Temperature field intelligent control method and system for magnetic material production

Abstract

This invention provides a method and system for intelligent temperature field control in the production of magnetic materials, relating to the field of production control technology. The method includes: constructing a sintering digital twin model and obtaining intermediate activation signals from multiple denoising steps; constructing a linear probe to analyze classification accuracy and obtain the distribution of the implicit decision layer; monitoring the physical field in real time and extracting real-time feature vectors of the implicit decision layer; calculating a guiding vector based on the real-time feature vectors and outputting the modification direction; triggering a matching layer activation guidance command to generate a superimposed basic temperature control increment, driving the mechanism to instantaneously compensate for temperature field deviations. This addresses the technical problem that traditional temperature field control methods can only perform hysteretic adjustments based on macroscopic temperature deviations, making it difficult to identify and intervene in the implicit formation tendency of microstructures during sintering. The method achieves the technical effect of compensating for instantaneous temperature field deviations and improving the consistency of microstructures and the stability of magnetic properties of magnetic materials through real-time detection and activation guidance of the implicit decision layer.

Description

Technical Field

[0001] This invention relates to the field of production control technology, specifically to a method and system for intelligent temperature field control in the production of magnetic materials. Background Technology

[0002] The final magnetic properties of magnetic materials are closely related to the evolution of their microstructure during sintering, especially key attributes such as grain size, phase composition, and magnetic domain alignment, all of which are determined by the spatiotemporal distribution history of the temperature field. However, traditional magnetic material sintering control mainly relies on preset process instructions, which is essentially a macroscopic and coarse open-loop or semi-closed-loop control. In actual production, even when executing identical temperature-time curves, microscopic disturbances such as airflow disturbances within the sintering furnace, heating element aging, and differences in furnace loading can cause unpredictable implicit tendencies in the local temperature field inside the furnace. This can lead to abnormal grain growth, uneven phase distribution, or disordered magnetic domain alignment, ultimately resulting in poor consistency of magnetic properties and low yield among products in the same batch. Existing industrial control methods, such as PID and cascade control, are essentially passive adjustments, only correcting after temperature sensors detect deviations. They cannot actively predict or intervene in the generation logic of the microstructure, let alone implement feedforward compensation during critical phase transitions or grain growth. Therefore, how to construct a method that can actively guide temperature field changes during the sintering process of magnetic materials has become an urgent problem to be solved in this field. Summary of the Invention

[0003] This application provides a method and system for intelligent temperature field control in the production of magnetic materials, aiming to solve the technical problem that traditional temperature field control methods can only make hysteretic adjustments based on macroscopic temperature deviations, making it difficult to identify and intervene in the implicit tendency of microstructure formation during sintering.

[0004] The first aspect disclosed in this application provides a method for intelligent temperature field control in the production of magnetic materials. The method includes: constructing a digital twin model of the magnetic material sintering process; acquiring multiple intermediate activation signals at multiple denoising time steps within the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal, and a third activation signal; performing attribute classification accuracy analysis on the multiple intermediate activation signals by constructing a linear probe set to obtain the implicit decision-making layer feature distribution; monitoring the physical field state within the sintering furnace in real time to obtain the real-time feature vector of the implicit decision-making layer; combining the real-time feature vector with the implicit decision-making layer feature distribution to perform guidance vector calculation and output a decision modification direction; triggering an activation guidance command of the matching layer according to the decision modification direction; using the activation guidance command to generate an adjustment increment signal superimposed on the basic temperature field control output to drive the underlying actuator to perform instantaneous deviation compensation of the temperature field.

[0005] The second aspect of this application discloses an intelligent temperature field control system for the production of magnetic materials. This intelligent temperature field control system is used in conjunction with the aforementioned intelligent temperature field control method for the production of magnetic materials. The intelligent temperature field control system includes: a signal acquisition module: constructing a digital twin model of the magnetic material sintering process and acquiring multiple intermediate activation signals at multiple denoising time steps within the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal, and a third activation signal; a signal classification module: performing attribute classification accuracy analysis on the multiple intermediate activation signals by constructing a linear probe set to obtain the implicit decision layer feature distribution; a real-time monitoring module: monitoring the physical field state within the sintering furnace in real time and acquiring the real-time feature vector of the implicit decision layer; a vector calculation module: combining the real-time feature vector and performing guided vector calculation based on the implicit decision layer feature distribution to output the decision modification direction; and a deviation compensation module: triggering an activation guidance command of the matching layer according to the decision modification direction, using the activation guidance command to generate an adjustment increment signal superimposed on the basic temperature field control output, driving the underlying actuator to perform instantaneous deviation compensation for the temperature field.

[0006] One or more technical solutions provided in this application have at least the following beneficial effects:

[0007] First, a digital twin model of the sintering process is established to extract multi-stage and multi-type intermediate state signals within the furnace. Linear probes are then used to determine the correlation between these signals and the target magnetic properties, thereby identifying key implicit decision features affecting material microstructure formation. Next, in actual production, real-time physical field data within the furnace is collected to calculate the deviation direction of the current state relative to the optimal magnetic property target. This deviation is then converted into an executable temperature field adjustment increment, which is superimposed on the original temperature control output. This allows the heating unit to quickly and directionally compensate for local temperature deviations, thereby proactively intervening in the material microstructure evolution process during critical stages of grain growth and phase transformation, ultimately improving the consistency of magnetic properties.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of a method for intelligent temperature field control in the production of magnetic materials, provided in an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the intelligent temperature field control system for magnetic material production provided in an embodiment of this application.

[0011] Explanation of reference numerals in the attached diagram: Signal acquisition module 11, Signal classification module 12, Real-time monitoring module 13, Vector calculation module 14, Deviation compensation module 15. Detailed Implementation

[0012] This application provides a method and system for intelligent temperature field control in the production of magnetic materials, which solves the technical problem that traditional temperature field control methods can only make hysteretic adjustments based on macroscopic temperature deviations, making it difficult to identify and intervene in the implicit tendency of microstructure formation during sintering.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, a method for intelligent temperature field control in the production of magnetic materials is provided, the method comprising: A digital twin model of the sintering process of magnetic materials is constructed to obtain multiple intermediate activation signals at multiple noise reduction time steps in the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal, and a third activation signal.

[0015] In one embodiment, a digital twin model is first constructed using the magnetic material sintering furnace and its sintering process as the object. Specifically, a furnace spatial coordinate system is established according to the actual equipment structure of the sintering furnace. The furnace inner wall, insulation layer, firing plate, material tray, material loading area, and multiple heating units are mapped to spatial nodes in the digital twin model, and the relative positional relationships between each heating unit and temperature measurement point, infrared sampling point, and material loading position are recorded. Subsequently, process setting parameters, furnace structure parameters, heating unit arrangement parameters, temperature sensor data, infrared array sampling data, current response data, vacuum degree data, and sintering result data of the corresponding batch of magnetic materials are collected during the historical production process of the sintering furnace. These data are then subjected to timestamp unification, outlier removal, missing value completion, sampling frequency alignment, and spatial position calibration. Data from different sources are synchronized to the same process time according to the sintering time axis to form a multi-source time-series sample corresponding to the actual sintering process.

[0016] After forming multi-source time-series samples, the sintering process is divided into heating, holding, phase transformation sensitive, and cooling stages according to temperature curves and material state changes. A corresponding state identifier is assigned to each stage. Typically, the heating stage is identified based on the set temperature change rate; the holding stage is identified based on the stable holding range near the target temperature; the phase transformation sensitive stage is identified based on vacuum fluctuations, sudden changes in temperature change rate, abnormal current response, or historical magnetic transition point ranges; and the cooling stage is identified based on the cooling curve. Subsequently, the temperature field data, power input data, vacuum data, and material result labels for each stage are input into a digital twin model for training. This allows the model to learn the correspondence between the output of different heating units, furnace heat conduction, changes in heat absorption and release of materials, and the final magnetic properties.

[0017] The digital twin model comprises an input encoding layer, a self-attention layer, a denoising and diffusion layer, and a state output layer. The input encoding layer converts process setting parameters, furnace spatial node information, heating unit power data, temperature spatial distribution data, vacuum level data, and sintering stage identifiers into feature embedding vectors of a unified dimension. The self-attention layer is positioned between the input encoding layer and the denoising and diffusion layer, or embedded in an intermediate layer within the denoising and diffusion layer, and is used to calculate the correlation weights between different time steps, different spatial nodes, and different physical field features. Specifically, the self-attention mechanism module maps each spatial node and various physical field features into query vectors, key vectors, and value vectors, respectively. It calculates attention weights based on the correlation between query vectors and key vectors, and then weights and fuses the value vectors based on these attention weights, thereby obtaining attention features that can simultaneously characterize temporal correlations, spatial correlations, and multi-physics coupling relationships. Thus, the model can identify the impact of a heating unit power fluctuation on adjacent temperature measurement areas, the impact of a local temperature anomaly on the heat flux distribution in the material loading area, and the correlation between vacuum level fluctuations and material phase transition sensitive stages. Next, the actual physical field states in historical sintering samples are used as the target state. During training, perturbation noise of varying intensities is introduced to form multiple noise state samples. The digital twin model takes these noise state samples, process setting parameters, furnace space node information, and real-time physical field signals as input. First, multi-source feature embedding is formed through the input encoding layer. Then, spatiotemporal correlation and multi-physics coupling feature extraction are performed through the self-attention layer. Subsequently, the denoising diffusion module gradually predicts the noise residual or the recovered furnace state within multiple denoising time steps, thereby learning the evolution law from the perturbation state to the stable sintering state. Each denoising time step corresponds to a latent variable update process for the current sintering state. During this update process, different levels of the self-attention layer form corresponding intermediate activation outputs, attention weight matrices, and attention maps. The model not only outputs observable temperature field prediction results but also forms intermediate activation signals at the self-attention level to characterize the latent states within the furnace.

[0018] During model operation, the entire sintering evolution process is divided into multiple denoising time steps. Each denoising time step corresponds to an evolutionary stage in which the material microstructure gradually approaches the target magnetic property state from a disordered state. The system extracts intermediate activation signals within the model at each denoising time step to reflect the implicit characteristic state of the sintering process at the current stage. These intermediate activation signals include a first activation signal, a second activation signal, and a third activation signal. The first activation signal reflects the operating state characteristics such as heating system load changes, thermal inertia, and power output stability. The second activation signal characterizes the consistency of the heat flow field in different regions and the characteristics of local temperature difference changes. The third activation signal reflects the state changes during the volatilization, phase transformation, and microstructure evolution processes within the material. Through this method, without directly disrupting the sintering process, the furnace heating response, temperature spatial distribution, and phase transformation-related perturbations can be transformed into calculable intermediate feature vectors, providing a data foundation for subsequent identification of implicit decision layers, judgment of magnetic property generation tendencies, and generation of temperature field compensation control quantities.

[0019] Furthermore, the first activation signal is a current response feature vector based on the power fluctuation of the sintering furnace, used to characterize the latent variable of the operating state of the heating system; the second activation signal is a temperature spatial distribution feature vector based on infrared array sampling, used to characterize the consistency latent variable of the heat flow field in the furnace; and the third activation signal is a vacuum degree feature vector based on the pressure fluctuation of the magnetic transition point, used to characterize the phase transition latent variable of the material microstructure evolution.

[0020] Preferably, the first, second, and third activation signals are all extracted by the digital twin model in the intermediate layer of multiple denoising time steps. Specifically, the current sampling values, infrared array temperature sampling values, and furnace vacuum sampling values ​​of each heating unit of the sintering furnace are aligned according to the same sampling time and input into the feature encoding layer of the digital twin model to obtain current response encoding vector, temperature space encoding vector, and vacuum degree encoding vector. These three types of encoding vectors are then input into the self-attention layer of the digital twin model. The self-attention layer performs correlation calculations on the current change, temperature space distribution, and vacuum degree change under the same denoising time step to generate corresponding intermediate layer latent variables. Specifically, the current sampling sequence of each heating unit of the sintering furnace within the current sampling period is used as input, converted into a current response encoding vector by the feature encoding layer, and then input into the self-attention layer. The hidden state vector of the current channel corresponding to the output of the self-attention layer is extracted by weighted fusion and used as the first activation signal. The furnace temperature matrix obtained by the infrared array in the current sampling period is used as input, converted into a temperature spatial encoding vector by the feature encoding layer, and then input into the self-attention layer for spatial correlation calculation to extract the hidden state vector of the temperature channel corresponding to the output of the self-attention layer as the second activation signal. The vacuum degree sampling sequence of the furnace in the temperature range corresponding to the magnetic transition point is used as input, converted into a vacuum degree encoding vector by the feature encoding layer, and then input into the self-attention layer for temporal correlation calculation to extract the hidden state vector of the vacuum degree channel corresponding to the output of the self-attention layer as the third activation signal.

[0021] By constructing a linear probe set, the attribute classification accuracy of the multiple intermediate activation signals is analyzed to obtain the feature distribution of the implicit decision layer.

[0022] In one embodiment, after obtaining the first, second, and third activation signals at multiple denoising time steps, a set of linear probes is constructed to determine the characterization ability of different intermediate layer activation signals in the digital twin model for the final magnetic performance target attribute. Specifically, firstly, preset magnetic performance target attribute labels are established based on the detection results of historical sintering batches, including grain size level, phase composition ratio, and magnetic domain alignment orientation. The detection results of each historical sintering batch after sintering are converted into corresponding classification labels. Subsequently, after the digital twin model completes training, the model parameters of the digital twin model are fixed and no longer updated. Only a linear classifier is connected after each self-attention layer of the digital twin model to form a set of linear probes corresponding one-to-one with different self-attention layers. For any denoising time step, the first, second, and third activation signals extracted at that denoising time step are concatenated to form the intermediate activation fusion vector corresponding to that denoising time step. The intermediate activation fusion vector is then input into the linear probe after the corresponding self-attention layer, and the linear probe outputs the predicted classification results of grain size level, phase composition ratio, and magnetic domain alignment orientation, respectively. Next, the predicted classification results output by the linear probe are compared with the actual magnetic property target attribute labels corresponding to historical sintering batches, and the attribute classification accuracy at each self-attention layer and each denoising time step is calculated. If the classification accuracy of a certain self-attention layer for the preset magnetic property target attribute label is higher than the preset threshold in multiple consecutive samples, it indicates that the intermediate activation signal of the self-attention layer has contained effective information related to the final magnetic property formation result, and the self-attention layer is identified as an implicit decision layer. Then, the weight matrix and attention map feature vector spatial density of the implicit decision layer at the corresponding denoising time step are extracted to construct the implicit decision layer feature distribution, which is used to characterize the magnetic property generation tendency formed within the digital twin model during the sintering process, and to provide a benchmark for subsequent real-time feature vector deviation judgment and guiding vector calculation, thereby realizing early identification and targeted intervention of key stages of magnetic property generation.

[0023] Furthermore, by constructing a linear probe set to perform attribute classification accuracy analysis on the multiple intermediate activation signals, the feature distribution of the implicit decision layer is obtained, including: Obtain preset magnetic property target attribute labels, wherein the preset magnetic property target attribute labels include grain size level, phase composition ratio, and magnetic domain alignment orientation; use multiple intermediate activation signals under multiple denoising time steps as input to train a linear probe set to perform posterior classification of the preset magnetic property target attribute labels; select a specific layer with a classification accuracy greater than a preset threshold and located at the self-attention level in the digital twin model to obtain an implicit decision layer, wherein the implicit decision layer is a network layer that intervenes in the magnetic property generation path; construct the implicit decision layer feature distribution by extracting the weight matrix, attention map, and feature vector space density of the implicit decision layer under the denoising time step, wherein the implicit decision layer feature distribution is used to characterize the spontaneous generation tendency of the current sintering process.

[0024] Preferably, the magnetic materials after historical sintering are first subjected to microstructure and magnetic property testing to obtain the grain size, phase composition ratio, and magnetic domain arrangement state of the corresponding batch. Then, the test results are labeled according to the preset grading rules. Specifically, the average grain size is divided into multiple grain size level labels, the ratio of main phase to secondary phase content is divided into multiple phase composition ratio labels, and the consistency of magnetic domain arrangement direction is divided into multiple magnetic domain arrangement orientation labels. The above labels are used as preset magnetic property target attribute labels for the corresponding historical sintering batches. The preset magnetic property target attribute labels are bound to the multi-source time-series samples of the corresponding sintering batches to form a supervised dataset for subsequent linear probe training.

[0025] After obtaining the preset magnetic property target attribute labels, multiple intermediate activation signals at multiple denoising time steps are used as inputs to train a linear probe set to perform posterior classification of the preset magnetic property target attribute labels. Specifically, after the digital twin model is trained, the network parameters in the digital twin model are kept constant, and corresponding linear probes are connected to the output of each attention layer of the digital twin model. These linear probes can be constructed based on logistic regression classifiers, single-layer fully connected classifiers, or linear support vector classifiers. Taking a single-layer fully connected classifier as an example, the linear probe includes an input layer and an output layer. The dimension of the input layer is consistent with the dimension of the intermediate activation fusion vector, and the output layer includes three parallel linear classification heads, corresponding to grain size level, phase composition ratio, and magnetic domain alignment orientation, respectively. If the first, second, and third activation signals are all 128-dimensional, then the concatenated intermediate activation fusion vector is 384-dimensional, and the input dimension of the linear probe is set to 384-dimensional. The output dimension of the grain size classification head is set to the number of grain size categories, the output dimension of the phase composition ratio classification head is set to the number of phase composition ratio categories, and the output dimension of the magnetic domain alignment orientation classification head is set to the number of magnetic domain alignment orientation categories. Each classification head consists of a linear mapping layer and a Softmax output layer, used to output the category probability of the corresponding attribute label.

[0026] For any historical sintering sample, the first, second, and third activation signals extracted from multiple denoising time steps are concatenated to form an intermediate activation fusion vector for the corresponding denoising time step. This intermediate activation fusion vector is then input into the linear probe following the corresponding self-attention layer. The linear probe outputs the predicted grain size, phase composition ratio, and magnetic domain orientation. During training, only the weights and bias parameters of the linear mapping layer in the linear probe are updated; the parameters of the digital twin model are not updated. Then, the three attribute predictions output by the linear probe are compared with the true grain size label, the true phase composition ratio label, and the true magnetic domain orientation label, respectively. Three cross-entropy losses are calculated, and the three cross-entropy losses are weighted and summed according to preset weights to obtain the total loss of the linear probe. Gradient descent optimization is then used to iteratively update the linear probe parameters until the classification accuracy of the validation set reaches the preset convergence condition. After training, the classification accuracy of each linear probe at the corresponding self-attention layer and the corresponding denoising time step is recorded for subsequent implicit decision layer selection. Next, the classification accuracy of each self-attention layer for grain size level, phase composition ratio and magnetic domain arrangement orientation at multiple denoising time steps was statistically analyzed. Self-attention layers with classification accuracy higher than a preset threshold were marked as candidate layers. The changes in classification accuracy of candidate layers at different denoising time steps were then analyzed. The self-attention layer with the highest classification accuracy and the largest gradient of classification accuracy change was selected as the implicit decision layer. This implicit decision layer indicates that the digital twin model has formed a latent variable expression that is strongly correlated with the final magnetic performance generation result. Its corresponding network layer can reflect the evolution trend of the magnetic performance generation path during sintering.

[0027] After determining the implicit decision layer, the attention weight matrix of the implicit decision layer at the denoising time step is read, and the corresponding attention map is generated according to the attention weight matrix to characterize the correlation strength between different spatial regions, different time steps and different physical field features. Then, the intermediate activation fusion vector output by the implicit decision layer is extracted, and the intermediate activation fusion vectors corresponding to multiple historical sintering samples are mapped to a unified feature space. Then, cluster statistics are performed on the intermediate activation fusion vectors according to different magnetic property target attribute labels. That is, for any magnetic property target attribute label category, all intermediate activation fusion vectors belonging to that category are formed into a category sample set. The mean of each intermediate activation fusion vector in the category sample set is calculated, and the mean vector is used as the feature center position of that category. The Euclidean distance between each intermediate activation fusion vector in that category and the feature center position is calculated, and the vectors are sorted in ascending order of distance. The distance value corresponding to a preset coverage ratio is selected as the feature boundary radius of that category. For example, the distance value that can cover 95% of the samples in that category is selected as the feature boundary radius. The feature boundary range of that category is determined by the feature center position and the feature boundary radius. Using the feature center position as the kernel density estimation benchmark, Gaussian kernel density is calculated on the intermediate activation fusion vectors in the category sample set to obtain the feature vector spatial density distribution of that category in the feature space. By summarizing the calculated feature center locations, feature boundary ranges, and feature vector spatial density distributions corresponding to each category, an implicit decision-making layer feature distribution is formed. This implicit decision-making layer feature distribution is used to characterize the spontaneous generation tendency of the current sintering process within the digital twin model, thereby providing a basis for subsequent real-time feature deviation analysis and temperature field guided compensation.

[0028] Furthermore, this includes: The linear probe set is used to traverse the entire denoising time step sequence of the digital twin model, including the multiple denoising time steps; the classification accuracy of the self-attention layer activation signal for the preset magnetic property target attribute label at each denoising time step in the entire denoising time step sequence is detected to obtain the classification accuracy sequence; the gradient value of the classification accuracy sequence as a function of time steps is calculated, and the time step interval with the largest absolute value of the gradient value is determined as the screening denoising time step.

[0029] Optionally, the trained set of linear probes can be used to perform a comprehensive analysis of the entire denoising time step sequence in the digital twin model. Specifically, during the inference process of a sintering sample completed by the digital twin model, the diffusion denoising process is divided into a series of entire denoising time steps according to a preset number of time steps, for example, denoised as the 1st denoising time step to the Tth denoising time step. For any denoising time step, the activation signal output by each attention layer of the digital twin model at that denoising time step is read, and the activation signal is input into the linear probe corresponding to the self-attention layer. The linear probe outputs the classification prediction results of grain size level, phase composition ratio, and magnetic domain alignment orientation. Subsequently, the classification prediction results output by the linear probe are compared with the preset magnetic property target attribute labels, and the classification accuracy rates of grain size level, phase composition ratio, and magnetic domain alignment orientation are calculated respectively. These are then weighted and summed according to preset weights to obtain the comprehensive classification accuracy rate corresponding to the self-attention layer at that denoising time step. Following the same method, the comprehensive classification accuracy is calculated sequentially for each denoising time step in the entire denoising time step sequence, and the results are arranged in order to obtain a classification accuracy sequence. Then, the comprehensive classification accuracy of the previous denoising time step is subtracted from the comprehensive classification accuracy of the next denoising time step to obtain the accuracy change between adjacent denoising time steps. This change is used as the gradient value of the corresponding time step interval. The absolute values ​​of the gradient values ​​for all time step intervals are then taken, and the time step interval with the largest absolute value is selected as the interval with the most drastic change in classification accuracy. This interval is then designated as the selected denoising time step. This selected denoising time step represents the key stage in the digital twin model most relevant to the rapid formation or transformation of magnetic property generation tendency. It can serve as the preferred time position for subsequent extraction of implicit decision-making layer feature distribution and execution of activation-guided intervention, thereby improving the timeliness and accuracy of subsequent temperature field compensation control.

[0030] Real-time monitoring of the physical field state inside the sintering furnace to obtain the real-time feature vector of the implicit decision layer.

[0031] In one embodiment, after determining the implicit decision-making layer and the denoising time step, the physical field state inside the sintering furnace is monitored in real time, and the real-time monitoring results are input into the digital twin model. The feature encoding layer of the digital twin model generates current response encoding vector, temperature space encoding vector, and vacuum degree encoding vector, respectively. The three types of encoding vectors are input into the self-attention layer for correlation calculation to generate the real-time feature vector of the implicit decision-making layer under the current sintering state. This vector is used to characterize the comprehensive state of the current furnace heating response, temperature space distribution, and phase transition disturbance in the implicit decision-making layer, thereby providing a real-time data basis for subsequent judgment on whether the current sintering state deviates from the preset magnetic property generation direction.

[0032] Furthermore, real-time monitoring of the physical field state within the sintering furnace is used to obtain the real-time feature vector of the implicit decision layer, including: The physical field state inside the sintering furnace is monitored in real time to obtain the current response feature vector corresponding to the first activation signal, the temperature spatial distribution feature vector corresponding to the second activation signal, and the vacuum degree feature vector corresponding to the third activation signal. The current response feature vector, temperature spatial distribution feature vector, and vacuum degree feature vector are input into the digital twin model, and the real-time feature vector in the implicit decision layer is updated through the computational flow of the self-attention mechanism.

[0033] Preferably, when the sintering furnace is in operation, the physical field state inside the furnace is monitored in real time according to a preset sampling period. The current value of each heating unit, the infrared array temperature matrix, and the furnace vacuum value are collected separately, and then the three types of data are synchronized and aligned according to the same sampling timestamp. Subsequently, the current values ​​of each heating unit are arranged in order of heating unit number to form a current sampling sequence, and after normalization, the current response feature vector corresponding to the first activation signal is obtained. The infrared array temperature matrix is ​​expanded into a temperature space matrix according to the furnace spatial coordinates, and after spatial position encoding, the temperature spatial distribution feature vector corresponding to the second activation signal is obtained. The furnace vacuum value is arranged into a vacuum time series according to a continuous sampling window, and after rate of change calculation and normalization, the vacuum feature vector corresponding to the third activation signal is obtained. Then, the current response feature vector, temperature spatial distribution feature vector, and vacuum feature vector are input into a pre-trained digital twin model. The feature encoding layer converts the three types of feature vectors into a unified-dimensional embedding vector, and the embedding vector is sent to the self-attention layer. The self-attention layer generates attention weights based on the correlation between current response characteristics, temperature spatial distribution characteristics, and vacuum degree characteristics. These three types of features are then weighted and fused based on these attention weights to obtain the fused latent variable corresponding to the current sampling time. Next, the digital twin model updates its state according to a predetermined filtering and denoising time step. When the computation flow reaches the implicit decision layer, the hidden state vector at the output of this implicit decision layer is read and used as the real-time feature vector corresponding to the current physical field state of the sintering furnace. This allows the real-time physical field changes within the furnace to be mapped into the implicit decision layer space related to the magnetic property generation tendency, providing a real-time basis for subsequent deviation judgment and guiding vector calculation.

[0034] Based on the real-time feature vector, the guiding vector is calculated according to the feature distribution of the implicit decision layer, and the direction of decision modification is output.

[0035] In one embodiment, after obtaining the real-time feature vector of the implicit decision layer, the real-time feature vector obtained at the current sampling moment is input into the unified feature space where the feature distribution of the implicit decision layer is located, so that its real-time state is in the same coordinate system as the feature center, feature boundary range, and feature vector spatial density distribution corresponding to different magnetic performance target attributes in historical samples. Subsequently, the optimal performance benchmark region corresponding to the preset magnetic performance target attribute label is selected as the target region, and the directional difference between the real-time feature vector and the feature center of the target region is calculated. Then, the guiding vector is determined by combining the distance from the real-time feature vector to the feature boundary of the target region, which serves as the decision deviation vector to represent the direction and magnitude of the current sintering state's movement in the implicit decision layer feature space. Finally, the decision deviation vector is normalized to obtain the decision modification direction output, which is used to characterize the trend of the latent variable change that needs to be corrected in the current sintering process. This provides a basis for subsequently triggering the matching layer activation guidance command and generating the temperature field adjustment increment signal, thereby improving the consistency and stability of the final magnetic properties of the magnetic material.

[0036] Furthermore, combining the real-time feature vector, a guiding vector is calculated based on the feature distribution of the implicit decision layer to output the direction of decision modification, including: The real-time feature vector is mapped to the feature space where the implicit decision layer feature distribution is located; the direction cosine distance between the real-time feature vector and the preset optimal performance benchmark manifold in the implicit decision layer feature distribution is calculated to obtain the decision deviation vector; the decision deviation vector is normalized and inversely projected to generate a decision modification direction pointing to the preset magnetic performance target attribute label.

[0037] Optionally, after obtaining the real-time feature vector corresponding to the implicit decision layer, the feature space coordinate reference, feature center matrix, and spatial distribution parameters saved when constructing the feature distribution of the implicit decision layer are read. The current real-time feature vector is then input into a unified mapping function for spatial coordinate transformation according to the same feature dimension order, so that the real-time feature vector and the intermediate activation fusion vector corresponding to the historical sintering sample are in the same latent variable feature space. Subsequently, the corresponding optimal performance benchmark manifold is read according to the preset magnetic performance target attribute label. This optimal performance benchmark manifold is a continuous high-density distribution region formed by the set of feature vectors that meet the target magnetic performance requirements in the historical sintering sample within the feature space of the implicit decision layer. Then, taking the location of the real-time feature vector as the starting point and the nearest manifold feature point in the optimal performance benchmark manifold as the target point, a real-time deviation direction vector is constructed. The cosine angle between the real-time deviation direction vector and the local tangent vector of the optimal performance benchmark manifold is calculated, and the direction cosine distance is obtained based on the cosine angle. If the direction cosine distance is greater than the preset deviation threshold, it is determined that the evolution direction of the current real-time feature vector has deviated from the target magnetic performance generation path. Then, the real-time deviation direction vector and the direction cosine distance are weighted and combined to generate a decision deviation vector, where the direction cosine distance characterizes the degree of deviation, and the magnitude of the real-time deviation direction vector characterizes the magnitude of deviation. The decision deviation vector is then normalized to limit its magnitude within a preset range to eliminate the influence of feature scale differences at different sintering stages. Based on the pre-established feature mapping relationship between the implicit decision layer and the digital twin model's input control variables, the normalized decision deviation vector is inversely projected to map the deviation direction in the latent variable space to the corresponding heating region, temperature field direction, and power change trend control direction. Finally, the inversely projected control direction is determined as the decision modification direction, characterizing the temperature field correction trend that needs adjustment in the current sintering process, thus providing a basis for subsequently generating activation guidance commands and temperature field compensation control quantities.

[0038] The activation guidance instruction of the matching layer is triggered according to the decision modification direction. The activation guidance instruction is used to generate an adjustment increment signal superimposed on the basic control output of the temperature field, which drives the underlying actuator to perform instantaneous deviation compensation of the temperature field.

[0039] In one embodiment, after obtaining the decision modification direction, the change value of the feature dimension in the decision modification direction is read, and the corresponding compensation position is determined according to the pre-established mapping relationship between the feature dimension and the physical coordinates of the heating unit. This mapping relationship is established during the training phase of the digital twin model and is used to record the correspondence between each feature dimension of the implicit decision layer and different heating areas of the sintering furnace. For example, when the feature dimension offset of the corresponding left side of the furnace in the decision modification direction is the largest, the heating unit on the left side of the furnace is determined to be the current priority compensation area. Subsequently, the compensation degree is calculated according to the vector magnitude of the decision modification direction, and the compensation degree is converted into the power correction amount of the corresponding heating unit. When the current temperature change rate is lower than the preset change rate threshold, the compensation amount is increased; when the current temperature change rate is higher than the preset change rate threshold, the compensation amount is decreased, thereby avoiding temperature field overshoot. Afterwards, the compensation position is encoded as the target heating unit address, the compensation degree is encoded as the corresponding power correction value, and both are encapsulated into an activation guidance command and sent to the underlying controller. Upon receiving the activation guidance command, the bottom-level controller superimposes an adjustment increment signal onto the existing PID temperature control output, forming a compensation pulse signal for instantaneous temperature field correction. This compensation pulse signal is output to the corresponding heating unit with a millisecond-level cycle. By briefly increasing or decreasing the local heating power, it preemptively corrects implicit deviation trends in the temperature field without altering the overall sintering process curve. Finally, the bottom-level actuator drives the corresponding heating unit to output instantaneous compensation heat according to the adjustment increment signal, adjusting the local temperature field towards the evolution direction corresponding to the preset magnetic performance target. This suppresses implicit formation tendencies such as abnormal grain growth, phase shift, or uneven magnetic domain orientation, achieving real-time guidance and control of the microstructure evolution during the sintering process.

[0040] Furthermore, based on the decision modification direction, an activation guidance instruction for the matching layer is triggered, and the activation guidance instruction is used to generate an adjustment increment signal superimposed on the temperature field basic control output, including: Extract the feature dimension index with the highest contribution from the decision modification direction, match the corresponding physical coordinates of the underlying heating unit through a preset spatial mapping table, and determine the compensation location information; calculate the power compensation amount required to correct the implicit generation tendency based on the modulus of the decision modification direction in the feature space and the real-time temperature change rate, and determine the compensation degree information; trigger the activation guidance instruction of the matching layer based on the compensation location information and the compensation degree information.

[0041] Preferably, after obtaining the decision modification direction, the decision modification direction is represented as a direction vector composed of multiple feature dimension components. For example, multiple feature dimensions related to temperature field, heat flow field, and tissue evolution in the feature space of the implicit decision layer, such as the current response dimension corresponding to the power fluctuation of the heating unit, the hot spot offset dimension corresponding to the local overheating tendency, etc., are calculated respectively. Then, the feature dimension with the largest absolute value is determined as the feature dimension index with the highest contribution. Subsequently, a preset spatial mapping table is called to query the corresponding underlying heating unit number, heating unit physical coordinates, and control channel address based on the feature dimension index. This preset spatial mapping table is established during the digital twin model calibration stage and is used to record the correspondence between the feature dimensions of the implicit decision layer and the heating area in the sintering furnace. Through the above query results, the location of the heating unit that needs to be temperature-compensated is determined, and the heating unit number, physical coordinates, and control channel address are used as the compensation location information. Next, the square root of the sum of squares of each dimension component of the decision modification direction is calculated to obtain the vector magnitude of the decision modification direction. Then, the real-time temperature change rate of the heating area corresponding to the compensation location information within the current sampling period is obtained. The vector magnitude is used as the implicit deviation intensity, and the real-time temperature change rate is used as the temperature response state correction factor. The power compensation amount is calculated according to the preset power mapping coefficient. If the decision modification direction indicates that the current area needs to enhance the heat input, the power compensation amount is determined as a positive compensation amount; if the decision modification direction indicates that the current area needs to weaken the heat input, the power compensation amount is determined as a negative compensation amount. Thus, compensation degree information including compensation direction, compensation amplitude, and compensation duration is obtained. Finally, the compensation location information is mapped to the neuron activation index, the compensation degree information is converted into a weight offset, and encapsulated to form an activation guidance instruction that the matching layer can recognize. After receiving the activation guidance instruction, the matching layer sends it to the corresponding bottom controller, enabling the bottom controller to generate a corresponding temperature control compensation pulse in addition to the basic temperature control output, thereby realizing instantaneous deviation compensation for the target heating area.

[0042] Furthermore, triggering the activation guidance instruction of the matching layer based on the compensation location information and the compensation degree information includes: The compensation location information is mapped to the neuron activation index of the corresponding implicit decision layer in the digital twin model; the compensation degree information is converted into a weight offset pointing to a preset magnetic performance target attribute label, and the weight offset and the neuron activation index are encapsulated to generate an activation guidance instruction.

[0043] Optionally, after obtaining the compensation location information and compensation degree information, the heating unit number, physical coordinates, and control channel address in the compensation location information are read, and a preset spatial mapping table is called to query the furnace space node number, self-attention layer number, and implicit decision layer level number corresponding to the heating unit in the digital twin model. Subsequently, the furnace space node number is used as the spatial channel identifier, the self-attention channel number as the feature channel identifier, and the implicit decision layer level number as the level identifier. These three are combined to form a neuron activation index. This neuron activation index is used to determine the specific level and spatial channel for which activation guidance needs to be implemented, so that subsequent guidance operations can act on the implicit decision layer neuron position corresponding to the target heating area in the digital twin model. Afterwards, the power compensation amount, compensation direction, and compensation duration in the compensation degree information are read, and the power compensation amount is converted into a weight adjustment amplitude in the implicit decision layer according to a preset weight mapping coefficient. When the compensation direction is to enhance the heat input direction, the weight adjustment amplitude is set to a positive weight offset; when the compensation direction is to weaken the heat input direction, the weight adjustment amplitude is set to a negative weight offset. Then, based on the target feature center corresponding to the preset magnetic property target attribute label, the weight adjustment amplitude is directionally corrected to point to the feature distribution region corresponding to the target grain size level, target phase composition ratio, and target magnetic domain arrangement orientation, thus obtaining the weight offset. Finally, the neuron activation index is written into the index field of the activation guidance instruction, the weight offset is written into the offset field, the preset magnetic property target attribute label is written into the target field, and the compensation duration is written into the duration field, forming an activation guidance instruction containing the action level, spatial channel, offset direction, offset amplitude, and target label. Through this activation guidance instruction, the location and guidance intensity of the implicit decision layer that needs to be guided in the digital twin model can be clearly defined, providing a basis for converting the activation offset at the model level into temperature control compensation pulses of the underlying heating unit, thereby improving the matching accuracy between temperature field compensation control and the target magnetic property generation direction.

[0044] Furthermore, this includes: The weight offset in the activation guidance instruction is analyzed and mapped to the current duty cycle correction value of the corresponding underlying heating unit; a high-frequency pulse current is generated at the actuator address pointed to by the compensation position information according to the duty cycle correction value to obtain the adjustment increment signal.

[0045] Optionally, after receiving the activation guidance command, the underlying controller first reads the neuron activation index, weight offset, target attribute label, and timeliness field from the activation guidance command, and determines the corresponding compensation position information based on the neuron activation index. Then, it converts the weight offset into the current duty cycle correction value for the corresponding underlying heating unit according to a preset offset-duty cycle mapping coefficient. If the weight offset is positive, the current duty cycle correction value is used as a positive increment to increase the conduction ratio of the target heating unit in the current sampling period; if the weight offset is negative, the current duty cycle correction value is used as a negative increment to decrease the conduction ratio of the target heating unit in the current sampling period. After obtaining the current duty cycle correction value, the underlying controller reads the base duty cycle corresponding to the temperature field's basic control output and superimposes the current duty cycle correction value with the base duty cycle to obtain the pulse control duty cycle of the target heating unit. To avoid compensation overshoot, upper and lower limits are constrained for the pulse control duty cycle, ensuring it does not exceed a preset maximum duty cycle and is not lower than a preset minimum duty cycle. Subsequently, the underlying controller determines the pulse duration based on the timing field in the activation guidance command and outputs the corresponding high-frequency pulse control signal at the actuator address pointed to by the compensation position information. This high-frequency pulse control signal is used to drive the target heating unit to generate a high-frequency pulse current within a millisecond-level control cycle. The pulse amplitude is determined by the rated current of the target heating unit, the pulse conduction ratio is determined by the pulse control duty cycle, and the pulse duration is determined by the timing field in the activation guidance command. By superimposing this high-frequency pulse current on top of the basic temperature control output, an adjustment increment signal is obtained. This adjustment increment signal is used to temporarily enhance or weaken the heat input to the target heating area, thereby instantaneously compensating for local deviations in the temperature field without changing the overall sintering temperature curve, improving the stability of the sintering process and the quality of the magnetic material product.

[0046] Example 2, based on the same inventive concept as the intelligent temperature field control method for magnetic material production in the foregoing examples, such as... Figure 2 As shown in the embodiment of this application, a temperature field intelligent control system for the production of magnetic materials is provided. The temperature field intelligent control system for the production of magnetic materials includes: Signal acquisition module 11: Constructs a digital twin model of the magnetic material sintering process, acquires multiple intermediate activation signals at multiple denoising time steps within the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal, and a third activation signal; Signal classification module 12: Performs attribute classification accuracy analysis on the multiple intermediate activation signals by constructing a linear probe set, and obtains the implicit decision layer feature distribution; Real-time monitoring module 13: Monitors the physical field state within the sintering furnace in real time, and acquires the real-time feature vector of the implicit decision layer; Vector calculation module 14: Combines the real-time feature vector with the implicit decision layer feature distribution to perform guided vector calculation and outputs the decision modification direction; Deviation compensation module 15: Triggers the activation guidance command of the matching layer according to the decision modification direction, and uses the activation guidance command to generate an adjustment increment signal superimposed on the temperature field basic control output, driving the underlying actuator to perform instantaneous deviation compensation for the temperature field.

[0047] Furthermore, the signal acquisition module 11 is used to perform the following operation steps: The first activation signal is a current response feature vector based on the power fluctuation of the sintering furnace, used to characterize the latent variable of the operating state of the heating system; the second activation signal is a temperature spatial distribution feature vector based on infrared array sampling, used to characterize the consistency latent variable of the heat flow field in the furnace; the third activation signal is a vacuum degree feature vector based on the pressure fluctuation of the magnetic transition point, used to characterize the phase transition latent variable of the material microstructure evolution.

[0048] Furthermore, the signal classification module 12 is used to perform the following operation steps: Obtain preset magnetic property target attribute labels, wherein the preset magnetic property target attribute labels include grain size level, phase composition ratio, and magnetic domain alignment orientation; use multiple intermediate activation signals under multiple denoising time steps as input to train a linear probe set to perform posterior classification of the preset magnetic property target attribute labels; select a specific layer with a classification accuracy greater than a preset threshold and located at the self-attention level in the digital twin model to obtain an implicit decision layer, wherein the implicit decision layer is a network layer that intervenes in the magnetic property generation path; construct the implicit decision layer feature distribution by extracting the weight matrix, attention map, and feature vector space density of the implicit decision layer under the denoising time step, wherein the implicit decision layer feature distribution is used to characterize the spontaneous generation tendency of the current sintering process.

[0049] Furthermore, the signal classification module 12 is used to perform the following operation steps: The linear probe set is used to traverse the entire denoising time step sequence of the digital twin model, including the multiple denoising time steps; the classification accuracy of the self-attention layer activation signal for the preset magnetic property target attribute label at each denoising time step in the entire denoising time step sequence is detected to obtain the classification accuracy sequence; the gradient value of the classification accuracy sequence as a function of time steps is calculated, and the time step interval with the largest absolute value of the gradient value is determined as the screening denoising time step.

[0050] Furthermore, the real-time monitoring module 13 is used to perform the following operation steps: The physical field state inside the sintering furnace is monitored in real time to obtain the current response feature vector corresponding to the first activation signal, the temperature spatial distribution feature vector corresponding to the second activation signal, and the vacuum degree feature vector corresponding to the third activation signal. The current response feature vector, temperature spatial distribution feature vector, and vacuum degree feature vector are input into the digital twin model, and the real-time feature vector in the implicit decision layer is updated through the computational flow of the self-attention mechanism.

[0051] Furthermore, the vector calculation module 14 is used to perform the following operation steps: The real-time feature vector is mapped to the feature space where the implicit decision layer feature distribution is located; the direction cosine distance between the real-time feature vector and the preset optimal performance benchmark manifold in the implicit decision layer feature distribution is calculated to obtain the decision deviation vector; the decision deviation vector is normalized and inversely projected to generate a decision modification direction pointing to the preset magnetic performance target attribute label.

[0052] Furthermore, the deviation compensation module 15 is used to perform the following operation steps: Extract the feature dimension index with the highest contribution from the decision modification direction, match the corresponding physical coordinates of the underlying heating unit through a preset spatial mapping table, and determine the compensation location information; calculate the power compensation amount required to correct the implicit generation tendency based on the modulus of the decision modification direction in the feature space and the real-time temperature change rate, and determine the compensation degree information; trigger the activation guidance instruction of the matching layer based on the compensation location information and the compensation degree information.

[0053] Furthermore, the deviation compensation module 15 is used to perform the following operation steps: The compensation location information is mapped to the neuron activation index of the corresponding implicit decision layer in the digital twin model; the compensation degree information is converted into a weight offset pointing to a preset magnetic performance target attribute label, and the weight offset and the neuron activation index are encapsulated to generate an activation guidance instruction.

[0054] Furthermore, the deviation compensation module 15 is used to perform the following operation steps: The weight offset in the activation guidance instruction is analyzed and mapped to the current duty cycle correction value of the corresponding underlying heating unit; a high-frequency pulse current is generated at the actuator address pointed to by the compensation position information according to the duty cycle correction value to obtain the adjustment increment signal.

[0055] Through the foregoing detailed description of the intelligent temperature field control method for magnetic material production, those skilled in the art can clearly understand the intelligent temperature field control system for magnetic material production in this embodiment. Since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0056] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent control of temperature field for production of magnetic materials, characterized in that, The method includes: A digital twin model of the sintering process of magnetic materials is constructed to obtain multiple intermediate activation signals at multiple noise reduction time steps in the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal and a third activation signal; By constructing a linear probe set, the attribute classification accuracy of the multiple intermediate activation signals is analyzed to obtain the feature distribution of the implicit decision layer; Real-time monitoring of the physical field state inside the sintering furnace to obtain the real-time feature vector of the implicit decision layer; Combining the real-time feature vector, a guiding vector is calculated based on the feature distribution of the implicit decision layer, and the direction of decision modification is output. The activation guidance instruction of the matching layer is triggered according to the decision modification direction. The activation guidance instruction is used to generate an adjustment increment signal superimposed on the basic control output of the temperature field, which drives the underlying actuator to perform instantaneous deviation compensation of the temperature field.

2. The temperature field intelligent control method for magnetic material production of claim 1, wherein, The first activation signal is a current response feature vector based on the power fluctuation of the sintering furnace, which is used to characterize the latent variables of the operating state of the heating system; The second activation signal is a temperature spatial distribution feature vector based on infrared array sampling, used to characterize the consistency latent variable of the heat flow field inside the furnace; The third activation signal is a vacuum degree feature vector based on the pressure fluctuation at the magnetic transition point, used to characterize the phase transition latent variable of material microstructure evolution.

3. The temperature field intelligent control method for magnetic material production of claim 1, wherein, By constructing a linear probe set to perform attribute classification accuracy analysis on the multiple intermediate activation signals, the feature distribution of the implicit decision layer is obtained, including: Obtain preset magnetic performance target attribute tags, wherein the preset magnetic performance target attribute tags include grain size level, phase composition ratio and magnetic domain arrangement orientation; Using multiple intermediate activation signals from multiple denoising time steps as input, a linear probe set is trained to perform posterior classification of the preset magnetic property target attribute label; A specific layer with a classification accuracy greater than a preset threshold and located at the self-attention level in the digital twin model is selected to obtain the implicit decision layer, wherein the implicit decision layer is a network layer that intervenes in the magnetic property generation path; By extracting the weight matrix, attention map, and feature vector space density of the implicit decision layer at the screening and denoising time step, the feature distribution of the implicit decision layer is constructed, wherein the feature distribution of the implicit decision layer is used to characterize the spontaneous generation tendency of the current sintering process.

4. The temperature field intelligent control method for magnetic material production of claim 3, wherein, include: The linear probe set is used to traverse the entire denoised time step sequence of the digital twin model, including the multiple denoised time steps. The classification accuracy of the self-attention layer activation signal at each denoising time step in the entire denoising time step sequence for the preset magnetic property target attribute label is detected, and a classification accuracy sequence is obtained. Calculate the gradient value of the classification accuracy sequence as a function of time steps, and determine the time step interval with the largest absolute value of the gradient value as the filtering and denoising time step.

5. The temperature field intelligent control method for magnetic material production of claim 1, wherein, Real-time monitoring of the physical field state within the sintering furnace to obtain the real-time feature vector of the implicit decision layer, including: Real-time monitoring of the physical field state inside the sintering furnace yields the current response feature vector corresponding to the first activation signal, the temperature spatial distribution feature vector corresponding to the second activation signal, and the vacuum degree feature vector corresponding to the third activation signal. The current response feature vector, temperature spatial distribution feature vector, and vacuum degree feature vector are input into the digital twin model, and the real-time feature vector in the implicit decision layer is updated through the computational flow of the self-attention mechanism.

6. The temperature field intelligent control method for magnetic material production of claim 5, wherein, Combining the real-time feature vector, a guiding vector is calculated based on the feature distribution of the implicit decision layer, and the direction of decision modification is output, including: The real-time feature vector is mapped to the feature space where the implicit decision layer feature distribution is located; Calculate the direction cosine distance between the real-time feature vector and the preset optimal performance benchmark manifold in the feature distribution of the implicit decision layer to obtain the decision bias vector; The decision deviation vector is normalized and inversely projected to generate a decision modification direction pointing to a preset magnetic property target label.

7. The temperature field intelligent control method for magnetic material production of claim 1, wherein, The activation guidance instruction of the matching layer is triggered according to the decision modification direction, and the adjustment increment signal superimposed on the temperature field basic control output is generated using the activation guidance instruction, including: Extract the feature dimension index with the highest contribution from the decision modification direction, and match the corresponding physical coordinates of the underlying heating unit through a preset spatial mapping table to determine the compensation location information; Based on the magnitude of the decision modification direction in the feature space, combined with the real-time temperature change rate, the power compensation required to correct the implicit generation tendency is calculated, and the compensation degree information is determined. The activation guidance instruction of the matching layer is triggered based on the compensation location information and the compensation degree information.

8. The temperature field intelligent control method for magnetic material production of claim 7, wherein, The activation guidance instruction for the matching layer is triggered based on the compensation location information and the compensation degree information, including: The compensation location information is mapped to the neuron activation index of the corresponding implicit decision layer in the digital twin model; The compensation level information is converted into a weight offset pointing to a preset magnetic performance target attribute label, and the weight offset is encapsulated with the neuron activation index to generate an activation guidance instruction.

9. The temperature field intelligent control method for magnetic material production of claim 8, wherein, include: The weight offset in the activation guidance instruction is analyzed and mapped to the current duty cycle correction value of the corresponding underlying heating unit. Based on the duty cycle correction value, a high-frequency pulse current is generated at the actuator address indicated by the compensation position information to obtain the adjustment increment signal.

10. A temperature field intelligent control system for magnetic material production, characterized in that, For implementing the intelligent temperature field control method for magnetic material production according to any one of claims 1-9, the system comprises: Signal acquisition module: Constructs a digital twin model of the magnetic material sintering process and acquires multiple intermediate activation signals at multiple noise reduction time steps in the sintering furnace, wherein the intermediate activation signals include a first activation signal, a second activation signal and a third activation signal; Signal classification module: By constructing a linear probe set, the module performs attribute classification accuracy analysis on the multiple intermediate activation signals to obtain the feature distribution of the implicit decision layer; Real-time monitoring module: Real-time monitoring of the physical field state inside the sintering furnace to obtain the real-time feature vector of the implicit decision layer; Vector calculation module: Combines the real-time feature vector, performs guided vector calculation based on the feature distribution of the implicit decision layer, and outputs the direction of decision modification; Deviation Compensation Module: Based on the decision modification direction, the matching layer is triggered to activate the guidance instruction. The activation guidance instruction is used to generate an adjustment increment signal superimposed on the basic control output of the temperature field, which drives the underlying actuator to perform instantaneous deviation compensation for the temperature field.

Images