A digital twin method for high-temperature annealing of silicon steel

Abstract

This invention relates to the field of digital twin technology and discloses a digital twin method for high-temperature annealing of silicon steel. The method includes: collecting temporal characteristic data of the temperature field and projecting it onto a manifold space to construct a feature vector; calculating transient evolution residual data using a state machine based on the state transition weight matrix of the previous cycle; transitioning to a mechanism drift state when the temporal evolution gradient of the residual data meets set conditions; calling a boundary conservation projection optimization operator; generating a physical constraint boundary factor based on the product mapping of the cumulative power scalar and the spatial divergence of the feature vector; thereby smoothly correcting the state transition weight matrix to reconstruct the digital twin model and output feedforward power compensation data. This invention effectively suppresses control overshoot and power oscillations through state discretization decoupling and physical boundary constraint interlocking, ensuring uniform temperature distribution of the steel coil and improving model tracking accuracy.

Description

Technical Field

[0001] This invention relates to a digital twin method for high-temperature annealing of silicon steel, belonging to the field of digital twin technology. Background Technology

[0002] Currently, in the control framework of industrial furnaces and kilns, establishing a digital image model of the physical production line to complete closed-loop control of the temperature field distribution in the production space is a common control method. Conventional technical paths usually rely on steady heat transfer equations or static topology structures, constructing a simulation channel at the virtual control terminal to calculate the temperature state during the heating process.

[0003] However, when this scheme is applied to the continuous annealing process, the furnace heat transfer coefficient, furnace wall heat storage inertia, and steel coil emissivity drift with the progress of the operation. The inherent structural characteristics of the static model cannot adaptively fit this nonlinear evolution, causing a trend of disconnect between the predicted parameters and the actual physical state. The mapped parameters exhibit cumulative distortion. The temperature measuring components have thermal inertia and spatial limitations, making it difficult to intuitively and in real-time obtain the multidimensional dynamic characteristics inside the furnace. Existing digital monitoring and control methods in the control software layer have shortcomings. For example, Chinese invention patent CN113930600B discloses a monitoring and control method for the annealing process of a bell-type furnace based on digital twin technology. This method relies on physical monitoring devices to collect data and uses conventional convection heat transfer models and flow velocity calculation equations to infer the flow field distribution inside the furnace. This method is based on the premise that physical parameters such as channel equivalent roughness are idealized and steady. It does not construct physical boundary constraints to cope with the long-period evolution of physical property parameters. When facing highly nonlinear evolution conditions such as high-temperature annealing of silicon steel, it lacks thermodynamics. Energy conservation boundary constraints cause structural errors in the model to diverge numerically during time-series iterations, leading to cumulative distortion of mapping parameters in nonlinear evolution. To maintain tracking accuracy, parameter calibration is usually performed empirically or with fixed-value open-loop compensation. However, such adjustment methods are insufficient to reveal the potential risks of temperature field evolution. Furthermore, due to the lag in calibration parameters, overshooting of the adjustment components causes frequent oscillations in output power. To mitigate deviations, conventional improvement approaches tend to increase network weights or introduce full error feedback. However, the data collected by the physical circuit is accompanied by random temperature measurement noise. If the feedback gain is excessively increased, high-frequency transient thermal disturbances and low-frequency trend mechanism drift become confused, causing the filter to overfit to random noise. When updating the mapping network, transient disturbances trigger unnecessary matrix reconstruction, causing numerical singularities and solution jumps in the weight matrix during iteration, resulting in model numerical divergence. Therefore, achieving causal decoupling between trend mechanism drift and random noise under high-temperature conditions is a difficult contradiction to overcome in the existing architecture.

[0004] Therefore, the technical problem to be solved by this invention is how to design a method that can accurately identify mechanism drift in nonlinear evolution, complete the in-situ reconstruction of the state transition weight matrix using energy conservation boundaries while isolating random noise, and provide feedforward control parameters to adjust the output power of the heating component. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A digital twin method for high-temperature annealing of silicon steel, comprising the following steps:

[0006] Step S1: Obtain the time series characteristic data of the annealing furnace temperature field and project it onto the low-dimensional manifold space to construct the feature vector. Set up the time series residual data transformation state machine in the standard tracking state inside the control layer and pre-store the state transition weight matrix in the register.

[0007] Step S2: Multiply the state transition weight matrix of the previous period with the feature vector of the previous sample to calculate the predicted feature vector of the current period. Calculate the difference between the predicted feature vector and the feature vector of the current sample by subtraction to obtain the transient evolution residual data.

[0008] Step S3: When the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, control the temporal residual data transition state machine to migrate from the standard tracking state or transient disturbance state to the mechanism drift state.

[0009] Step S4: In the mechanism drift state, the boundary conservation projection optimization operator is invoked to extract the cumulative power scalar based on the historical power data stream. The cumulative power scalar is then multiplied and mapped with the spatial divergence of the feature vector to generate a dimensionless physical constraint boundary factor.

[0010] Step S5: Construct a structural correction increment matrix based on the gradient distribution of transient evolution residual data, use physical constraint boundary factors to smoothly scale the structural correction increment matrix, update the state transition weight matrix through matrix addition to reconstruct the digital twin model, and output feedforward power compensation data to the thyristor power tuning circuit.

[0011] Preferably, when thermal shock causes a single increase in transient evolution residual data without a sustained monotonic gradient, since the set convergence threshold of 0.05 is not exceeded for three consecutive sampling periods, the time-series residual data transformation state machine is kept in the transient disturbance state and the state transition weight matrix is ​​frozen to isolate temperature measurement interference noise. When the absolute value of the temporal evolution gradient of the transient evolution residual data falls below the set convergence threshold of 0.05, the time-series residual data transformation state machine is released from the state lock and returns to the standard tracking state.

[0012] Preferably, step S4 includes the following sub-steps: Step S41, collect the radiation heat transfer coefficient data inside the furnace body and the thermal inertia parameter data of the furnace wall refractory material through the data interface, and perform weighted integration calculation on the radiation heat transfer coefficient data and thermal inertia parameter data with the historical power data stream to obtain the cumulative power scalar; Step S42, calculate the degree of partial differential divergence of the eigenvector in the low-dimensional manifold space to determine the spatial divergence; Step S43, multiply the cumulative power scalar with the spatial divergence to determine the physical constraint boundary factor, and use the physical constraint boundary factor to constrain the value of the state transition weight matrix within the energy conservation boundary interval.

[0013] Preferably, after step S5, the following steps are also included: Step S6, predicting the temperature field trend inside the furnace body in the next sampling period based on the updated state transition weight matrix, and calculating the feedforward power compensation data for offsetting the drift of the time-varying mechanism; Step S7, converting the feedforward power compensation data into a power adjustment pulse width modulation command through the network bus interface and sending it to the thyristor power tuning circuit to adjust the output power of the heating component and control the spatial temperature field deviation between the state simulation curve and the physical state to be within 1%.

[0014] Preferably, step S1 includes the following sub-steps: Step S11, performing standardized filtering and denoising processing on the acquired annealing furnace temperature field time series feature data, which includes the furnace space temperature sequence and heating power feedback sequence collected from multiple points; Step S12, using a manifold mapping algorithm to map the denoised annealing furnace temperature field time series feature data from a high-dimensional physical parameter space to a low-dimensional manifold space with a low-dimensional topological structure; Step S13, extracting basis vectors that can characterize the thermal degradation of refractory materials and the time-varying characteristics of workpiece surface emissivity in the low-dimensional manifold space, and combining them to obtain feature vectors.

[0015] Preferably, step S2 includes the following sub-steps: step S21, constructing a sliding temporal buffer queue inside the temporal residual data transformation state machine, the length of which is fixed at 5; step S22, inputting the feature vectors in adjacent sampling periods into the temporal buffer queue for Euclidean distance calculation to obtain the original residual data; step S23, performing a first-order temporal difference operation on the original residual data to extract the residual change rate to generate transient evolution residual data.

[0016] Preferably, step S3 includes the following sub-steps: Step S31, the temporal residual data conversion state machine dynamically calculates the temporal evolution gradient index of the transient evolution residual data in the sliding temporal buffer queue, with the sampling time interval fixed at 100ms; Step S32, when the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, it is determined that the system has entered the mechanism drift state caused by long-period nonlinear evolution, and the temporal residual data conversion state machine is controlled to migrate to the mechanism drift state and release the state lock to trigger the weight in-situ correction operator.

[0017] Preferably, after updating the state transition weight matrix, the following steps are also included: Step S8, controlling the time-series residual data transition state machine to revert from the mechanism drift state to the standard tracking state, and restoring the adaptive real-time mapping of the spatial temperature field distribution; Step S9, when the digital twin model encounters an external sudden power disturbance, controlling the digital twin model to complete convergence within 2 sampling periods.

[0018] Preferably, the following initial benchmark calibration steps are included before step S1: Step S01, acquiring the initial parameters of the furnace structure and the initial ambient temperature data in the empty furnace state, and constructing the basic geometric topology of the digital twin space; Step S02, measuring the initial response rate of the furnace wall temperature field by injecting a calibration power sequence, and storing the initial response rate as a priori benchmark value, which is used to perform initial deviation elimination on the transient evolution residual data calculated by the time-series residual data conversion state machine in subsequent sampling periods.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. In the digital twin of high-temperature annealing of silicon steel, multiple thermocouples synchronously collect high-dimensional physical temperature data under high-temperature radiation conditions. The manifold mapping module receives the data and separates the temperature measurement noise, extracting the low-dimensional manifold topological feature vector that characterizes the core evolution trend of the physical temperature field. The forward state prediction vector and the feature vector of the previous period are calculated by the subtraction operator to generate the transient evolution residual vector. The historical evolution residual first-in-first-out buffer queue dynamically maintains the residual vector for 5 periods, so that the transient heat conduction characteristics are transformed into the residual gradient distribution of the continuous time window within adjacent sampling periods, eliminating the bottleneck of frequent overshoot and power oscillation of the control loop caused by the interweaving of random noise and mechanism drift.

[0021] 2. The temporal residual resonance state machine dynamically calculates the temporal evolution gradient index of the residuals in the buffer queue and migrates between three mutually exclusive discrete control states. When a transient thermal shock causes a single sudden increase in the residuals, the convergence threshold is not exceeded due to the lack of a continuous monotonic gradient. The state machine remains in the transient disturbance state and freezes the state transition weight matrix to isolate temperature measurement noise. When the degradation of refractory materials or a sudden change in emissivity causes the residuals to accumulate continuously in the same direction and exceed the threshold for three consecutive sampling periods, the state machine migrates to the mechanism drift state and initiates weight correction to avoid model divergence and prevent the system from misidentifying trend mechanism drift as transient random fluctuations.

[0022] 3. When the state machine enters the mechanism drift state, the thermodynamic boundary conservation projection operator is synchronously invoked. Based on the historical output power sequence, the cumulative heat injection power scalar is extracted and multiplied with the spatial divergence of the low-dimensional manifold topological eigenvector to complete the inner product mapping, generating a dimensionless manifold physical constraint boundary factor. The manifold physical constraint boundary factor is multiplied with the structural correction increment matrix by boundary smoothing, and then the state transition weight matrix in the register is updated by matrix addition. This rigidly anchors the matrix iterative evolution in the long-period nonlinear evolution within the boundary range of the energy conservation law, eliminating the weight numerical singularities and non-physical computational solution jumps caused by extreme perturbations. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the reconstruction process of the digital twin model of the annealing temperature field in this invention.

[0024] Figure 2 This is a graph showing the changes in transient evolution residuals and temporal gradient indices of the present invention.

[0025] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0026] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0027] A digital twin method for high-temperature annealing of silicon steel includes the following steps:

[0028] Step S1: Obtain the time series characteristic data of the annealing furnace temperature field and project it onto the low-dimensional manifold space to construct the feature vector. Set up the time series residual data transformation state machine in the standard tracking state inside the control layer and pre-store the state transition weight matrix in the register.

[0029] Step S2: Multiply the state transition weight matrix of the previous period with the feature vector of the previous sample to calculate the predicted feature vector of the current period. Calculate the difference between the predicted feature vector and the feature vector of the current sample by subtraction to obtain the transient evolution residual data.

[0030] Step S3: When the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, control the temporal residual data transition state machine to migrate from the standard tracking state or transient disturbance state to the mechanism drift state.

[0031] Step S4: In the mechanism drift state, the boundary conservation projection optimization operator is invoked to extract the cumulative power scalar based on the historical power data stream. The cumulative power scalar is then multiplied and mapped with the spatial divergence of the feature vector to generate a dimensionless physical constraint boundary factor.

[0032] Step S5: Construct a structural correction increment matrix based on the gradient distribution of transient evolution residual data, use physical constraint boundary factors to smoothly scale the structural correction increment matrix, update the state transition weight matrix through matrix addition to reconstruct the digital twin model, and output feedforward power compensation data to the thyristor power tuning circuit.

[0033] Preferably, when thermal shock causes a single increase in transient evolution residual data without a sustained monotonic gradient, since the set convergence threshold of 0.05 is not exceeded for three consecutive sampling periods, the time-series residual data transformation state machine is kept in the transient disturbance state and the state transition weight matrix is ​​frozen to isolate temperature measurement interference noise. When the absolute value of the temporal evolution gradient of the transient evolution residual data falls below the set convergence threshold of 0.05, the time-series residual data transformation state machine is released from the state lock and returns to the standard tracking state.

[0034] Preferably, step S4 includes the following sub-steps: Step S41, collect the radiation heat transfer coefficient data inside the furnace body and the thermal inertia parameter data of the furnace wall refractory material through the data interface, and perform weighted integration calculation on the radiation heat transfer coefficient data and thermal inertia parameter data with the historical power data stream to obtain the cumulative power scalar; Step S42, calculate the degree of partial differential divergence of the eigenvector in the low-dimensional manifold space to determine the spatial divergence; Step S43, multiply the cumulative power scalar with the spatial divergence to determine the physical constraint boundary factor, and use the physical constraint boundary factor to constrain the value of the state transition weight matrix within the energy conservation boundary interval.

[0035] Preferably, after step S5, the following steps are also included: Step S6, predicting the temperature field trend inside the furnace body in the next sampling period based on the updated state transition weight matrix, and calculating the feedforward power compensation data for offsetting the drift of the time-varying mechanism; Step S7, converting the feedforward power compensation data into a power adjustment pulse width modulation command through the network bus interface and sending it to the thyristor power tuning circuit to adjust the output power of the heating component and control the spatial temperature field deviation between the state simulation curve and the physical state to be within 1%.

[0036] Preferably, step S1 includes the following sub-steps: Step S11, performing standardized filtering and denoising processing on the acquired annealing furnace temperature field time series feature data, which includes the furnace space temperature sequence and heating power feedback sequence collected from multiple points; Step S12, using a manifold mapping algorithm to map the denoised annealing furnace temperature field time series feature data from a high-dimensional physical parameter space to a low-dimensional manifold space with a low-dimensional topological structure; Step S13, extracting basis vectors that can characterize the thermal degradation of refractory materials and the time-varying characteristics of workpiece surface emissivity in the low-dimensional manifold space, and combining them to obtain feature vectors.

[0037] Preferably, step S2 includes the following sub-steps: step S21, constructing a sliding temporal buffer queue inside the temporal residual data transformation state machine, the length of which is fixed at 5; step S22, inputting the feature vectors in adjacent sampling periods into the temporal buffer queue for Euclidean distance calculation to obtain the original residual data; step S23, performing a first-order temporal difference operation on the original residual data to extract the residual change rate to generate transient evolution residual data.

[0038] Preferably, step S3 includes the following sub-steps: Step S31, the temporal residual data conversion state machine dynamically calculates the temporal evolution gradient index of the transient evolution residual data in the sliding temporal buffer queue, with the sampling time interval fixed at 100ms; Step S32, when the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, it is determined that the system has entered the mechanism drift state caused by long-period nonlinear evolution, and the temporal residual data conversion state machine is controlled to migrate to the mechanism drift state and release the state lock to trigger the weight in-situ correction operator.

[0039] Preferably, after updating the state transition weight matrix, the following steps are also included: Step S8, controlling the time-series residual data transition state machine to revert from the mechanism drift state to the standard tracking state, and restoring the adaptive real-time mapping of the spatial temperature field distribution; Step S9, when the digital twin model encounters an external sudden power disturbance, controlling the digital twin model to complete convergence within 2 sampling periods.

[0040] Preferably, the following initial benchmark calibration steps are included before step S1: Step S01, acquiring the initial parameters of the furnace structure and the initial ambient temperature data in the empty furnace state, and constructing the basic geometric topology of the digital twin space; Step S02, measuring the initial response rate of the furnace wall temperature field by injecting a calibration power sequence, and storing the initial response rate as a priori benchmark value, which is used to perform initial deviation elimination on the transient evolution residual data calculated by the time-series residual data conversion state machine in subsequent sampling periods.

[0041] Example 1: In a specific application of a digital twin method for high-temperature annealing of silicon steel, when the control layer adaptively tracks the low-frequency trend drift caused by the long-period nonlinear evolution inside the continuous annealing furnace, the radiation heat transfer coefficient inside the furnace, the thermal inertia of the furnace wall refractory material, and the emissivity of the workpiece surface undergo unpredictable time-varying drifts due to the continuous advancement of the annealing cycle. Traditional technical approaches, which assume that the structural thermodynamic parameters inside the furnace are quasi-static within a single annealing cycle, are prone to misidentifying the low-frequency trend drift as transient random fluctuations, thereby causing frequent overshoot and power oscillations in the control loop, which manifest as thermal blind zones and local overheating in the physical space, ultimately leading to structural defects in the steel coil with excessive electromagnetic core losses.

[0042] Before acquiring the time-series characteristic data of the annealing furnace temperature field, the core control layer acquires the initial parameters of the furnace structure and the initial ambient temperature data under empty furnace conditions. This data is used to construct the basic geometric topology of the digital twin space. The initial response rate of the furnace wall temperature field is measured by inputting a calibration power sequence. This initial response rate is stored as a priori reference value to eliminate initial biases in the transient evolution residual data calculated in subsequent sampling periods. The core control layer synchronously reads the physical temperature parameters of multiple thermocouples through the sensor interface at a fixed sampling period of 100ms. Using a manifold mapping algorithm, the high-dimensional physical temperature data after normalization and denoising is mapped from the high-dimensional physical parameter space to a low-dimensional manifold space with a low-dimensional topology. In this low-dimensional manifold space, basis vectors characterizing the time-varying features of refractory material thermal degradation and workpiece surface emissivity are extracted. These basis vectors are then combined to construct the current sampling period. eigenvectors .

[0043] The macroscopic temperature field time-series characteristic data contains the superposition effect of multiple thermophysical evolutions at different time scales within the system. Among these, the thermal degradation caused by long-term high-temperature operation of the refractory material manifests as a slow, monotonically drifting of the overall thermal inertia of the system, while the time-varying emissivity of the workpiece surface caused by high-temperature oxidation manifests as periodic nonlinear fluctuations in radiative heat transfer efficiency. This invention, through intrinsic orthogonal decomposition in the manifold mapping algorithm, successfully separates a feature subspace with orthogonal characteristics in a low-dimensional topological structure. The evolution trajectory of its first principal component basis vector is positively correlated with the long-term cumulative operating time, thus mapping the thermal degradation trend of the refractory material. The second principal component basis vector is used to respond to thermal radiation feedback fluctuations to characterize the emissivity variation of the workpiece surface, thereby achieving decoupling and quantitative characterization of the thermodynamic response to the physical state drift. Based on the state transition weight matrix of the previous period... Compared with the feature vector of the previous sample Matrix multiplication is performed to calculate the predicted feature vector for the current period, and then a subtraction operator is used to calculate its difference from the currently sampled feature vector. The difference is calculated, and the feature vectors within adjacent sampling periods are input into a sliding temporal buffer queue to calculate the Euclidean distance, obtaining the original residual data. Then, the first-order temporal difference is calculated on the original residual data to generate transient evolution residual data. At this time, it is in standard tracking mode. The state machine for transitioning time-series residual data internally constructs a sliding time-series buffer queue with a fixed length of 5, which dynamically receives transient evolution residual data. .

[0044] In the specific implementation of this step, based on the temporal trend characteristics of the state machine state transition determination and the spatial multi-channel image characteristics of the subsequent matrix reconstruction, the transient evolution residual data is carried in the control program using a composite data structure containing multi-dimensional spatial deviation components and temporal scalar change rates. Specifically, the difference between the current period prediction feature vector calculated by the matrix subtraction operator and the current sampled feature vector constitutes the spatial deviation component, used to completely preserve the spatial location reference features of the distributed multi-point sensor array. Meanwhile, the feature vectors within adjacent sampling periods are input into the historical cache queue and subjected to Euclidean distance calculation and first-order temporal... Differential operations extract the residual change rate, reflecting the overall anomaly intensity of the temperature field, as the time-series scalar change rate. The input of this composite data structure comes entirely from the real-time features of the current sampling and the state of the buffer queue. Its processing rules call the time-series scalar change rate to perform state transition threshold comparison in the state determination stage, and call the spatial deviation component as the basis for the spatial quantitative allocation of each temperature measurement channel in the matrix reconstruction stage, so that the time-series state machine can perform state transitions and perform in-situ correction of the multi-dimensional spatial weight matrix. The time-series residual data conversion state machine dynamically calculates the time-series evolution gradient index of the transient evolution residual data within the sliding time-series buffer queue. It satisfies the formula ,in, For time-series evolution gradient index, This is the transient evolution residual data for the current period. This is the transient evolution residual data for the fourth cycle. The sampling time interval is fixed at 100ms; when the furnace experiences a transient thermal shock due to a step response in the physical heating circuit, resulting in a single increase in the transient evolution residual data without a sustained monotonic gradient, the time-series evolution gradient index... If the absolute value of the time-series residual data does not exceed the set convergence threshold of 0.05 for three consecutive sampling periods, the control time-series residual data conversion state machine remains in the transient disturbance state. And freeze the state transition weight matrix This isolates temperature measurement interference noise, and when the absolute value of the temporal evolution gradient of the transient evolution residual data falls below the set convergence threshold of 0.05, the control mechanism releases the state lock and returns to the standard tracking state of the temporal residual data transition state machine. When the refractory material of the furnace wall undergoes thermal inertia drift due to long-term continuous high-temperature operation, or when a characteristic oxide film forms on the surface of the steel coil, causing a trend change in emissivity, resulting in continuous accumulation of residuals in the same direction, and the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, the system is determined to have entered a mechanism drift state caused by long-period nonlinear evolution, and the temporal residual data conversion state machine is controlled to migrate to the mechanism drift state. And release the state lock, thereby triggering the weight in-situ correction operator.

[0045] In the mechanism drift state To prevent the state transition weight matrix from generalizing and diverging without physical meaning during long-period nonlinear evolution, the time-series residual data transformation state machine synchronously calls the internally maintained boundary conservation projection optimization operator. This operator belongs to the scalar algebraic product mapping module embedded in the main control chip of the control layer. It contains a historical power retrieval unit, a spatial divergence receiving register, and a boundary factor output unit. Its specific execution action is to retrieve the calculated cumulative power scalar and the spatial divergence of the eigenvector and input them into the product mapping update register. Through pure algebraic multiplication, it outputs a dimensionless scalar coefficient between 0.12 and 0.85. This scalar coefficient is used as the limiting boundary for the disordered divergence of the contraction matrix. Then, through the data interface, it collects the radiation heat transfer coefficient data inside the furnace body and the thermal inertia parameter data of the furnace wall refractory material. The radiation heat transfer coefficient data and the thermal inertia parameter data are weighted and integrated with the historical power data stream to obtain the cumulative power scalar of the physical control loop in the current sampling period. Simultaneously calculate the feature vector The degree of partial differential divergence in the low-dimensional manifold space is used to determine the eigenvectors. Spatial divergence, and cumulative power scalar Mapping with the spatial divergence product generates a dimensionless physical constraint boundary factor. The state transition weight matrix is ​​pre-stored in the register. In actual operation, the radiation heat transfer coefficient data inside the furnace is obtained through online optical measurement using a bicolor radiation pyrometer pre-placed on the inner wall of the furnace, combined with dynamic correction based on the pre-set calibration curvature of the steel surface. The thermal inertia parameter data of the furnace wall refractory material is obtained by real-time acquisition of multi-point temperature gradients using pairs of thermocouples embedded at different depths inside the refractory lining, and by multiplying the material heat capacity and thermal conductivity calculated using the known thickness and a one-dimensional unsteady thermal conduction inversion model. These two types of thermal physical parameters are synchronously uploaded to the static storage area of ​​the main control system in real time every 100 milliseconds via the underlying fieldbus data interface, thus providing accurate basic data for subsequent weighted integration. The weight in-situ correction operator constructs a structural correction increment matrix based on the gradient distribution of the transient evolution residual data. and utilize physical constraint boundary factors Structural correction increment matrix Smooth scaling is multiplied by boundary smoothing, and finally the result is obtained through matrix addition formula. The state transition weight matrix is ​​updated, and the absolute value of each matrix element is rigidly constrained to a discrete numerical range greater than or equal to 0.01 and less than or equal to 1.50. This numerical range corresponds to the energy conservation dynamic equilibrium boundary between the internal heat input power of the annealing furnace and the total heat storage of the refractory material. If the value of any element output by the matrix addition calculation exceeds the upper limit of 1.50 or the lower limit of 0.01 of this range, the main control chip of the control layer automatically activates the numerical truncation device, forcibly resetting the value of that element to the corresponding boundary critical value of 1.50 or 0.01 to reconstruct the digital twin model. This is the updated state transition weight matrix. This is the state transition weight matrix for the previous period. A preset dynamic relaxation factor, set to 0.15, is used to smooth out numerical jumps during the matrix update process. For physical constraint boundary factors, The structural correction increment matrix; based on the heat transfer principle that thermal resistance is proportional to spatial distance, the physical spatial distribution of the temperature measurement channel determines the heat transfer rate in adjacent areas, and the structural correction increment matrix... Based on spatial relationships and numerical conversion algorithms, the matrix is ​​specifically defined as a 5x5 two-dimensional array. Its row and column numbers correspond to five longitudinally evenly distributed thermocouple temperature measurement channels within the annealing furnace. During construction, a 5x5 zero matrix is ​​first initialized in a register. The 20 off-diagonal elements are used to calculate the physical spatial straight-line distance between the corresponding row and column channels using the three-dimensional Cartesian coordinates. The reciprocal of this physical spatial straight-line distance is taken as the dimensionless original conduction coefficient. Then, the original conduction coefficients of the remaining four off-diagonal positions within the same row are summed and normalized to generate spatial thermal conduction correlation coefficients. These coefficients are then multiplied by the corresponding diagonal principal element value, and the product is directly written to the corresponding off-diagonal register unit to precisely define the topological constraints. Therefore, when calculating diagonal elements, the main control unit... The chip extracts the temporal evolution gradient index of each temperature measurement channel from the historical evolution residual cache queue, calculates the sum of the absolute values ​​of this index for all channels, divides the absolute value of the index of the current channel by the sum to determine the preset primary and secondary distribution ratio, and then multiplies it with the transient evolution residual data of the current channel. The product is then filled into the corresponding diagonal position to achieve quantitative spatial allocation of residual gradient amplitude. When calculating off-diagonal elements, the three-dimensional Euclidean distance between any two channels is calculated based on the installation coordinates of the multi-point distributed thermocouples. The reciprocal of the three-dimensional Euclidean distance is taken as the original conduction coefficient, and the sum within the same row is calculated. The original conduction coefficients are divided by the sum to generate spatial thermal conduction correlation coefficients, which are then multiplied by the values ​​of the diagonal elements at the corresponding positions. The product is then filled into the off-diagonal position to represent the thermal conduction correlation of adjacent temperature measurement areas. A structural correction increment matrix with the same dimension as the state transition weight matrix of the previous cycle is constructed. To avoid matrix numerical divergence; the updated digital twin model releases the state lock and controls the state machine transition mechanism for time-series residual data to change states. Revert to standard tracking status The adaptive real-time mapping of the spatial temperature field distribution is restored, and the core control layer is based on the updated state transition weight matrix. The invention predicts the temperature field trend inside the furnace body in the next sampling period, calculates the feedforward power compensation data to offset the drift of the time-varying mechanism, and converts the feedforward power compensation data into power adjustment pulse width modulation commands through the network bus interface and sends them to the thyristor power tuning circuit to adjust the output power of the heating component. Specifically, when constructing the structural correction increment matrix, the invention extracts the time-series evolution gradient index of each channel in the historical evolution residual buffer queue, determines the sign direction of the gradient, and when the gradient sign is positive, it indicates that the simulation temperature is lower than the physical temperature of the entity, and the correction direction of the corresponding matrix element is based on the positive value. When the gradient sign is negative, it is based on the negative value. According to the magnitude of the absolute value of the gradient index, the contribution weight of each channel to the state transition weight is determined. The larger the magnitude, the larger the absolute value of the correction increment. Thus, the gradient magnitude of each channel is linearly mapped to the target matrix with the corresponding number of rows and columns according to the preset primary and secondary distribution ratio. The off-diagonal elements are filled with weights according to the spatial thermal conduction correlation coefficient of adjacent thermocouples, thereby generating a structural correction increment matrix with the same dimension as the state transition weight matrix of the previous period.

[0046] Through the rigid boundary constraint interlocking of the discrete-time control state switching and the boundary conservation projection optimization operator completed by the state machine of the time-series residual data transformation, the digital twin model can converge within 2 sampling periods and smooth the power oscillation of the control loop when facing external sudden power disturbances. The spatial temperature field deviation between the state simulation curve and the physical state of the entity is reduced to less than 1%, eliminating the local overheating phenomenon and thermal blind zone in the high-temperature annealing process to ensure uniform temperature distribution of the steel coil, improving the model tracking accuracy and avoiding the defect of excessive electromagnetic core loss caused by uneven annealing. The weight iterative evolution of the digital image model is rigidly anchored within the boundary range of the law of energy conservation, realizing the two-way interlocking of information layer state reconstruction and physical layer total energy throughput.

[0047] Example 2: When the system faces the low-frequency trend mechanism drift caused by the long-period nonlinear evolution inside the continuous annealing furnace, the heat transfer variation at the physical level leads to a decrease in temperature field control accuracy. To test the operating characteristics of the method claimed in this invention under industrial annealing conditions full of random interference, a continuous annealing furnace thermal test bench with a full radiation heating loop was established as a physical verification platform. The physical verification platform collects high-temperature temperature field characteristic data by arranging distributed multi-point thermocouples around the heating components. To evaluate the anti-interference performance of the digital twin model under complex electromagnetic conditions in actual industry, Gaussian white noise with a signal-to-noise ratio of 20dB was synchronously superimposed on the raw sampling data of the distributed multi-point thermocouples at the signal input terminal to simulate the impact of multi-source random thermal noise on temperature measurement accuracy in the production line. In the experimental design, the sampling time interval of the core system parameters is... The determination is achieved through topological response analysis, which affects the sampling time interval. The main technical factors for determining the values ​​include the transient response characteristic frequency of heat flow conduction within the furnace and the computational load of the central processing unit of the control host. Since capturing time-varying temperature characteristics requires millisecond-level responses to high-frequency disturbances, a low sampling frequency will fail to separate transient thermal shocks, leading to overshoot in the heating circuit. However, an excessively tight sampling time interval will cause frequent refreshes of the sliding time-series buffer queue, resulting in a monotonically increasing memory read / write addressing load. To balance the real-time performance of temperature field time-series characteristic data acquisition with the computational load of the control system, the system establishes a Nyquist sampling boundary relationship model in the control program. Specifically, when the maximum dynamic response spectral width of the temperature field caused by the radiative heat transfer coefficient inside the annealing furnace reaches the upper limit of the low-dimensional topology transformation bandwidth of the preset low-dimensional manifold space, the sampling time interval is controlled. Close to the lower limit of the set time domain window, 100ms.

[0048] After the physical verification platform is in the annealing stage and the initial parameters of the furnace structure are stored, the system starts the tracking program and controls the operation of multiple different control and experimental groups. The control and experimental groups include the first sample group, the second sample group, the third sample group, the fourth sample group, and the sample group of the present invention. The first sample group maintains the overall characteristics of the present invention and adjusts the set convergence threshold parameter to 0.02, thus serving as an out-of-range control group below the lower limit of the defined range. The second sample group maintains the overall characteristics of the present invention and adjusts the set convergence threshold parameter to 0.12, thus serving as an out-of-range control group above the upper limit of the defined range. The third sample group selectively removes the temporal residual data transformation state machine and places it in its mechanism drift state. The weight adjustment is stopped at the next step, and this is used as a control group for partial feature loss. The fourth sample group uses a conventional industrial twin architecture with a fixed thermodynamic analytical equation model as a benchmark for comparison. The sampling time interval of the sample group in this invention is as follows. The convergence threshold within the time-series residual data transformation state machine is set to 0.05, with a fixed 100ms interval, and the dynamic relaxation factor is also set. The value was set to 0.15. Meanwhile, in order to verify the gradient response law of the technical effect, the sample group of this invention was configured to run in three time-varying mechanism drift gradient environments: low-intensity thermodynamic variation conditions, medium-intensity thermodynamic variation conditions, and high-intensity thermodynamic variation conditions, according to the problem intensity gradient control system.

[0049] During the 240th sampling cycle of continuous operation of the furnace, distributed multi-point thermocouples collected raw low-frequency physical temperature data including noise, and in standard tracking mode. The following steps involve converting the data into low-dimensional feature vectors using a manifold mapping algorithm to construct the current sampling period. eigenvectors When faced with sudden thermal shock temperature measurement interference, the transient evolution residual data of the sample group in the 242nd sampling period of this invention... The value spiked to 0.082, then dropped back to 0.012 in cycle 243 due to the dissipation of the thermal shock. The first-order difference gradient of the sliding temporal buffer queue dynamically calculated by the temporal residual data transformation state machine did not show persistence, and the calculated temporal evolution gradient index... The absolute value is 0.021, due to the time-series evolution gradient index. If the absolute value does not exceed the set convergence threshold of 0.05 for three consecutive sampling periods, the state machine transitions to the transient disturbance state. And freeze the state transition weight matrix This isolates thermal noise errors at the register level. When the furnace wall refractory material undergoes thermal inertia variation due to long-term continuous high-temperature operation, the transient evolution residual data of the continuous intermediate feature variables extracted by the system shows a monotonically accumulating trend. Under the high-intensity thermodynamic variation conditions of the sample group of this invention, the transient evolution residual data of the 500th cycle... The transient evolution residual data for the 501st period is 0.068. The transient evolution residual data for the 502nd period is 0.074. The value is 0.081, representing the temporal evolution gradient index calculated over three consecutive sampling periods based on five consecutive periods of transient evolution residual data within a sliding temporal buffer queue. The absolute values ​​are 0.062, 0.065, and 0.071, respectively, representing the time-series evolution gradient index. The calculation satisfies the formula ,in, For time-series evolution gradient index, This is the transient evolution residual data for the current period. This is the transient evolution residual data for the fourth cycle. For the sampling time interval, the temporal evolution gradient index for the above three consecutive periods all exceeded the set convergence threshold of 0.05, triggering the transition of the temporal residual data state machine to the mechanism drift state. The in-situ weight correction operator is activated, and the boundary conservation projection optimization operator calculates the current cumulative power scalar based on the integration of the historical power data stream read in the current period. The power is 45.3 kW. The manifold physical constraint boundary factor is obtained by combining low-dimensional space divergence calculation. The value is 0.84, thus the structural incremental matrix is ​​modified using the manifold physical constraint boundary factor. Smooth scaling and updating the state transition weight matrix, at the end of the heating cycle, the final state indicators of each sample group when completing the data change analysis are compared. The spatial temperature field tracking deviation of the sample group under low-intensity thermodynamic variation conditions, medium-intensity thermodynamic variation conditions and high-intensity thermodynamic variation conditions are maintained at 0.42%, 0.65% and 0.88% respectively, all less than the preset 1% limit boundary. Moreover, when the operating condition changes, its digital twin state space recovers and converges within 2 sampling cycles. The power adjustment pulse width modulation command output by the thyristor power tuning circuit of the subsequent stage does not produce divergent oscillation.

[0050] In contrast, the third sample group, which removed the weight update mechanism, could not resist the drift caused by long-term thermal inertia. Its spatial temperature field tracking deviation monotonically increased to 4.85% in the later stages of annealing, and the digital twin model exhibited significant drift divergence during long-term evolution. The first sample group, with its low convergence threshold of 0.02, was susceptible to interference from random power frequency electromagnetic noise, resulting in frequent control state switching errors and miscorrection of the state transition weight matrix. This increased the maximum tracking residual between the simulated temperature field and the actual furnace to 3.12%, and caused frequent overshooting of the heating output in the subsequent thyristor power tuning circuit. The second sample group, with its high convergence threshold of 0.12, exhibited significant system response hysteresis and was unable to activate the drift mechanism state within the predetermined period when facing medium-to-high intensity thermal inertia drift. The temperature field prediction residual accumulated to 2.96% in the 600th cycle, increasing beyond the allowable value for temperature control accuracy. The numerical change law confirms that the 0.05 convergence threshold set by this invention constitutes an optimization working window that synergistically takes into account noise resistance, fault tolerance and drift correction. When the system model leaves this window, it will produce negative effects of misadjustment or omission at both ends. There is a causal interlock relationship between the logic control state switching and the physical boundary optimization operator in each step of this method. By applying manifold projection constraints to the evolution of high-dimensional information flow through the cumulative power scalar, the reconstructed digital twin space accurately reproduces the low-frequency mechanism drift in the continuous annealing process, eliminating the defect of model divergence caused by the lack of physical boundary constraints in conventional mapping algorithms, providing high-fidelity prediction for heating power tuning, and the final electromagnetic core loss test value of the workpiece is stably within the technical specifications.

[0051] Example 3: This example combines Figures 1 to 2 A digital twin method for high-temperature annealing of silicon steel is described, as follows: Figure 1As shown, step S1 acquires the time-series characteristic data of the annealing furnace temperature field and projects it onto a low-dimensional manifold space to construct a feature vector. A time-series residual data transition state machine in standard tracking state is set up inside the control layer, and the state transition weight matrix is ​​pre-stored in the register. Step S2 calculates the predicted feature vector for the current period by multiplying the previous period's state transition weight matrix with the previously sampled feature vector. The difference between the predicted feature vector and the currently sampled feature vector is calculated by subtraction to obtain the transient evolution residual data. Step S3, when the absolute value of the time-series evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, the control time... Step S4: The residual data transformation state machine transitions from the standard tracking state or transient disturbance state to the mechanism drift state. In the mechanism drift state, the boundary conservation projection optimization operator is called, and the cumulative power scalar is extracted based on the historical power data stream. The cumulative power scalar is multiplied and mapped with the spatial divergence of the eigenvector to generate a dimensionless physical constraint boundary factor. Step S5: The structure correction increment matrix is ​​constructed based on the gradient distribution of the transient evolution residual data. The structure correction increment matrix is ​​smoothly scaled using the physical constraint boundary factor. The state transition weight matrix is ​​updated by matrix addition to reconstruct the digital twin model. The feedforward power compensation data is output to the thyristor power tuning circuit.

[0052] like Figure 2 As shown, the coordinate system consists of the vertical axis residuals and gradient values, and the horizontal axis sampling period. The data processing system operates and acquires the numerical distribution and flow patterns of transient evolution residual data, the absolute value of the temporal evolution gradient index, and the set convergence threshold within sampling periods of 240, 241, 242, 243, and 244. In the 242nd sampling period, the transient evolution residual data shows a single increase and the value exceeds the set convergence threshold. In the 243rd sampling period, the value falls back below the set convergence threshold, and the absolute value of the temporal evolution gradient index does not exceed the set convergence threshold for three consecutive sampling periods within the sampling period range. The system thus remains in a transient disturbance state and freezes the state transfer weight matrix to isolate temperature measurement interference noise.

[0053] Example 4: When the system is in the continuous annealing furnace operation cycle, the temperature field inside the furnace body will experience low-frequency trend drift due to the change in thermal inertia of the furnace wall refractory material and the time-varying emissivity of the workpiece surface. At the same time, there is random noise interference in the high-dimensional physical temperature data collected by multiple thermocouples, which causes the digital image model of the spatial temperature field calculated by the control host to diverge. This leads to control overshoot and power oscillation in the thyristor power tuning circuit, ultimately causing the output power of the heating component to deviate from the target process curve and causing the steel coil to have excessive electromagnetic core loss.

[0054] To determine the spatial temperature field distribution and separate random noise, the system utilizes the main control processing chip and static random access memory to synchronously read the physical temperature data vectors of multiple thermocouples every 100ms sampling period, and then continuously... The physical temperature data vector for each period is input into the first buffer in the static random access memory for mean calculation to output a mean matrix vector. A subtraction operator is used to subtract the mean matrix vector from the currently sampled physical temperature data vector to output a centered temperature matrix. The covariance matrix is ​​calculated from the centered temperature matrix and input into an algebraic eigenvalue solver for characteristic polynomial solving, thus outputting the corresponding eigenvalue sequence and a set of mutually orthogonal unit eigenvectors. Then, the eigenvalue sequence is sorted according to its monotonically decreasing property from largest to smallest, and the top... The unit eigenvectors corresponding to the largest eigenvalues ​​are combined to construct a topological projection matrix, thereby orthogonally projecting the centered temperature matrix into the low-dimensional manifold space to generate the current sampling period characterizing the temperature field evolution trend. eigenvectors Its calculation process satisfies the formula ,in, The feature vector of the current sampling period, For the reason before Constructed by combining unit feature vectors with dimensions satisfying The topological projection matrix of the order relation, For the current sample containing A vector of physical temperature data for each channel. The contents calculated in the first buffer The mean matrix vector of each channel, with superscript This represents the matrix transpose operation.

[0055] The main control processing chip will calculate the feature vector The input is placed into the first-in-first-out history evolution residual buffer queue allocated in the static random access memory, according to the formula... Real-time calculation of transient evolution residual data for the current period, where, This is the transient evolution residual data vector for the current period. The feature vector of the current sampling period, The state transition weight matrix of the previous cycle is pre-stored in the register. The feature vector of the previous cycle; the length of the first-in-first-out historical evolution residual buffer queue is fixed at 5 and dynamically receives transient evolution residual data, set within the control program to be in standard tracking state. The temporal residual data transformation state machine dynamically calculates the first-in-first-out historical evolution residual cache queue temporal evolution gradient index. Its calculation satisfies the formula ,in, For time-series evolution gradient index, This is the transient evolution residual data vector for the current period. This is the transient evolution residual data vector for the fourth cycle. Set a fixed sampling time interval of 100ms; set a convergence threshold. Satisfy the calibration relation ,in, To set a convergence threshold, The first self-inspection under the initial cold state condition of the furnace. Working state residual data vector for each sampling period This represents the expected value of the residual data vector of the working state within the first 50 sampling periods; when a sudden local thermal shock occurs, causing the time-series evolution gradient index to... The absolute value did not exceed the set convergence threshold for three consecutive sampling periods. At that time, the time-series residual data transformation state machine transitions from the standard tracking state. Migration to transient disturbance state And control the state lock to freeze the state transition weight matrix. To isolate noise interference until the time-series evolution gradient index The absolute value falls back to the set convergence threshold. The control state machine returns to the standard tracking state at the following time. When refractory materials exhibit thermal inertia and time-varying behavior, residuals accumulate monotonically in the same direction, and the time-series evolution gradient index... The absolute value exceeds the set convergence threshold for three consecutive sampling periods. When a low-frequency trend mechanism drift occurs in the system, the control time-series residual data state machine is transitioned to the mechanism drift state. And release the state lock to activate the in-situ weight correction operator, in the mechanism drift state. Next, the time-series residual data transformation state machine calls the internal boundary conservation projection optimization operator, collects the radiation heat transfer coefficient data inside the furnace body and the thermal inertia parameter data of the furnace wall refractory material through the data interface, and performs a weighted integration operation on the radiation heat transfer coefficient data and thermal inertia parameter data in the historical power data stream to output the cumulative power scalar of the physical control loop in the current sampling period. Simultaneously calculate the feature vector of the current sample. The partial differential gradient in the low-dimensional manifold space is used to output the spatial divergence, and the cumulative power scalar is used to... Multiplying with spatial divergence to generate a dimensionless physical constraint boundary factor. The adaptive dynamic relaxation factor is calculated using the real-time residual control relation. It satisfies the formula ,in, It is an adaptive dynamic relaxation factor. The baseline relaxation scalar is fixed at 0.15. The smooth attenuation adjustment coefficient is fixed at 0.25. The Euclidean norm of the transient evolution residual data vector for the current period; the in-situ weight correction operator constructs a structural correction increment matrix based on the direction and magnitude of the temporal evolution gradient index. And update the formula through the matrix. Calculate and update the state transition weight matrix in the register to correct the digital twin model, where, This is the updated state transition weight matrix. This is the state transition weight matrix for the previous period. For physical constraint boundary factors, The incremental matrix is ​​corrected for structural modification; the reconstructed digital twin model releases the state lock and controls the state machine's transition mechanism to control the drift state of the time-series residual data. Revert to standard tracking status A new real-time mapping of the spatial temperature field distribution is established, and the main control processing chip uses the updated state transition weight matrix. The system predicts the temperature field change trend inside the furnace during the next sampling period, calculates feedforward power compensation data to offset the drift of the time-varying mechanism, and converts the feedforward power compensation data into power regulation pulse width modulation commands via a network bus interface. These commands are then sent to the thyristor power tuning circuit to adjust the output power of the heating components, ensuring that the spatial temperature field deviation between the state simulation curve and the physical state remains within 1%. This eliminates overshoot oscillations in the heating circuit caused by the divergence of the digital imaging model, guaranteeing uniform temperature distribution of the steel coil during the continuous annealing long-cycle process. Specifically, the aforementioned weighted integral operation completes the calculation of power and transmission within the time domain window. After the thermal characteristic data is accumulated and summed, in order to ensure that the final output cumulative power scalar is absolutely consistent with the control dimension of the subsequent thyristor circuit, this invention, after integration, forcibly divides the total energy scalar obtained by integration by the total time span corresponding to the historical power data stream, thus completing the time averaging normalization process. Through this cross-dimensional conversion interface, the time-domain cumulative value of the energy scale is converted back into a power dimension parameter that can directly characterize the current equivalent heat input level of the furnace. Therefore, the final output cumulative power scalar is still expressed in power units in terms of physical dimensions, so as to eliminate the dimensional conflict between time-domain integration and transient power control.

[0056] Example 5: When the system faces the deployment of a newly built furnace, due to the unknown impedance variation of thermocouple circuits and the heat transfer constant of insulation materials, the digital imaging model is prone to initial matrix value divergence. Before starting the heating cycle, the core control layer initiates the initial setting program. The main control processing chip continuously sends a calibration pulse sequence with a duration of 60 seconds and a fixed power of 10kW to the thyristor power tuning circuit, simultaneously triggering multiple thermocouples to collect the initial feedback temperature scalar of the temperature field. The main control processing chip transmits the initial feedback temperature scalar to the static random access memory to calculate the channel temperature rise rate. When the temperature rise rate is greater than 5.0℃ / s, the power limiter is triggered to halve the calibration power to control thermal shock. Finally, the matrix subtraction operator is used to subtract the ambient temperature from the equilibrium temperature of each channel, outputting the initial response feature vector. ,in Identify the initial response feature vector, subscript Indicates the initial state.

[0057] The main control processing chip controls the matrix multiplication operator to multiply the initial response feature vector. Project the matrix onto a low-dimensional manifold, calculate its magnitude, and then transfer the initial state to the weight matrix. The diagonal elements are assigned the reciprocal of the modulus, thereby achieving a low-dimensional curvature reconstruction mapping of the static topological baseline and boundary conditions of the digital twin space; the system releases the initial state lock and switches to the standard tracking state. Enable real-time temperature field mapping with a sampling period of 100ms, utilizing the initial state transition weight matrix. Eliminating the start-up prediction residual caused by environmental wiring impedance, controlling the residual between the simulation state space and the actual temperature field of the annealing furnace during the initial heating stage to be kept within 1%, eliminating the overshoot of the thyristor power tuning circuit heating power caused by the inaccuracy of the initial state of the system, and ensuring that the scalar output of the annealing furnace in the entire control layer smoothly tracks the target process curve.

[0058] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A digital twin method for high-temperature annealing of silicon steel, characterized in that, Includes the following steps: Step S1: Obtain the time series characteristic data of the annealing furnace temperature field and project it onto the low-dimensional manifold space to construct the feature vector. Set up the time series residual data transformation state machine in the standard tracking state inside the control layer and pre-store the state transition weight matrix in the register. Step S2: Multiply the state transition weight matrix of the previous period with the feature vector of the previous sample to calculate the predicted feature vector of the current period. Calculate the difference between the predicted feature vector and the feature vector of the current sample by subtraction to obtain the transient evolution residual data. Step S3: When the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, control the temporal residual data transition state machine to migrate from the standard tracking state or the transient disturbance state to the mechanism drift state. Step S4: In the mechanism drift state, the boundary conservation projection optimization operator is invoked to extract the cumulative power scalar based on the historical power data stream. The cumulative power scalar is then multiplied and mapped with the spatial divergence of the feature vector to generate a dimensionless physical constraint boundary factor. Step S5: Construct a structural correction increment matrix based on the gradient distribution of transient evolution residual data, use physical constraint boundary factors to smoothly scale the structural correction increment matrix, update the state transition weight matrix through matrix addition to reconstruct the digital twin model, and output feedforward power compensation data to the thyristor power tuning circuit.

2. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, When thermal shock causes a single increase in transient evolution residual data without a sustained monotonic gradient, the time-series residual data transition state machine is kept in transient disturbance state and the state transition weight matrix is ​​frozen to isolate temperature measurement interference noise because the set convergence threshold of 0.05 is not exceeded for three consecutive sampling periods. When the absolute value of the transient evolution residual data's temporal evolution gradient falls below the set convergence threshold of 0.05, the time-series residual data transition state machine releases the state lock and returns to the standard tracking state.

3. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, Step S4 includes the following sub-steps: Step S41, collect the radiation heat transfer coefficient data inside the furnace body and the thermal inertia parameter data of the furnace wall refractory material through the data interface, and perform weighted integration calculation on the radiation heat transfer coefficient data and thermal inertia parameter data with the historical power data stream to obtain the cumulative power scalar; Step S42, calculate the degree of partial differential divergence of the eigenvector in the low-dimensional manifold space to determine the spatial divergence; Step S43, multiply the cumulative power scalar with the spatial divergence to determine the physical constraint boundary factor, and use the physical constraint boundary factor to constrain the value of the state transition weight matrix within the energy conservation boundary interval.

4. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, The steps following step S5 are as follows: Step S6, predict the temperature field trend inside the furnace body in the next sampling period based on the updated state transition weight matrix, and calculate the feedforward power compensation data to offset the drift of the time-varying mechanism; Step S7, convert the feedforward power compensation data into a power adjustment pulse width modulation command through the network bus interface and send it to the thyristor power tuning circuit to adjust the output power of the heating component and control the spatial temperature field deviation between the state simulation curve and the physical state to be within 1%.

5. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11, performing standardized filtering and denoising processing on the acquired annealing furnace temperature field time series feature data, which includes the furnace space temperature sequence and heating power feedback sequence collected from multiple points; Step S12, using a manifold mapping algorithm to map the denoised annealing furnace temperature field time series feature data from a high-dimensional physical parameter space to a low-dimensional manifold space with a low-dimensional topological structure; Step S13, extracting basis vectors that can characterize the thermal degradation of refractory materials and the time-varying characteristics of workpiece surface emissivity in the low-dimensional manifold space, and combining them to obtain feature vectors.

6. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21, construct a sliding time-series buffer queue inside the time-series residual data transformation state machine, with the length of the time-series buffer queue fixed at 5; Step S22, input the feature vectors within adjacent sampling periods into the time-series buffer queue for Euclidean distance calculation to obtain the original residual data; Step S23, perform first-order time-series difference operation on the original residual data to extract the residual change rate to generate transient evolution residual data.

7. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, Step S3 includes the following sub-steps: Step S31, the temporal residual data conversion state machine dynamically calculates the temporal evolution gradient index of the transient evolution residual data in the sliding temporal buffer queue, with the sampling time interval fixed at 100ms; Step S32, when the absolute value of the temporal evolution gradient of the transient evolution residual data exceeds the set convergence threshold of 0.05 for three consecutive sampling periods, it is determined that the system has entered the mechanism drift state caused by long-period nonlinear evolution, and the temporal residual data conversion state machine is controlled to migrate to the mechanism drift state and release the state lock to trigger the weight in-situ correction operator.

8. The digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, After updating the state transition weight matrix, the following steps are also included: Step S8, control the time-series residual data transformation state machine to fall back from the mechanism drift state to the standard tracking state, and restore the adaptive real-time mapping of the spatial temperature field distribution; Step S9, when the digital twin model encounters an external sudden power disturbance, control the digital twin model to complete convergence within 2 sampling periods.

9. A digital twin method for high-temperature annealing of silicon steel according to claim 1, characterized in that, Before step S1, the following initial benchmark calibration steps are also included: Step S01, obtain the initial parameters of the furnace body structure and the initial ambient temperature data in the empty furnace state, and construct the basic geometric topology of the digital twin space; Step S02 involves measuring the initial response rate of the furnace wall temperature field by injecting a calibration power sequence, and storing the initial response rate as a priori reference value for initial deviation elimination of the transient evolution residual data calculated by the time-series residual data conversion state machine in subsequent sampling periods.

Images