A laser heating temperature measurement calibration system

Abstract

The present application relates to the technical field of semiconductor laser heating and infrared temperature measurement calibration, and particularly relates to a laser heating temperature measurement calibration system, which comprises a device and a control device, and synchronously extracts infrared radiation original data, real temperature data and environmental space pose data through a cooperative acquisition module. A model evolution module embeds a thermal radiation distance attenuation constraint equation and an angle reflectivity penalty function in a physical information neural network hidden layer to generate a converged model and a confidence boundary tensor. When the confidence is lower than a preset physical effective threshold, an active optimization module drives an infrared temperature measurement element to move to a target position to avoid a measurement blind area. A precise calibration module extracts dynamic compensation weights in combination with updated position data, eliminates errors from the infrared radiation original data, and outputs a calibrated temperature. A heating control module calculates an energy deviation according to the calibrated temperature, regulates and controls the power of a laser heating device, and realizes high-precision closed-loop constant temperature control.

Description

Technical Field

[0001] This invention relates to the field of semiconductor laser heating and infrared temperature measurement and calibration technology, and in particular to a laser heating temperature measurement and calibration system. Background Technology

[0002] Laser heating requires precise temperature control. Infrared thermometry, as a non-contact temperature measurement method, is unaffected by laser heating and is widely used in laser heating applications. However, the accuracy of infrared thermometry is influenced by a combination of external factors, including measurement distance, measurement angle, the material and surface condition of the object being measured, and ambient temperature. This can cause the measured temperature to deviate from the actual temperature, directly affecting the laser heating effect. To ensure measurement accuracy and heating effect, infrared thermometry must be calibrated.

[0003] Existing linear calibration methods struggle to accurately handle complex nonlinear errors involving multiple variables. Artificial neural networks, however, possess powerful nonlinear data fitting capabilities, leading researchers to introduce machine learning algorithms to construct temperature compensation calibration models. In practice, contact temperature measurement methods such as thermistors or thermocouples are used to acquire real-world temperature data, which is then used as training labels for the neural network model. Simultaneously, infrared temperature measurement data from an infrared thermometer and environmental parameters serve as input features for the model. The trained neural network model can calibrate the emissivity and compensation values ​​of infrared temperature measurements, bringing the data to the true temperature. Applying deep learning methods to infrared temperature calibration systems effectively addresses the inaccuracies of non-contact temperature measurement. By using calibrated infrared temperature data for temperature control, the object being heated can be precisely heated to its true temperature. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a laser heating temperature measurement calibration system. This invention solves the technical problem that non-contact temperature measurement deviates from the actual temperature due to the influence of measurement distance, measurement angle, and ambient temperature on infrared temperature measurement.

[0005] To solve the above-mentioned technical problems, the specific contents of the present invention are as follows: This invention provides a laser heating temperature measurement and calibration system, including an equipment device and a control device that establishes a communication connection with the equipment device. The control device is used to control the equipment device. The control device includes a collaborative acquisition module, a model evolution module, an active optimization module, a precision calibration module, and a heating control module. Collaborative acquisition module: Connects to the device to synchronously extract raw infrared radiation data, real temperature data, and environmental spatial pose data. It sends the real temperature data to the model evolution module and the raw infrared radiation data and environmental spatial pose data to the model evolution module and the precision calibration module. Model Evolution Module: Receives real temperature data, raw infrared radiation data, and environmental spatial pose data. Inputs the raw infrared radiation data and environmental spatial pose data into a preset physical information neural network model for iterative training to generate a converged neural network model and a confidence boundary tensor. Sends the converged neural network model to the precision calibration module and the confidence boundary tensor to the active optimization module. Active optimization module: Receives the confidence boundary tensor, and when the confidence boundary tensor is lower than the preset physical effective threshold, generates a kinematic command and sends it to the position adjustment mechanism to drive the infrared temperature measuring element to move to the target spatial position. This triggers the collaborative acquisition module to extract and update the position data at the target spatial position and send it to the precision calibration module. Precision calibration module: Receives the converged neural network model, raw infrared radiation data, and updated position data. Inputs the updated position data into the converged neural network model to extract dynamic compensation weights. Uses the dynamic compensation weights to eliminate errors in the raw infrared radiation data and outputs calibration temperature data. Sends the calibration temperature data to the heating control module. Heating control module: Receives calibration temperature data, calculates energy deviation margin based on calibration temperature data, generates power modulation command and sends it to laser heating equipment to control output energy.

[0006] Furthermore, the laser heating temperature measurement and calibration system of the present invention includes a constant temperature heating platform for carrying the object to be measured, a contact temperature measuring element, a position adjustment mechanism, an infrared temperature measuring element, an ambient temperature sensing element, and a laser heating device. The constant temperature heating platform carries the object being tested; the contact temperature measuring element is attached to the surface of the object being tested; the position adjustment mechanism is arranged in the space above and to the side of the object being tested; the infrared temperature measuring element is installed at the end of the position adjustment mechanism; the ambient temperature sensing element is arranged in the space around the optical path between the infrared temperature measuring element and the object being tested; the laser emitting end of the laser heating equipment is aligned with the surface of the object being tested; the environmental spatial pose data includes environmental thermal disturbance data, temperature measurement distance data, and temperature measurement angle data.

[0007] Furthermore, the laser heating temperature measurement and calibration system of the present invention includes a collaborative acquisition module comprising a clock synchronization unit, an analog-to-digital conversion unit, and a protocol parsing unit. The clock synchronization unit generates trigger pulses according to a set period and sends them to the analog-to-digital conversion unit; the analog-to-digital conversion unit receives the trigger pulses, synchronously reads the analog electrical signals of the infrared temperature measuring element, the contact temperature measuring element, and the ambient temperature sensing element, converts the analog electrical signals into digital signals, and sends them to the protocol parsing unit; The protocol parsing unit receives digital signals, reads the temperature measurement distance data and temperature measurement angle data fed back by the position adjustment mechanism, and adds a common time sequence timestamp tag to the digital signals, temperature measurement distance data and temperature measurement angle data to generate multi-dimensional spatiotemporal synchronization sequence data. The multi-dimensional spatiotemporal synchronization sequence data includes raw infrared radiation data, real temperature data and environmental spatial pose data. The protocol parsing unit sends the multi-dimensional spatiotemporal synchronization sequence data to the model evolution module and the precision calibration module.

[0008] Furthermore, in the laser heating temperature measurement and calibration system of the present invention, the model evolution module includes a feature fusion unit and an equation embedding unit; The feature fusion unit receives multi-dimensional spatiotemporal synchronization sequence data, extracts the real temperature data from the multi-dimensional spatiotemporal synchronization sequence data as supervision labels, and extracts the original infrared radiation data, environmental thermal disturbance data, temperature measurement distance data and temperature measurement angle data from the multi-dimensional spatiotemporal synchronization sequence data to construct a multimodal input tensor and send it to the equation embedding unit. The equation embedding unit receives a multimodal input tensor, embeds the thermal radiation distance attenuation constraint equation and the angular reflectivity penalty function into the hidden layer of the physical information neural network model, and outputs a fused feature tensor.

[0009] Furthermore, in the laser heating temperature measurement and calibration system of the present invention, the model evolution module further includes a network training unit and a confidence evaluation unit; The network training unit receives the fused feature tensor output by the equation embedding unit, inputs the fused feature tensor into the physical information neural network model for forward propagation to obtain the output result, calculates the loss function value between the output result and the supervision label, and uses the backpropagation algorithm to update the model weights to generate the converged neural network model. The confidence assessment unit extracts the information entropy change data of the fused feature tensor under physical constraints, calculates the information entropy change data, and outputs the confidence boundary tensor.

[0010] Furthermore, the laser heating temperature measurement and calibration system of the present invention includes an active optimization module comprising a threshold determination unit and a spatial gradient calculation unit. The threshold determination unit receives the confidence boundary tensor and compares it with the preset physical effective threshold. When the confidence boundary tensor is lower than the preset physical effective threshold, the threshold determination unit outputs an optimization trigger signal to the spatial gradient calculation unit. The spatial gradient calculation unit receives the optimization trigger signal and starts the confidence maximization optimization algorithm to calculate the spatial gradient compensation vector pointing to the high confidence region.

[0011] Furthermore, the laser heating temperature measurement and calibration system of the present invention further includes an instruction generation unit in its active optimization module; The instruction generation unit receives the spatial gradient compensation vector, converts the spatial gradient compensation vector into position adjustment instructions and pitch and yaw angle adjustment instructions adapted to the three-dimensional coordinate system as kinematic instructions, and sends them to the position adjustment mechanism. The position adjustment mechanism receives position adjustment commands and pitch and yaw angle adjustment commands to move the physical position and orientation of the infrared temperature measuring element.

[0012] Furthermore, in the laser heating temperature measurement and calibration system of the present invention, the precision calibration module includes a status monitoring unit and a nonlinear mapping unit; The status monitoring unit receives multi-dimensional spatiotemporal synchronization sequence data containing updated position data in real time, determines that the current data pose is within a preset high-confidence physical range, generates a target verification signal, and sends the multi-dimensional spatiotemporal synchronization sequence data with the target verification signal to the nonlinear mapping unit. The nonlinear mapping unit receives multi-dimensional spatiotemporal synchronization sequence data with target verification signals, loads a synchronously updated converged neural network model, and extracts output feature values ​​by inputting the multi-dimensional spatiotemporal synchronization sequence data with target verification signals into the converged neural network model.

[0013] Furthermore, in the laser heating temperature measurement and calibration system of the present invention, the precision calibration module also includes a compensation output unit; The compensation output unit receives the output feature value, decodes the output feature value to obtain the dynamic compensation weight, which includes the dynamic emissivity compensation weight and the temperature compensation value. The compensation output unit uses the dynamic emissivity compensation weight and temperature compensation value in the dynamic compensation weight to perform truncation error elimination on the raw infrared radiation data and outputs the calibration temperature data.

[0014] Furthermore, the laser heating temperature measurement and calibration system of the present invention is characterized in that the heating control module includes a deviation calculation unit, a strategy calculation unit, and a power modulation unit; The deviation calculation unit receives calibration temperature data and preset target heating temperature data, performs difference calculation between calibration temperature data and preset target heating temperature data to obtain energy deviation margin, and sends the energy deviation margin to the strategy calculation unit. The strategy calculation unit receives the energy deviation margin, substitutes the energy deviation margin into the proportional-integral-derivative control algorithm matrix to perform numerical calculation to obtain the laser output power adjustment amount, and sends the laser output power adjustment amount to the power modulation unit. The power modulation unit receives the laser output power adjustment amount and generates a pulse width modulation control signal as a power modulation command, which is then sent to the laser heating equipment to adjust the output energy.

[0015] Beneficial effects of this invention: This invention provides a laser heating temperature measurement and calibration system that overcomes the shortcomings of traditional linear calibration methods in handling complex nonlinear errors. A collaborative acquisition module synchronously extracts raw infrared radiation data, real temperature data, and environmental spatial pose data, providing a unified multimodal data source for underlying computation. A model evolution module embeds thermal radiation distance attenuation constraint equations and angular reflectivity penalty functions into the hidden layers of the physical information neural network model. The introduction of physical and mechanical constraints enables the neural network to accurately fit the nonlinear thermal radiation relationship and output a confidence boundary tensor. An active optimization module actively drives the infrared temperature measurement element to a target spatial position with a high signal-to-noise ratio when the confidence boundary tensor is determined to be below a preset physical effective threshold. This hardware-software linkage optimization mechanism avoids measurement blind zone interference caused by strong attenuation and high reflectivity from the physical acquisition source. A precise calibration module inputs the extracted updated position data into the converged neural network model to extract dynamic compensation weights, which are then used to perform truncation error elimination on the raw infrared radiation data. The error elimination calculation logic effectively solves the problem of non-contact temperature measurement distortion caused by a combination of external factors, including measurement distance, measurement angle, and ambient temperature, thereby outputting high-precision calibration temperature data. The heating control module calculates the energy deviation margin based on the calibration temperature data, generates a power modulation command based on the energy deviation margin, and sends it to the laser heating equipment, realizing high-precision closed-loop constant temperature control of the energy output to the heated object. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0017] Figure 1 This is a system architecture diagram of a laser heating temperature measurement and calibration system according to the present invention. Detailed Implementation

[0018] To make the technical solution of the present invention clearer, the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. To better understand the purpose of the present invention, the present invention will be described in further detail below.

[0019] Please see Figure 1The present invention provides a laser heating temperature measurement and calibration system, including an equipment device and a control device that establishes a communication connection with the equipment device. The control device is used to control the equipment device. The control device includes a collaborative acquisition module, a model evolution module, an active optimization module, a precision calibration module, and a heating control module. Collaborative acquisition module: Connects to the device to synchronously extract raw infrared radiation data, real temperature data, and environmental spatial pose data. It sends the real temperature data to the model evolution module and the raw infrared radiation data and environmental spatial pose data to the model evolution module and the precision calibration module. Model Evolution Module: Receives real temperature data, raw infrared radiation data, and environmental spatial pose data. Inputs the raw infrared radiation data and environmental spatial pose data into a preset physical information neural network model for iterative training to generate a converged neural network model and a confidence boundary tensor. Sends the converged neural network model to the precision calibration module and the confidence boundary tensor to the active optimization module. Active optimization module: Receives the confidence boundary tensor, and when the confidence boundary tensor is lower than the preset physical effective threshold, generates a kinematic command and sends it to the position adjustment mechanism to drive the infrared temperature measuring element to move to the target spatial position. This triggers the collaborative acquisition module to extract and update the position data at the target spatial position and send it to the precision calibration module. Precision calibration module: Receives the converged neural network model, raw infrared radiation data, and updated position data. Inputs the updated position data into the converged neural network model to extract dynamic compensation weights. Uses the dynamic compensation weights to eliminate errors in the raw infrared radiation data and outputs calibration temperature data. Sends the calibration temperature data to the heating control module. Heating control module: Receives calibration temperature data, calculates energy deviation margin based on calibration temperature data, generates power modulation command and sends it to laser heating equipment to control output energy.

[0020] This system achieves high-frequency data interaction between the control device and the equipment via a communication bus. The collaborative acquisition module utilizes a hardware triggering mechanism to synchronously extract multi-dimensional basic parameters from the physical environment. Raw infrared radiation data originates from the attenuation signal of thermal radiation energy on the surface of the heated object captured by the infrared thermography element. Actual temperature data is obtained by directly measuring the surface of the heated object using a contact thermography element via heat conduction. Environmental spatial pose data consists of environmental thermal disturbance data, temperature measurement distance data, and temperature measurement angle data, reflecting the physical boundary conditions during the infrared measurement process. The collaborative acquisition module performs time-series alignment of the extracted multi-dimensional basic parameters, providing the underlying data source for subsequent model calculations.

[0021] The collaborative acquisition module distributes the time-aligned multi-dimensional basic parameters in separate streams. Real temperature data, raw infrared radiation data, and environmental spatial pose data are sent in parallel to the model evolution module. The model evolution module uses real temperature data as a supervision label and raw infrared radiation data and environmental spatial pose data as multimodal input features to drive the pre-defined physical information neural network model through iterative training of forward and backward propagation. The hidden layers of the physical information neural network model embed thermal radiation distance attenuation constraint equations and angular reflectivity penalty functions, ensuring that the model evolution process follows thermodynamic physical laws. As the number of iterations increases, the weight parameters of the neural network gradually converge, generating a converged neural network model. The model evolution module synchronously calculates the confidence boundary tensor based on the information entropy change of the multimodal input features under physical constraints. The confidence boundary tensor characterizes the reliability of the current spatial physical pose measurement signal-to-noise ratio. The model evolution module sends the converged neural network model to the precision calibration module and the confidence boundary tensor to the active optimization module.

[0022] The active optimization module receives the confidence boundary tensor and compares it with a preset physical effective threshold. The preset physical effective threshold represents the minimum signal-to-noise ratio (SNR) requirement for the infrared thermometer to acquire a valid thermal radiation signal. When the confidence boundary tensor is determined to be lower than the preset physical effective threshold, it indicates that the current measurement angle or distance of the infrared thermometer is significantly affected by environmental factors, leading to distorted temperature data. In this case, the active optimization module initiates a confidence maximization optimization algorithm, calculating a spatial gradient compensation vector pointing towards a high SNR region based on the gradient descent principle. The active optimization module parses the spatial gradient compensation vector into kinematic commands to control the motor's movement and sends these commands to the position adjustment mechanism control terminal. Upon receiving the kinematic commands, the position adjustment mechanism executes physical movements, driving the mounted infrared thermometer to adjust its pitch, yaw, or linear displacement in three-dimensional space until it reaches a target spatial position with high confidence.

[0023] After the infrared temperature sensing element moves to the target spatial position, the change in physical pose triggers the collaborative acquisition module to re-execute the data extraction action. The collaborative acquisition module extracts updated position data containing the new pose information at the target spatial position and sends the updated position data to the precision calibration module. Simultaneously, the precision calibration module receives the updated position data, retrieves the converged neural network model from the model evolution module, and receives the raw infrared radiation data transmitted in real time from the collaborative acquisition module. The precision calibration module inputs the updated position data into the converged neural network model and extracts the dynamic compensation weights corresponding to the current target spatial position through high-dimensional nonlinear mapping. The dynamic compensation weights specifically include dynamic emissivity compensation weights and temperature compensation values. The precision calibration module uses the dynamic emissivity compensation weights to perform emissivity correction on the raw infrared radiation data and combines the temperature compensation values ​​to perform truncated error elimination on the raw infrared radiation data, outputting high-precision calibration temperature data. The precision calibration module then sends the calibration temperature data to the heating control module.

[0024] The heating control module receives calibration temperature data and simultaneously retrieves the system's preset target heating temperature data. It then calculates the difference between the calibration temperature data and the preset target heating temperature data to determine the instantaneous energy deviation margin of the heated object. This energy deviation margin is used as a feedback variable in the proportional-integral-derivative (PID) control algorithm matrix for numerical calculation, deriving the required adjustment amount of the laser output power. Based on this adjustment, the heating control module generates a pulse-width modulation (PWM) power modulation command and sends it to the driver of the laser heating device. The laser heating device then adjusts the laser output power in real time according to the power modulation command, achieving closed-loop control of the output energy.

[0025] In the laser thermal processing of semiconductor silicon wafers, the underlying hardware layout directly determines the physical acquisition status of multi-source data. A constant-temperature heating platform serves as the basic support surface, placing the object under test and providing a globally stable background heat source. Contact temperature sensing elements are tightly attached to the surface of the object, utilizing the thermoelectric effect of the material to directly measure the physical heat conduction in the contact area and output true temperature data. A position adjustment mechanism positioned above and to the side of the object changes the absolute coordinates of the execution end in three-dimensional space based on received control signals from the underlying layer, ensuring that the infrared temperature sensing element installed at the end is aligned with the object surface to receive non-contact thermal radiation energy. Simultaneously, environmental temperature sensing elements distributed around the optical path monitor environmental thermal disturbance data caused by the airflow. The laser emitting end of the laser heating equipment outputs a high-energy beam vertically downwards to irradiate the object surface. The environmental thermal disturbance data, combined with the temperature measurement distance and angle data fed back by the position adjustment mechanism, constitute environmental spatial pose data, fully mapping the physical attenuation boundary conditions in the non-contact temperature measurement link.

[0026] To avoid data timing misalignment caused by multimodal sensors in high-speed laser heating, a hardware-level synchronous timing mechanism is deployed within the collaborative acquisition module. The clock synchronization unit, relying on an internal high-frequency oscillator, continuously sends square wave trigger pulses to the analog-to-digital converter (ADC) at a set period. Upon receiving the rising edge of the trigger pulse, multiple sampling channels of the ADC synchronously latch the continuously changing analog electrical signals fed back by the infrared temperature sensing element, the contact temperature sensing element, and the ambient temperature sensing element. The internal quantization and encoding circuit then maps the discretized analog voltage amplitude into a discrete digital signal readable by the underlying digital register. The digital signal, after ADC processing, is sent to the protocol parsing unit via a parallel bus, and then enters the higher-level data unpacking and alignment processing stage.

[0027] The protocol parsing unit undertakes the underlying data processing tasks of heterogeneous data fusion and spatiotemporal alignment. It reads the encoder pulse values ​​fed back in real time from the underlying driver of the position adjustment mechanism via a serial communication interface, and uses kinematic forward solving to obtain the current temperature measurement distance and angle data. The protocol parsing unit extracts the controller system's master clock value to generate a unified timestamp tag. It then concatenates and encapsulates the digital signal containing thermal radiation intensity, temperature measurement distance data, and temperature measurement angle data according to the timestamp tag, forming a multi-dimensional spatiotemporal synchronization sequence data containing the time dimension. This multi-dimensional spatiotemporal synchronization sequence data fully encapsulates the raw infrared radiation data, the real temperature data representing the reference value, and the environmental spatial pose data characterizing external disturbances. It is then distributed in parallel to the model evolution module and the precision calibration module via the system's high-speed communication bus.

[0028] The feature fusion unit within the model evolution module reconstructs the tensor dimension of the input multi-dimensional spatiotemporal synchronization sequence data, extracting real temperature data with benchmark reference value as supervision labels according to the set data slicing rules. The feature fusion unit extracts raw infrared radiation data, environmental thermal disturbance data, temperature measurement distance data, and temperature measurement angle data, stacking and splicing feature channels in a memory matrix to form a multimodal input tensor containing amplitude features and spatial geometric features. This multimodal input tensor integrates scalar temperature values ​​and vector spatial coordinates into a unified mathematical computation space, providing the neural network with a low-level input vector containing complete physical boundary conditions, and pushing it into the equation embedding unit's internal computation graph.

[0029] To compensate for the fitting deficiencies of purely data-driven algorithms, the equation embedding unit introduces a physical constraint mechanism into the feedforward computation network of the neural network. This mechanism connects a thermal radiation distance attenuation constraint equation and an angle reflectivity penalty function in series within the hidden layer computation nodes of the physical information neural network model. The thermal radiation distance attenuation constraint equation characterizes the physical law that radiation energy decreases inversely with the square of the temperature measurement distance data. The angle reflectivity penalty function dynamically increases the network weight attenuation penalty term based on the angle at which the temperature measurement angle data deviates from the normal, suppressing erroneous induction of output features by the raw infrared radiation data at large angles. In practical applications of silicon wafer laser heating, when the infrared temperature measuring element moves to a very small grazing angle, the angle reflectivity penalty function generates a large negative gradient, causing the network to reduce the activation level of nonlinear mapping nodes. Finally, the equation embedding unit performs nonlinear transformation calculations with physical equation constraints, mapping the multimodal input tensor into a fused feature tensor that includes physical attenuation characteristics and internal data correlations. In the underlying architecture of the physical information neural network model, the network layers specifically include an input layer, at least three fully connected hidden layers, and an output layer. To clarify the intrinsic relationships between multi-source heterogeneous data, the input layer is configured with a linear normalization operator to eliminate the dimensional differences between the raw infrared radiation data (containing energy levels) and the environmental spatial pose data (containing geometric levels), mapping the values ​​of the multimodal input tensors to a preset normalization interval. Each fully connected hidden layer uses the hyperbolic tangent function as a nonlinear activation function to maintain the second-order continuous differentiability of the physical constraint partial derivatives during calculation. The intrinsic relationship between the input and output data of the physical information neural network model is as follows: the normalized temperature measurement distance data, temperature measurement angle data, and raw infrared radiation data are used as independent variables for spatial forward propagation; the real temperature data obtained by the contact temperature measurement element is used as a hard constraint supervision label; and the distance attenuation penalty feature value and angle reflectivity penalty term generated by the equation embedding unit are introduced as soft constraint regularization terms into the computational graph at the back end of the model. The constraint neural network prioritizes searching for weighted solutions that conform to thermodynamic physical laws within the solution space.

[0030] In the laser thermal processing of semiconductor silicon wafers, the physical information neural network model needs to map complex nonlinear thermal radiation relationships. The network training unit receives the fused feature tensor output by the equation embedding unit and inputs it into the multilayer perceptron architecture of the physical information neural network model for forward propagation. The forward propagation calculation maps the predicted temperature value through layer-by-layer nonlinear activation functions as the output. Simultaneously, the network training unit extracts real temperature data as supervision labels and calculates the mean squared error between the output and the supervision labels to obtain the loss function value. Guided by the goal of minimizing the loss function value, the network training unit iteratively adjusts the connection weights between hidden layer nodes using the gradient descent backpropagation algorithm. As the iteration cycle progresses, the loss function value gradually converges and stabilizes, allowing the network weights to reach an optimal state, generating a converged neural network model capable of handling nonlinear measurement errors. The specific training steps and parameter settings for the network training unit to perform model parameter iteration are as follows: Step 1, assign initial connection weights to each fully connected hidden layer node in the physical information neural network model using a uniform distribution initialization algorithm; Step 2, during the training period, extract multi-dimensional spatiotemporal synchronization sequence data according to a batch size of sixty-four samples and perform forward propagation matrix operations; Step 3, construct a joint optimization loss function, setting the mean square error between the predicted temperature value output by the forward propagation and the actual temperature data as the data-driven loss term and assigning it a basic weight coefficient of one, assigning the distance attenuation penalty feature value calculated by the thermal radiation distance attenuation constraint equation a weight coefficient of zero. The angle reflectivity penalty term calculated by the angle reflectivity penalty function is assigned a weight coefficient of 0.05. The three are algebraically summed to generate the global loss function value. Step four: The adaptive moment estimation optimization algorithm is called to calculate the target gradient. The initial learning rate parameter is set to 0.001. With the exponential decay learning rate scheduling strategy based on the iteration rounds, the weights of all hidden layer nodes are updated using the backpropagation mechanism. Step five: The above forward propagation and backpropagation operations are executed repeatedly until the global loss function value is lower than the set convergence accuracy threshold of 0.001 or the maximum iteration round of 10,000 is reached. The training action ends and the current network connection parameters are fixed. Finally, the converged neural network model is output.

[0031] In sync with the network weight iteration, the control device needs to assess the reliability of the current spatial measurement pose. The confidence assessment unit extracts the information entropy change data of the fused feature tensor under physical constraints. The information entropy change data characterizes the degree of dispersion and uncertainty of the input features in a high-dimensional spatial distribution mapping when constrained by the thermal radiation distance attenuation constraint equation and the angular reflectivity penalty function. When the silicon wafer surface is in a highly reflective state, scattered interference light will increase the uncertainty of the thermal radiation signal, causing a significant increase in the information entropy change data. The confidence assessment unit inversely maps the information entropy change data into a continuous probability distribution matrix, and outputs the confidence boundary tensor through numerical calculation. The confidence boundary tensor serves as a quantitative indicator characterizing the signal-to-noise ratio of the current spatial physical pose measurement, providing an objective mathematical basis for subsequent physical position adjustment decisions.

[0032] The active optimization module uses the confidence boundary tensor as the underlying input for logical judgments. The threshold determination unit receives the confidence boundary tensor and compares it numerically with a preset physical effective threshold. The preset physical effective threshold defines the minimum measurement signal-to-noise ratio limit necessary to maintain stable operation of the temperature control closed loop. When the comparison result indicates that the confidence boundary tensor is lower than the preset physical effective threshold, the threshold determination unit confirms that the current measurement pose of the infrared temperature measuring element is in a strong attenuation or high reflectivity physical blind zone, and then outputs an optimization trigger signal to the spatial gradient calculation unit. The spatial gradient calculation unit receives the optimization trigger signal and starts the confidence maximization optimization algorithm. The confidence maximization optimization algorithm establishes a local spatial derivative model with the confidence boundary tensor as the objective function. By solving the partial derivative of the confidence with respect to three-dimensional spatial coordinates, it searches for the gradient direction with the fastest increase in signal-to-noise ratio and calculates the spatial gradient compensation vector pointing to the high confidence region.

[0033] The instruction generation unit receives the spatial gradient compensation vector and performs parameter analysis on the vector based on the inverse kinematics model of the underlying hardware mechanism. Through a coordinate transformation matrix, the instruction generation unit converts the spatial gradient compensation vector into position adjustment instructions and pitch / yaw angle adjustment instructions adapted to the three-dimensional coordinate system. The position adjustment instruction defines the target coordinate values ​​for linear translation along orthogonal spatial coordinate axes, while the pitch / yaw angle adjustment instruction defines the target rotation angle parameters for attitude deflection around spatial rotation axes. The instruction generation unit packages the position adjustment instructions and pitch / yaw angle adjustment instructions into kinematic instructions and sends them to the position adjustment mechanism control terminal. The position adjustment mechanism receives the position adjustment instructions and pitch / yaw angle adjustment instructions, uses its internal servo drive module to control the mechanical transmission structure's movements, and executes multi-degree-of-freedom spatial linkages according to the instruction parameters, directly changing the physical pose of the infrared temperature sensing element installed at the end.

[0034] In the laser thermal processing of semiconductor silicon wafers, after the infrared temperature sensing element is physically repositioned, the status monitoring unit receives multi-dimensional spatiotemporal synchronization sequence data containing the updated position data in real time. The updated position data consists of absolute coordinate values ​​in a three-dimensional spatial coordinate system, reflecting the new physical pose after avoiding the high reflectivity angle of the silicon wafer surface. The status monitoring unit performs low-level data quality screening on the input multi-dimensional spatiotemporal synchronization sequence data, determining that the current data pose is within a preset high-confidence physical range based on the signal-to-noise ratio distribution model. Being within the preset high-confidence physical range means that the intensity of the thermal radiation signal acquired by the system is much greater than the ambient background shot noise, meeting the boundary requirements of the neural network model inference. The status monitoring unit appends a target verification signal to the header of the data packet and sends the multi-dimensional spatiotemporal synchronization sequence data with the target verification signal to the nonlinear mapping unit, establishing a secure data channel for subsequent nonlinear calculations.

[0035] The nonlinear mapping unit receives multi-dimensional spatiotemporal synchronization sequence data with target verification signals and synchronously loads the converged neural network model, which has completed weight parameter iteration, from system memory. The nonlinear mapping unit inputs the multi-dimensional spatiotemporal synchronization sequence data, containing amplitude and spatial dimensions, into the converged neural network model and uses the network's internal nonlinear activation matrix to perform layer-by-layer spatial transformation mapping on the multimodal input features. In the high-dimensional latent space, the neural network separates the feature components affected by temperature measurement distance attenuation and temperature measurement angle reflection interference, extracting output feature values ​​that characterize the absolute magnitude of the physical error. The output feature values ​​exist in the form of a high-dimensional vector, integrating the underlying spatial geometric constraint information and thermal radiation attenuation law captured by the multimodal sensor.

[0036] The compensation output unit receives the extracted output feature values ​​and uses its built-in decoding layer array to perform dimensionality reduction decoding operations on the output feature values, restoring the high-dimensional vector to dynamic compensation weights with actual physical meaning. The decoding layer array inside the compensation output unit contains a multilayer perceptron decoding network. The multilayer perceptron decoding network consists of a first fully connected layer, a nonlinear activation layer, and a second fully connected layer connected in series. The multilayer perceptron decoding network receives the output feature values ​​extracted by the nonlinear mapping unit. The first fully connected layer reduces the high-dimensional output feature values ​​to a two-dimensional feature vector. The nonlinear activation layer uses the Sigmoid function to establish a mapping relationship. The nonlinear activation layer nonlinearly maps the first dimension value of the two-dimensional feature vector to a standardized continuous interval from 0.1 to 1.0. The mapped values ​​within the standardized continuous interval are directly output as dynamic emissivity compensation weights. The second fully connected layer performs a linear scaling operation on the second dimension value of the two-dimensional feature vector. The second fully connected layer outputs the amplified values ​​as temperature compensation values. The multilayer perceptron decoding network constructs a complete mathematical calculation path from high-dimensional abstract output feature values ​​to specific physical compensation parameters.

[0037] The dynamic compensation weight is specifically decomposed into dynamic emissivity compensation weight and temperature compensation value in the data structure. The dynamic emissivity compensation weight reflects the ratio of the attenuation of the true emissivity of the silicon wafer surface to that of an ideal blackbody at a specific measurement angle, while the temperature compensation value quantifies the absolute loss of thermal radiation energy caused by the measurement distance. Based on the physical calibration equation, the compensation output unit uses the dynamic emissivity compensation weight to perform divisional scaling correction on the raw infrared radiation data and combines it with the temperature compensation value to perform additive compensation calculations. The compensation output unit uses a truncated error elimination method to remove abnormal fluctuation signals outside the error tolerance range and calculates and outputs calibration temperature data characterizing the true thermodynamic state of the silicon wafer.

[0038] The closed-loop control logic is activated upon receiving feedback signals. The deviation calculation unit within the heating control module receives real-time calibration temperature data and preset target heating temperature data. The preset target heating temperature represents the nominal thermodynamic temperature threshold required for the semiconductor silicon wafer to reach the ideal lattice annealing state. The deviation calculation unit subtracts the real-time feedback calibration temperature data from the preset target heating temperature data to calculate the energy deviation margin characterizing the current system's thermal surplus / deficit state. The strategy calculation unit receives the energy deviation margin and substitutes the continuously changing energy deviation margin into the proportional-integral-derivative (PID) control algorithm matrix for numerical calculation. The PID matrix comprehensively evaluates the system's thermal inertia response trend based on the current deviation amplitude, historical cumulative deviation, and deviation change rate, calculating the laser output power adjustment amount used to eliminate dynamic thermal deviation.

[0039] The power modulation unit receives the laser output power adjustment and generates a pulse width modulation (PWM) control signal with the corresponding pulse duty cycle based on the underlying hardware electrical response characteristics. This PWM control signal, acting as a power modulation command to change the laser excitation current, is sent to the drive end of the laser heating equipment via a high-speed industrial fieldbus. The laser heating equipment responds to the power modulation command by rapidly adjusting the drive current of the semiconductor pump source, thereby changing the laser output power in real time and achieving microsecond-level closed-loop constant temperature control of the energy output to the heated area of ​​the silicon wafer.

[0040] When the equation embedding unit performs tensor feature processing in the hidden layer, it injects thermal radiation physical and mechanical constraints into the physical information neural network model. The thermal radiation distance attenuation constraint equation is used to constrain the physical attenuation effect of temperature measurement distance data in multi-dimensional spatiotemporal synchronization sequence data on the original infrared radiation data. The specific calculation formula of the thermal radiation distance attenuation constraint equation is expressed as follows:

[0041] In the formula, This represents the distance attenuation penalty eigenvalue calculated from the thermal radiation distance attenuation constraint equation. This represents the distance constraint weight coefficient set by the system. This represents the raw infrared radiation data extracted by the collaborative acquisition module. This represents the reference radiation constant used for thermodynamic calibration of the material being tested. This represents the temperature measurement distance data included in the multi-dimensional spatiotemporal synchronous sequence data extracted by the collaborative acquisition module. The equation embedding unit synchronously processes the temperature measurement angle data, applying an angle reflectivity penalty function to impose an exponential weight attenuation on the raw infrared radiation data over a large angle range to eliminate high-intensity specular reflection interference caused by large incident angles. The specific calculation formula for the angle reflectivity penalty function is expressed as follows:

[0042] In the formula, This represents the angle reflectivity penalty term calculated using the angle reflectivity penalty function. This represents the pre-set angle penalty attenuation coefficient within the system. This refers to the temperature measurement angle data included in the multi-dimensional spatiotemporal synchronization sequence data extracted by the collaborative acquisition module. An angle sensitivity index that indicates the steepness of the decay curve.

[0043] The network training unit receives a fused feature tensor containing distance attenuation penalty features and angular reflectivity penalty terms. The network training unit inputs this fused feature tensor into the physical information neural network model for forward propagation calculation, yielding the output for the prediction phase. The network training unit calculates the loss function value between the output and the supervision label; the specific formula for calculating the loss function value is as follows:

[0044] In the formula, This represents the loss function value calculated by the network training unit. This represents the total number of samples in the multi-dimensional spatiotemporal synchronization sequence data within the current training batch. This represents the sequence number of the sample data within the current input matrix. This represents the output result obtained from the forward propagation calculation of the physical information neural network model. This indicates that the contact temperature sensing element acquires and extracts the real temperature data as a supervision label by the model evolution module. This represents the distance attenuation penalty eigenvalue calculated from the thermal radiation distance attenuation constraint equation. This represents the angle reflectivity penalty term calculated using the angle reflectivity penalty function. The confidence assessment unit extracts the information entropy change data of the fused feature tensor under physical constraints and calculates and outputs the confidence boundary tensor, which characterizes the signal-to-noise ratio of spatial physical pose measurement. The specific calculation formula for the confidence boundary tensor is expressed as follows:

[0045] In the formula, This represents the confidence boundary tensor calculated and output by the confidence assessment unit. The exponentiation operator represents the exponentiation base based on the natural constant. This represents the system's preset signal-to-noise ratio smoothing scaling factor. This represents the information entropy change data of the fusion feature tensor extracted by the confidence assessment unit under physical constraints.

[0046] Upon receiving the optimization trigger signal, the spatial gradient calculation unit within the active optimization module initiates the confidence maximization optimization algorithm to calculate the spatial gradient compensation vector. The specific formula for calculating the spatial gradient compensation vector is as follows:

[0047] In the formula, This represents the spatial gradient compensation vector calculated by the spatial gradient calculation unit. The partial differential variable representing the confidence boundary tensor calculated by the confidence assessment unit. Represents the partial differential spatial variable along the horizontal X-axis in a three-dimensional coordinate system. Represents the partial differential spatial variables along the vertical Y-axis in a three-dimensional coordinate system. This represents the partial differential spatial variable along the vertical Z-axis in a three-dimensional coordinate system. The compensation output unit within the precision calibration module utilizes dynamic emissivity compensation weights and temperature compensation values ​​to numerically reconstruct and eliminate truncation errors from the raw infrared radiation data, outputting the calibration temperature data. The specific calculation formula for the calibration temperature data is as follows:

[0048] In the formula, This represents the calibration temperature data calculated and output by the precision calibration module. This represents the radiation energy to temperature conversion coefficient established during system calibration. This represents the raw infrared radiation data extracted by the collaborative acquisition module. This represents the dynamic emissivity compensation weights of the nonlinear mapping unit, derived from the decoding of the converged neural network model. This represents the temperature compensation value obtained by the nonlinear mapping unit based on the converged neural network model.

[0049] The deviation calculation unit inside the heating control module performs algebraic subtraction to obtain the energy deviation margin. The strategy calculation unit substitutes the energy deviation margin into the proportional-integral-derivative (PID) control algorithm matrix for floating-point numerical calculation to obtain the laser output power adjustment amount. The specific calculation formula for the laser output power adjustment amount is expressed as follows:

[0050] In the formula, This represents the laser output power adjustment calculated by the strategy operation unit. This represents the proportional gain coefficient in the proportional-integral-derivative (PID) control algorithm matrix. This represents the calibration temperature data calculated and output by the precision calibration module. This indicates the preset target heating temperature data synchronously acquired by the heating control module. This represents the integral gain coefficient in the proportional-integral-derivative (PID) control algorithm matrix. This represents the cumulative temperature deviation calculated by integrating the temperature deviation from the initial running time to the current time point. This represents the differential gain coefficient in the proportional-integral-derivative (PID) control algorithm matrix. This represents the rate of change of temperature deviation calculated by differentiation over a continuous time dimension.

[0051] To verify the data processing path and technical feasibility of the underlying multi-level mathematical calculation model in industrial processing equipment, system technicians performed laser annealing processing tests on a semiconductor silicon carbide wafer as the test object. The collaborative acquisition module used internal sensors to extract temperature measurement distance data of 150 mm, temperature measurement angle data of 60 degrees, and raw infrared radiation data of 3200 watts per square meter. The system's underlying control program set the reference radiation constant to 72000 and the distance constraint weight coefficient to 0.10. The equation embedding unit read the temperature measurement distance data and raw infrared radiation data, substituted them into the thermal radiation distance attenuation constraint equation, and calculated that the distance attenuation penalty characteristic value was 0. The control program set the angle penalty attenuation coefficient to 0.50 and the angle sensitivity index to 2. The cosine physical constant corresponding to the 60-degree temperature measurement angle data was 0.50. The equation embedding unit read the relevant values, substituted them into the angle reflectivity penalty function, and calculated that the angle reflectivity penalty term was 0.125. The network training unit uses forward and backward propagation operations to process the fused feature tensor containing relevant penalty terms, and iteratively outputs the converged neural network model.

[0052] The nonlinear mapping unit then reads the updated position data and inputs it into the converged neural network model. After high-dimensional tensor decoding, it outputs a dynamic emissivity compensation weight of 0.82 and a temperature compensation value of 15 degrees Celsius. The system's calibrated radiation energy to temperature conversion coefficient is specifically 0.25. The compensation output unit reads the original infrared radiation data, dynamic emissivity compensation weight, temperature compensation value, and radiation energy to temperature conversion coefficient and substitutes them into the calibration calculation formula. The compensation output unit multiplies 0.25 by 3200 to get 800, divides 800 by 0.82 to get 975.6 degrees Celsius, adds the 15-degree Celsius temperature compensation value to 975.6 degrees Celsius, and finally outputs a calibration temperature of 990.6 degrees Celsius. The preset target heating temperature data obtained by the heating control module is specifically 1000 degrees Celsius. The deviation calculation unit subtracts the preset target heating temperature data of 1000 degrees Celsius from the calibration temperature data of 990.6 degrees Celsius to calculate the energy deviation margin, which is -9.4 degrees Celsius. The proportional gain coefficient set in the strategy operation unit is specifically 5.0. Within the current millisecond-level sampling control cycle, the strategy operation unit directly uses the pure proportional control branch for amplification calculation, multiplying the energy deviation margin of -9.4 degrees Celsius by the proportional gain coefficient to obtain a specific absolute increase of 47 watts in laser output power adjustment. The power modulation unit generates a corresponding pulse width modulation control signal as a power modulation command based on the absolute power requirement of 47 watts. After the power modulation command is sent to the laser heating equipment, it precisely increases the thermal output energy, completing the closed-loop calibration and temperature control data path.

[0053] The system's underlying memory pre-programs numerical physical thresholds to support the operation of the logic judgment module. The specific value of the preset effective physical threshold is set to 0.75. A confidence boundary tensor value of 0.75 represents the ratio of the thermal radiation signal intensity captured by the infrared thermometer to the ambient background noise energy reaching the lower limit of the physical measurement signal-to-noise ratio. The specific value of the convergence accuracy threshold is set to 0.001. A value of 0.001 represents the global loss function value calculated by the network training unit reaching the allowable limit of small errors. The preset high-confidence physical interval is specifically defined as a closed interval where the confidence boundary tensor value is between 0.85 and 1.00. A closed interval where the confidence boundary tensor value is between 0.85 and 1.00 represents the infrared thermometer being in the optimal spatial measurement pose with low reflection and low attenuation. The numerical physical thresholds provide a precise mathematical basis for the system's underlying data flow.

[0054] Embodiment 1 of this invention: In the annealing process of silicon carbide wafers for semiconductors, the operator places the silicon carbide wafer to be processed on a constant-temperature heating platform. The constant-temperature heating platform activates its working mechanism to provide an initial base heat of 400 degrees Celsius. The contact temperature sensing element attached to the surface of the silicon carbide wafer uses a K-type thermocouple, utilizing the thermoelectric effect of the metal conductor to directly obtain the actual physical temperature value of the physical contact area as the true temperature data. The clock synchronization unit inside the collaborative acquisition module relies on an independent hardware crystal oscillator to continuously send square wave trigger pulses to the analog-to-digital conversion unit at a fixed operating frequency of 5000 Hz. At the moment the analog-to-digital conversion unit captures the rising edge of the square wave trigger pulse, it opens multiple synchronous sampling channels, latches the raw infrared radiation data collected by the infrared temperature sensing element and the environmental thermal disturbance data acquired by the environmental temperature sensing element. The protocol parsing unit merges the discretized digital signal with the 150 mm temperature measurement distance data and 30 degree temperature measurement angle data fed back by the position adjustment mechanism, and adds a unified timestamp tag from the system's underlying layer to generate multi-dimensional spatiotemporal synchronization sequence data. Multi-dimensional spatiotemporal synchronous sequence data is fed into the internal memory matrix of the model evolution module. The feature fusion unit extracts the real temperature data measured by the contact temperature sensing element from the multi-dimensional spatiotemporal synchronous sequence data as supervision labels. The hidden layer computing nodes of the physical information neural network model call the thermal radiation distance attenuation constraint equation to calculate the spatial attenuation of radiation energy at a distance of 150 mm. At the same time, based on the 30-degree temperature measurement angle data, an angle reflectivity penalty function is introduced to increase the network connection weight penalty term. The network training unit uses the constructed multimodal input tensor to perform forward and backward propagation matrix operations. After the operation terminates, it outputs the converged neural network model and a confidence boundary tensor with a quantization value of 0.92.

[0055] Embodiment 2 of the present invention: During local laser heat treatment of a semiconductor silicon wafer with a high-reflectivity metal film coating, the high-reflectivity metal film induces a strong specular reflection physical effect. The raw infrared radiation data collected by the infrared thermometer shows significant high-frequency amplitude fluctuations, and the ambient temperature sensing element simultaneously detects thermal disturbance airflow at a velocity of 3 meters per second caused by the injection of process cooling gas in the local space. The confidence assessment unit within the model evolution module extracts the information entropy change data of the fused feature tensor under physical constraints, performs algebraic operations, and outputs a confidence boundary tensor with a quantized value of 0.65. The threshold determination unit within the active optimization module receives the confidence boundary tensor with a quantized value of 0.65 and compares it with a pre-set physical threshold of 0.80. The threshold determination unit determines that the current input value is lower than the set safe operating physical limit and outputs a level-flipping optimization trigger signal to the spatial gradient calculation unit. The spatial gradient calculation unit initiates the confidence maximization optimization algorithm, establishes a partial differential equation containing the derivative of the three-dimensional spatial coordinate system, and calculates the spatial gradient compensation vector pointing to the low-reflection interference region. The command generation unit converts the spatial gradient compensation vector into a position adjustment command with a target coordinate value of 120 mm and a pitch and yaw angle adjustment command with a target yaw angle of 15 degrees. These kinematic commands are then packaged and sent to the servo drive controller of the position adjustment mechanism. The position adjustment mechanism drives the onboard infrared temperature sensing element to move to the updated physical space coordinates, triggering the collaborative acquisition module to extract updated position data containing the 120 mm and 15 degree parameters, which is then sent to the precision calibration module. The nonlinear mapping unit inputs the updated position data into the converged neural network model, calculating and extracting a dynamic emissivity compensation weight of 0.85 and a temperature compensation value of 12 degrees Celsius. The compensation output unit performs numerical reconstruction calculations on the original infrared radiation data based on the dynamic emissivity compensation weight and the temperature compensation value, and outputs the calibration temperature data.

[0056] Embodiment 3 of the present invention: In the laser-assisted welding and packaging process of microelectromechanical system (MEMS) chips, the physical temperature fluctuation range of the welding molten pool area is subject to strict thermodynamic boundary constraints. The compensation output unit inside the precision calibration module continuously outputs high-frequency sampled calibration temperature data, and the calibration temperature data output in the current system operating cycle is 850 degrees Celsius. The deviation calculation unit inside the heating control module receives the input temperature value of 850 degrees Celsius and simultaneously retrieves the preset target heating temperature data of 880 degrees Celsius from the register address. The deviation calculation unit performs an algebraic difference operation between the calibration temperature data and the preset target heating temperature data to solve for an energy deviation margin of 30 degrees Celsius. The strategy calculation unit receives the 30-degree Celsius energy deviation margin and substitutes it into the proportional-integral-derivative (PID) control algorithm matrix, which includes the proportional coefficient, integral time constant, and derivative time constant, for floating-point numerical calculation. The PID control algorithm matrix determines that the current physical stage is continuous heating based on the integral accumulation term set by the system and derives the laser output power adjustment amount that needs to be increased by 150 watts of energy output. Based on the 150-watt power increase requirement, the power modulation unit generates a pulse width modulation control signal with a 65% duty cycle as a power modulation command. This command is sent directly to the semiconductor pump source drive control board of the laser heating equipment via a high-speed industrial fieldbus network. In response to the power modulation command, the laser heating equipment synchronously increases its internal drive current amplitude, raising the energy of the actual high-energy laser beam output to a physical radiation energy level matching the target set power.

[0057] This invention implemented a comparative verification experiment to support the beneficial effects of the system. A parallel silicon wafer heating test was conducted between a control group system using a conventional linear calibration algorithm and the system of this invention using a physical information neural network model. The invention set 1000 degrees Celsius as the preset target heating temperature and controlled both systems to run continuously for 500 hours. When the temperature measurement angle was 45 degrees and the measurement distance was 200 mm, the average absolute error between the measured temperature value output by the control group system and the actual temperature data reached 15 degrees Celsius, and its maximum temperature overshoot under closed-loop control reached 25 degrees Celsius. In contrast, the system of this invention, driven by the active optimization module, automatically avoided the aforementioned 45-degree interference angle and adjusted the physical pose of the infrared temperature measuring element to 15 degrees. Under this pose, the average absolute error between the calibrated temperature data output by this system and the actual temperature data was reduced to 0.5 degrees Celsius; simultaneously, thanks to the precise closed-loop control of the heating control module, the maximum temperature overshoot during operation was strictly limited to within 2 degrees Celsius. The above comparative data demonstrates that the combination of the physical information neural network model and the spatial dynamic optimization mechanism can effectively overcome the shortcomings of conventional linear calibration algorithms in handling nonlinear errors, and achieve the technical effect of high-precision closed-loop constant temperature control.

Claims

1. A laser heating temperature measurement and calibration system, characterized in that, It includes equipment and control devices that establish communication connections with the equipment; the control devices are used to control the equipment. The control device includes; Collaborative acquisition module: Connects to the device to synchronously extract raw infrared radiation data, real temperature data, and environmental spatial pose data. It sends the real temperature data to the model evolution module and the raw infrared radiation data and environmental spatial pose data to the model evolution module and the precision calibration module. Model Evolution Module: Receives real temperature data, raw infrared radiation data, and environmental spatial pose data. Inputs the raw infrared radiation data and environmental spatial pose data into a preset physical information neural network model for iterative training to generate a converged neural network model and a confidence boundary tensor. Sends the converged neural network model to the precision calibration module and the confidence boundary tensor to the active optimization module. Active optimization module: Receives the confidence boundary tensor, and when the confidence boundary tensor is lower than the preset physical effective threshold, generates a kinematic command and sends it to the position adjustment mechanism to drive the infrared temperature measuring element to move to the target spatial position. This triggers the collaborative acquisition module to extract and update the position data at the target spatial position and send it to the precision calibration module. Precision calibration module: Receives the converged neural network model, raw infrared radiation data, and updated position data. Inputs the updated position data into the converged neural network model to extract dynamic compensation weights. Uses the dynamic compensation weights to eliminate errors in the raw infrared radiation data and outputs calibration temperature data. Sends the calibration temperature data to the heating control module. Heating control module: Receives calibration temperature data, calculates energy deviation margin based on calibration temperature data, generates power modulation command and sends it to laser heating equipment to control output energy.

2. The laser heating temperature measurement and calibration system according to claim 1, characterized in that, The equipment includes a constant temperature heating platform that carries the object to be tested, a contact temperature measuring element, a position adjustment mechanism, an infrared temperature measuring element, an ambient temperature sensing element, and a laser heating device; The constant temperature heating platform carries the object being tested; the contact temperature measuring element is attached to the surface of the object being tested; the position adjustment mechanism is arranged in the space above and to the side of the object being tested; the infrared temperature measuring element is installed at the end of the position adjustment mechanism; the ambient temperature sensing element is arranged in the space around the optical path between the infrared temperature measuring element and the object being tested; the laser emitting end of the laser heating equipment is aligned with the surface of the object being tested; the environmental spatial pose data includes environmental thermal disturbance data, temperature measurement distance data, and temperature measurement angle data.

3. The laser heating temperature measurement and calibration system according to claim 2, characterized in that, The collaborative acquisition module includes a clock synchronization unit, an analog-to-digital conversion unit, and a protocol parsing unit; The clock synchronization unit generates trigger pulses according to a set period and sends them to the analog-to-digital conversion unit; The analog-to-digital conversion unit receives the trigger pulse, synchronously reads the analog electrical signals of the infrared temperature measuring element, the contact temperature measuring element, and the ambient temperature sensing element, converts the analog electrical signals into digital signals, and sends them to the protocol parsing unit. The protocol parsing unit receives digital signals, reads the temperature measurement distance data and temperature measurement angle data fed back by the position adjustment mechanism, and adds a common time sequence timestamp tag to the digital signals, temperature measurement distance data and temperature measurement angle data to generate multi-dimensional spatiotemporal synchronization sequence data. The multi-dimensional spatiotemporal synchronization sequence data includes raw infrared radiation data, real temperature data and environmental spatial pose data. The protocol parsing unit sends the multi-dimensional spatiotemporal synchronization sequence data to the model evolution module and the precision calibration module.

4. The laser heating temperature measurement and calibration system according to claim 3, characterized in that, The model evolution module includes a feature fusion unit and an equation embedding unit; The feature fusion unit receives multi-dimensional spatiotemporal synchronization sequence data, extracts the real temperature data from the multi-dimensional spatiotemporal synchronization sequence data as supervision labels, and extracts the original infrared radiation data, environmental thermal disturbance data, temperature measurement distance data and temperature measurement angle data from the multi-dimensional spatiotemporal synchronization sequence data to construct a multimodal input tensor and send it to the equation embedding unit. The equation embedding unit receives a multimodal input tensor, embeds the thermal radiation distance attenuation constraint equation and the angular reflectivity penalty function into the hidden layer of the physical information neural network model, and outputs a fused feature tensor.

5. The laser heating temperature measurement and calibration system according to claim 4, characterized in that, The model evolution module also includes a network training unit and a confidence evaluation unit; The network training unit receives the fused feature tensor output by the equation embedding unit, inputs the fused feature tensor into the physical information neural network model for forward propagation to obtain the output result, calculates the loss function value between the output result and the supervision label, and uses the backpropagation algorithm to update the model weights to generate the converged neural network model. The confidence assessment unit extracts the information entropy change data of the fused feature tensor under physical constraints, calculates the information entropy change data, and outputs the confidence boundary tensor.

6. The laser heating temperature measurement and calibration system according to claim 5, characterized in that, The active optimization module includes a threshold determination unit and a spatial gradient calculation unit; The threshold determination unit receives the confidence boundary tensor and compares it with the preset physical effective threshold. When the confidence boundary tensor is lower than the preset physical effective threshold, the threshold determination unit outputs an optimization trigger signal to the spatial gradient calculation unit. The spatial gradient calculation unit receives the optimization trigger signal and starts the confidence maximization optimization algorithm to calculate the spatial gradient compensation vector pointing to the high confidence region.

7. The laser heating temperature measurement and calibration system according to claim 6, characterized in that, The active optimization module also includes an instruction generation unit; The instruction generation unit receives the spatial gradient compensation vector, converts the spatial gradient compensation vector into position adjustment instructions and pitch and yaw angle adjustment instructions adapted to the three-dimensional coordinate system as kinematic instructions, and sends them to the position adjustment mechanism. The position adjustment mechanism receives position adjustment commands and pitch and yaw angle adjustment commands to move the physical position and orientation of the infrared temperature measuring element.

8. The laser heating temperature measurement and calibration system according to claim 7, characterized in that, The precision calibration module includes a condition monitoring unit and a nonlinear mapping unit; The status monitoring unit receives multi-dimensional spatiotemporal synchronization sequence data containing updated position data in real time, determines that the current data pose is within a preset high-confidence physical range, generates a target verification signal, and sends the multi-dimensional spatiotemporal synchronization sequence data with the target verification signal to the nonlinear mapping unit. The nonlinear mapping unit receives multi-dimensional spatiotemporal synchronization sequence data with target verification signals, loads a synchronously updated converged neural network model, and extracts output feature values ​​by inputting the multi-dimensional spatiotemporal synchronization sequence data with target verification signals into the converged neural network model.

9. The laser heating temperature measurement and calibration system according to claim 8, characterized in that, The precision calibration module also includes a compensation output unit; The compensation output unit receives the output feature value, decodes the output feature value to obtain the dynamic compensation weight, which includes the dynamic emissivity compensation weight and the temperature compensation value. The compensation output unit uses the dynamic emissivity compensation weight and temperature compensation value in the dynamic compensation weight to perform truncation error elimination on the raw infrared radiation data and outputs the calibration temperature data.

10. The laser heating temperature measurement and calibration system according to claim 1, characterized in that, The heating control module includes a deviation calculation unit, a strategy calculation unit, and a power modulation unit; The deviation calculation unit receives calibration temperature data and preset target heating temperature data, performs difference calculation between calibration temperature data and preset target heating temperature data to obtain energy deviation margin, and sends the energy deviation margin to the strategy calculation unit. The strategy calculation unit receives the energy deviation margin, substitutes the energy deviation margin into the proportional-integral-derivative control algorithm matrix to perform numerical calculation to obtain the laser output power adjustment amount, and sends the laser output power adjustment amount to the power modulation unit. The power modulation unit receives the laser output power adjustment amount and generates a pulse width modulation control signal as a power modulation command, which is then sent to the laser heating equipment to adjust the output energy.

Images