A magnetic ring heat treatment parameter prediction optimization method and device and storage medium

Abstract

The application discloses a magnetic ring heat treatment parameter prediction optimization method and device and a storage medium, and relates to the field of magnetic ring heat treatment optimization. First, the aerodynamic characteristic fingerprint representing the turbulent state in the furnace is extracted from the nitrogen pressure pulsation signal and the fan current perturbation signal; then, the aerodynamic characteristic is inverted into the spatial distribution of the dynamic convection heat transfer coefficient through a physical embedded deep neural network, so that the quantitative mapping between the flow field disturbance and the heat exchange is established; on this basis, the transient temperature field of the magnetic ring stack is tomographically estimated and the thermal shock risk is synchronously evaluated in combination with the unsteady porous medium heat conduction equation; finally, the temperature field tracking deviation is taken as the target, the thermal shock risk is taken as the punishment constraint, and the fan rotating speed, the nitrogen filling valve opening degree and the heating power are collaboratively and rolling optimized and fed forward through a nonlinear model prediction controller. In this way, the vicious circle caused by the flow-heat coupling instability is eliminated.

Description

Technical Field

[0001] This application relates to the field of magnetic ring heat treatment optimization, and more specifically, to a method, apparatus and storage medium for predicting and optimizing magnetic ring heat treatment parameters. Background Technology

[0002] Magnetic rings, as core magnetic components in power electronics, communication filtering, and electromagnetic compatibility, rely heavily on precise temperature control during heat treatment for their final performance. In a typical magnetic ring heat treatment process, the rings are densely stacked in multiple layers within a protective atmosphere furnace, undergoing heating, holding, and controlled cooling processes within a nitrogen-filled, slightly positive pressure environment, ranging from 380°C to 560°C, to achieve grain recrystallization and directional optimization of the magnetic domain structure. However, engineering practice shows that even when the macroscopic temperature parameters of the furnace meet the standards, key performance indicators such as brittleness and permeability within the same batch of magnetic rings still exhibit significant dispersion, making it difficult to guarantee product quality consistency. The fundamental reason lies in the fact that the densely stacked magnetic rings constitute a complex porous medium space with high wind resistance. When the circulating fan operates under the critical state of 0.3 kPa nitrogen micro-positive pressure, the dynamic viscosity and density of the gas change drastically with temperature. This generates a Bernoulli effect locally, causing a sudden drop in regional pressure, forming eddy stagnant zones or high-speed throughflow. This causes the convective heat transfer coefficient attached to the surface of the magnetic rings to fluctuate drastically in both spatial and temporal dimensions, leading to localized instantaneous overheating or undercooling, which in turn exacerbates thermal stress concentration and causes the magnetic rings to become embrittled. Therefore, there is an urgent engineering need to develop a method that can deeply sense the fluid-thermal coupling state and achieve accurate temperature field prediction and active optimization.

[0003] In existing technologies, temperature control in magnetic ring heat treatment generally employs a traditional PID control scheme. This involves collecting temperature signals from multiple temperature sensors within the furnace and then adjusting the electric heating power using a single-variable feedback method after proportional-integral-derivative (PI-DI) calculations. This scheme completely separates temperature control from the flow field state within the furnace. It fails to perceive the coupling interference between circulating fan disturbances and the nitrogen micro-pressure field, and it lacks the ability to predict the three-dimensional temperature distribution and thermal stress state within the magnetic ring stack. When transient turbulent purging causes abrupt changes in sensor readings, the PID controller misinterprets local disturbances as overall temperature deviations, incorrectly issuing full-load heating or rapid cooling commands. This further exacerbates local thermal oscillations, creating a vicious cycle where wind disturbances induce temperature illusions, trigger erroneous compensation, and ultimately lead to temperature field fragmentation. Furthermore, due to the lack of modeling and prediction mechanisms for the hidden interrelationships between micro-pressure gas dynamics, dense porous media heat conduction, and nonlinear fluctuations in convective heat transfer, existing schemes cannot achieve multi-variable synergistic optimization between fan speed, nitrogen valve opening, and heating power. Consequently, they struggle to fundamentally eliminate the impact of flow-thermal coupling instability on the quality of magnetic ring heat treatment. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this application provides a method for predicting and optimizing the heat treatment parameters of a magnetic ring, comprising: Step 1, performing frequency domain decomposition and time-series filtering feature extraction on the acquired multi-source sensor data stream to obtain an aerodynamic feature fingerprint matrix and a calibrated temperature sequence vector; Step 2, based on a physically embedded deep neural network, performing fluid-solid interface heat transfer coefficient inversion and reconstruction on the aerodynamic feature fingerprint matrix to obtain a dynamic convection heat transfer coefficient mapping matrix; Step 3, performing three-dimensional thermal shock risk and temperature tomography prediction on the calibrated temperature sequence vector and the dynamic convection heat transfer coefficient mapping matrix under unsteady flow field disturbance to obtain a transient temperature field tensor and a local thermal shock risk assessment matrix for the magnetic ring; Step 4, using the transient temperature field tensor of the magnetic ring as the tracking target and configuring the local thermal shock risk assessment matrix as a penalty constraint term, performing multivariate collaborative rolling optimization on it through a nonlinear model predictive controller to obtain a collaborative execution instruction set; Step 5, performing feedforward collaborative execution of the variable frequency speed of the circulating fan, the opening degree of the nitrogen filling valve, and the electric heating power based on the collaborative execution instruction set.

[0005] This application also provides a magnetic ring heat treatment parameter prediction and optimization device, comprising: a sensor data decomposition and filtering module, used to perform frequency domain decomposition and time-series filtering feature extraction on the acquired multi-source sensor data stream to obtain an aerodynamic feature fingerprint matrix and a calibrated temperature sequence vector; a heat transfer coefficient inversion module, used to perform fluid-solid interface heat transfer coefficient inversion and reconstruction on the aerodynamic feature fingerprint matrix based on a physically embedded deep neural network to obtain a dynamic convection heat transfer coefficient mapping matrix; and a magnetic ring temperature prediction module, used to analyze the calibrated temperature sequence vector and the dynamic convection heat transfer coefficient mapping matrix. Three-dimensional thermal shock risk and temperature tomography prediction of the magnetic ring stack under unsteady flow field disturbance are performed to obtain the transient temperature field tensor and local thermal shock risk assessment matrix of the magnetic ring. The collaborative rolling optimization module is used to take the transient temperature field tensor of the magnetic ring as the tracking target and configure the local thermal shock risk assessment matrix as the penalty constraint term. The multivariate collaborative rolling optimization is performed on it by the nonlinear model predictive controller to obtain the collaborative execution instruction set. The collaborative execution module is used to feedforward collaboratively issue and execute the variable frequency speed of the circulating fan, the opening degree of the nitrogen filling valve and the electric heating power based on the collaborative execution instruction set.

[0006] According to another aspect of this application, a storage medium is also provided that stores computer instructions for causing a computer to perform the methods described above.

[0007] Compared with the prior art, this application provides a method, device and storage medium for predicting and optimizing parameters of magnetic ring heat treatment, which addresses the technical problem that the coupling interference between the circulating fan disturbance and the nitrogen micro-pressure field during the magnetic ring heat treatment process causes drastic fluctuations in the local convective heat transfer coefficient, thereby leading to temperature field runaway. This solution transforms the traditional single-variable temperature feedback control paradigm into a closed-loop predictive optimization paradigm of "flow field perception - heat transfer reconstruction - temperature field prediction - multi-variable collaborative optimization". First, aerodynamic fingerprints characterizing the turbulent state within the furnace are extracted from nitrogen pressure pulsation signals and fan current perturbation signals, overcoming the limitation of existing schemes that cannot sense flow field disturbances. Then, a physically embedded deep neural network is used to invert the aerodynamic features into spatially distributed dynamic convective heat transfer coefficients, establishing a quantitative mapping between flow field disturbances and heat exchange. Based on this, the transient temperature field of the magnetic ring stack is tomographically predicted using the unsteady porous medium heat conduction equation, and thermal shock risk is simultaneously assessed. Finally, with temperature field tracking deviation as the target and thermal shock risk as the penalty constraint, a nonlinear model predictive controller performs coordinated rolling optimization and feedforward distribution of fan speed, nitrogen filling valve opening, and heating power. This breaks down the information barrier between flow field perception and temperature control, enabling the control system to proactively compensate for turbulent disturbances before they deteriorate into temperature deviations, eliminating the vicious cycle caused by flow-thermal coupling instability. Attached Figure Description

[0008] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0009] Figure 1 This is a flowchart of a method for predicting and optimizing heat treatment parameters of a magnetic ring according to an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the data flow of the magnetic ring heat treatment parameter prediction and optimization method according to an embodiment of this application.

[0011] Figure 3 This is a flowchart of step four in the magnetic ring heat treatment parameter prediction and optimization method according to an embodiment of this application.

[0012] Figure 4 This is a block diagram of a magnetic ring heat treatment parameter prediction and optimization device according to an embodiment of this application. Detailed Implementation

[0013] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0014] To address the shortcomings in the aforementioned technical field, this application proposes a method for predicting and optimizing the heat treatment parameters of magnetic rings. For example... Figure 1 and Figure 2As shown, Figure 1 This is a flowchart of a method for predicting and optimizing heat treatment parameters of a magnetic ring according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow of the magnetic ring heat treatment parameter prediction and optimization method according to an embodiment of this application.

[0015] Specifically, in step one, the acquired multi-source sensor data stream is subjected to frequency domain decomposition and time-series filtering feature extraction to obtain the aerodynamic feature fingerprint matrix and the calibrated temperature sequence vector. In this application, the multi-source sensor data stream includes a nitrogen micro-positive pressure high-frequency pulsating signal, a circulating fan motor current perturbation signal, and a multi-zone furnace temperature measurement sequence. It should be understood that during the magnetic ring heat treatment process, the furnace atmosphere is in a critical state of 0.3 kPa nitrogen micro-positive pressure. When the high-speed rotation of the circulating fan drives the airflow through the complex porous medium space composed of multiple densely stacked magnetic rings, the dynamic viscosity and density of the gas change drastically with the heating process from 380°C to 560°C. Due to the Bernoulli effect, the pressure in local areas drops instantaneously, forming irregular vortex dead water zones and high-speed penetrating flows, causing the convective heat transfer coefficient attached to the surface of the magnetic ring to exhibit violent nonlinear fluctuations in the spatiotemporal dimension. This kind of hidden flow-thermal coupling disturbance cannot be directly captured by conventional temperature sensors. Traditional PID controllers misinterpret the sudden changes in sensor readings caused by flow field disturbances as overall furnace temperature deviations, and thus issue incorrect compensation commands, exacerbating temperature field distortion. In order to perceive this flow field disturbance state from the source and provide reliable input for subsequent heat transfer coefficient inversion and temperature field prediction, this application first performs step one.

[0016] In one operable embodiment of this application, step one includes: performing frequency domain energy spectrum decomposition on the nitrogen micro-positive pressure high-frequency pulsation signal to obtain the eddy current energy feature vector; performing high-frequency detail coefficient stripping on the circulating fan motor current perturbation signal through multi-scale discrete wavelet transform to obtain the high-frequency jitter feature sequence; performing multi-modal fusion compression on the eddy current energy feature vector, the high-frequency jitter feature sequence, and the hardware execution log to obtain the aerodynamic feature fingerprint matrix; and performing adaptive sliding filtering on the multi-zone furnace temperature measurement sequence based on the aerodynamic feature fingerprint matrix to obtain the calibrated temperature sequence vector.

[0017] In the physical structure of the magnetic ring heat treatment furnace, a high-sensitivity micro-pressure sensor is installed at the end of the nitrogen supply pipeline. This sensor, with a sampling frequency set to 2000 Hz, is used to collect high-frequency pulsating signals of the micro-positive pressure of nitrogen gas inside the furnace in real time. This signal is essentially a micro-amplitude high-frequency fluctuation superimposed on a 0.3 kPa reference pressure within the furnace, with an amplitude within ±15 Pa and a frequency range covering 0.5 Hz to 800 Hz. These high-frequency pulsating components directly reflect the periodic suction effect caused by the rotation of the circulating fan and the eddy current shedding phenomenon generated when the airflow passes through the porous medium of the magnetic ring stack. The micro-perturbation signal of the circulating fan motor current is collected by a high-precision current transformer installed at the output of the frequency converter, with a sampling frequency set to 1000 Hz. This signal has a small jitter component superimposed on the rated current baseline of the fan, with an amplitude of approximately 0.5% to 3% of the full-load current. These micro-perturbation components originate from the dynamic aerodynamic load changes experienced by the fan impeller in a non-uniform flow field and can indirectly characterize the instantaneous state of the turbulence intensity and spatial distribution within the furnace. The multi-zone furnace temperature measurement sequence is provided by armored thermocouple arrays distributed in the upper, middle, and lower zones of the furnace. Three thermocouples are arranged in each zone, totaling nine temperature measurement points. The sampling period is 500 milliseconds, the measurement range covers ambient temperature to 650℃, and the resolution is 0.1℃. The hardware execution log is recorded in real time by the programmable logic controller, including the current frequency conversion command value of the circulating fan, the opening percentage of the nitrogen charging solenoid valve, and the conduction angle of the electric heating thyristors in each zone. The recording period is synchronized with the temperature measurement sequence at 500 milliseconds.

[0018] The specific process of frequency domain energy spectrum decomposition for nitrogen micro-positive pressure high-frequency pulsating signals is as follows: In this application, 1024 sampling points traced back from the current time are taken as an analysis time window. A 1024-point discrete Fourier transform is performed on the nitrogen pulsating signal sequence within this time window to convert the time-domain pulsating signal into a frequency-domain complex spectrum. The square of the amplitude of each frequency component of the frequency-domain complex spectrum is taken to obtain the power spectral density estimate. The frequency range from 0 to 1000 Hz is divided into 8 sub-bands, each with a width of 125 Hz. The power spectral density values ​​in each sub-band are integrated and accumulated to obtain the energy concentration value of each sub-band. The energy values ​​of these 8 sub-bands are arranged in ascending order of frequency to form an 8-dimensional column vector, which is the eddy current energy characteristic vector. Taking specific values ​​as an example, when the turbulence inside the furnace is at a moderate intensity, the energy value of the first sub-band (0 to 125 Hz) is 0.42, the energy value of the second sub-band (125 to 250 Hz) is 1.87, the energy value of the third sub-band (250 to 375 Hz) is 3.15, the energy value of the fourth sub-band (375 to 500 Hz) is 2.63, and the energy values ​​of the fifth to eighth sub-bands are 1.04, 0.38, 0.12, and 0.05 respectively. This constitutes the eddy current energy characteristic vector. The energy peak of the third sub-band corresponds to the main vortex shedding frequency caused by the passing frequency of the wind turbine blades.

[0019] The specific process for high-frequency detail coefficient stripping of the circulating fan motor current perturbation signal is as follows: 512 current sampling points traced back from the current moment are selected to form an analysis window. The current signal within this window is decomposed using multi-scale discrete wavelet transform, with the fourth-order Daobesi wavelet as the mother wavelet and a decomposition level of 5. After 5 levels of decomposition, one approximate coefficient sequence and five detail coefficient sequences are obtained. The frequency ranges corresponding to the 4th and 5th level detail coefficients are 31.25 to 62.5 Hz and 15.625 to 31.25 Hz, respectively. These two levels of high-frequency detail coefficients carry the torque pulsation information generated by the non-uniform aerodynamic load excitation of the fan impeller. The root mean square value, kurtosis value, and crest factor of the 4th level detail coefficient sequence, as well as the root mean square value, kurtosis value, and crest factor of the 5th level detail coefficient sequence, are extracted, and these six features are arranged sequentially to form a high-frequency jitter feature sequence. If the root mean square value of the detail coefficients in layer 4 is 0.023 amperes, the kurtosis is 4.7, and the crest factor is 3.2, and the root mean square value of the detail coefficients in layer 5 is 0.015 amperes, the kurtosis is 3.9, and the crest factor is 2.8, then the high-frequency jitter feature sequence is: .

[0020] The specific process of multimodal fusion compression of eddy current energy feature vector, high-frequency jitter feature sequence, and hardware execution log is as follows: First, the five values ​​in the hardware execution log—circulating fan frequency conversion command value, nitrogen filling valve opening percentage, and heating conduction angle of each zone—are normalized to the 0-1 interval to form a 5-dimensional hardware state vector. The 8-dimensional eddy current energy feature vector, 6-dimensional high-frequency jitter feature sequence, and 5-dimensional hardware state vector are sequentially concatenated to form a 19-dimensional input vector. This input vector is fed into a fully connected network structure with two hidden layers. The first hidden layer has 32 neurons, the second hidden layer has 16 neurons, and the output layer has 12 neurons. Each hidden layer uses a hyperbolic tangent activation function. The weight matrix and bias vector of this fully connected network are obtained through offline collection of historical operating data, using mean squared error as the loss function and an adaptive moment estimation optimizer iteratively trained with a learning rate. The network then performs a linear transformation on the 19-dimensional input vector using the first hidden layer. Linear transformation of the second hidden layer Then, the output layer generates 12-dimensional feature codes. ,in It is a 32×19 weight matrix. It is a 16×32 weight matrix. It is a 12×16 weight matrix. , , These are the bias vectors for the corresponding dimensions. The 12-dimensional feature codes from six consecutive sampling times are stacked vertically in chronological order to form a 6-row, 12-column aerodynamic feature fingerprint matrix. .

[0021] The specific process of adaptive sliding filtering of the temperature measurement sequence of a multi-zone furnace based on the aerodynamic feature fingerprint matrix is ​​as follows: First, calculate the Frobenius norm of the aerodynamic feature fingerprint matrix at the current time. This norm value characterizes the overall intensity of the current turbulent disturbance within the furnace. Substituting this norm value into the negative exponential mapping function, the dynamic adaptive smoothing coefficient is calculated using the following formula:

[0022]

[0023] in is the dynamic adaptive smoothing coefficient at time t, with a value ranging from 0 to 1; The decay rate constant is calibrated according to the furnace type and loading amount. In this embodiment, it is set as follows: =0.15; Let be the Frobenius norm of the aerodynamic fingerprint matrix at time t. The norm increases with severe turbulent disturbances, and after negative exponential mapping... A value approaching 0 indicates that the weight of the current raw temperature measurement is significantly reduced, and the filter retains more of the smoothed result from the previous moment to suppress spurious fluctuations; when the turbulent disturbance is gentle, the norm value decreases. As the value approaches 1, the filter quickly follows changes in actual temperature. Specifically, when... When =3.8, =exp(-0.15×3.8)=0.565. Using this dynamic smoothing coefficient, a first-order exponential low-pass moving average is performed on each of the nine temperature measurement channels:

[0024]

[0025] in Let t be the calibrated temperature sequence vector at time t, with a dimension of 9×1, corresponding to the filtered temperature values ​​of the 9 temperature measurement points; Let be the original multi-zone furnace temperature measurement sequence vector at time t; This represents the calibrated temperature value from the previous moment. The formula indicates that the current output of each temperature measurement point is a weighted combination of its original reading and the filtered output from the previous moment, with the weights dynamically determined by the turbulence intensity represented by the aerodynamic fingerprint matrix. After the above adaptive sliding filter processing, the output calibrated temperature sequence vector effectively eliminates spurious temperature changes caused by transient turbulent purging, preserving the gradual trend of the furnace's true thermal state.

[0026] Specifically, in step two, based on a physically embedded deep neural network, the aerodynamic feature fingerprint matrix is ​​reconstructed by inverting the fluid-solid interface heat transfer coefficient to obtain a dynamic convective heat transfer coefficient mapping matrix. Correspondingly, after the frequency domain decomposition and time-series filtering feature extraction in step one, an aerodynamic feature fingerprint matrix capable of characterizing the turbulent disturbance state within the furnace has been obtained. However, this matrix is ​​only an indirect digital encoding of the degree of wind field turbulence and has not yet established a quantitative physical correlation between flow field disturbance and the actual heat exchange intensity on the magnetic ring surface. Inside the magnetic ring heat treatment furnace, the densely stacked magnetic rings constitute a complex porous medium. When nitrogen flows through different spatial locations, phenomena such as abrupt changes in the flow channel cross-section, eddy shedding, and local backflow cause spatial differences in the convective heat transfer coefficient at various locations, ranging from several times to orders of magnitude. This crucial physical quantity cannot be directly measured by sensors. Without a precise understanding of the point-by-point convective heat transfer coefficient in three-dimensional space, subsequent prediction of the magnetic ring stack temperature field will lack core boundary condition input, making it impossible to accurately locate and predict the risks of locally overheated or undercooled areas. Therefore, step two is executed.

[0027] In one operable embodiment of this application, step two includes: performing nonlinear dimensionality reduction projection on the aerodynamic feature fingerprint matrix based on a high-dimensional manifold encoder in a physically embedded deep neural network to obtain a latent space flow field feature tensor; performing scaling and matching and multi-layer sensing decoding on the latent space flow field feature tensor and three-dimensional preset grid coordinates to obtain a microscopic wind speed profile tensor; and performing grid-point-by-grid point convective heat transfer coefficient calculation and spatial reconstruction on the microscopic wind speed profile tensor to obtain a dynamic convective heat transfer coefficient mapping matrix.

[0028] The physically embedded deep neural network used in this step is a deep learning architecture that embeds the fluid dynamics conservation equations as regularization constraints into the network loss function. Its overall structure consists of three cascaded functional modules: a high-dimensional manifold encoder, a self-attention scaling matching decoder, and a point-by-point heat transfer coefficient physical calculation layer. The network is trained offline, using training data obtained from multiple sets of three-dimensional flow field and corresponding heat transfer coefficient distribution data obtained through computational fluid dynamics numerical simulations under different fan speeds, nitrogen flow rates, and furnace temperatures. Simultaneously, the momentum and energy conservation residuals of the Navier-Stokes equations are added as penalty terms to the mean squared error loss function. An adaptive moment estimation optimizer is used for iterative training until convergence with an initial learning rate, ensuring that the network output strictly adheres to fluid physics laws while meeting data fitting accuracy requirements.

[0029] The specific process of nonlinear dimensionality reduction projection of the aerodynamic feature fingerprint matrix by the high-dimensional manifold encoder is as follows: First, the aerodynamic feature fingerprint matrix is... The rows are flattened into 72-dimensional column vectors as the encoder input. The high-dimensional manifold encoder consists of four fully connected hidden layers, with the number of neurons in each layer being 64, 48, 32, and 16 respectively. After performing an affine transformation, each layer applies a linear unit activation function with leakage correction. The first hidden layer performs the computation. ,in This is the first layer weight matrix. For bias vectors, The activation function is a linear unit with leakage correction, and the slope of the negative half-axis is set to 0.01. After layer-by-layer compression transformation, the fourth hidden layer outputs a 16-dimensional vector, which is then passed through the output projection layer. Mapped to the final latent space flow field feature tensor ,in This is the projection layer weight matrix. This 16-dimensional latent space flow field feature tensor condenses the core topological pattern determining the turbulent state of the flow field within the furnace into an abstract semantic form detached from the original physical coordinates. Each weight matrix and bias vector is obtained through iterative updates by the optimizer during the aforementioned offline training phase. Taking a specific numerical example, when the aerodynamic feature fingerprint matrix output in step one corresponds to a medium turbulence intensity condition, the latent space flow field feature tensor obtained after dimensionality reduction projection by the encoder is: .

[0030] The specific process of scaling and matching the latent space flow field feature tensor with the three-dimensional pre-set grid coordinates and performing multi-layer sensing decoding is as follows: The three-dimensional pre-set grid is a discretized spatial coordinate system pre-constructed based on the internal geometry of the magnetic ring heat treatment furnace and the magnetic ring stacking method. It covers the length, width, and height of the effective heating area of ​​the furnace. The grid has 20 nodes in the length direction, 15 nodes in the width direction, and 10 nodes in the height direction, for a total of 3000 spatial discrete points. Each node is represented by a three-dimensional coordinate vector. This represents its spatial position within the furnace, with coordinate values ​​normalized to the [-1, 1] interval. The spacing between grid nodes is determined based on the effective heating area of ​​the furnace. In this embodiment, the effective furnace area is 600 mm × 450 mm × 400 mm, so the node spacing in the three directions is approximately 31.6 mm, 32.1 mm, and 44.4 mm, respectively. The decoding process employs a self-attention scaling matching mechanism, which sets the coordinates of each grid node... Projected into a query vector through linear transformation ,in To query the projection matrix; the latent space flow field feature tensor Generate key vectors through linear transformations respectively Sum value vector ,in and These are the key projection matrix and the value projection matrix, respectively. A scaled dot product is performed on the query vector and the key vector. Obtain the attention weight scalar, then multiply the scalar by the value vector to obtain the attention output. The attention output is then fed into a three-layer multilayer perceptron for decoding, with the number of neurons in each layer being 32, 16, and 3, respectively. The final output is a 3D vector representing the instantaneous wind speed components in the three directions at that grid point. After performing the scaling and matching and multilayer sensing decoding operations sequentially on all 3000 grid nodes, the three-dimensional wind speed vectors of all nodes are assembled into a microscopic wind speed profile tensor according to their spatial coordinates. For example, the microscopic wind speed at the grid point located at coordinates [0.0, 0.0, 0.0] in the middle of the furnace is [1.24, -0.37, 0.82] m / s after decoding, while the microscopic wind speed at the bottom edge of the magnetic ring stack at coordinates [-0.8, -0.7, -0.9] is [0.15, 0.08, -0.04] m / s due to flow obstruction. This spatial difference reflects the non-uniform distribution of airflow when passing through the dense magnetic ring stack. The query projection matrix, key projection matrix, value projection matrix, and the weights and bias parameters of the multilayer perceptron are all obtained through joint optimization with the encoder during the offline training phase.

[0031] The specific process of solving and spatially reconstructing the convective heat transfer coefficient of the microscopic wind speed profile tensor on a grid-by-grid basis is as follows: Traversing each grid node in the microscopic wind speed profile tensor Extract the three-dimensional wind speed vector at that point and calculate its absolute scalar wind speed. Simultaneously, the thermophysical properties of nitrogen under the current temperature conditions are read: the dynamic density of nitrogen at 450℃ and 0.3 kPa micro-positive pressure. =0.47 kg / m³, dynamic viscosity Pascal second, thermal conductivity =0.045 watts per meter Kelvin, Prandtl number =0.71. The fluid equivalent diameter of the magnetic ring as a porous medium. The equivalent diameter is determined based on the hydraulic diameter formula for the outer diameter of a single magnetic ring and the stacking gap. In this embodiment, the outer diameter of the magnetic ring is 25 mm, the inner diameter is 15 mm, the height is 10 mm, and the stacking gap between layers is 3 mm. =0.0058 meters. Based on the above parameters, the following three physical calculations are performed sequentially for each grid node. Step 1: Calculate the local Reynolds number:

[0032]

[0033] in For grid nodes The local Reynolds number at that location, The dynamic density of nitrogen. The absolute scalar wind speed at that point. The equivalent diameter of the fluid. The dynamic viscosity is used. The second step involves calculating the local Nusselt number using an empirical correlation formula for forced convection in porous media.

[0034]

[0035] in , , This is an empirical fitting constant, calibrated based on heat transfer experimental data of densely stacked porous media. In this embodiment, it is taken as... =0.35、 =0.6、 =0.36, applicable to forced convection conditions in porous media with Reynolds numbers ranging from 50 to 5000. The third step is to convert the Nusselt number to the local convective heat transfer coefficient:

[0036]

[0037] in For nodes The local convective heat transfer coefficient at a given location is expressed in watts per square kelvin. Taking a point in the middle of the furnace as an example, the absolute scalar wind speed at that point is... meters per second, substituting into the equation... ,and then ,final Kelvin per square meter. The absolute scalar wind speed at the coordinate point at the bottom edge of the magnetic ring stack is... meters per second, substituting into the equation... , , The convective heat transfer coefficients at the two points mentioned above differ by nearly 3.7 times, directly reflecting the significant impact of non-uniformity of the flow field within the furnace on local heat transfer intensity. After completing point-by-point calculations for all 3000 grid nodes, all local convective heat transfer coefficients are reassembled according to the original three-dimensional grid spatial coordinates, ultimately forming a dynamic convective heat transfer coefficient mapping matrix. This matrix fully describes the distribution of convective heat transfer intensity at each location in the furnace at the current moment in a three-dimensional spatial resolution.

[0038] Specifically, in the aforementioned basic implementation, the calculation of the convective heat transfer coefficient for the microscopic wind speed profile tensor on a grid-by-grid basis employs a semi-empirical Nusselt number correlation based on conventional porous media. This correlation calculates the Reynolds number using the macroscopic scalar wind speed and then reduces it to the convective heat transfer coefficient using the Nusselt number. Its implicit premise is that the gas is in a completely continuous flow state and the wall surface satisfies no-slip boundary conditions, while treating nitrogen physical properties as constants related only to the macroscopic temperature of the furnace. However, in the magnetic ring heat treatment furnace involved in this application, the furnace maintains a nitrogen micro-positive pressure of 0.3 kPa (absolute pressure approximately 101.6 kPa). Although the gas as a whole remains in a continuous medium state, the equivalent characteristic pore size of the gap channels formed by the densely stacked magnetic rings is only on the order of millimeters. Under high-temperature conditions, the dynamic viscosity of the gas increases, leading to a significant expansion of the thermal boundary layer thickness adhering to the wall. The thermal flow boundary layer on adjacent magnetic ring surfaces is prone to spatial interference and superposition effects within microscopic slits, causing local airflow stagnation and resulting in an actual heat exchange efficiency significantly lower than the theoretical value under the assumption of a wide flow channel. Simultaneously, the periodic pulsating airflow generated by the high-speed rotation of the circulating fan, when traversing the porous medium of the magnetic ring stack, creates strong shear flow at local flow channel contractions and the edges of the magnetic ring surface. This shearing action can peel off and thin the thermal boundary layer, thereby enhancing local convective heat transfer. However, the Reynolds number calculated solely based on macroscopic scalar wind speed cannot quantitatively characterize the compression and thinning effect of this three-dimensional shear strain rate on the thermal resistance boundary layer. Furthermore, within a wide temperature range of 380℃ to 560℃, there is a significant difference between the gas temperature inside the thermal boundary layer adhering to the magnetic ring wall and the mainstream temperature, causing the gas viscosity and thermal conductivity within the boundary layer to deviate from the mainstream values, resulting in a radial drift effect of temperature-viscosity properties. To address this, a grid-by-grid dynamic convection heat transfer coefficient calculation method based on thermal boundary layer interference hindrance correction and shear strain rate enhancement is employed. This method aims to integrate the three mechanisms of boundary layer spatial overlap interference correction, three-dimensional shear strain rate tensor enhancement, and wall temperature-viscosity property adaptive correction into the heat transfer coefficient calculation, so that the dynamic convection heat transfer coefficient mapping matrix can more accurately reflect the actual heat transfer state at each location within the furnace.

[0039] Based on this, in a preferred embodiment of this application, the convective heat transfer coefficient is calculated and spatially reconstructed point by point on the micro wind speed profile tensor to obtain a dynamic convective heat transfer coefficient mapping matrix, including: calculating the dynamic convective heat transfer coefficient point by point on the micro wind speed profile tensor based on micro-pressure slip flow correction and shear strain rate enhancement to obtain a dynamic convective heat transfer coefficient mapping matrix.

[0040] The specific process is as follows: Traverse each grid node in the microscopic wind speed profile tensor. First, extract the three-dimensional wind speed vector at that node. And calculate its absolute scalar wind speed. Simultaneously, the current instantaneous temperature of the node is read from the transient temperature field tensor of the magnetic ring. The thermophysical property parameters of nitrogen are updated in real time based on this temperature value: dynamic density. Using the ideal gas law Calculate the dynamic viscosity The results are updated in real time using the Sutherland formula. Calculated at 450℃ (723 Kelvin). =0.47 kg per cubic meter, Pascal second, thermal conductivity =0.045 watts per meter Kelvin, Prandtl number =0.71. Equivalent diameter of the porous medium fluid in the magnetic ring. =0.0058 meters, determined by geometric parameters of the magnetic ring: outer diameter 25 mm, inner diameter 15 mm, height 10 mm, and interlayer stacking gap 3 mm.

[0041] After real-time updates of physical property parameters are completed, three pre-calculations are performed sequentially: calculation of the local boundary layer interferometry number, calculation of the symmetric strain rate tensor, and calculation of the local Reynolds number. The formula for calculating the local boundary layer interferometry number is:

[0042]

[0043] in For nodes The local boundary layer interference number at a given location characterizes the degree of spatial interference and compression between the thermal flow boundary layer thickness and the channel size within the micropores. Pascal-Mi / Kelvin is the empirical constant for thermal interference in high-temperature porous media; The instantaneous absolute temperature at this node is expressed in Kelvin. =101600 Pascals is the absolute pressure in the furnace, which is the standard atmospheric pressure plus 0.3 kPa gauge pressure; The pore size is the microscopic equivalent characteristic aperture of the gap channel formed by the stacking of multiple magnetic rings, determined by the harmonic average of the radial and axial gaps between adjacent magnetic rings. Taking a node temperature of 723 Kelvin as an example, the calculation yields... This value falls within the region of significant boundary layer interference overlap, indicating that the superposition interference effect of wall thermal resistance cannot be ignored.

[0044] The formula for calculating the symmetric strain rate tensor is:

[0045]

[0046] in For nodes The 3×3 symmetric strain rate tensor at the location; The spatial first-order gradient tensor of the microscopic wind speed profile tensor at this node is obtained by dividing the wind speed difference between adjacent nodes by the grid spacing using the central difference method. The square root of the second invariant is taken for this tensor. As a scalar measure characterizing the shear deformation intensity of the flow field at a given point. Taking the coordinate point in the middle of the furnace as an example, the wind speed in the x-direction at this point is 1.24 m / s, the wind speeds in the x-direction of adjacent nodes are 1.08 and 1.37 m / s respectively, and the grid spacing in the x-direction is 0.0316 m, then... By calculating the remaining velocity gradient components per second and assembling the symmetric strain rate tensor, the following can be obtained: Every second. The bottom edge coordinates are low due to the overall low wind speed and gentle spatial gradient. Every second.

[0047] The formula for calculating the local Reynolds number is:

[0048]

[0049] in The local dynamic Reynolds number at this node is calculated based on real-time updated physical property parameters; The dynamic density of nitrogen at this node temperature; For absolute scalar wind speed; The equivalent diameter of the fluid; Let be the dynamic viscosity at this node temperature. The absolute scalar wind velocity at the coordinate point in the middle of the furnace is... meters per second, substituting into the equation... The absolute scalar wind speed at the bottom edge coordinate point is... meters per second, substituting into the equation... .

[0050] Based on the above preliminary calculation results, the dynamic convection heat transfer coefficient of each grid node is calculated using a triple correction mechanism, and the main formula is as follows:

[0051]

[0052] in For nodes The dynamic local convective heat transfer coefficient at the location, expressed in watts per square kelvin; The nitrogen thermal conductivity is adaptively updated based on the instantaneous temperature of this node; =0.0058 meters is the fluid equivalent diameter; =0.9 is the tangential momentum fit coefficient of the solid wall of the magnetic ring metal, which characterizes the degree of complete momentum exchange after the collision between gas molecules and the solid wall. It is determined based on the surface roughness characteristics of the nanocrystalline alloy. =1.4 is the specific heat ratio of nitrogen; The local Knudsen number is obtained from prior calculation; =0.71 is the Prandtl number; =0.35、 =0.6、 =0.36 is the empirical fitting constant for the Nusselt number of forced convection in porous media, which is consistent with the value of the basic implementation method; =0.14 is the coupling calibration constant for the shear strain rate enhancement term. =0.65 is the shear enhancement index, and the two were determined by regression fitting of high-resolution computational fluid dynamics simulation data and heat transfer experimental data. This is the temperature of the outer wall of the solid skeleton on the surface of the magnetic ring at this node; The mainstream nitrogen temperature at this node is obtained by interpolation of the calibrated temperature sequence vector. =0.14 is the temperature-viscosity correction index, which is determined with reference to the empirical correlation of high-temperature gas heat transfer.

[0053] The formula consists of four cascaded functional factors. The first factor is the temperature-adaptive heat transfer dimension conversion coefficient, which converts the dimensionless Nusselt number into a convective heat transfer coefficient with physical dimensions. At 450℃, its value is 0.045 / 0.0058 = 7.76 W / m² Kelvin. The second factor is the gap boundary layer interference attenuation factor. When the equivalent interference number increases, i.e., severe overlap and interference occur in the thermal boundary layers at the slit, this factor decreases to below 1, applying a reasonable attenuation correction to the local heat transfer coefficient. This physically simulates the decrease in actual heat transfer efficiency caused by the viscous flow restriction of gas in small gaps. The third factor is the main heat transfer intensity factor, which is composed of the contribution of the conventional Nusselt number and the contribution of the enhanced shear strain rate. The latter fully characterizes the enhanced shear flow effect on the stripping and thinning of the thermal boundary layer by introducing the second invariant of the three-dimensional symmetric strain rate tensor. The fourth factor... As a wall temperature-viscosity correction factor, when the wall temperature is higher than the mainstream gas temperature during the heating stage, this factor is greater than 1 and has an enhancing effect; when the wall temperature is lower than the gas temperature during the cooling stage, this factor is less than 1 and has an attenuating effect, thus achieving symmetrical correction of the radial drift effect of gas properties in the boundary layer under large temperature difference conditions.

[0054] A complete solution is performed using the coordinates of the center of the furnace as a specific numerical example. =0.073、 =130.3、 per second. First factor = 0.045 / 0.0058 = 7.76. Second factor, interference attenuation factor = The third factor, the regular part = The third factor shear enhancement component = The third factor totals 5.75 + 0.455 = 6.205. At this point, the magnetic ring wall temperature is slightly higher than the mainstream gas temperature during the heating phase. Therefore, we take... =726 / 723=1.004, fourth factor= .final The heat transfer coefficient is reduced by approximately 31.9% after boundary layer interference correction, compared to 44.6 W / m² Kelvin in the basic implementation. This truly reflects the substantial reduction in heat transfer efficiency caused by the thermal resistance compression of the boundary layer between the flow gaps. At the same time, the shear enhancement term contributes approximately 7.9% of the positive compensation.

[0055] The second set of numerical examples uses the coordinates of the bottom edge of the magnetic ring stack. =0.074、 =14.9、 Every second. The second factor, namely the interference attenuation factor = The third factor, the regular part = The third factor shear enhancement component The third factor totals 1.56 + 0.188 = 1.748. The fourth factor is set to 1.0. Final... The heat transfer coefficient is 12.1 W / m² Kelvin. Compared to the 12.1 W / m² Kelvin in the basic implementation, the corrected coefficient is reduced by approximately 29%, which is mainly due to the attenuation caused by boundary layer interference. The shear enhancement term in this low-velocity region only increases the local contribution from 1.56 to approximately 1.748. The ratio after the two corrections is 30.38 / 8.50 = 3.57 times, which is consistent with the 44.6 / 12.1 = 3.69 times spatial non-uniformity of the basic method. This indicates that the triple correction mechanism maintains the spatial distribution pattern while providing physically consistent correction to the absolute value.

[0056] After performing the above point-by-point calculations on all 3000 grid nodes, all local dynamic convection heat transfer coefficients are reassembled according to the original three-dimensional grid spatial coordinates to finally form a dynamic convection heat transfer coefficient mapping matrix. This matrix incorporates three physical mechanisms: boundary layer interference hindrance correction, shear strain rate enhancement, and temperature-viscosity property correction. Compared with the basic implementation method, the overall heat transfer coefficient is corrected by about 20% to 25%, which more realistically reflects the actual heat transfer characteristics of micro-positive pressure nitrogen in the millimeter-scale magnetic ring gap channel. It provides a fluid-solid interface thermal boundary condition input with higher physical fidelity for the three-dimensional transient temperature field tomographic prediction of the magnetic ring stack in step three.

[0057] Specifically, in step three, the three-dimensional thermal shock risk and temperature tomography prediction of the magnetic ring stack under unsteady flow field disturbance are performed on the calibrated temperature sequence vector and the dynamic convective heat transfer coefficient mapping matrix to obtain the transient temperature field tensor and local thermal shock risk assessment matrix of the magnetic ring. It is understandable that after the fluid-solid interface heat transfer coefficient inversion and reconstruction in step two, a dynamic convective heat transfer coefficient mapping matrix describing the point-by-point convective heat transfer intensity distribution in the three-dimensional space inside the furnace has been obtained. Simultaneously, the calibrated temperature sequence vector output in step one provides filtered and corrected true temperature information for each zone of the furnace. However, these two sets of data describe the heat transfer boundary conditions from the flow field side and the heat source side, respectively, and have not yet been transformed into the actual temperature state of each microscopic location inside the magnetic ring stack. Furthermore, they cannot directly reveal which spatial regions are experiencing thermal stress exceeding the material's bearing capacity and facing the risk of brittle fracture. During the heat treatment of the magnetic ring, the yield strength of the nanocrystalline magnetic material softens significantly with increasing temperature. Once the thermal expansion distortion stress induced by the localized severe temperature gradient approaches or even exceeds the yield limit at the current temperature, the magnetic ring will develop irreversible microcracks, leading to brittle failure and deterioration of magnetic properties. Therefore, step three is performed.

[0058] In one operable embodiment of this application, step three includes: using the calibrated temperature sequence vector as the far-field temperature boundary condition and the dynamic convection heat transfer coefficient mapping matrix as the fluid-solid interface source term, performing time-domain numerical derivation and solution of the partial differential equation for heat conduction in the unsteady porous medium to obtain the transient temperature field tensor of the magnetic ring; based on the elastic Young's modulus and the coefficient of thermal expansion, performing temperature spatial tomography and local thermal distortion stress stripping on the transient temperature field tensor of the magnetic ring to obtain the multidimensional structural tensile stress matrix; and based on the material temperature-related yield strength threshold, performing nonlinear thermal shock risk normalization mapping based on the yield critical limit on the multidimensional structural tensile stress matrix to obtain the local thermal shock risk assessment matrix.

[0059] Step 3 is illustrated here using the basic numerical values ​​from Step 2 as an example. The specific process of using the calibrated temperature sequence vector as the far-field temperature boundary condition and the dynamic convection heat transfer coefficient mapping matrix as the fluid-solid interface source term for time-domain numerical derivation and solution is as follows: This is processed using a three-dimensional unsteady porous medium heat conduction partial differential evolution solver. This solver is based on a numerical calculation module constructed using the energy conservation control volume method. Its spatial discrete grid is consistent with the three-dimensional pre-set grid from Step 2, i.e., 3000 control volumes in total, with 20 nodes in the length direction, 15 nodes in the width direction, and 10 nodes in the height direction. Each control volume represents an equivalent porous medium micro-element containing a magnetic ring solid skeleton and interstitial nitrogen gas. The solver iterates step-by-step along the time axis in an explicit time-progression format, with a time step size set to 0.5 seconds and a look-ahead prediction time-domain window set to 30 seconds, i.e., 60 time steps. Within each time step, the discretization solution of the energy balance equation is performed simultaneously on all 3000 grid nodes. The calibrated temperature sequence vector contains filtered temperature values ​​from nine temperature measurement points, distributed across the upper, middle, and lower zones of the furnace, with three points in each zone. These nine temperature values ​​are mapped to 3000 grid nodes using three-dimensional linear interpolation, with each node... Configure the far-field ambient temperature corresponding to its spatial location. This temperature represents the mainstream nitrogen temperature far from the surface of the magnetic ring. For example, when the temperatures at the three points in the upper region of the calibrated temperature sequence vector are 462.3℃, 458.7℃, and 460.1℃, the middle region is 455.8℃, 453.2℃, and 456.5℃, and the lower region is 448.6℃, 445.9℃, and 447.3℃, after three-dimensional linear interpolation, the far-field temperature at the center of the furnace (coordinates [0.0, 0.0, 0.0]) is 454.8℃, and the far-field temperature at the bottom edge (coordinates [-0.8, -0.7, -0.9]) is 446.2℃. For each grid node... At each time step, the discretized solution of the partial differential equation for heat conduction in the unsteady porous medium is performed. The equation is as follows:

[0060]

[0061] in The equivalent density of the porous medium of the magnetic ring is calculated by weighting the density and porosity of the solid material of the magnetic ring. In this embodiment, the magnetic ring material is an iron-based nanocrystalline alloy with a solid density of 7200 kg / m³ and a stacked porosity of 0.42. Therefore, the equivalent density is... =7200×(1-0.42)=4176 kg per cubic meter; The equivalent specific heat capacity is taken as 520 joules per kilogram Kelvin; Grid nodes exist The internal temperature of the magnetic ring at any given time; The equivalent thermal conductivity of the porous medium, taking into account both the thermal conductivity of the solid skeleton and the thermal conductivity of the interstitial gas, is taken as 8.5 W / m Kelvin; This is the heat conduction and diffusion term, which characterizes the heat conduction between adjacent grid nodes due to temperature differences. The dynamic convection heat transfer coefficient at this node is directly taken from the value at the corresponding position in the dynamic convection heat transfer coefficient mapping matrix output in step two. The specific surface area per unit volume of the porous medium is calculated based on the geometric parameters of the magnetic ring. In this embodiment, it is taken as... =180 square meters per cubic meter; This represents the far-field ambient temperature at this node. The first term on the right-hand side of the equation is the heat conduction-diffusion term, describing the heat conduction process between the solid skeletons inside the magnetic ring stack; the second term is the convective heat transfer source term, describing the forced convective heat transfer process when nitrogen flows over the surface of the magnetic ring, the intensity of which is determined by the dynamic convective heat transfer coefficient. Taking the node in the middle of the furnace as an example, this point... =44.6 watts per square meter Kelvin, far-field temperature 454.8℃, current internal temperature of the magnetic ring 449.2℃, the convection source term contribution is 44.6 × 180 × (454.8 - 449.2) = 44956 watts per cubic meter, while the bottom edge node =12.1 watts per square meter Kelvin, far-field temperature 446.2℃, and current temperature 443.8℃, the convection source term is only 12.1×180×(446.2-443.8)=5227 watts per cubic meter, a difference of nearly 8.6 times, demonstrating the significant impact of flow field non-uniformity on local heating rates. The solver iterates through 60 time steps at 0.5-second intervals, assembling the temperature values ​​of all 3000 nodes over those 60 time steps into a transient temperature field tensor for the magnetic ring. .

[0062] The specific process of performing temperature space tomography and local thermal distortion stress stripping of the magnetic ring based on the transient temperature field tensor of the magnetic ring using the elastic Young's modulus and the coefficient of thermal expansion is as follows: The elastic Young's modulus is a mechanical constant characterizing the stress-strain ratio of a material during elastic deformation, and the coefficient of thermal expansion is a physical parameter characterizing the relative change in linear dimensions of a material per unit temperature increase. In this embodiment, the magnetic ring material is an iron-based nanocrystalline alloy, and the elastic Young's modulus at 450℃ is... =155 gigapascals, coefficient of thermal expansion per Kelvin, Poisson's ratio =0.28, the above parameter was obtained from the high-temperature mechanical property test report provided by the material supplier. For each grid node in the transient temperature field tensor of the magnetic ring, the temperature gradient between it and its adjacent nodes is calculated along the three-dimensional orthogonal coordinate direction. To avoid smoothing out the real thermal abrupt changes between adjacent porous media grids due to the excessively long span of the traditional central difference method, a one-way finite difference method is used here to maintain the high-frequency spatial sensitivity of the discrete calculation. (The last sentence appears to be incomplete and possibly refers to a separate process.) For example, using first-order forward differential discretization, its local temperature gradient in the x-direction is approximately calculated as follows: ,in The grid spacing for this direction is 31.6 mm. The gradient components in the three directions are synthesized into the Euclidean norm of the temperature gradient vector. Based on the constitutive relations of thermoelasticity, the temperature gradient norm is transformed into local thermal distortion stress. The calculation formula is as follows:

[0063]

[0064] in For grid nodes The local thermal distortion stress value at time t, in Pascals; It is the elastic Young's modulus; The coefficient of thermal expansion; Poisson's ratio; This represents the temperature gradient norm at that point, expressed in Kelvin per meter. When the temperature gradient between adjacent regions is greater, the constraint stress generated by the different thermal expansion amounts of different parts of the material within the equivalent constraint distance is greater. The existence of Poisson's ratio reflects the stress concentration amplification effect caused by the restricted lateral deformation under three-dimensional constraint conditions. For example, at a node in the middle of the furnace, the temperature of its forward adjacent point is 451.6℃, and its own temperature is 447.8℃, with a distance of 31.6 mm. Considering only the unidirectional differential gradient in the x-direction... Kelvin per meter, substituting, we get Pascal is approximately 298 megapascals. After performing the above calculations on all 3000 nodes over 60 time steps, a multidimensional structural tensile stress matrix is ​​constructed. .

[0065] The specific process of nonlinear thermal shock risk normalization mapping of the multidimensional structural tensile stress matrix based on the material temperature-dependent yield strength threshold is as follows: The yield strength of iron-based nanocrystalline alloys decreases with increasing temperature, and its temperature-dependent yield strength... Based on the current temperature values ​​of each node in the transient temperature field tensor of the magnetic ring, retrieved from the material property database, in this embodiment, the yield strength is 680 MPa at 400℃, 620 MPa at 450℃, 540 MPa at 500℃, and 450 MPa at 550℃. Intermediate temperatures are obtained through linear interpolation. A nonlinear exponential mapping operation is performed on each node in the multidimensional structural tensile stress matrix, and its calculation formula is as follows:

[0066]

[0067] in For nodes The local thermal shock risk assessment value at time t; This is a risk amplification factor, used to control the exponential growth rate of the risk value as the stress ratio approaches the yield limit. In this embodiment, it is set as follows: =5; The stress ratio is the ratio of the current thermal stress to the yield strength at that temperature. When the stress ratio is less than 1, the exponential value is negative, and the mapping result approaches 0, indicating that the node is in a safe state. When the stress ratio is equal to 1, the mapping result is 1, indicating that the yield criticality has just been reached. When the stress ratio is greater than 1, the mapping result increases exponentially, with extremely high values ​​indicating that the node has entered the dangerous brittle failure range. Taking the aforementioned middle node as an example, the actual internal temperature of this point is 449.2℃. Based on the precise linear interpolation derivation in the 400℃ to 450℃ range, its corresponding yield strength is 621 MPa, and the corresponding generated thermal stress is 298 MPa. The stress ratio is 298 / 621≈0.48. Substituting this into the equation yields the local thermal shock risk value. =exp(5×(0.48-1))=0.074, indicating an extremely low risk. However, when a bottom edge node experiences a drastic increase in local temperature gradient to 380 Kelvin per meter due to a dead zone in the flow field, corresponding to a local thermal stress of 940 MPa, and the actual temperature at that point is 530℃ due to weak convective heat transfer, according to strict linear interpolation calculations within the 500℃ to 550℃ range, the corresponding true yield strength should be 486 MPa after degradation. In this case, the actual stress ratio jump is 940 / 486≈1.934. Substituting these values ​​into the exponential formula yields the local thermal shock risk assessment value. =exp(5×(1.934-1))=106.7, indicating that the mapping result of this node shows an exponential increase, with an extremely high value indicating that the node has entered the dangerous brittle failure range. After mapping all nodes and time steps, a local thermal shock risk assessment matrix is ​​assembled. .

[0068] Specifically, in step four, the transient temperature field tensor of the magnetic ring is used as the tracking target, and the local thermal shock risk assessment matrix is ​​configured as a penalty constraint. A nonlinear model predictive controller is then used for multivariate collaborative rolling optimization to obtain a collaborative execution instruction set. It should be understood that after the time-domain numerical extrapolation in step three, the transient temperature field tensor describing the three-dimensional transient temperature distribution inside the magnetic ring stack and the local thermal shock risk assessment matrix quantifying the degree of brittle fracture risk at each spatial node have been obtained. However, these two prediction results are only passive observations of future temperature evolution trends and potential hazardous areas under current operating conditions, and have not yet been transformed into control instructions capable of driving active intervention by the actuators. During the magnetic ring heat treatment process, there is a strong coupled nonlinear relationship between the three control variables: the circulating fan speed, the nitrogen filling valve opening, and the electric heating power. Independent adjustment of any single variable may trigger a chain reaction of fluctuations in the corresponding physical quantities of other variables through a flow-thermal coupling mechanism. Traditional PID controllers, due to their single-variable feedback nature, cannot handle this multi-input, multi-output coupled optimization problem. Meanwhile, there is an inherent contradiction between temperature tracking accuracy and thermal shock safety. Aggressive heating strategies can improve temperature tracking speed but exacerbate local thermal stress, while conservative strategies lead to a shift in the phase transition time window. Therefore, step four is performed.

[0069] Figure 3 This is a flowchart of step four in the magnetic ring heat treatment parameter prediction and optimization method according to an embodiment of this application. Figure 3 As shown, in one operable embodiment of this application, step four includes: step four-one, using the weighted sum of squares of the deviation between the transient temperature field tensor of the magnetic ring and the ideal phase transition tracking curve as the tracking cost term, and configuring the local thermal shock risk assessment matrix as an exponential impedance penalty constraint term to construct a constrained prediction cost function equation; step four-two, performing a second-order Taylor expansion and rolling time-domain convexity iteration on the constrained prediction cost function equation to obtain the full-scale optimal control sequence; step four-three, based on the rolling implementation strategy, performing a first-step time-domain slice stripping and industrial communication protocol encoding on the full-scale optimal control sequence to obtain a cooperative execution instruction set.

[0070] The core decision variable of a nonlinear model predictive controller is the control increment vector. ,in For the variable frequency speed increment of the circulating fan, For the nitrogen charging valve opening increment, For the increment of electric heating power, three variables control the furnace flow field intensity, atmosphere pressure, and heat source input, respectively. Prediction in the time domain. Set 60 steps to correspond to 30 seconds, control the time domain. The setting is 10 steps corresponding to 5 seconds, that is, the optimal incremental sequence for the next 10 control steps is solved within each control cycle.

[0071] The specific process of constructing the constrained prediction cost function equation by using the weighted sum of squares of the deviations between the transient temperature field tensor of the magnetic ring and the ideal phase transition tracking curve as the tracking cost term and configuring the local thermal shock risk assessment matrix as an exponential impedance penalty constraint term is as follows: Ideal phase transition tracking curve This is a pre-set target temperature-time trajectory based on the crystallization kinetics of iron-based nanocrystalline alloys. The criteria are: a uniform heating rate of 2.5℃ per minute in the 380℃ to 460℃ heating range to avoid thermal shock caused by excessive heating; precise temperature control at a slow rate of 1.2℃ per minute in the phase transformation core range of 460℃ to 520℃ to ensure uniform nucleation of nanocrystals; and a heating rate of 1.8℃ per minute in the 520℃ to 560℃ range to complete the grain growth process. This curve is plotted by process engineers based on differential scanning calorimetry experiments, determining the material's phase transformation initiation temperature, peak temperature, and termination temperature, and stored in the controller parameter table. Taking the current furnace temperature as an example, which is in the phase transformation core range, the ideal tracking temperature is... =485.0℃, and increasing at a rate of 1.2℃ per minute within a 30-second prediction time domain, then =485.0 + 1.2 × 0.5 = 485.6℃. Constrained prediction cost function equation. The complete form is:

[0072]

[0073] The first item is the temperature tracking cost. Let be the magnetic ring temperature field at time t, predicted in the k-th step based on the current control increment. The ideal phase transition tracking temperature for the k-th step. The diagonal matrix for temperature tracking weights is configured such that its diagonal elements are assigned different weight values ​​according to their spatial location. The weight of the central region of the magnetic ring stack is set to 1.5, while that of the edge region is set to 0.8 to emphasize the temperature accuracy of the core area. The weighted L2 norm squared is... The second item is the cost of controlling energy consumption. This is the control increment vector for the (k-1)th step. To control the incremental penalty weights of the diagonal matrix, its three diagonal elements are set as follows: =0.01、 =0.05 and =0.02, used to suppress drastic changes in control quantities to protect mechanical actuators. The third item is the thermal shock risk penalty cost. This is the local thermal shock risk assessment value predicted in step k. The global risk penalty weighting coefficient is set to 50. The risk space weight matrix is ​​set to an identity matrix. Because... The stress ratio itself already possesses the characteristic of rapidly amplifying when it exceeds the yield limit through the exponential mapping in step three. Therefore, squaring it and multiplying it by a larger factor... The value constitutes an impedance penalty constraint with veto power; when the thermal shock risk value of any node spikes, it will dominate the entire cost function, forcing the controller to prioritize safety. Simultaneously, the physical envelope constraint of the control increment is set as follows: Rotation per minute , kilowatt.

[0074] The specific process of solving the constrained prediction cost function equation through second-order Taylor expansion and rolling time-domain convexification iteration to obtain the full-scale optimal control sequence is as follows: Since the cost function contains a nonlinear temperature field prediction model and an exponential risk penalty term, directly solving for the global optimal solution is computationally infeasible. Therefore, a sequential quadratic programming iteration strategy is adopted to gradually convexify the non-convex problem. At the current operating point, the cost function is... Perform a second-order Taylor expansion to calculate the first-order gradient vector. and second-order Hessian matrix superscript This represents the l-th iteration. Since the control time domain has 10 steps with 3 control variables per step, then... It is a 30-dimensional vector, corresponding to the Hessian matrix. gradient vector The expanded quadratic approximation problem can be formulated as a standard convex quadratic programming subproblem:

[0075]

[0076] in The optimal control increment sequence to satisfy the constraints. To ensure the convexity of the subproblem, a positive definite Hessian matrix is ​​required. The first-order gradient vector indicates the steepest descent direction. During the solution process, Lagrange multipliers are introduced to handle the physical envelope inequality constraints controlling the increment. The effective set method is used to identify active constraints in each iteration and solve for the optimal solution under the equality constraints. The maximum number of inner loop iterations is set to 15, and the convergence criterion is that the rate of change of the objective function value between two adjacent iterations is less than... Taking a specific numerical value as an example, the initial iteration point is... The first iteration yielded... =28 revolutions per minute =-2.1%, =1.8 kW corresponds to a cost function value of =342.6, converged to after 8 iterations. =22 revolutions per minute =-1.6%, =1.3 kW corresponding cost function value =215.3. The optimal increments of the 10 control steps are arranged in chronological order to form the full-scale optimal control sequence. .

[0077] The specific process of performing initial time-domain slicing and industrial communication protocol encoding on the full-scale optimal control sequence based on the rolling implementation strategy to obtain the cooperative execution instruction set is as follows: According to the rolling time-domain implementation principle of model predictive control, only the first set of control increments corresponding to the current time moment in the full-scale optimal control sequence is extracted. Discard the remaining 9 sets of subsequent predicted redundant increments. Add the first set of increments to the actual operating baseline value of the equipment from the previous control cycle, and calculate the absolute control command to be applied in the current cycle:

[0078]

[0079] in The set of cooperative execution instructions output at time t; This is the basis vector of the equipment's actual operation in the previous cycle, which includes three absolute values: the current speed of the fan, the current opening degree of the valve, and the current power of the heater. The first step is to extract the matrix from the time-domain slice. The first 3 elements are extracted from the 3-dimensional identity matrix, which is the initial increment. Subsequent elements are set to zero. This is an industrial communication protocol encoding function that scales and encapsulates floating-point absolute instruction values ​​according to the data frame format of the industrial fieldbus protocol. For example, if the base values ​​of the previous cycle are a fan speed of 1450 rpm, a nitrogen filling valve opening of 35%, and a heating power of 28 kW, after adding the optimal increment from the first step, the absolute instruction values ​​are a fan speed of 1472 rpm, a nitrogen filling valve opening of 33.4%, and a heating power of 29.3 kW. After protocol encoding, a set of coordinated execution instructions is generated, which includes the target speed scaling integer, the valve opening scaling integer, and the power scaling integer.

[0080] Specifically, in step five, the circulating fan inverter speed, nitrogen filling valve opening, and electric heating power are fed forward and collaboratively executed based on the collaborative execution instruction set. That is, after the multivariate collaborative rolling optimization of the nonlinear model predictive controller in step four, a collaborative execution instruction set containing the three-dimensional optimal control instructions for fan speed, nitrogen filling valve opening, and heating power has been obtained. However, this instruction set still exists in the controller's internal floating-point data frame format, belonging to the abstract decision results of the digital domain, and has not yet been converted into standard industrial electrical signals that can be recognized and responded to by the physical actuators. In the actual hardware architecture of the magnetic ring heat treatment furnace, the circulating fan inverter, servo nitrogen filling valve, and heating thyristor, the three types of actuators, use different signal interface specifications and response characteristics. If the three control instructions are sent directly in a unified data packet format, each actuator cannot correctly parse its required action parameters, resulting in the control decision not being translated into physical actions. Therefore, step five is finally executed.

[0081] In one operable embodiment of this application, step five includes: parsing the polymorphic communication protocol and decoupling the actuation signal of the cooperative execution instruction set to obtain a three-dimensional linkage actuation signal vector; and sending the three-dimensional linkage actuation signal vector to the circulating fan inverter, the servo nitrogen charging valve and the heating thyristor to perform physical actions.

[0082] The specific process of parsing the polymorphic communication protocol and decoupling the actuation signal of the collaborative execution instruction set to obtain the three-dimensional linkage actuation signal vector is as follows: After receiving the data frame of the collaborative execution instruction set, the underlying device protocol parsing middleware first performs frame header verification and cyclic redundancy check to confirm data integrity. Then, it performs byte unpacking operation on the payload segment, restoring the scalable quantized integer to a floating-point absolute instruction value. The unpacked three-dimensional instruction value is decomposed according to the device address code for addressing and routing, and separated into the circulating fan frequency converter reference value. =1472 revolutions per minute, servo nitrogen charging valve duty cycle =33.4% and heating thyristor conduction angle =29.3 kW three independent operating modes. Then, the electrical gain coefficient of each actuator is introduced for range linear mapping, converting the pure digital domain modal values ​​into standard industrial analog electrical signals. The circulating fan frequency converter accepts 0 to 10 volt analog inputs corresponding to a speed range of 0 to 3000 rpm, and its electrical gain coefficient is... =10 / 3000=0.00333 volts per revolution per minute, therefore the fan control voltage is =1472 × 0.00333 = 4.90 volts. The servo-controlled nitrogen charging valve accepts a 4 to 20 mA current signal corresponding to a 0% to 100% opening range, and its electrical gain coefficient is... =6 / 100=0.16 mA per percent, therefore the valve control current is =4 + 33.4 × 0.16 = 9.34 mA. The heating thyristor accepts analog inputs from 0 to 10 volts, corresponding to a power range of 0 to 60 kilowatts, and its electrical gain coefficient is... =10 / 60=0.1667 volts per kilowatt, therefore the heating control voltage is =29.3 × 0.1667 = 4.88 volts. The above electrical gain coefficient is determined by the factory nameplate parameters of each actuator and is stored in the configuration register of the protocol parsing middleware. The three converted standard industrial electrical signals are sequentially assembled into a three-dimensional linkage actuation signal vector. .

[0083] The specific process of sending the three-dimensional linkage actuation signal vector to the circulating fan inverter, servo nitrogen charging valve, and heating thyristor for physical action is as follows: The three electrical signals in the three-dimensional linkage actuation signal vector are synchronously sent within the same control cycle through the digital-to-analog converter output port in a feedforward coordinated manner. The 4.90-volt voltage signal is transmitted to the analog input terminal of the circulating fan inverter through a shielded cable. The inverter adjusts the output frequency accordingly to smoothly transition the fan motor speed to the target value of 1472 revolutions per minute within 0.5 seconds. The 9.34-mA current signal is transmitted to the current loop input terminal of the servo nitrogen charging valve. The valve actuator adjusts the valve core displacement accordingly to stably output nitrogen flow at 33.4% opening. The 4.88-volt voltage signal is transmitted to the trigger control terminal of the heating thyristor. The thyristor adjusts the conduction angle accordingly to make the output power of the electric heating tube reach 29.3 kilowatts. The synchronized action of the three actuators achieves coordinated adjustment of fan speed, nitrogen flow rate, and heating power at the physical level. This allows the flow field intensity, atmosphere pressure, and heat source input inside the furnace to change in a coordinated manner at the same time according to the optimal strategy obtained in step four. As a result, when the next sampling cycle arrives, the flow-thermal coupling state inside the furnace has evolved along the expected direction, laying a better initial operating condition for subsequent sensor data acquisition and a new round of prediction optimization.

[0084] In summary, the method for predicting and optimizing the parameters of magnetic ring heat treatment based on the embodiments of this application has been clarified. This method addresses the technical problem that the coupling interference between the circulating fan disturbance and the nitrogen micro-pressure field during the magnetic ring heat treatment process leads to drastic fluctuations in the local convective heat transfer coefficient, which in turn causes temperature field runaway. This solution transforms the traditional single-variable temperature feedback control paradigm into a closed-loop predictive optimization paradigm of "flow field perception - heat transfer reconstruction - temperature field prediction - multi-variable collaborative optimization". First, aerodynamic fingerprints characterizing the turbulent state within the furnace are extracted from nitrogen pressure pulsation signals and fan current perturbation signals, overcoming the limitation of existing schemes that cannot sense flow field disturbances. Then, a physically embedded deep neural network is used to invert the aerodynamic features into spatially distributed dynamic convective heat transfer coefficients, establishing a quantitative mapping between flow field disturbances and heat exchange. Based on this, the transient temperature field of the magnetic ring stack is tomographically predicted using the unsteady porous medium heat conduction equation, and thermal shock risk is simultaneously assessed. Finally, with temperature field tracking deviation as the target and thermal shock risk as the penalty constraint, a nonlinear model predictive controller performs coordinated rolling optimization and feedforward distribution of fan speed, nitrogen filling valve opening, and heating power. This breaks down the information barrier between flow field perception and temperature control, enabling the control system to proactively compensate for turbulent disturbances before they deteriorate into temperature deviations, eliminating the vicious cycle caused by flow-thermal coupling instability.

[0085] Figure 4 This is a block diagram of a magnetic ring heat treatment parameter prediction and optimization device according to an embodiment of this application. Figure 4 As shown, the magnetic ring heat treatment parameter prediction and optimization device 100 according to an embodiment of this application includes: a sensor data decomposition and filtering module 110, used to perform frequency domain decomposition and time-series filtering feature extraction on the acquired multi-source sensor data stream to obtain an aerodynamic feature fingerprint matrix and a calibrated temperature sequence vector; a heat transfer coefficient inversion module 120, used to perform fluid-solid interface heat transfer coefficient inversion and reconstruction on the aerodynamic feature fingerprint matrix based on a physically embedded deep neural network to obtain a dynamic convection heat transfer coefficient mapping matrix; and a magnetic ring temperature prediction module 130, used to analyze the calibrated temperature sequence vector and the dynamic convection heat transfer coefficient... The mapping matrix is ​​used to predict the three-dimensional thermal shock risk and temperature tomography of the magnetic ring stack under unsteady flow field disturbance to obtain the transient temperature field tensor and local thermal shock risk assessment matrix of the magnetic ring; the collaborative rolling optimization module 140 is used to take the transient temperature field tensor of the magnetic ring as the tracking target and configure the local thermal shock risk assessment matrix as the penalty constraint term, and perform multivariate collaborative rolling optimization on it through the nonlinear model predictive controller to obtain the collaborative execution instruction set; the collaborative execution module 150 is used to perform feedforward collaborative execution of the variable frequency speed of the circulating fan, the opening degree of the nitrogen filling valve and the electric heating power based on the collaborative execution instruction set.

[0086] Here, those skilled in the art will understand that the specific operations of each step in the above-described magnetic ring heat treatment parameter prediction and optimization device have been referenced above. Figures 1 to 3The method for predicting and optimizing the heat treatment parameters of magnetic rings has been described in detail, and therefore, its repeated description will be omitted.

[0087] According to another aspect of this application, a storage medium is also provided, which stores computer instructions for causing a computer to perform the methods described above. That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

Claims

1. A method for predicting and optimizing heat treatment parameters of a magnetic ring, characterized in that, include: Step 1: Perform frequency domain decomposition and time-series filtering feature extraction on the acquired multi-source sensor data stream to obtain the aerodynamic feature fingerprint matrix and the calibrated temperature sequence vector. Step 2: Based on the physical embedded deep neural network, the aerodynamic feature fingerprint matrix is ​​inverted and reconstructed to obtain the dynamic convection heat transfer coefficient mapping matrix. Step 3: Perform three-dimensional thermal shock risk and temperature tomography prediction of the magnetic ring stack under unsteady flow field disturbance on the calibrated temperature sequence vector and dynamic convection heat transfer coefficient mapping matrix to obtain the transient temperature field tensor and local thermal shock risk assessment matrix of the magnetic ring. Step 4: The transient temperature field tensor of the magnetic ring is used as the tracking target and the local thermal shock risk assessment matrix is ​​configured as the penalty constraint. The multivariable collaborative rolling optimization is performed on it by the nonlinear model predictive controller to obtain the collaborative execution instruction set. Step 5: Based on the collaborative execution instruction set, feedforward collaborative execution is performed on the variable frequency speed of the circulating fan, the opening degree of the nitrogen filling valve, and the electric heating power.

2. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 1, characterized in that, The multi-source sensor data stream includes high-frequency pulsating signals of nitrogen micro-positive pressure, micro-perturbation signals of circulating fan motor current, and multi-zone furnace temperature measurement sequences.

3. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 2, characterized in that, Step one includes: Frequency domain energy spectrum decomposition was performed on the high-frequency pulsating signal of nitrogen micro-positive pressure to obtain the eddy current energy feature vector; High-frequency jitter feature sequence is obtained by stripping high-frequency detail coefficients from the micro-perturbation signal of the circulating fan motor current through multi-scale discrete wavelet transform. Multimodal fusion compression is performed on the eddy current energy feature vector, high-frequency jitter feature sequence, and hardware execution log to obtain the aerodynamic feature fingerprint matrix; Based on the aerodynamic feature fingerprint matrix, an adaptive sliding filter is applied to the temperature measurement sequences of the multi-zone furnace to obtain the calibrated temperature sequence vector.

4. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 1, characterized in that, Step two includes: Based on a high-dimensional manifold encoder in a physically embedded deep neural network, a nonlinear dimension reduction projection is performed on the aerodynamic feature fingerprint matrix to obtain the latent space flow field feature tensor. The microscopic wind speed profile tensor is obtained by scaling and matching the feature tensor of the latent space flow field with the coordinates of the three-dimensional preset grid and performing multi-layer sensing decoding. The convective heat transfer coefficient is solved and spatially reconstructed point by point on the micro wind speed profile tensor to obtain the dynamic convective heat transfer coefficient mapping matrix.

5. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 1, characterized in that, Step three includes: The calibrated temperature sequence vector is used as the far-field temperature boundary condition and the dynamic convection heat transfer coefficient mapping matrix is ​​used as the fluid-solid interface source term. The time-domain numerical derivation and solution of the partial differential equation of heat conduction in unsteady porous media are performed to obtain the transient temperature field tensor of the magnetic ring. Based on the elastic Young's modulus and the coefficient of thermal expansion, the transient temperature field tensor of the magnetic ring is subjected to temperature space tomography and local thermal distortion stress stripping to obtain the multidimensional structural tensile stress matrix. Based on the material temperature-dependent yield strength threshold, a nonlinear thermal shock risk normalization mapping based on the yield critical limit is performed on the multidimensional structural tensile stress matrix to obtain the local thermal shock risk assessment matrix.

6. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 1, characterized in that, Step four includes: The weighted sum of squares of the deviations between the transient temperature field tensor of the magnetic ring and the ideal phase transition tracking curve is used as the tracking cost term, and the local thermal shock risk assessment matrix is ​​configured as an exponential impedance penalty constraint term to construct a constrained prediction cost function equation. The constrained prediction cost function equation is solved by second-order Taylor expansion and rolling time-domain convexity iteration to obtain the full-scale optimal control sequence. Based on the rolling implementation strategy, the optimal control sequence at all scales is first stripped into time domain slices and encoded with industrial communication protocols to obtain a collaborative execution instruction set.

7. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 1, characterized in that, Step five includes: Polymorphic communication protocol parsing and actuation signal decoupling are performed on the cooperative execution instruction set to obtain a three-dimensional linkage actuation signal vector; The three-dimensional linkage actuation signal vector is sent down to the circulating fan frequency converter, servo nitrogen charging valve and heating thyristor to perform physical actions.

8. The method for predicting and optimizing heat treatment parameters of a magnetic ring according to claim 4, characterized in that, The dynamic convective heat transfer coefficient mapping matrix is ​​obtained by performing grid-by-grid point dynamic convective heat transfer coefficient calculation and spatial reconstruction on the micro wind speed profile tensor. This includes: performing grid-by-grid point dynamic convective heat transfer coefficient calculation on the micro wind speed profile tensor based on micro-pressure slip flow correction and shear strain rate enhancement to obtain the dynamic convective heat transfer coefficient mapping matrix.

9. A device for predicting and optimizing heat treatment parameters of a magnetic ring, characterized in that, include: The sensor data decomposition and filtering module is used to perform frequency domain decomposition and time-series filtering feature extraction on the acquired multi-source sensor data stream to obtain the aerodynamic feature fingerprint matrix and the calibrated temperature sequence vector. The heat transfer coefficient inversion module is used to reconstruct the fluid-solid interface heat transfer coefficient based on the aerodynamic feature fingerprint matrix using a physically embedded deep neural network to obtain the dynamic convection heat transfer coefficient mapping matrix. The magnetic ring temperature prediction module is used to perform three-dimensional thermal shock risk and temperature tomography prediction of the magnetic ring stack under unsteady flow field disturbance on the calibrated temperature sequence vector and dynamic convection heat transfer coefficient mapping matrix to obtain the transient temperature field tensor and local thermal shock risk assessment matrix of the magnetic ring. The collaborative rolling optimization module is used to take the transient temperature field tensor of the magnetic ring as the tracking target and configure the local thermal shock risk assessment matrix as the penalty constraint. It uses a nonlinear model predictive controller to perform multivariable collaborative rolling optimization to obtain a collaborative execution instruction set. The collaborative execution module is used to feedforward and collaboratively issue commands to the variable frequency speed of the circulating fan, the opening degree of the nitrogen filling valve, and the power of the electric heating based on the collaborative execution instruction set.

10. A storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method as described in any one of claims 1 to 8.

Images