A real-time control method for the temperature of a release agent curing process

Abstract

The application provides a kind of real-time control method for release agent curing process temperature, the control method includes: collecting real-time temperature in curing furnace, combining preset target temperature to solve temperature deviation, deviation rate of change, after standardization, combined with the control law gradient norm of last control cycle to form fuzzy neural network input vector, control law gradient norm is zero in initial stage, get prediction reasoning result and control reasoning result respectively by fuzzy neural network with prediction output end and control output end, construct dimensionless prediction temperature deviation square and dimensionless control law gradient square index weighted joint cost function;Rely on fuzzy rule to calculate each rule confidence weight, establish the nonlinear positive correlation between weight and membership function width, optimize membership width parameter, obtain adjustment component by control reasoning result demisting, obtain steady-state feedforward component by target temperature mapping, superimpose both and limit safely, output heating control signal.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to a method for real-time temperature control during the curing process of release agents. Background Technology

[0002] Release agents play a crucial role in numerous industrial manufacturing sectors, with the curing process determining the surface quality and release effect of the product. In the release agent curing process, the temperature within the curing oven is the most critical process parameter. Excessive temperature can lead to thermal degradation and loss of activity of the release agent, while insufficient temperature results in incomplete curing, affecting the density and adhesion of the film. Real-time temperature control within the curing oven is essential for ensuring product yield. The industry widely employs conventional PID controllers or basic fuzzy logic controllers to regulate the curing temperature. While PID controllers rely on empirical parameter tuning and are simple in structure, they often fail to achieve ideal tracking performance when dealing with curing thermal systems exhibiting highly nonlinear, large inertia, and hysteresis characteristics.

[0003] While basic fuzzy logic controllers do not rely on mathematical models, they lack self-learning capabilities in actual operation. Membership functions and control rules typically require repeated manual trial and error, resulting in limited adaptability under complex operating conditions. Existing network algorithms often focus solely on the current temperature deviation during optimization, neglecting the drastic changes in the control commands themselves. This leads to abrupt outputs of heating control signals in the face of external disturbances, causing frequent actuator actions and system thermodynamic oscillations. Existing fuzzy neural networks, when adjusting internal parameters, fail to provide a refined assessment of the actual performance contribution of each fuzzy rule. This results in significant blind updates to key parameters such as membership functions, making it difficult for the network to reliably identify and strengthen the most advantageous rules for current control during inference, thus hindering improvements in temperature control accuracy. Current temperature control methods typically rely solely on error feedback adjustment, failing to integrate system steady-state characteristics into the generation of control signals. Even when the system state approaches equilibrium, adjustment lag persists, making it difficult to meet the stringent requirements of high-precision, smooth temperature control in the release agent curing process. Summary of the Invention

[0004] To improve the response performance and steady-state control accuracy of the temperature control system for release agent curing ovens, this disclosure provides a method for real-time temperature control during the release agent curing process.

[0005] In a first aspect, embodiments of this disclosure provide a method for real-time temperature control during the curing process of a release agent, comprising the following steps:

[0006] The current temperature is obtained by the sensor in the curing oven. The temperature deviation and the rate of change of the deviation are calculated based on the current temperature and the preset target temperature. The current input vector of the fuzzy neural network is obtained based on the standardized result of the temperature deviation and the rate of change of the deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero. Based on the current input vector, reasoning and adjustment are performed through the fuzzy neural network with prediction output and control output to obtain prediction inference results and control inference results. A joint cost function is obtained by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square of the gradient of the dimensionless control law. The confidence weight is obtained by calculating the comprehensive contribution of the fuzzy rules to reduce the joint cost function. The adjusted membership function width parameter is obtained by establishing a nonlinear positive correlation between the confidence weight and the membership function width that constitutes the premise of the rule. Based on the control inference result, the adjustment component is obtained by defuzzification, and the steady-state feedforward component is obtained by mapping according to the preset target temperature. The adjustment component and the steady-state feedforward component are summed and subjected to safety limiting to generate a heating control signal.

[0007] Secondly, this disclosure also provides a real-time temperature control system for the release agent curing process, including the following modules: The acquisition module is used to acquire the current temperature based on the sensor in the curing oven, calculate the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature, and obtain the current input vector of the fuzzy neural network based on the standardized result of the temperature deviation and the rate of change of deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero. The calculation module is used to perform reasoning and adjustment based on the current input vector through the fuzzy neural network with a prediction output and a control output to obtain prediction reasoning results and control reasoning results. It obtains a joint cost function by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square index of the dimensionless control law gradient. It calculates the confidence weight based on the comprehensive contribution of fuzzy rules to reducing the joint cost function. It obtains the adjusted membership function width parameter by establishing a nonlinear positive correlation between the confidence weight and the membership function width that constitutes the premise of the rule. The control module is used to defuzzify the control inference result to obtain the adjustment component, map the preset target temperature to obtain the steady-state feedforward component, sum the adjustment component and the steady-state feedforward component and perform safety limiting to generate a heating control signal.

[0008] This invention incorporates the control law gradient squared index from the previous control cycle into the input vector of a fuzzy neural network and introduces a control law gradient regularization term into the joint cost function. This suppresses the oversensitivity of the control output to the neuron's response, reduces drastic fluctuations in the control signal, and improves the operational stability of the temperature control system. A joint cost function integrating dimensionless predicted temperature deviation and the dimensionless control law gradient squared index is constructed. Based on fuzzy rules, confidence weights are calculated for the comprehensive contribution of reducing this cost function, establishing a nonlinear positive correlation between the weights and the membership function width, thus achieving online optimization of network parameters. This mechanism not only balances temperature tracking accuracy and smooth actuator movement but also enhances the system's anti-interference capability and stability in complex environments. Combining steady-state feedforward components with adjustment components and implementing safety limiting ensures the speed and accuracy of the heating process, guarantees the safe operation of the hardware, and improves the curing quality of the release agent and the product yield. Attached Figure Description

[0009] Figure 1 A flowchart of a method for real-time temperature control during the curing process of a release agent; Figure 2 A schematic diagram of the hardware signal link for multi-point temperature acquisition in the curing oven; Figure 3 This is a schematic diagram illustrating the update of confidence weights and membership widths for fuzzy rules. Detailed Implementation

[0010] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] Reference Figure 1 This invention proposes a method for real-time temperature control during the curing process of a release agent, comprising the following steps: S1. The current temperature is obtained based on the sensor in the curing oven. The temperature deviation and the rate of change of the deviation are calculated based on the current temperature and the preset target temperature. The current input vector of the fuzzy neural network is obtained based on the standardized result of the temperature deviation and the rate of change of the deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero.

[0012] A PT100 temperature sensor array is evenly distributed across multiple temperature zones in the curing oven. The analog voltage signals collected by the sensors are converted into digital signals using an analog-to-digital converter, and then subjected to median filtering and Kalman filtering algorithms to reduce high-frequency electromagnetic interference and random noise. The arithmetic mean of the digital signals from each sensor is taken as the current temperature within the curing oven. The temperature deviation is obtained by subtracting the current temperature from the target temperature required for release agent curing, which is preset in the controller register. Figure 2 The diagram shows the hardware link for acquiring the temperature of the curing oven. The link includes, in sequence, a PT100 array for the curing oven, a signal conditioning unit, an A / D conversion unit, and a median and Kalman filter unit, which outputs the average temperature signal of the curing oven.

[0013] The deviation rate of change is calculated by subtracting the temperature deviation of the previous control cycle from the temperature deviation of the current control cycle and then dividing by the sampling period of the control system. A basic universe of discourse for the fuzzy neural network is defined. Using proportional mapping and amplitude limiting, the temperature deviation is multiplied by the first standardization coefficient, and the deviation rate of change is multiplied by the second standardization coefficient, mapping to the standard normalized universe of discourse for standardization. The partial derivative vectors of the heating control signal output by the controller in the previous control cycle with respect to the outputs of neurons in each layer of the fuzzy neural network are extracted. The sum of the squares of all elements in this partial derivative vector is calculated to obtain the squared norm of the control law gradient in the previous control cycle.

[0014] Determine if the control system is currently in its first sampling period after startup. If so, set the elements of the dimension corresponding to the control law gradient square norm to zero when constructing the input matrix. Concatenate the standardized temperature deviation, the standardized deviation change rate, and the dimensionless gradient input obtained by dividing the control law gradient square norm by the preset gradient reference value in a preset dimension order to form a 1x3 matrix, which serves as the current input vector of the fuzzy neural network.

[0015] In one possible embodiment, the step of calculating the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature includes: The feedback voltage values ​​of each channel of the PT100 temperature sensor array distributed in the curing oven are read according to a sampling period of 1000ms after being converted by the signal conditioning circuit. The feedback voltage values ​​of each channel are converted into digital temperature signals by an analog-to-digital converter and a calibration curve, and the current digital temperature signal is obtained by averaging the digital temperature signals of each channel. The temperature deviation is obtained by subtracting the current temperature digital signal from the preset target temperature. The temperature deviation is subtracted from the temperature deviation of the previous sampling period, and the difference is divided by the duration corresponding to 1000ms to obtain the deviation change rate. When the system is in the first sampling period after power-on and there is no temperature deviation from the previous sampling period, the deviation change rate is set to zero, or the temperature deviation of the previous sampling period is assigned as the current temperature deviation.

[0016] A timer interrupt is set to execute every 1000ms to read the temperature signals from four PT100 temperature sensors distributed within the curing oven. The resistance signals from these sensors are amplified and filtered by a signal conditioning circuit, then converted into feedback voltage values ​​for each of the 0-5V range. Assume that one of the voltage measurements acquired in the current cycle is 2.52V. A 12-bit high-precision analog-to-digital converter is driven to digitize this analog voltage and interpolate it according to a calibration curve to convert it into a digital signal of the current temperature. Let's assume the actual temperature extracted by the system at this point is 120.5℃. After acquiring the current temperature, given a preset target temperature of 130.0℃, the floating-point arithmetic logic unit performs a subtraction operation, obtaining a result of 9.5, which is the temperature deviation.

[0017] Extract the temperature deviation data from the previous sampling period, assuming it to be 9.8℃. Subtract the historical temperature deviation of 9.8℃ from the current temperature deviation of 9.5℃; the difference is -0.3℃. Since the sampling interval is 1000ms (1 second), divide the difference by the elapsed time to obtain a result of -0.3℃ / s. Write the calculated deviation change rate parameter into the controller's designated floating-point variable register, and then send this parameter along with the temperature deviation variable to the next stage algorithm.

[0018] In one possible embodiment, the step of obtaining the current input vector of the fuzzy neural network based on the standardized result of the temperature deviation and the rate of change of deviation and the control law gradient square norm of the previous control cycle, and replacing the control law gradient square norm with zero when the system is in the initial stage, includes: The dimensionless numerical constraint obtained by multiplying the temperature deviation by the first standardization coefficient using the limiting function is then transformed into a standardized temperature deviation within the universe of discourse interval of -3 to 3. The dimensionless numerical constraint obtained by multiplying the deviation change rate by the second standardization coefficient using the amplitude limiting function is then transformed into a standardized deviation change rate within the universe of discourse interval of -3 to 3. Extract the sum of squared partial derivatives of the heating control signal with respect to the outputs of neurons in each layer of the fuzzy neural network in the previous control cycle, and calculate the squared norm of the control law gradient: ; in, The squared norm of the gradient of the control law. This is the heating control signal from the previous control cycle. The output of each layer of neurons in the fuzzy neural network; The standardized temperature deviation, the rate of change of the standardized deviation, and the dimensionless gradient input obtained by dividing the square norm of the control law gradient by the preset gradient reference value are concatenated into a 1-row, 3-column matrix to construct the current input vector.

[0019] A fuzzy neural network with prediction and control outputs is constructed. The fuzzy neural network includes a shared input layer, a fuzzification layer, a rule-based preamble layer, and two consequent output branches. The prediction output outputs the predictive inference results related to the temperature deviation in the next control cycle, and the control output outputs the activation values ​​of each node participating in the defuzzification. After obtaining the environmental variables, a data normalization process is initiated. Assuming the current temperature deviation is 9.5℃, the first normalization coefficient is set to 5.0℃. - ¹, Multiply the deviation by the first standardization factor to obtain a dimensionless value of 47.5. Call the clipping function to clip the out-of-bounds value: 47.5 exceeds the upper limit of the universe of discourse interval of 3.0, so the clipping algorithm truncates 47.5 to 3.0; this is the standardized temperature deviation. For the rate of change of deviation, assuming the measured current value is -0.3℃ / s, and the second standardization factor is 2.5s / ℃, multiply the current value by the second standardization factor to obtain a dimensionless value of -0.75. After clipping, this value is within the interval of -3 to 3, so no truncation operation is performed, generating a standardized deviation rate of change of -0.75.

[0020] After completing the static proportional constraints, the derived parameters from the end of the previous operation are extracted using DMA: assuming the neural network structure has 5 nodes, the computational core extracts the partial derivative values ​​of the gradients of the control instructions of the previous cycle on the independent output responses of these 5 hidden layer neurons, assumed to be 0.10, -0.20, 0.15, 0.05, and -0.10. The bottom-level multiplication unit performs element-wise self-square operation, generating the sequence 0.010, 0.040, 0.0225, 0.0025, and 0.010, which are then summed to obtain a total of 0.085. This sum is the square norm of the control law gradient. Dividing this square norm of the control law gradient by a preset gradient benchmark value yields the dimensionless gradient input. Following a specified matrix filling order, 3.0 is placed in the first column, -0.75 in the second column, and the dimensionless gradient input is filled in the third column, constructing a 1-row, 3-column matrix variable.

[0021] S2, based on the current input vector, reasoning and adjustment are performed through the fuzzy neural network with prediction output and control output to obtain prediction inference results and control inference results. A joint cost function is obtained by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square index of the dimensionless control law gradient. A confidence weight is obtained by calculating the comprehensive contribution of the fuzzy rules to reduce the joint cost function. An adjusted membership function width parameter is obtained by establishing a nonlinear positive correlation between the confidence weight and the membership function width constituting the premise of the rule.

[0022] The current input vector is fed into the input layer of the fuzzy neural network. In the fuzzification layer, the Gaussian function is used to calculate the membership degree of each element of the input vector to different fuzzy linguistic variables. The trigger strength of each fuzzy rule is obtained by multiplying the membership degrees through a rule antecedent layer. The consequent layer of the fuzzy neural network is configured with two branches: a prediction output and a control output. The prediction output combines the trigger strength of each rule with the consequent parameters of the prediction branch to obtain the prediction inference result, while the control output combines the trigger strength of each rule with the consequent parameters of the control branch to obtain the control inference result used for defuzzification.

[0023] To obtain the predicted temperature deviation, a system identification model is pre-calibrated using historical measured temperature data. The prediction output includes the system identification model calibrated with historical measured temperature data. The inputs of the system identification model include the current temperature deviation, the rate of change of standardized deviation, and candidate heating control signals. The output is the predicted temperature deviation for the next control cycle. The candidate heating control signal is obtained by preliminary defuzzification of the control inference results from the control output. The system identification model predicts the temperature response for the next control cycle based on this candidate heating control signal, thus ensuring that the joint cost function corresponds to the temperature change under the actual heating control action.

[0024] When constructing the joint cost function, the predicted temperature deviation for the next control cycle is divided by a preset temperature deviation benchmark value to obtain the dimensionless predicted temperature deviation, and the control law gradient square norm is divided by a preset gradient benchmark value to obtain the dimensionless control law gradient square index. The square of the dimensionless predicted temperature deviation is calculated as the first-generation value, and the dimensionless control law gradient square index is multiplied by a regularization factor pre-written into memory to obtain the second-generation value. The first-generation value and the second-generation value are algebraically added to construct a joint cost function that balances temperature tracking accuracy and control stability.

[0025] Within each weight update cycle, the activation intensity of each fuzzy rule in each control cycle is recorded, along with the actual decrease in the dimensionless joint cost function within the corresponding control cycle. The instantaneous contribution is obtained by multiplying the activation intensity of a single rule by its corresponding cost decrease. The instantaneous contributions of the same rule within each weight update cycle are accumulated to obtain the comprehensive contribution. A Softmax operation is then performed on the comprehensive contribution of each rule to obtain the confidence weight. This confidence weight calculation process is based on the rule activation intensity and the cost function decrease, and does not use the reciprocal of the partial derivative of the cost function with respect to the rule trigger intensity.

[0026] The natural logarithm function is used as the positive mapping relationship. The confidence weights are added to the base of the natural logarithm, e, and then used as the input to the natural logarithm function to calculate the scaling adjustment coefficient. This scaling adjustment coefficient is then used to multiply the initial width parameter of the original Gaussian function in the fuzzy layer to complete the online update and adjustment of the membership function width parameter. (Refer to...) Figure 3 .

[0027] In one possible embodiment, obtaining the joint cost function by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the predicted output and the squared gradient index of the dimensionless control law includes: Based on the current temperature deviation, the rate of change of standardized deviation, and candidate heating control signals, the predicted temperature deviation for the next control cycle is obtained by a system identification model set at the prediction output and calibrated with historical measured temperature data. The dimensionless predicted temperature deviation is obtained by dividing the predicted temperature deviation of the next control cycle by the preset temperature deviation benchmark value, and the square of the dimensionless predicted temperature deviation is used as the first generation value. Divide the squared norm of the control law gradient by the preset gradient benchmark to obtain the dimensionless squared control law gradient index, and multiply the squared control law gradient index by the regularization factor to obtain the second generation value. Add the first generation value and the second generation value to obtain the joint cost function. The updated network connection weights are obtained using the gradient descent method. The update formula for the network connection weights is as follows: ; in, For the updated network connection weights, Here, 'a' represents the current connection weights, 'a' represents the set learning rate parameter, and 'J' represents the joint cost function. The partial derivative of the calculated joint cost function with respect to each of the current connection weights.

[0028] Assuming the predicted temperature deviation for the next control cycle is 2.0℃, and the preset temperature deviation benchmark is 10℃, the dimensionless predicted temperature deviation is 0.2. Squaring this value yields 0.04. This result of 0.04 is set as the first-generation value. The control law gradient square norm calculated in the pre-preparation stage is retrieved, assumed to be 0.085, and the preset gradient benchmark is 1.0. Therefore, the dimensionless control law gradient square index is 0.085. Multiplying 0.085 by the regularization factor of 0.02 yields 0.0017. Adding 0.04 to 0.0017 gives a comprehensive evaluation metric of 0.0417, which is established as the dimensionless joint cost function.

[0029] After establishing the optimization direction, the error backpropagation module is initiated to update the weights. The weights corresponding to a specific node are used as the optimization target: assuming the current connection weight of this node is extracted to be 0.55, and the partial derivative of the joint cost function with respect to this node is -0.20. A descent update is performed according to the fine-tuning iteration law with a step-size compensation coefficient, and the control logic calls the preset learning rate parameter 0.01. 0.01 is multiplied by the partial derivative value -0.20 to generate a weight correction amount of -0.002. The current connection weight of 0.55 is subtracted from the weight correction amount -0.002, resulting in 0.552. The updated variable is calculated and written back to achieve replacement. In specific implementations, the partial derivative of the joint cost function with respect to the current connection weight can be obtained using chain rule differentiation, automatic differentiation, or finite difference approximation methods; this application does not limit this method.

[0030] In one possible embodiment, the step of calculating the confidence weight based on the comprehensive contribution of fuzzy rules to reducing the joint cost function includes: Set every 100 control cycles as one weight update cycle; Record the activation intensity of each fuzzy rule in each control cycle within the current weight update cycle; Based on historical activation data, the updated confidence weights for each rule are calculated using the following formulas: , ; ; in, The instantaneous contribution of a single rule within a single control period t. The activation strength of the corresponding record. This corresponds to the actual decrease in the dimensionless joint cost function within the control period, and , The cumulative extracted dimensionless comprehensive contribution, Here, e represents the updated confidence weights, where e is the base of the natural logarithm. This represents the summation of the natural exponent terms of all fuzzy rules in the fuzzy rule base.

[0031] During implementation, a weight update cycle is defined as every 100 control cycles. Assuming we are currently in the 50th control cycle, the activation strength of the 3rd fuzzy rule is read, assumed to be 0.8. Simultaneously, the actual decrease in the dimensionless joint cost function within this micro-cycle is calculated, assuming... =0.05. Multiplying the activation intensity of 0.8 by the actual decrease of 0.05 yields 0.04. This result is set as the instantaneous contribution of the rule during this week.

[0032] When the system loop count reaches 100, the instantaneous contribution of the rule within this weight update cycle is accumulated to obtain the cumulative comprehensive contribution value. Taking the third fuzzy rule as an example, the instantaneous contributions stored in these 100 control cycles are added together. Assume the total cumulative result is 4.20. This value is the cumulative comprehensive contribution value of the rule. Assume the fuzzy rule base contains 5 fuzzy rules, all of which have completed their respective comprehensive contribution calculations. Then, the main calculation kernel performs exponential calculations on each scoring parameter, taking the exponent of 4.2, resulting in approximately 66.69. Simultaneously, the exponents of the 5 evaluation parameters are summed, assuming a result of 104.08. A percentage calculation is then performed, dividing 66.69 by the global total of 104.08, yielding 0.641. This result is the confidence weight index.

[0033] In one possible embodiment, the step of establishing a nonlinear positive correlation between the confidence weight and the membership function width constituting the rule premise to obtain the adjusted membership function width parameter includes: The Gaussian membership function that constitutes the premise of a single fuzzy rule is set to a base standard deviation of 1.0; The association is established using the following positive mapping formula, and the updated membership function width parameter is calculated, extracted, and applied: , ; in, The base standard deviation of the Gaussian membership function is set. The current confidence weights for each fuzzy rule are obtained, where ln represents the natural logarithm of the values ​​within the parentheses. To calculate the corresponding adjustment coefficient, This is the width parameter of the updated membership function that is derived and used in the application.

[0034] The initial standard deviation of the Gaussian distribution surface function is set to a constant of 1.0. The new confidence weight of rule 3, 0.641, is read from the variable access area. The instruction flow is called via a scaling-based transformation correction function. The confidence weight 0.641 is added to the constant e (approximately 2.71828), and the natural logarithm is performed. The new sum is imported into the kernel for calculation. The corresponding floating-point result is obtained through logarithmic transformation, generating a scaling adjustment coefficient of approximately 1.2117. The original base standard deviation of 1.0 is extracted by the control allocation code, and multiplied by the scaling adjustment coefficient 1.2117. The result is 1.2117. The updated value is written to the storage unit as the new membership function width parameter, completing the replacement operation and putting it into use.

[0035] S3, based on the control inference result, the adjustment component is obtained by defuzzification, and the steady-state feedforward component is obtained by mapping according to the preset target temperature. The adjustment component and the steady-state feedforward component are summed and subjected to safety limiting to generate a heating control signal.

[0036] The centroid method is used to defuzzify and perform integral and division operations on the control inference results at the output of the fuzzy neural network control, mapping the control inference results back to the actual power percentage range to generate adjustment components for compensating for temperature fluctuations. Based on the heating mechanism model of the curing oven and the oven body heat loss coefficient, a mapping table between the target temperature and the basic heating power required to maintain that temperature is pre-established. The steady-state feedforward component required to maintain thermal equilibrium is calculated based on the preset target temperature using a lookup table method combined with a linear interpolation algorithm.

[0037] An addition operation instruction is invoked to add the adjustment component and the steady-state feedforward component to obtain the initial control signal. The maximum safe drive duty cycle and the minimum protection duty cycle that the curing oven heater can withstand are set as the upper and lower thresholds of the safe limiting range, respectively. A saturation limiting algorithm is used to compare the initial control signal. If the initial control signal is greater than the upper threshold, the upper threshold is output; if it is less than the lower threshold, the lower threshold is output; if it is within the range, the original value is output, thus generating the heating control signal. This heating control signal is converted into a pulse-width modulated digital pulse with a continuously adjustable duty cycle, driving the solid-state relay in the curing oven's electrical control cabinet to control the on / off heating of the three-phase heating elements.

[0038] In one possible embodiment, the process of obtaining an adjustment component by defuzzifying the control inference result, obtaining a steady-state feedforward component based on the preset target temperature mapping, and summing the adjustment component and the steady-state feedforward component and applying a safety limit to generate a heating control signal includes: The centroid coordinates of the horizontal axis corresponding to the control output terminal are extracted as the adjustment component, and the unit is limited to power percentage. The centroid method used in the defuzzification calculation is mathematically expressed as follows: ; in, The extracted horizontal axis centroid coordinates are also known as the adjustment components. The activation values ​​of each node at the control output terminal are collected. The horizontal axis scalar coordinates representing the power percentage corresponding to each node; Based on the preset target temperature, a pre-built static thermal balance mapping table is queried to obtain the required basic percentage of heating power as the steady-state feedforward component. Perform control calculations and safety limiting to generate numerical results: ; in, The numerical results of the heating control signal after safety limiting processing. The mapping table outputs the corresponding steady-state feedforward components, where min represents the minimum value function and max represents the maximum value function; The result The duty cycle signal is loaded into the pulse width modulation generation module to control the on / off action of the solid-state relay and generate a heating control signal.

[0039] Three coordinate nodes carrying action intensity characteristics were set in the output plane of the control model structure, with horizontal axis coordinates of -10.0, 0.0, and 10.0, representing the heating power percentage. The activation values ​​of these three nodes at the control output end were collected as 0.2, 0.6, and 0.8, respectively. The calculation began: the activation intensity of the first node (0.2) multiplied by the horizontal axis (-10.0) equals -2.0; the activation intensity of the second node (0.6) multiplied by the horizontal axis (0.0) equals 0; the activation intensity of the third node (0.8) multiplied by the horizontal axis (10.0) equals 8.0. The sum of the numerators is 6.0. The cumulative sum of the activation values ​​of all nodes is 1.6, which is used as the denominator. A floating-point division operation is performed, dividing the total sum of numerators (6.0) by the total sum of denominators (1.6), resulting in 3.75. This is the extracted horizontal axis centroid coordinate value, i.e., the adjustment component.

[0040] The main control program reads the preset target temperature of 130.0℃, queries the thermal balance static mapping table to extract the required steady-state feedforward component, which is 45.0. The adjustment component 3.75 is added to the steady-state feedforward component 45.0, resulting in a sum of 48.75. An extreme value judgment is performed: 48.75 is less than the upper limit of 100, so it remains unchanged. A lower limit judgment is then performed: 48.75 is greater than the lower limit of 0, so it is also unaffected. The resulting value is 48.75. The control program loads this data as the duty cycle into the drive register of the PWM underlying hardware, outputting a square wave with a duty cycle of 48.75% to control the on / off state of the solid-state relay, executing the heating control signal.

[0041] This invention also proposes a real-time temperature control system for the curing process of release agents, comprising the following modules: The acquisition module is used to acquire the current temperature based on the sensor in the curing oven, calculate the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature, and obtain the current input vector of the fuzzy neural network based on the standardized result of the temperature deviation and the rate of change of deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero. The calculation module is used to perform reasoning and adjustment based on the current input vector through the fuzzy neural network with a prediction output and a control output to obtain prediction reasoning results and control reasoning results. It obtains a joint cost function by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square index of the dimensionless control law gradient. It calculates the confidence weight based on the comprehensive contribution of fuzzy rules to reducing the joint cost function. It obtains the adjusted membership function width parameter by establishing a nonlinear positive correlation between the confidence weight and the membership function width that constitutes the premise of the rule. The control module is used to defuzzify the control inference result to obtain the adjustment component, map the preset target temperature to obtain the steady-state feedforward component, sum the adjustment component and the steady-state feedforward component and perform safety limiting to generate a heating control signal.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for real-time temperature control during the curing process of a release agent, characterized in that, Includes the following steps: The current temperature is obtained by the sensor in the curing oven. The temperature deviation and the rate of change of the deviation are calculated based on the current temperature and the preset target temperature. The current input vector of the fuzzy neural network is obtained based on the standardized result of the temperature deviation and the rate of change of the deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero. Based on the current input vector, reasoning and adjustment are performed through the fuzzy neural network with prediction output and control output to obtain prediction inference results and control inference results. A joint cost function is obtained by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square of the gradient of the dimensionless control law. The confidence weight is obtained by calculating the comprehensive contribution of the fuzzy rules to reduce the joint cost function. The adjusted membership function width parameter is obtained by establishing a nonlinear positive correlation between the confidence weight and the membership function width that constitutes the premise of the rule. Based on the control inference result, the adjustment component is obtained by defuzzification, and the steady-state feedforward component is obtained by mapping according to the preset target temperature. The adjustment component and the steady-state feedforward component are summed and subjected to safety limiting to generate a heating control signal.

2. The method according to claim 1, characterized in that, The step of calculating the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature includes: The feedback voltage values ​​of each channel of the PT100 temperature sensor array distributed in the curing oven are read according to a sampling period of 1000ms after being converted by the signal conditioning circuit. The feedback voltage values ​​of each channel are converted into digital temperature signals by an analog-to-digital converter and a calibration curve, and the current digital temperature signal is obtained by averaging the digital temperature signals of each channel. The temperature deviation is obtained by subtracting the current temperature digital signal from the preset target temperature. The temperature deviation is subtracted from the temperature deviation of the previous sampling period, and the difference is divided by the duration corresponding to 1000ms to obtain the deviation change rate. When the system is in the first sampling period after power-on and there is no temperature deviation from the previous sampling period, the deviation change rate is set to zero, or the temperature deviation of the previous sampling period is assigned as the current temperature deviation.

3. The method according to claim 1, characterized in that, The current input vector of the fuzzy neural network is obtained based on the standardized result of the temperature deviation and the rate of change of deviation, and the squared norm of the control law gradient from the previous control cycle. When the system is in the initial stage, the squared norm of the control law gradient is replaced with zero, including: The dimensionless numerical constraint obtained by multiplying the temperature deviation by the first standardization coefficient using the limiting function is then transformed into a standardized temperature deviation within the universe of discourse interval of -3 to 3. The dimensionless numerical constraint obtained by multiplying the deviation change rate by the second standardization coefficient using the amplitude limiting function is then transformed into a standardized deviation change rate within the universe of discourse interval of -3 to 3. Extract the sum of squared partial derivatives of the heating control signal with respect to the outputs of neurons in each layer of the fuzzy neural network in the previous control cycle, and calculate the squared norm of the control law gradient: ； in, The squared norm of the gradient of the control law. This is the heating control signal from the previous control cycle. The output of each layer of neurons in the fuzzy neural network; The standardized temperature deviation, the rate of change of the standardized deviation, and the dimensionless gradient input obtained by dividing the square norm of the control law gradient by the preset gradient reference value are concatenated into a 1-row, 3-column matrix to construct the current input vector.

4. The method according to claim 2, characterized in that, The joint cost function is obtained by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the predicted output terminal and the squared gradient index of the dimensionless control law, including: Based on the current temperature deviation, the rate of change of standardized deviation, and candidate heating control signals, the predicted temperature deviation for the next control cycle is obtained by a system identification model set at the prediction output and calibrated with historical measured temperature data. The dimensionless predicted temperature deviation is obtained by dividing the predicted temperature deviation of the next control cycle by the preset temperature deviation benchmark value, and the square of the dimensionless predicted temperature deviation is used as the first generation value. Divide the squared norm of the control law gradient by the preset gradient benchmark to obtain the dimensionless squared control law gradient index, and multiply the squared control law gradient index by the regularization factor to obtain the second generation value. Add the first generation value and the second generation value to obtain the joint cost function. The updated network connection weights are obtained using the gradient descent method. The update formula for the network connection weights is as follows: ； in, For the updated network connection weights, Here, 'a' represents the current connection weights, 'a' represents the set learning rate parameter, and 'J' represents the joint cost function. The partial derivative of the calculated joint cost function with respect to each of the current connection weights.

5. The method according to claim 1, characterized in that, The calculation of the confidence weight based on fuzzy rules to reduce the overall contribution of the joint cost function includes: Set every 100 control cycles as one weight update cycle; Record the activation intensity of each fuzzy rule in each control cycle within the current weight update cycle; Based on historical activation data, the updated confidence weights for each rule are calculated using the following formulas: , ; ； in, The instantaneous contribution of a single rule within a single control period t. The activation strength of the corresponding record. This corresponds to the actual decrease in the dimensionless joint cost function within the control period, and , The cumulative extracted dimensionless comprehensive contribution, Here, e represents the updated confidence weights, where e is the base of the natural logarithm. This represents the summation of the natural exponent terms of all fuzzy rules in the fuzzy rule base.

6. The method according to claim 1, characterized in that, The step of establishing a nonlinear positive correlation based on the confidence weight and the membership function width of the constitutive rule premise to obtain the adjusted membership function width parameter includes: The Gaussian membership function that constitutes the premise of a single fuzzy rule is set to a base standard deviation of 1.0; The association is established using the following positive mapping formula, and the updated membership function width parameter is calculated, extracted, and applied: , ; in, The base standard deviation of the Gaussian membership function is set. The current confidence weights for each fuzzy rule are obtained, where ln represents the natural logarithm of the values ​​within the parentheses. To calculate the corresponding adjustment coefficient, This is the width parameter of the updated membership function that is derived and used in the application.

7. The method according to claim 1, characterized in that, The process of obtaining an adjustment component by defuzzifying the control inference result, obtaining a steady-state feedforward component based on the preset target temperature mapping, and summing the adjustment component and the steady-state feedforward component and applying a safety limit to generate a heating control signal includes: The centroid coordinates of the horizontal axis corresponding to the control output terminal are extracted as the adjustment component, and the unit is limited to power percentage. The centroid method used in the defuzzification calculation is mathematically expressed as follows: ； in, The extracted horizontal axis centroid coordinates are also known as the adjustment components. The activation values ​​of each node at the control output terminal are collected. The horizontal axis scalar coordinates representing the power percentage corresponding to each node; Based on the preset target temperature, a pre-built static thermal balance mapping table is queried to obtain the required basic percentage of heating power as the steady-state feedforward component. Perform control calculations and safety limiting to generate numerical results: ; in, The numerical results of the heating control signal after safety limiting processing. The mapping table outputs the corresponding steady-state feedforward components, where min represents the minimum value function and max represents the maximum value function; The result The duty cycle signal is loaded into the pulse width modulation generation module to control the on / off action of the solid-state relay and generate a heating control signal.

8. A real-time temperature control system for the curing process of a release agent, characterized in that, Includes the following modules: The acquisition module is used to acquire the current temperature based on the sensor in the curing oven, calculate the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature, and obtain the current input vector of the fuzzy neural network based on the standardized result of the temperature deviation and the rate of change of deviation and the control law gradient square norm of the previous control cycle. When the system is in the initial stage, the control law gradient square norm is replaced by zero. The calculation module is used to perform reasoning and adjustment based on the current input vector through the fuzzy neural network with a prediction output and a control output to obtain prediction reasoning results and control reasoning results. It obtains a joint cost function by weighted summation of the square of the dimensionless predicted temperature deviation corresponding to the prediction output and the square index of the dimensionless control law gradient. It calculates the confidence weight based on the comprehensive contribution of fuzzy rules to reducing the joint cost function. It obtains the adjusted membership function width parameter by establishing a nonlinear positive correlation between the confidence weight and the membership function width that constitutes the premise of the rule. The control module is used to defuzzify the control inference result to obtain the adjustment component, map the preset target temperature to obtain the steady-state feedforward component, sum the adjustment component and the steady-state feedforward component and perform safety limiting to generate a heating control signal.

9. The system according to claim 8, characterized in that, The step of calculating the temperature deviation and the rate of change of deviation based on the current temperature and the preset target temperature includes: The feedback voltage values ​​of each channel of the PT100 temperature sensor array distributed in the curing oven are read according to a sampling period of 1000ms after being converted by the signal conditioning circuit. The feedback voltage values ​​of each channel are converted into digital temperature signals by an analog-to-digital converter and a calibration curve, and the current digital temperature signal is obtained by averaging the digital temperature signals of each channel. The temperature deviation is obtained by subtracting the current temperature digital signal from the preset target temperature. The temperature deviation is subtracted from the temperature deviation of the previous sampling period, and the difference is divided by the duration corresponding to 1000ms to obtain the deviation change rate. When the system is in the first sampling period after power-on and there is no temperature deviation from the previous sampling period, the deviation change rate is set to zero, or the temperature deviation of the previous sampling period is assigned as the current temperature deviation.

10. The system according to claim 8, characterized in that, The current input vector of the fuzzy neural network is obtained based on the standardized result of the temperature deviation and the rate of change of deviation, and the squared norm of the control law gradient from the previous control cycle. When the system is in the initial stage, the squared norm of the control law gradient is replaced with zero, including: The dimensionless numerical constraint obtained by multiplying the temperature deviation by the first standardization coefficient using the limiting function is then transformed into a standardized temperature deviation within the universe of discourse interval of -3 to 3. The dimensionless numerical constraint obtained by multiplying the deviation change rate by the second standardization coefficient using the amplitude limiting function is then transformed into a standardized deviation change rate within the universe of discourse interval of -3 to 3. Extract the sum of squared partial derivatives of the heating control signal with respect to the outputs of neurons in each layer of the fuzzy neural network in the previous control cycle, and calculate the squared norm of the control law gradient: ； in, The squared norm of the gradient of the control law. This is the heating control signal from the previous control cycle. The output of each layer of neurons in the fuzzy neural network; The standardized temperature deviation, the rate of change of the standardized deviation, and the dimensionless gradient input obtained by dividing the square norm of the control law gradient by the preset gradient reference value are concatenated into a 1-row, 3-column matrix to construct the current input vector.

Images