficient is lower than the threshold, it is confirmed that the material parameter decoupling is successful. The vectors that have passed the dual verification of mesh dependence and material decoupling are marked as the final spectrum feature vectors that can independently characterize the stress state, and output to the spectrum matching engine for real-time decoding. Through the above continuous processing method, the results of the previous step are transformed into high-confidence feature data verified by mesh independence and material decoupling, achieving the expected technical effect of driving stress trend inference without physical field reconstruction.

[0012] like Figure 3 As shown, step S4 involves: constructing a map-stress trend mapping knowledge base based on historical real cooling process data, consisting of map feature vectors and subsequent key region stress envelope annotations; inputting the currently measured map feature vectors into the map-stress trend mapping knowledge base to perform nearest neighbor search and local weighted interpolation operations, generating initial stress trend projection results. Specifically, this includes: S4.1: Clean and align the multiple sets of map feature vectors and corresponding subsequent key area stress envelope annotation data stored in the historical real cooling process dataset to eliminate timestamp drift and sensor noise interference, and generate a standardized map-stress association sample pair set as the basic data source for building the map-stress trend mapping knowledge base; S4.2: Based on the standardized graph-stress correlation sample set, perform a high-dimensional space index structure construction operation, use the ball tree algorithm or locality sensitive hashing mechanism to cluster and optimize the graph feature vectors, generate a graph-stress trend mapping knowledge base with millisecond-level retrieval capability, and realize the construction of a fast mapping channel from feature space to stress trend space. A high-dimensional spatial index structure is constructed on the cleaned and aligned standardized map-stress correlation sample pair set to ensure that the most relevant historical samples can be quickly located in subsequent retrieval processes; Based on the graph feature vector data in this set, a hierarchical spatial partitioning structure is established using the ball tree algorithm. The feature vectors are recursively segmented according to the high-dimensional Euclidean distance similarity, thereby forming a tree-like index network with multi-level nodes and leaf nodes. Locality-Sensitive Hash (LSH) encoding is performed on the segmented feature vector nodes to map high-dimensional features to multiple low-dimensional hash bucket spaces in order to capture key topological features and reduce the retrieval range. Path optimization is performed based on the feature distribution of samples within the hash bucket. By constructing a list of shortest traversal paths, the number of index jumps and access latency can be minimized during the retrieval phase. For each leaf node, a mapping table from the feature space to the stress trend space is established, and the stored feature vector index is bound to the corresponding stress envelope label so that stress trend data can be obtained directly through index mapping during retrieval. The above structure is used to build a knowledge base for graph-stress trend mapping with millisecond-level retrieval capabilities, realizing a fast mapping channel from feature space to stress trend space; By combining ball tree spatial partitioning with local sensitive hashing, the standardized graph-stress correlation sample pair set from the previous step is transformed into a knowledge base structure that can complete feature-stress trend matching in milliseconds, thereby significantly reducing retrieval latency and greatly improving mapping efficiency. For example, in the scenario of predicting cooling of precision ceramic grinding wheels, the constructed standardized graph-stress correlation sample pair set contains 200 sets of samples, each with a vector dimension of 15 and a corresponding stress envelope time length of 30 seconds. When constructing the ball tree index, the leaf node capacity is set to 10 samples, Euclidean distance is used as the segmentation metric, and the radius threshold is set to 0.35. When applying locality-sensitive hashing, six hash function clusters are selected, each containing 20 hash functions, with a hash bucket count of 1024. During path optimization, the shortest path between leaf nodes and the jump cost matrix are calculated, and the path list with the minimum average jump cost is selected as the retrieval sequence. The following formula is used for Euclidean distance calculation: in, For Euclidean distance, Let i be the i-th eigenvalue of the current vector. Let be the i-th feature value of the historical sample. After completing the index structure construction and path optimization, in actual retrieval, the 10 historical samples with the highest similarity can be located within 1 millisecond, and the corresponding stress envelope can be directly mapped, realizing a fast response of the cooling rate control strategy; S4.3: Input the currently measured and generated map feature vector into the map-stress trend mapping knowledge base to perform Euclidean distance metric calculation, and filter out several neighboring samples with the highest similarity to the current map feature vector structure by traversing the index structure to generate a subset of candidate stress trend samples and lock in the historical evolution mode that best matches the current working condition. The currently measured and generated map feature vector is loaded as input parameter into the constructed map-stress trend mapping knowledge base index structure; The Euclidean distance metric is calculated between the input graph feature vector and the feature vectors of each historical sample in the knowledge base. The Euclidean distance metric results are then mapped to the sorting module based on the high-dimensional space index structure and arranged in ascending order of distance values ​​to ensure priority retrieval of highly similar samples. A similarity filtering operation is performed on the sorted sample sequence. The dynamic threshold is determined based on the historical working condition distribution statistics, and several neighboring samples with structural similarity higher than the preset threshold are retained. A subset of candidate stress trend samples is generated from the selected neighboring sample set. Each element in this set contains the corresponding spectral feature vector and the labeled data of the stress envelope of the subsequent key region. Through this chain-processing method, the graph feature vectors from the previous step are matched to the historical evolution patterns with the closest structure in the knowledge base, forming a candidate sample dataset that supports subsequent interpolation and fusion, and realizing an initial trend inference data source that accurately matches the current working conditions. For example, in the application scenario of a precision ceramic grinding wheel cooling condition dataset, the current measured spectrum feature vector contains fifteen dimensions of feature information, and the historical knowledge base stores 200 sets of standardized feature vectors. When performing Euclidean distance calculation, all 200 sets of feature vectors are traversed, a dynamic threshold is set for samples with a distance not exceeding 2.5, samples are sorted in ascending order of distance, and the five nearest samples with the highest structural similarity are selected. Assuming the first dimension of the current feature vector has a value of 0.12, and the first dimension of a certain historical sample has a value of 0.15, the contribution of this dimension is calculated according to the formula. The global distance value is obtained by summing all dimensions and taking the square root. The final output subset of candidate stress trend samples contains 5 stress envelopes. Subsequent interpolation and fusion can reconstruct the trend curve in milliseconds, significantly improving the rapid response capability of cooling rate decision-making and meeting the engineering requirements of low-latency closed-loop control. S4.4: Based on the reciprocal of the distance between each neighboring sample in the candidate stress trend sample subset and the current map feature vector, perform weight coefficient allocation calculation, and use the inverse distance weighted interpolation algorithm to perform linear fusion processing on the stress envelope of the subsequent key area corresponding to each neighboring sample to generate a smooth and continuous initial stress trend inference result, and complete the reconstruction from discrete historical data to continuous prediction curve. The weighting coefficients are allocated based on the reciprocal of the Euclidean distance between each neighboring sample vector in the candidate stress trend sample subset and the current spectral feature vector. Numerical distance reflection processing is performed on the input candidate sample data containing feature vectors and corresponding key region stress envelopes to ensure that the contribution of each sample is proportional to its proximity in the feature space. When performing inverse distance calculation on the Euclidean distance data of each neighboring sample, the following formula is used: in, y is the original weight value of the neighboring sample, x is the current map feature vector, and y is the candidate sample feature vector. The Euclidean distance is obtained by accumulating the squared differences and taking the square root. The original set of weight values ​​is normalized according to the sum of all weight values ​​to generate a relative weight coefficient sequence, so as to ensure that the sum of the weights of different samples is 1, thus achieving proportional distribution. The relative weight coefficient sequence is multiplied point by point with the stress envelope data of the subsequent key regions corresponding to each neighboring sample to obtain the stress envelope matrix adjusted according to the weight ratio, so as to ensure that the contribution of different historical patterns is taken into account when the stress values ​​of each time sampling point are fused. The weighted stress envelope matrix is ​​linearly superimposed along the time axis, and the stress values ​​at the corresponding time sampling points are summed to generate a smooth and continuous initial stress trend curve. During the fusion process, the stress values ​​at adjacent time points are filled with data gaps using a first-order polynomial interpolation method to avoid fluctuations in the prediction curve caused by local value jumps. By using an inverse distance weighted interpolation strategy, discrete historical data is reconstructed into a continuous prediction curve, thereby transforming the candidate sample subset obtained in the previous step into an initial stress trend deduction result that can be directly used as input for subsequent trend correction and control strategies, thus achieving the continuity and stability of the prediction curve. For example, during the cooling process of a batch of precision ceramic grinding wheels, the candidate stress trend sample subset contains five neighboring samples with Euclidean distances of 2.0, 1.5, 1.8, 2.2, and 1.6 to the current spectral feature vector. Based on the above formula, the original weight values ​​for each sample are calculated to be 0.5, 0.6667, 0.5556, 0.4545, and 0.625, respectively. These weight values ​​are then normalized to obtain relative weight coefficients of 0.173, 0.230, 0.191, 0.156, and 0.250. These coefficients are then multiplied by the stress envelopes corresponding to each neighboring sample to obtain a weighted stress matrix. This matrix is ​​then summed point-by-point along the time axis within a 20-second prediction time window to generate an initial stress trend projection result with significantly improved curve smoothness. In this embodiment, the waveform decay of the predicted curve at the maximum stress peak is smoother after interpolation and the time delay is shortened to less than 100ms, which greatly improves the response speed and stability of the downstream cooling rate controller. S4.5: Perform time sequence integrity verification and boundary constraint correction on the initial stress trend projection results. Based on the physical properties of the yield limit of the precision ceramic grinding wheel material, remove abnormal prediction points that exceed the reasonable range, and generate the final initial stress trend projection results that conform to physical laws and have engineering usability, providing a benchmark reference sequence for subsequent abnormal disturbance monitoring.

[0013] Step S5: Monitor whether the actual cooling curve experiences sudden changes in ambient airflow or abnormal disturbances caused by accidental opening of the furnace door. If an abnormal disturbance is detected, the graph residual analysis module is triggered to calculate the structural similarity deviation between the current graph feature vector and the knowledge base sample, and to generate an incremental fusion correction instruction for the neighboring sample with the largest deviation. Specifically, this includes: S5.1: Perform sliding window differential processing on the real-time collected cooling medium temperature curve and grinding wheel surface temperature distribution data to extract the temperature change rate sequence after high-frequency noise suppression, and use the temperature change rate sequence as the input primitive for determining abnormal disturbances such as sudden changes in ambient airflow or accidental opening of the furnace door. S5.2: Execute a multi-scale mutation point detection algorithm based on the input primitive to identify abnormal disturbance events that exceed a preset dynamic threshold, and mark the identified abnormal disturbance events as trigger signals to generate a disturbance trigger command for activating the map residual analysis module; Multi-scale abrupt change point detection operations are performed on the temperature change rate sequence obtained after sliding window differential processing to establish a perturbation feature distribution model at different time scales; The temperature change rate sequence is subjected to transient feature extraction in multiple preset time scale windows. The time scale combination includes three types: short-period local fluctuations, periodic medium-amplitude changes, and long-period low-frequency drifts, in order to cover the temporal performance of different types of abnormal disturbances such as sudden changes in environmental airflow and furnace door opening. Within each timescale window, the standardized residual vector of the rate of change sequence is calculated, and a dynamic threshold function is used to determine the abrupt change judgment boundary. The dynamic threshold function consists of the statistical mean and standard deviation of the reference sequence, and the judgment value is calculated using the following formula: in, The threshold for scale determination, This represents the mean of the temperature change rate sequence within this scale window. Standard deviation The coefficient is set based on historical disturbance sensitivity; Cross-scale validation is performed on the mutation candidate locations for each scale window, and only candidate points that simultaneously meet the mutation conditions in at least two time scale windows are retained to improve the robustness of the abnormal perturbation judgment. The mutation locations that pass cross-scale verification are processed by event merging to eliminate redundant information caused by short-term repeated detection, and a unique perturbation event identifier is assigned to the merged event set. The final set of disturbance events is encoded with trigger signals, and the event identifiers are converted into disturbance trigger command data packets, which serve as input control signals to activate the residual analysis module. By using multi-scale mutation point detection and cross-validation, the temperature change rate sequence from the previous step is transformed into a trigger command with the ability to locate abnormal disturbances in time series, thereby achieving accurate identification and signal output of environmental mutations. For example, during the cooling process of a precision ceramic grinding wheel, a temperature change rate sequence with a window length of 20 sampling points is acquired through sliding window differential processing. Transient features are extracted within three time scale windows: 5 points, 20 points, and 60 points. For each scale window, the mean and standard deviation of the change rate sequence are calculated to obtain the short-scale... mesoscale Long-scale .set up The calculated thresholds were 0.225℃ / s for short-scale, 0.19℃ / s for mesoscale, and 0.155℃ / s for long-scale. Among the candidate points where the rate of change exceeded the threshold detected in each window, only three abrupt change points that simultaneously met the conditions at both short and mesoscale were retained after cross-validation. These were merged to obtain a single event identifier E01, which was encoded as the trigger command "TRG-E01". This trigger command, input to the residual analysis module, can trigger the subsequent stress prediction dynamic calibration process in milliseconds, significantly improving the response efficiency of the cooling control system under external disturbances. S5.3: The constructed graph residual analysis module is invoked according to the disturbance triggering command. The graph residual analysis module is used to perform graph residual analysis, calculate the Euclidean distance between the graph feature vector generated at the current time and the historical sample vector in the graph stress trend mapping knowledge base, so as to quantify the structural similarity deviation between the current working condition and the historical typical working condition, and generate a spatial distance set containing the deviation values ​​of each neighboring sample. Upon receiving the generated disturbance trigger command, the constructed graph residual analysis module is invoked, and the graph feature vector generated at the current moment is used as the query vector and input into the index structure of the graph-stress trend mapping knowledge base; The knowledge base index nodes are traversed and retrieved. The historical sample vectors are sequentially compared with the current graph feature vectors using Euclidean distance measurement. The square root operation of the sum of squared vector differences is used to evaluate the structural similarity of each sample. When calculating Euclidean distance, feature dimension normalization is performed to eliminate the influence of different feature scales on the consistency of the distance metric, and the normalized feature values ​​are substituted into the formula: Where d is the Euclidean distance. This represents the normalized eigenvalue of the i-th dimension of the current eigenvector. The i-th normalized eigenvalue of the historical sample vector. The total dimension of the feature vector; The above calculation results are stored as a set of deviation values ​​according to the sample index, and a spatial distance set containing all deviation values ​​is constructed. Perform integrity checks on the spatial distance set, remove null values ​​or non-numerical anomalies that occur in the calculation, and ensure that the set content can be used for subsequent extreme value screening and deviation weight calculation. This processing method transforms the disturbance trigger signal from the previous step into quantitative deviation data that characterizes the difference between the current operating condition and the historical typical operating condition, thereby achieving the calibrability target of the thermal stress prediction model under abnormal disturbance conditions. For example, during the cooling process of a precision ceramic grinding wheel, the normalized feature vector of the current measured spectrum is [0.12, 0.35, 0.08, 0.22, 0.15, 0.27, 0.33, 0.11, 0.19, 0.25, 0.21, 0.14, 0.18, 0.29, 0.17], and the normalized vector of the historical samples is [0.10, 0.36, 0.07, 0.20, 0.16, 0.26, 0.31, 0.12, 0.20, 0.24, 0.22, 0.15, 0.17, 0.28, 0.18]. Substituting these values ​​into the Euclidean distance calculation formula yields a distance value of 0.054. This value is stored in the spatial distance set as the structural similarity deviation between the historical sample and the current working condition. After repeatedly calculating and removing outliers from multiple sets of historical samples, a spatial distance set containing 100 deviation values ​​is obtained. The maximum value in the set corresponds to the sample where environmental disturbances have the most significant impact on the prediction model. In subsequent steps, this sample will serve as the target reference for incremental fusion correction, significantly improving the calibration accuracy and stability of the predicted trend under anomalous disturbances. S5.4: Perform descending sorting and extreme value filtering operations on the spatial distance set to locate the three neighboring sample indices with the largest structural similarity deviation, and encapsulate the three neighboring sample indices and their corresponding deviation weight coefficients into a local correction target set to generate a sample selection basis for guiding stress trend correction; S5.5: Based on the sample selection criteria, a weighted fusion strategy is executed to calculate and linearly superimpose the historical stress trend segments corresponding to the three neighboring sample indices with the current initial stress trend projection results to generate an incremental fusion correction instruction that can offset the effects of abnormal disturbances, thus completing the dynamic calibration preparation for the thermal stress prediction results. Receive the sample selection criteria generated by the preceding sub-step S5.4 as input conditions, which include the indices of the three nearest samples with the largest structural similarity deviation and the corresponding deviation weight coefficients; Call the historical stress trend fragment data associated with these sample indices, and extract the complete stress envelope time series from the map-stress trend mapping knowledge base according to the index mapping relationship; The extracted historical stress trend segments and the current initial stress trend projection results are aligned with the data time axis to ensure that each trend curve maintains the same sampling step size and starting reference point in the standardized time domain, so as to eliminate the superposition error caused by the difference in time base. A weighted factor matrix is ​​constructed based on the deviation weight coefficient. Each row in the matrix corresponds to the weight distribution of a neighboring sample. The weight factors are dynamically adjusted in a time window sliding manner to adapt to the influence of different stages of stress trend. A linear superposition operation is performed on the three historical stress trend segments and the initial stress trend projection results using a weighted factor matrix, where the superposition formula is: in, This is the dynamically corrected stress value after superposition. Let be the weight coefficient of the i-th sample at time t. Let be the stress trend value of the i-th sample at time t; Outlier filtering is performed on the superposition results to remove numerical points that exceed the physical constraints of the material yield limit, in order to prevent the introduction of physically unrealizable stress states due to perturbation correction. The superimposed curve after filtering is encapsulated into an incremental fusion correction instruction, which includes a set of locally weighted modification parameters for the initial stress trend extrapolation results, in preparation for the subsequent sub-step S6 to dynamically calibrate the prediction results. Through the above weighted fusion and outlier filtering methods, the sample selection criteria calculated in the previous step are transformed into correction parameters that can be directly applied to the initial prediction curve, thereby achieving pre-emptive offsetting and accurate compensation of the thermal stress prediction results affected by sudden changes in environmental airflow or accidental opening of the furnace door. For example, in a real-time monitoring scenario of a precision ceramic grinding wheel cooling process, assume that the sample selection criteria include three neighboring sample indices: 12, 37, and 89, with corresponding deviation weight coefficients of 0.45, 0.35, and 0.20, respectively. Historical stress trend segments corresponding to the indices in the knowledge base are retrieved. The segment length is 30 seconds, and the sampling frequency is 10Hz. The initial stress trend projection results are sampled at the same frequency. Time axis alignment is performed on the four trend curves, using a standardized time axis from 0 to 30 seconds with a step size of 0.1 seconds. A weighted factor matrix is ​​constructed. In the initial cooling stage (0-10 seconds), the weights of the three samples remain unchanged according to the deviation coefficients. In the middle stage (10-20 seconds), the weights are dynamically adjusted, increasing the weight of sample number 12 to 0.50 to adapt to its performance in the middle cooling stage. In the final stage (20-30 seconds), the weight of sample number 37 is reduced to 0.25 to reflect its decreased matching degree in the final stage. The adjusted weight coefficients are applied to the formula. The corrected stress value at each time point is calculated, and outliers are identified using the three-σ principle, eliminating unreasonable stress values. The corrected curve obtained after fusion and filtering, compared to the initial predicted curve, shows that the stress change trajectory tends to be within a physically reasonable range and the fluctuation amplitude is significantly reduced when environmental disturbances are triggered, providing a highly reliable reference input for the dynamic calibration in the subsequent S6 sub-step.

[0014] Step S6: Based on the incremental fusion correction instruction, locally correct the stress trends corresponding to the three neighboring samples with the largest deviation in the initial stress trend projection result, and generate corrected stress trend output data with dynamic calibration capability. Specifically, this includes: S6.1: Receive the incremental fusion correction instruction and the initial stress trend inference result generated by the previous steps. Based on the structural similarity deviation threshold contained in the incremental fusion correction instruction, perform secondary screening on the set of neighboring samples retrieved in the map-stress trend mapping knowledge base, and extract the three target neighboring samples with the largest structural similarity deviation and their corresponding original stress envelope sequences to lock the reference benchmark data most significantly affected by abnormal environmental disturbances. S6.2: Obtain the original stress envelope sequence of the three target neighbor samples and the real-time spectrum feature vector obtained by decoding the current measured temperature field scanning data. Utilize the dynamic weight allocation algorithm based on time window sliding to calculate the local confidence coefficient of each target neighbor sample on the current standardized time axis. Use the local confidence coefficient as a quantitative indicator to measure the adaptability of each sample to the current working condition, and generate a weighted sample dataset containing weight factors. The input dataset consists of the original stress envelope sequence of three target neighbor samples obtained through secondary screening in S6.1 and the real-time spectrum feature vector generated by decoding the current measured temperature field scanning data. The stress envelope sequence of each target neighboring sample is processed by time window sliding slicing on the standardized time axis according to the corresponding sampling point relationship and the real-time spectrum feature vector, and the stress change slope segment and temperature curve change segment within the same time window are extracted as local comparison units. For each comparison unit, the difference between the slope of the sample stress change and the slope of the measured stress change is calculated and the absolute value is taken to form an instantaneous deviation measurement matrix; Based on the instantaneous deviation measurement matrix, an exponential decay weighting function is used to construct the matching score of each sampling point within the time window. The matching score is averaged according to the length of the time window to obtain the local confidence coefficient used to measure the sample fit. The formula for calculating the local confidence coefficient is: in, Let be the difference between the slope of the sample stress change and the slope of the measured stress change at the i-th sampling point. The attenuation coefficient is... This represents the number of sampling points within the time window. This is the local confidence coefficient; The local confidence coefficient corresponding to each sample is appended to the original stress envelope sequence as a weighting factor to form a weighted sample dataset containing the weighting factor. By using a dynamic weight allocation method based on time window sliding, the screening results of the previous step are transformed into a weighted data structure that can quantify the adaptability of samples, so as to achieve precise control of the adaptive ratio in the subsequent fusion stage. For example, during the cooling process of a precision ceramic grinding wheel, the sampling frequency is set to 50Hz, the standardized time window length is 2 seconds, and the number of sampling points within each window is... The attenuation coefficient is 100. The value was set to 0.05. The stress change slope sequences of the three target neighboring samples within this time window were [0.12, 0.15, ..., 0.10], [0.08, 0.11, ..., 0.09], and [0.14, 0.17, ..., 0.13], respectively, while the measured sequence was [0.10, 0.14, ..., 0.11]. Calculations were performed for each sampling point. The absolute values ​​were then used to calculate the local confidence coefficients for the three samples using the formula for local confidence coefficients. The calculated local confidence coefficients were 0.91, 0.86, and 0.94, respectively. These coefficients were then used as weighting factors to be added to the corresponding stress envelope sequences, forming a weighted sample dataset. In the subsequent fusion processing stage, these weighting factors significantly improved the adaptability of the fusion results to the current cooling conditions, resulting in a substantial improvement in the smoothness and reliability of the dynamic calibration stress trend. S6.3: Based on the local confidence coefficients in the weighted sample dataset, perform multi-source heterogeneous data fusion processing on the original stress envelope sequences corresponding to the three target neighboring samples. Adaptive local weighted interpolation strategy is used to linearly superimpose and nonlinearly smooth the stress evolution paths of each sample according to the weight ratio, eliminate the boundary abrupt changes caused by the differences in historical working conditions of a single sample, and generate a preliminary fused dynamic calibration stress trend curve. Based on the weighted sample dataset containing local confidence coefficients generated by the preceding sub-step S6.2, multi-source heterogeneous data fusion processing is performed on the original stress envelope sequences corresponding to the three target neighbor samples. The stress envelope sequences of the three target samples are time-synchronized and aligned according to the standardized time axis to ensure that the node features of each sequence are consistent at the corresponding time. Using the local confidence coefficient as a weighting factor, a weight matrix is ​​constructed for each time node, and a linear combination operation is performed on the multi-source stress data to generate a preliminary fused sequence of node stress values. An adaptive local weighted interpolation strategy is adopted, and a smoothing factor based on the difference between nodes is introduced into the weight matrix. The stress curve of continuous time period is nonlinearly smoothed through interpolation operation to suppress boundary abrupt changes caused by the difference of working conditions of a single sample. During the interpolation process, a functional relationship between nodal stress values ​​and time intervals is established, and a curvature constraint term is introduced. When the curvature exceeds a preset limit, the interpolation weights are automatically adjusted to maintain the stability of the curve transition. Through the above linear superposition and nonlinear smoothing process, the original stress envelope sequence in the previous step is transformed into a preliminary fusion dynamic calibration stress trend curve with multi-source information fusion characteristics and elimination of boundary mutations, thereby improving the structural consistency and local stability of the predicted curve. For example, during the cooling stage of the precision ceramic grinding wheel, the original stress envelopes of three adjacent target samples are synchronized at 300 sampling points on the normalized time axis, with local confidence coefficients of 0.6, 0.25, and 0.15, respectively. During time synchronization, each curve is interpolated to a uniform sampling frequency of 100Hz. When constructing the weight matrix, the fused stress value at a given time point is... Calculate using the following formula: in, This is the local confidence coefficient. This represents the stress value of the corresponding sample at that node. In nonlinear smoothing, the interpolated stress value for adjacent nodes... and Local weighted average with curvature constraints is used: in, The interpolated stress value. and These are the effective weighting coefficients adjusted based on curvature constraints. Calculations show that, under curvature constraints, the rate of change of the stress curve at all nodes remains within the allowable range, boundary abrupt changes are completely eliminated, and the generated preliminary fused dynamic calibration stress trend curve can significantly improve the stability of the final prediction in the subsequent residual compensation stage. S6.4: Obtain the real-time residual feedback signal of the preliminary fusion dynamic calibration stress trend curve and the current actual cooling curve, use the closed-loop error compensation mechanism to filter and suppress the high-frequency noise component in the preliminary fusion dynamic calibration stress trend curve, and perform phase alignment correction on the low-frequency drift component, and reconstruct the stress evolution data after phase alignment correction into continuous and smooth corrected stress trend output data. S6.5: Receive the corrected stress trend output data, perform integrity verification on the stress change slope segment and confidence interval within the next two-second time window, and if the verification passes, encapsulate the corrected stress trend output data into a standardized data stream format and transmit it as the final output result with dynamic calibration capability to the subsequent trend fragment stream encapsulation module to support sub-millisecond cooling rate decision execution.

[0015] Step S7: Encapsulate the corrected stress trend output data into a trend fragment stream containing only the slope segment and confidence interval of stress change within the next two seconds, generating a lightweight control command input source adapted to sub-millisecond decision cycles. Specifically, this includes: S7.1: Perform time window truncation processing on the corrected stress trend output data to extract the continuous stress change sequence in the next two seconds time domain and generate the stress time sequence segment to be compressed. S7.2: Perform sliding difference operation based on the time sequence segment of the stress to be compressed to calculate the stress change rate between adjacent sampling points and fit the local linear slope to generate a stress change slope segment; The system receives a time sequence of compressible stress within the next two seconds as input. The time sequence contains a stress value sequence arranged at a fixed sampling interval and a corresponding time index. Based on this sequence, a sliding window difference operation is performed to calculate the stress difference between adjacent sampling points within each preset length window, and the difference is divided by the corresponding time difference to obtain the instantaneous stress change rate. Local linear fitting is performed on the rate of change data within each window, and the slope parameter at the center of the window is estimated using the least squares method to characterize the local stress change trend. During the fitting process, a linear equation is established using the time index as the independent variable and the stress value as the dependent variable. The local fitting slope is calculated using the following formula: in, Sampling time, For the sampled stress value, This represents the average time within the window. The average stress within the window. The slope is the fitted slope; The calculated set of local slope parameters are sequentially arranged according to the original sampling order to form a stress change slope segment. By using the above sliding difference and local fitting processing methods, the time window stress data extracted in the previous step is transformed into slope segments that reflect the rate of change of local stress, thereby realizing the core dynamic features required for the construction of trend fragment flow. For example, in the scenario of cooling control of a precision ceramic grinding wheel, the length of the time sequence segment of the compressible stress is set to 2000ms, the sampling frequency is 100Hz, forming a sequence containing 200 data points per window. The sliding window length is set to 10 sampling points, and the step size is 1 sampling point. The fitting slope of each window is calculated according to the formula, using the time index millisecond value and stress value (unit MPa) of the window time sequence data as input, the mean of the independent variable is set to the time of the window center point, and the mean of the dependent variable is the corresponding stress average value. Taking a certain window data as an example, the time difference range is 9ms, the stress value difference is 0.45MPa, and the slope obtained by the formula is 0.05MPa / ms. The entire sequence is processed to generate a slope segment consisting of a total of 199 local slope values. Compared with the original stress curve, this segment significantly improves the instantaneity in capturing the stress change rate characteristics, ensuring that the downstream cooling rate decision module obtains highly agile input features in a sub-millisecond response cycle. The verification results show that the delay of stress prediction and adjustment actions is significantly reduced under environmental disturbance conditions, and the stability of the cooling process is greatly improved. S7.3: Confidence boundary extrapolation is performed by combining stress change slope segments with historical residual statistical distribution to determine the upper and lower fluctuation range of the current prediction results and generate stress prediction confidence intervals; S7.4: The stress change slope segment and the stress prediction confidence interval are structured and packaged into a code to remove redundant numerical points in the complete curve and retain the core decision features, generating a lightweight trend fragment stream. S7.5: Perform protocol parsing and format mapping on the received lightweight trend fragment stream to convert it into a standardized data interface format that can be recognized by the downstream controller, and generate a lightweight control command input source adapted to sub-millisecond decision cycles.

[0016] Step S8: Receive the trend fragment flow and analyze the stress change slope segment and confidence interval within it. Based on the analysis results, adjust the cooling rate execution parameters of the cooling medium in real time to complete the low-latency closed-loop control action between thermal stress prediction and control strategies. Specifically, this includes: S8.1: Perform protocol parsing on the received trend fragment stream to extract stress change slope fragments and confidence interval data within the next two seconds, and generate a real-time stress evolution feature vector to be processed. S8.2: Perform deviation quantization operation based on the stress change slope segment in the real-time stress evolution feature vector to calculate the instantaneous deviation between the current thermal stress state and the preset safety threshold, and generate the thermal stress dynamic deviation coefficient. S8.3: Utilize the aforementioned dynamic deviation coefficient of thermal stress combined with confidence interval data to execute an adaptive gain adjustment algorithm to determine the sensitivity parameter for adjusting the cooling medium flow rate and generate a cooling rate adjustment command pulse; The input parameter set of the adaptive gain adjustment algorithm is constructed based on the dynamic deviation coefficient of thermal stress and the confidence interval of stress prediction, and each parameter is mapped to the multi-dimensional feature coordinate space of the gain adjustment calculation unit. Normalization is performed on the dynamic deviation coefficients of thermal stress in the input parameter set to ensure dimensional consistency and eliminate absolute value shifts under different working conditions, so as to form a standardized deviation sequence. The covariance matrix is ​​calculated for the standardized deviation sequence and the corresponding upper and lower bounds of the confidence interval to evaluate the degree of linear correlation between deviation and prediction uncertainty, and this correlation index is used as the basis coefficient for calculating the gain adjustment weight. The correlation index and standardized deviation sequence are input into the adaptive adjustment function, and the gain calculation formula based on error sensitivity weighting is used: in, Adjust the sensitivity parameters to adjust the cooling medium flow rate. and The preset adjustment coefficient, This is the normalized dynamic deviation of thermal stress. This represents the current prediction confidence interval width. The width of the reference confidence interval; Boundary constraint processing is performed on the calculated sensitivity parameters to ensure that they are between the preset minimum and maximum gain thresholds, preventing the cooling medium flow rate adjustment from exceeding the safe range; The sensitivity parameters after boundary constraints are input into the command pulse generation module and encoded into a cooling rate adjustment command pulse through pulse width modulation (PWM) to achieve high-precision transmission of the cooling medium flow rate adjustment command. Through the above-mentioned multi-dimensional standardization, covariance correlation, error sensitivity weighting and boundary constraint processing, the real-time stress evolution feature vector of the previous step is transformed into a cooling rate adjustment command pulse that can directly drive the actuator, thereby realizing dynamic optimization of the cooling medium flow rate adjustment sensitivity. For example, during the cooling process of a precision ceramic grinding wheel, the dynamic deviation coefficient of thermal stress in the real-time stress evolution characteristic vector was measured to be 0.35, with a confidence interval width of 0.12 MPa and a reference confidence interval width of 0.10 MPa. The adjustment coefficient... Set β to 1.5 and β to 0.8. The normalized bias Δ = 0.35, and the confidence interval ratio is 0.12 / 0.10 = 1.2. Substituting these values ​​into the formula: The calculated value is G = 0.525 + 0.96 = 1.485. This sensitivity parameter is constrained to between 0.5 and 2.0 after boundary constraint processing, requiring no further correction. A PWM-encoded adjustment command pulse with a pulse width of 1.485 ms is generated and transmitted to the cooling medium valve control unit. This adjusts the valve opening to increase the cooling medium flow rate, achieving real-time compensation for the corresponding dynamic deviation. In this embodiment, the actuator response time is less than 80 ms. After the cooling rate adjustment, the slope of the thermal stress change in the target area tends to stabilize, the slope fluctuation amplitude is significantly reduced, and the system closed-loop control delay is greatly compressed. S8.4: The cooling rate execution parameters of the cooling medium are mapped and updated in real time according to the cooling rate adjustment command pulse to generate a target flow rate setting value of the cooling medium that is adapted to the current working conditions. S8.5: Based on the target flow rate set value of the cooling medium, drive the cooling system actuator to complete the valve opening adjustment action, so as to realize low-delay closed-loop control between thermal stress prediction and control strategy.

[0017] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0018] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cooling rate control method based on thermal stress prediction during the sintering cooling stage of a grinding wheel, characterized in that, Includes the following steps: S1: Acquire temperature field scanning data and surface strain synchronous acquisition data of precision ceramic grinding wheel under typical cooling conditions, and use the temperature field scanning data and surface strain synchronous acquisition data as the original input primitives; S2: Based on the original input primitives and the finite element simulation inversion results, identify key response nodes and generate thermal stress response fingerprint spectrum; S3: Encode the thermal stress response fingerprint spectrum into a one-dimensional structured vector and perform decoupling verification to generate a spectrum feature vector; S4: Based on historical real cooling process data, construct a map-stress trend mapping knowledge base, and input the currently generated map feature vector into the map-stress trend mapping knowledge base to perform nearest neighbor search and weight coefficient allocation, and generate initial stress trend inference results; S5: Monitor whether there is any abnormal disturbance in the actual cooling curve. If the abnormal disturbance is detected, perform graph residual analysis, calculate the structural similarity deviation between the current graph feature vector and the knowledge base sample, and generate an incremental fusion correction instruction. S6: According to the incremental fusion correction instruction, the stress trend corresponding to the neighboring sample with the largest deviation in the initial stress trend extrapolation result is locally corrected to generate corrected stress trend output data. S7: Encapsulate the corrected stress trend output data into a trend fragment stream containing only the stress change slope segment and confidence interval within the next two seconds, and generate a control command input source.

2. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, Following S7, the following also includes: S8: Receive the trend fragment flow and analyze the stress change slope segment and confidence interval in it. Based on the analysis results, adjust the cooling rate execution parameters of the cooling medium in real time to complete the low-latency closed-loop control action between thermal stress prediction and control strategy.

3. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, The specific steps for acquiring temperature field scanning data and surface strain synchronously collected data of precision ceramic grinding wheels under typical cooling conditions are as follows: The infrared thermal imaging array sensor and the fiber optic strain sensing network are time-referenced to generate a synchronous trigger signal with a unified timestamp. Based on the synchronous trigger signal, the infrared thermal imaging array sensor is used to sample the radiation energy distribution on the surface of the grinding wheel at high frequency to obtain a discrete temperature matrix sequence and generate a raw data stream of temperature field scanning. At the same time, the driving fiber optic strain sensing network is used to capture the micro-deformation of the grinding wheel surface in parallel, extract the wavelength offset sequence, and generate a raw data stream of surface strain synchronous acquisition.

4. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, The key response nodes include the inflection point of initial thermal relaxation, the peak time of the circumferential gradient, the point of reversal of radial stress polarity, the starting position of the saturation segment of surface compressive stress, and the precursor interval before the peak of core tensile stress.

5. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, S3 specifically includes: For the timestamp data of key response nodes in the thermal stress response fingerprint spectrum, normalization mapping is performed based on the total cooling time under typical cooling conditions to generate a standardized time axis relative position sequence. Based on the standardized time axis relative position sequence, the local slope of the temperature field scanning data in the neighborhood of each key response node is calculated to generate a set of local slope change rates. For the fluctuation data of the neighborhood of each node in the set of local slope change rates, the neighborhood fluctuation entropy value is calculated to generate the neighborhood fluctuation entropy value index. The standardized time axis relative position sequence, the local slope change rate set, and the neighborhood fluctuation entropy index are serially spliced ​​and vectorized according to a preset feature dimension sorting rule to generate a one-dimensional structured vector. The one-dimensional structured vector is subjected to meshless dependency verification and material parameter decoupling verification processing, and the resulting spectrogram feature vector is output.

6. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 5, characterized in that, The one-dimensional structured vector is 15-dimensional, and consists of the standardized time axis relative position sequence of five key response nodes, the set of local slope change rates, and the neighborhood fluctuation entropy index. The standardized time axis relative position sequence is placed in the first to fifth dimensions of the feature vector, the set of local slope change rates is placed in the sixth to tenth dimensions of the feature vector, and the neighborhood fluctuation entropy index is placed in the eleventh to fifteenth dimensions of the feature vector.

7. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, S4 specifically includes: The multiple sets of map feature vectors stored in the historical real cooling process dataset and the corresponding subsequent key area stress envelope annotation data are cleaned and aligned to generate a standardized set of map-stress correlation sample pairs. Based on the standardized graph-stress correlation sample set, a high-dimensional spatial index structure construction operation is performed, and the graph feature vectors are clustered and the path is optimized to generate a graph-stress trend mapping knowledge base. The current measured feature vector of the map is input into the map-stress trend mapping knowledge base to perform Euclidean distance metric calculation, and several neighboring samples with the highest structural similarity to the current map feature vector are selected to generate a subset of candidate stress trend samples. Based on the inverse distance between each neighboring sample in the candidate stress trend sample subset and the current map feature vector, a weight coefficient allocation operation is performed, and the stress envelopes of the subsequent key regions corresponding to each neighboring sample are linearly fused to generate the initial stress trend projection result.

8. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 7, characterized in that, S4 further includes: The initial stress trend projection results are subjected to time sequence integrity verification and boundary constraint correction. Based on the physical properties of the yield limit of the precision ceramic grinding wheel material, abnormal prediction points that exceed the reasonable range are eliminated to generate the final initial stress trend projection results.

9. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 1, characterized in that, S5 specifically includes: The temperature curve of the cooling medium and the temperature distribution data of the grinding wheel surface are collected in real time. The sliding window difference processing is performed to extract the temperature change rate sequence, and the temperature change rate sequence is used as the input primitive to determine the abnormal disturbance of sudden change in ambient airflow or accidental opening of furnace door. Multi-scale mutation point detection is performed based on the input primitives to identify abnormal disturbance events that exceed a preset dynamic threshold, and the identified abnormal disturbance events are marked as trigger signals to generate disturbance trigger commands. According to the disturbance triggering command, the constructed map residual analysis module is invoked to calculate the Euclidean distance between the map feature vector generated at the current moment and the historical sample vector in the map stress trend mapping knowledge base, and to generate a spatial distance set. Perform descending sorting and extreme value filtering operations on the spatial distance set to locate the three nearest sample indices with the largest structural similarity deviation, and encapsulate the three nearest sample indices and their corresponding deviation weight coefficients into a local correction target set to generate the sample selection criteria. Based on the sample selection criteria, a weighted fusion strategy is executed to calculate and linearly superimpose the historical stress trend segments corresponding to the three neighboring sample indices with the current initial stress trend projection results to generate an incremental fusion correction instruction.

10. The cooling rate control method based on thermal stress prediction during the sintering cooling stage of the grinding wheel according to claim 9, characterized in that, The spatial distance set includes the deviation values ​​of each neighboring sample.

Images