A smart temperature control method and system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification.

Abstract

This invention discloses an intelligent temperature control method and system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification. The method includes: collecting operational data and identifying thermal response parameters online; establishing a thermal risk budget account to determine the thermal risk budget balance; generating candidate control actions and using counterfactual prediction to calculate the thermal risk consumption value of each action; adding actions that meet preset disabling conditions to a blacklist; inputting actions not on the blacklist into multiple shadow models to predict future temperatures and calculating model divergence; and selecting either degradation protection based on whether the divergence exceeds a threshold or selecting the optimal action based on a comprehensive cost function to output to the heater. This invention can assess thermal risks in advance, disable high-risk actions, and proactively degrade using model divergence adjudication, solving control safety problems such as thermal inertia overshoot, parameter inaccuracies, lack of future risk assessment, and model uncertainty.

Description

Technical Field

[0001] This invention belongs to the field of intelligent temperature control technology for heat treatment equipment, and particularly relates to an intelligent temperature control method and system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification. Background Technology

[0002] Existing heat treatment equipment typically employs proportional-integral-derivative (PID) control, fuzzy control, model predictive control, or empirical curve control to regulate furnace cavity temperature. These methods generally calculate the heater output power or pulse width modulation duty cycle based on the deviation between the current temperature and the target temperature, thereby gradually bringing the furnace cavity temperature closer to the set temperature.

[0003] However, in actual heat treatment processes, the equipment exhibits significant thermal inertia. Even if the controller reduces or shuts off the heating power, the heat stored in the furnace cavity, heating elements, insulation materials, workpieces, and fixtures may continue to be released, causing the temperature to continue to rise and resulting in temperature overshoot.

[0004] Furthermore, variations in the loading quantity, material heat capacity, placement method, furnace door opening, ambient temperature, sensor delay, and heater aging status of different batches of workpieces can all lead to changes in thermal response. Existing control methods typically only focus on the temperature deviation itself and cannot adequately determine whether a particular control action will lead to overshoot risk in the future, nor can they proactively enter a safety degradation mode when model predictions are inconsistent. Therefore, the following problems are prone to occur: 1. Continuing to output high power when approaching the target temperature leads to temperature overshoot; 2. Changes in the loading rate render the original control parameters ineffective, leading to a decrease in control accuracy; 3. Aggressive control actions continue to be executed even when the model prediction error increases; 4. Relying solely on a single model or controller, lacking a risk adjudication mechanism; 5. There is a lack of mechanisms to screen out potential dangers of future control actions in advance; 6. After an abnormal disturbance occurs, the recovery process is prone to overshoot or oscillation again; 7. Existing control methods lack risk screening and safety adjudication mechanisms suitable for controllers with low computing power, making it difficult to make real-time judgments and safety constraints on the future overshoot risk, model uncertainty, and abnormal recovery process of candidate heating actions under limited computing resources.

[0005] Therefore, there is a need for a temperature control method for heat treatment equipment that can assess thermal risks in advance during the temperature control process, screen out high-risk control actions, and automatically degrade protection when the model is uncertain. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes an intelligent temperature control method and system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification, thereby resolving the issues existing in the prior art.

[0007] To achieve the above objectives, the present invention provides an intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification, comprising: Collect operating data of the heat treatment equipment, and identify the thermal response parameters of the heat treatment equipment online based on the operating data; A thermal risk budget account is established based on the operating data, the thermal response parameters, and the current process stage, and the thermal risk budget balance for the current control cycle is determined. Multiple candidate control actions are generated, and counterfactual prediction is performed on each candidate control action to calculate its heat risk consumption value; Candidate control actions that meet any preset disabling condition are added to the action blacklist. The preset disabling condition includes at least the thermal risk consumption value of the candidate control action being greater than the thermal risk budget balance. Candidate control actions not added to the action blacklist are input into multiple shadow models for future temperature prediction, and the model divergence degree is calculated based on the prediction results of each shadow model. If the model divergence exceeds a preset divergence threshold, a degradation protection control is executed, and a degradation control action is output; if the model divergence does not exceed the preset divergence threshold, an optimal candidate control action is selected from the candidate control actions that have never been included in the action blacklist, based on the comprehensive cost function, as the final control action. The final control action or the degraded control action is output to the heater for execution, thereby realizing intelligent temperature control of the heat treatment equipment based on thermal inertia dynamic identification and shadow model verification.

[0008] Optionally, the thermal response parameters can be identified online using a recursive least squares algorithm, a sliding window least squares algorithm, or a Kalman filter algorithm. The thermal response parameters are updated in real time based on the prediction error at the current moment. The thermal response parameters include the thermal inertia coefficient, the heating influence coefficient, and the environmental influence coefficient.

[0009] Optionally, establishing the thermal risk budget account includes: setting a baseline safety limit for the thermal risk budget account; obtaining the current temperature, target temperature, temperature rise rate, model prediction error, and sensor delay time based on the operating data; determining the corresponding process stage weight based on the current process stage; calculating the residual heat surge risk, temperature rise momentum risk, loading state uncertainty risk, model error risk, sensor delay risk, and process stage sensitivity risk based on the thermal response parameters, the difference between the current temperature and the target temperature, the temperature rise rate, the model prediction error, the sensor delay time, the process stage weight, and the loading state uncertainty, respectively; and deducting the calculated risks from the baseline safety limit to obtain the thermal risk budget balance.

[0010] Optionally, generating the plurality of candidate control actions includes: Obtain a preset set of candidate actions, calculate the maximum power allowed in the current control cycle based on the thermal risk budget balance, and eliminate candidate control actions whose power values ​​are greater than the maximum power from the set of candidate actions.

[0011] Optionally, when performing counterfactual prediction for each candidate control action, the thermal response parameters obtained through online identification are used as inputs, including the current temperature, the candidate control action, and the ambient temperature, to predict the temperature trajectory for multiple future sampling periods, and to extract the predicted maximum temperature and the predicted maximum temperature rise rate from it.

[0012] Optionally, calculating the thermal risk consumption value of the candidate control action includes: The heat risk consumption value is obtained by weighted summing the overshoot of the predicted maximum temperature and target temperature corresponding to the candidate action, the excess of the predicted maximum temperature rise rate and the safe temperature rise rate threshold, the loading state uncertainty calculated based on the operating data, the model prediction error, and the change range of the candidate action and the control action of the previous cycle.

[0013] Optionally, the preset disable conditions also include: The predicted maximum temperature corresponding to this candidate action exceeds the upper limit of the target temperature. The predicted maximum temperature rise rate corresponding to this candidate action exceeds the absolute safe rate limit; The power value of the candidate action exceeds the maximum power allowed in the current control cycle.

[0014] Optionally, when calculating the model divergence degree, for the same candidate control action, the temperature trajectory predicted by each shadow model for multiple future sampling periods is obtained; For each future sampling period, calculate the difference between the maximum and minimum values ​​of the predicted temperature in each shadow model, and take the maximum value of the difference in all future sampling periods as the model divergence degree; Alternatively, for each future sampling period, calculate the average value of the predicted temperature of each shadow model, then calculate the standard deviation of the predicted temperature of each shadow model relative to the average value, and take the average value of the standard deviations in all future sampling periods as the model divergence degree. The shadow model includes at least two of the following: fast response shadow model, slow response shadow model, conservative heat storage shadow model, historical batch matching shadow model, and lightweight neural network shadow model.

[0015] Optionally, it also includes: when the absolute value of temperature deviation is less than the temperature deviation threshold, the absolute value of temperature rise rate is less than the temperature rise rate threshold, the absolute value of model prediction error is less than the model error threshold, and the model divergence degree is less than the preset divergence threshold for L consecutive sampling periods, the current thermal risk budget balance is increased by a replenishment step size until the thermal risk budget balance reaches the benchmark safety limit. When the thermal risk consumption value of a candidate control action is less than the product of the safety release coefficient and the current thermal risk budget balance, and the model divergence degree of the candidate control action is less than the preset divergence threshold, the candidate control action is removed from the action blacklist.

[0016] The present invention also provides an intelligent temperature control system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification, for performing the above method, comprising: The temperature acquisition module, the target temperature input module, and the historical data storage module are used to acquire operational data, input the target temperature, and store historical data, respectively. The main controller, and the following modules respectively connected to the main controller: A thermal inertia dynamic identification module is used to identify thermal response parameters online based on the operating data; The loading state uncertainty calculation module is used to calculate the loading state uncertainty. The thermal risk budget module is used to establish thermal risk budget accounts and determine thermal risk budget balances; The candidate action generation module is used to generate multiple candidate control actions; The counterfactual prediction module is used to predict future temperatures for candidate control actions; The action risk calculation module is used to calculate the thermal risk consumption value of candidate control actions; The action blacklist module is used to add candidate control actions that meet any preset disabling conditions to the action blacklist. The multi-shadow model prediction module is used to predict future temperatures using multiple shadow models. The disagreement resolution module is used to calculate the model disagreement degree; Degradation protection module, used to output degradation control actions; The final action selection module is used to select the final control action; The risk replenishment module is used to replenish the remaining balance of the hot risk budget. The control output module is used to convert control actions into heater control signals.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: (1) It can predict the future risks brought about by heating action in advance and reduce temperature overshoot; (2) Limit control actions through thermal risk budget accounts to avoid continuing to output high-risk power when approaching the target temperature; (3) Prohibit dangerous control actions in advance through an action blacklist mechanism, rather than remedial action after overshooting; (4) Improve the safety of the control system under model uncertainty by using the multi-shadow model divergence judgment; (5) Through the degradation protection mechanism, stable control can still be maintained when there are changes in charge, sensor delays, or environmental disturbances; (6) Through the risk recovery mechanism, the equipment can be smoothly restored from the abnormal state, reducing oscillations; (7) It does not rely on high-cost and complex models and can be implemented on low-cost controllers; (8) Applicable to various heat treatment equipment such as box furnace, tube furnace, pit furnace, drying oven, muffle furnace, experimental furnace, sintering furnace, and drying furnace; (9) Compared with simple proportional-integral-derivative control or ordinary model prediction, this invention emphasizes safety adjudication, action disabling and model uncertainty protection, and has stronger practicality and auditability. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a system structure block diagram according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a thermal risk budget account according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the action blacklist adjudication method according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the multi-shadow model divergence degradation in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0021] Example 1 This embodiment provides an intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification. The overall process is as follows: Figure 2 As shown, it includes: First, define the following variables: k: Current sampling time; dt: sampling period; T(k): The current temperature of the furnace cavity at the kth sampling time; Ts(k): The target temperature at the kth sampling time; Ta(k): The ambient temperature at the k-th sampling time; u(k): The heating control output at the kth sampling time, with a value range of 0 to 1, representing the heater power ratio or pulse width modulation duty cycle; T_history(k): Historical temperature sequence; p(k): The current process stage identifier, used to indicate the stage of the heat treatment process curve; it can be represented by discrete numbers or weight values, for example, the heating stage is denoted as p1, the stage approaching the target temperature is denoted as p2, the holding stage is denoted as p3, the cooling stage is denoted as p4, and the material sensitive range is denoted as p5; different stages correspond to different process stage weights stage_weight(k).

[0022] q(k): Uncertainty regarding the charge state; ε(k): Model prediction error.

[0023] The temperature deviation is: e(k) = Ts(k) - T(k); The rate of temperature rise is: v(k) = (T(k) - T(k-1)) / dt; The acceleration due to temperature rise is: a(k) = (v(k) - v(k-1)) / dt; The historical temperature series can be represented as: T_history(k)={T(kn), T(k-n+1),.... T(K)}; The system state vector can be represented as: X(k)=[T(k), Ts(k), Ta(k), u(k-1), v(k), a(k), p(k), q(k)]; Where X(k) is used to describe the current temperature state, control state, process stage, and loading uncertainty of the heat treatment equipment.

[0024] S1, Run the data acquisition steps; The temperature acquisition module collects operating data of the heat treatment equipment according to a fixed sampling period dt. The collected data includes: 1. Current furnace temperature T(k); 2. Target temperature Ts(k); 3. Ambient temperature Ta(k); 4. Heating control output u(k-1) of the previous cycle; 5. Historical temperature sequence T_history(k); 6. Current process stage p(k); 7. Heater operating status; 8. Sensor status; 9. Furnace door status; 10. Historical control actions and their execution results.

[0025] After data collection, the system calculates the temperature deviation, temperature rise rate, and temperature rise acceleration: e(k) = Ts(k) - T(k); v(k) = (T(k) - T(k-1)) / dt; a(k) = (v(k) - v(k-1)) / dt; This step can be implemented using existing temperature sensors, thermocouples, resistance temperature detectors (RTDs), infrared temperature measurement modules, microcontrollers, programmable logic controllers (PLCs), or industrial computers, and belongs to existing data acquisition technologies.

[0026] The innovation of this invention lies not in the temperature acquisition hardware itself, but in the subsequent use of the acquired data for thermal risk budget accounts, counterfactual action blacklists, multi-shadow model disagreement adjudication, and degradation protection control.

[0027] S2, Thermal inertia dynamic identification steps; The system constructs an input vector and calculates the prediction error based on the historical temperature sequence T_history(k), the current and historical control output u(k), the ambient temperature Ta(k), and the actual temperature feedback T(k+1) at the next sampling time. It then updates the equivalent thermal response parameters of the heat treatment equipment online using a recursive least squares algorithm, a sliding window least squares algorithm, or a Kalman filter algorithm.

[0028] A first-order heat dissipation model can be used: T(k+1)=α(k)T(k)+β(k)u(k)+γ(k)Ta(k)+ε(k); in: α(k): Thermal inertia coefficient, used to represent the heat storage effect of furnace cavity, heater, insulation material, workpiece and fixture; β(k): Heating influence coefficient, used to represent the effect of heating output on the temperature change of the furnace cavity; γ(k): Environmental influence coefficient, used to represent the influence of ambient temperature on furnace cavity temperature; ε(k): Model perturbation term or residual term; during online identification, the model prediction error is defined as ε(k+1)=T(k+1)-T_hat(k+1).

[0029] To facilitate online updates, the parameter vector is defined as: θ(k) = [α(k), β(k), γ(k)]^T; Define the input vector as: φ(k) = [T(k), u(k), Ta(k)]^T; The predicted temperature is: T_hat(k+1)=φ^T(k)θ(k); The prediction error is: ε(k+1)=T(k+1)-T_hat(k+1); The system can use a recursive least squares algorithm to update parameters: K(k)=P(k)φ(k) / [λ+φ^T(k)P(k)φ(k)]; θ(k+1)=θ(k)+K(k)ε(k+1); P(k+1)=λ^{-1}[P(k)-K(k)φ^T(k)P(k)]; Where: K(k): parameter correction gain; P(k): covariance matrix; lambda: forgetting factor, which can range from 0.90 to 0.99; φ(k)^T: transpose of φ(k).

[0030] The recursive least squares algorithm is an existing parameter identification technique. This invention can directly use the existing recursive least squares algorithm, or it can use the existing sliding window least squares algorithm, Kalman filter algorithm, or empirical parameter update algorithm.

[0031] The innovation of this invention lies not in the recursive least squares algorithm itself, but in using the α(k), β(k), and γ(k) obtained online for subsequent hot risk budgeting, counterfactual action prediction, action blacklist establishment, and multi-shadow model disagreement adjudication.

[0032] S3, Steps for establishing a thermal risk budget account; The system establishes a thermal risk budget account based on the current temperature, target temperature, temperature rise rate, model error, loading uncertainty, and process stage.

[0033] The thermal risk budget account represents the amount of risk that can be consumed during the current control cycle without temperature overshoot, severe oscillations, or breaches of process safety boundaries.

[0034] The thermal risk budget balance is defined as: B(K)=BO-Rh(K)-Rv(k)-RI(k)-Rm(k)-Rd(k)-Rs(k) Where: B(k): current thermal risk budget balance; BO: baseline safety allowance; Rh(k): residual heat surge risk; Rv(k): temperature rise momentum risk; Rl(k): loading status uncertainty risk; Rm(k): model error risk; Rd(k): sensor delay risk; Rs(k): process stage sensitive risk.

[0035] 3.1 Risk of residual heat rising upwards; The risk of residual heat rising indicates that even if heating is currently turned off, the heat already accumulated in the furnace cavity, heating element, insulation layer, workpiece, and fixture may still cause the temperature to continue to rise.

[0036] Assuming heating is immediately turned off, i.e., u(k) = 0, the system predicts the temperature for the next H sampling periods: T_off(k+1)=α(k)T(k)+β(k)·0+γ(k)Ta(k); T_off(k+j)=α(k)T_off(k+j-1)+γ(k)Ta(k); Where j = 1 to H.

[0037] The predicted highest temperature after the power outage is: T_off_max(k)=max{T_off(k+1),T_off(k+2),...,T_off(k+H)}; The risk of residual heat rising upwards is: Rh(k) = wh max(0, T_off_max(k)- Ts(k)) Where: wh: residual heat risk weight; T_off_max(k): the predicted maximum temperature after power failure.

[0038] When T_off_max(k) is greater than the target temperature Ts(k), it indicates that overshoot may occur even if heating is turned off immediately. In this case, Rh(k) increases and the thermal risk budget balance B(k) decreases.

[0039] 3.2 Risk of temperature rise momentum; Temperature rise momentum risk indicates the risk of continued upward movement as the current temperature continues to rise rapidly.

[0040] Rv(k) = wv max(O, v(k)-v_safe) Where: wv: temperature rise momentum risk weight; v(k): current temperature rise rate; v_safe: safe temperature rise rate threshold.

[0041] When the rate of temperature rise exceeds the safety threshold, it indicates that the system is in a state of overly aggressive heating, and the allowable heating power should be reduced.

[0042] 3.3 Model error risk; Model error risk is used to reflect whether the current prediction model is reliable.

[0043] Rm(k) = wm abs(ε(k)); Where: wm: model error risk weight; abs(ε(k)): absolute value of model prediction error.

[0044] When the actual temperature differs significantly from the predicted temperature, it indicates that the current model's description of the equipment's thermal state is inaccurate, and the system should reduce the level of aggressive control.

[0045] 3.4 Sensor delay risk; Sensor delay risk is used to represent the control risk caused by lag in temperature measurement signals.

[0046] Rd(k) = wd tau_d abs(v(k)) Where: wd: sensor delay risk weight; tau_d: sensor response delay time; abs(v(k)): absolute value of temperature rise rate.

[0047] The faster the temperature rises, the greater the deviation between the actual temperature and the measured temperature caused by sensor delay, and therefore the higher the risk.

[0048] 3.5 Sensitive risks in the process stage; Different stages of the process are sensitive to temperature overshoot to varying degrees. For example, faster heating is permissible in the initial stage of heating, but more stringent control is required when approaching the target temperature, during the holding stage, in the phase transition range, or in the material's sensitive range.

[0049] The sensitive risks in the process stage are: Rs(k) = ws stage_weight(k) Where: ws: risk weight of process stage; stage_weight(k): weight of the current process stage.

[0050] For example: stage_weight(k) = 0.2 in the initial heating stage; stage_weight(k) = 0.8 in the stage approaching the target temperature stage; stage_weight(k) = 1.0 in the heat preservation stage; stage_weight(k) = 1.2 in the material sensitive range.

[0051] The above values ​​can be set according to different heat treatment processes. Figure 3 It illustrates the composition of the thermal risk budget account and the deduction relationship between various risks.

[0052] S4, Steps for calculating the uncertainty of the loading status; The loading condition uncertainty is used to characterize the differences between the loading quantity, material heat capacity, and placement method of the current batch of workpieces and historical standard operating conditions. This invention does not require precise identification of workpiece quality, but rather calculates the loading condition uncertainty indirectly through temperature response.

[0053] The temperature rise response per unit power is defined as: ρ(k) = v(k) / (u(k) + δ) Where: ρ(k) is the current unit power temperature rise response; δ: A small constant to prevent the denominator from being zero; it can be taken as 0.001. v(k): Current rate of temperature rise; u(k): Current control output.

[0054] The historical standard unit power temperature rise response is ρ0.

[0055] The uncertainty of the charge condition is: q(k)= min(1, abs(ρ(k)- ρ0) / ( ρ0 +δl); The risk of uncertainty in the loading status is: RI(K)=wl q(k) Where: wl: Weight of uncertainty risk in loading.

[0056] The larger q(k) is, the greater the difference between the current batch and historically stable batches. The system automatically increases the risk penalty and reduces the allowable heating power. This method does not rely on a weighing device and does not require the user to input accurate loading amounts, making it suitable for low-cost heat treatment equipment.

[0057] S5, the current control period allows for the calculation of risk limits; Based on the thermal risk budget account balance B(k), the system calculates the maximum thermal risk allowance that can be consumed in the current cycle.

[0058] The maximum allowable power can be set as follows: u_max(k) = min(1, B(K) / B0) When B(k) is close to BO, it indicates that the current risk is low and higher power output is allowed; When B(k) is low, it indicates that the current risk is high, and the system automatically reduces the maximum allowable power. When B(k) is less than or equal to 0, it indicates that the current risk has exceeded the limit, and the system prohibits further heating or enters the downgrade protection.

[0059] You can also set a minimum safety threshold B_safe: If B(k) <= B_safe, then u_max(k) = 0; Where: B_safe: security risk threshold.

[0060] S6, Candidate control action generation step; The candidate action generation module generates multiple candidate control actions.

[0061] The candidate action set can be represented as: U(K) = {u1, u2, ..., N}; For example: U(k)={0,0.2,0.4,0.6,0.8,1.0}; Where: 0 indicates heating off; 0.2 indicates 20% power; 0.4 indicates 40% power; 0.6 indicates 60% power; 0.8 indicates 80% power; 1.0 indicates 100% power.

[0062] Candidate actions can also be dynamically shrunk based on u_max(k): Only retain candidate actions that satisfy u_i <= u_max(k).

[0063] In other words: If u_i > u_max(k), then u_i will not be included in the allowed candidate set. In this way, when the risk is high, the system does not need to wait until the final decision stage to limit power, but can eliminate high-power actions during the candidate action generation stage.

[0064] S7, Counterfactual Action Prediction Steps; For each candidate control action u_i, the system assumes that the action will be executed in the current period and predicts the temperature change over the next H sampling periods. This prediction is not an actual execution, but rather a simulation of "what would happen if u_i were executed," hence it is called counterfactual action prediction.

[0065] For candidate action u_i, the future temperature prediction is: T_i(k+1)=α(k) T(k)+β(k) u_i+γ(k) Ta(k) T_i(k+j)=α(k) T_i(k+j-1)+β(k) u_i+γ(k) Ta(k); in: j=1 to H; T_i(k+j): The predicted temperature of the j-th future period after executing candidate action u_i; H: Pre-step length, such as 10 seconds, 20 seconds, or 30 seconds.

[0066] The predicted highest temperature corresponding to candidate action u_i is: T_i_max(k)= max(T_i(k+1), T_i(k+2),.... T_i(k+H) The predicted maximum temperature rise rate corresponding to candidate action u_i is: v_i_max(k)= maxllT_i(k+j)-T_i(k+j-1}) / dt) Where j = 1 to H.

[0067] This step allows the system to determine in advance whether an action might cause overshoot, rapid heating, or oscillation before the action is actually output.

[0068] S8, Steps for calculating the thermal risk consumption value of candidate actions; The system calculates the thermal risk consumption value Ci(k) for each candidate action u_i.

[0069] Ci(k)=c1·max(0, T_i_max(k)-Ts(k)-T_allow) + c2·max(0, v_i_max(k)-v_safe) + c3·q(k) + c4·abs(ε(k)) + c5·abs(u_i - u(k-1)); Where: Ci(k): thermal risk consumption value of candidate action u_i; T_i_max(k): the predicted maximum temperature corresponding to candidate action u_i; Ts(k): target temperature; T_allow: allowable temperature surge; v_i_max(k): predicted maximum temperature rise rate corresponding to candidate action u_i; v_safe: safe temperature rise rate threshold; q(k): loading status uncertainty; ε(k): model prediction error; abs(u_i- u(k-1)): control action change amplitude; c1, c2, c3, c4, c5: risk weight coefficients.

[0070] Among them, c1 corresponds to temperature overshoot risk, c2 corresponds to rapid heating risk, c3 corresponds to charging uncertainty risk, c4 corresponds to model error risk, and c5 corresponds to sudden change risk in control action.

[0071] S9, Steps for establishing a counterfactual actions blacklist The action blacklist is used to store candidate control actions that are prohibited from being executed in the current cycle.

[0072] If a candidate action u_i satisfies any of the following conditions, it is added to the action blacklist BlackList(k).

[0073] Condition one: Ci(k)>B(k) This means that the risk consumption value of the candidate action is greater than the current hot risk budget balance.

[0074] Condition two: T_i_max(k)>Ts(k)+ T_limit This means that the predicted maximum temperature exceeds the upper limit of the target temperature.

[0075] Condition three: v_i_max(k)>v_limit This means that the predicted rate of temperature rise exceeds the upper limit of the absolute safe rate.

[0076] Condition four: u_i>u_max(k) This means that the power of the candidate action exceeds the current maximum allowed power.

[0077] Condition five: This action has caused overshoot or oscillation under similar historical conditions.

[0078] It can be represented as: if Ci(k)>B(k), then u_i belongs to BlackList(k) if T_i_max(k)>Ts(k)+ T_limit, then u_i belongs to BlackList(k) if v_i_max(k)>v_limit, then u_i belongs to BlackList(k) if u_i>u_max(k), then u_i belongs to BlackList(k) Actions on the action blacklist must not be executed during the current control cycle.

[0079] The allowed set of candidate actions is: U_allow(k) = U(k) - BlackList(k) Where: U_allow(k): the set of allowed candidate actions; BlackList(k): the action blacklist.

[0080] Figure 4 A schematic diagram illustrating the adjudication logic of the action blacklist is provided.

[0081] S10, Multi-Shadow Model Prediction Step; This invention sets up multiple shadow models to predict the same candidate action in parallel. The shadow models do not directly control the heater, but are used to determine future temperature trends and model reliability.

[0082] Shadow models can include the following categories: 10.1 Fast Response Shadow Model; Fast response models are used to simulate situations with small charge volumes, rapid thermal response, or strong heater effects.

[0083] T_fast(k+1)=a_fast T(k) + b_fast u_i + g_fast Ta(k); Where: a_fast: fast response thermal inertia coefficient; b_fast: fast response heating coefficient; g_fast: fast response environmental coefficient.

[0084] 10.2 Slow Response Shadow Model; Slow response models are used to simulate situations with large charge volumes, slow thermal response, or large heat capacity.

[0085] T_slow(k+1)=a_slow T(k) + b_slow u_i + g_slow Ta(k); Where: a_slow: slow response thermal inertia coefficient; b_slow: slow response heating coefficient; g_slow: slow response environmental coefficient.

[0086] 10.3 Conservative thermal storage shadow model; The conservative heat storage model is used to simulate situations where there is strong residual heat release and the temperature may continue to rise even after the heating is turned off.

[0087] T_safe(k+1)=a_safe T(k) + b_safe u_i + g_safe Ta(k) + η abs(v(k)) Where: η abs(v(k)) is used to represent the effect of current temperature rise momentum or residual heat accumulation on future temperature.

[0088] 10.4 Historical Batch Matching Shadow Model; The system can select historical batches that are similar to the current temperature, target temperature, loading response, and process stage from historical data and establish a historical batch matching model.

[0089] Similarity can be expressed as: Sim(r,k)= d1abs(T_r-T(k))+ d2abs(v_r-v(k))+d3abs(q_r-q(k))+ d4abs(p_r-p(k)) in: r: Historical batch number; T_r: Temperature corresponding to the historical batch; v_r: Temperature rise rate corresponding to the historical batch; q_r: Loading uncertainty corresponding to the historical batch; p_r: Process stage corresponding to the historical batch; d1, d2, d3, d4: Similarity weights.

[0090] The historical batch with the smallest Sim(rk) is selected as a reference to obtain the historical batch matching prediction results.

[0091] 10.5 Lightweight Neural Network Shadow Model; If the controller's computing power allows, a lightweight neural network can be used as a shadow model.

[0092] This neural network can adopt a multilayer perceptron structure, as follows: Input layer: Input X(k), including T(K), Ts(k), Ta(k), u(k), v(k), a(k), q(k), p(k); First hidden layer: 16 neurons; Second hidden layer: 8 neurons; Output layer: Outputs the predicted temperature values ​​for the next H sampling periods.

[0093] The activation function is a linear rectified function: relu(x) = max(O,x). The neural network calculation process is as follows: H1=relu(W1 X(K)+ b1) H2 = relu(W2 H1+ b2) Y_hat(k)=W3 H2+b3 Where: W1, W2, W3: neural network weight matrices; b1, b2, b3: bias terms; H1: output of the first hidden layer; H2: output of the second hidden layer; Y_hat(k): temperature prediction sequence for the next H sampling periods.

[0094] Multilayer perceptron neural networks belong to existing neural network technologies.

[0095] The parameters of this invention can be obtained using existing training methods.

[0096] The innovation of this invention lies not in the neural network structure itself, but in using a lightweight neural network as one of multiple shadow models and using the prediction divergence between multiple shadow models to trigger degradation protection.

[0097] S11, Steps for calculating model divergence; For candidate action u_i, different shadow models output the predicted trajectory of future temperature.

[0098] For example: T_fast_i(k+j): The predicted temperature of candidate action u_i by the fast response model; T_slow_i(k+j): The predictivity of the slow response model for candidate action u_i; T_safe_i(k+j): The predicted temperature of candidate action u_i by the conservative heat storage model; T_nn_i(k+j): The predicted temperature of candidate action u_i by the neural network model.

[0099] The model divergence can be calculated using the maximum difference method: D_i(k)= max over j [ max(T_fast_i(k+j), T_slow_i(k+j), T_safe_i(k+j),T_nn_i(k+j))- min(T_fast_i(k+j), T_slow_i(k+j), T_safe_i(k+j), T_nn_i(k+j)) ]; Where j = 1 to H.

[0100] Alternatively, the average degree of divergence can be used for calculation: T_avg_i(k+j)=average(T_fast_i(k+j), T_slow_i(k+j), T_safe_i(k+j), T_nn_i(k+j)) D_i(k)=average over j [ sqrt( average( (T_model_i(k+j) - T_avg_i(k+j))^2 ) ) ] Where: D_i(k): model divergence degree of candidate action u_i; T_model_i(k+j): the predicted temperature of a certain shadow model in the j-th future period; T_avg_i(k+j): the average predicted temperature of all shadow models.

[0101] When D_i(k) is small, it indicates that multiple models have consistent judgments on future temperature trends, and the prediction reliability is high. When D_i(k) is large, it indicates that the current thermal state is uncertain, and the system cannot continue to execute aggressive control actions.

[0102] If: D_i(k)>D_th Then the candidate action u_i is restricted or prohibited from execution.

[0103] Where: D_th: Model divergence threshold.

[0104] Figure 5 The decision-making path for predicting divergence and downgrade protection using a multi-shadow model is demonstrated.

[0105] S12, Degradation protection control steps; The system triggers degradation protection when any of the following conditions occur: 1. All candidate actions are added to the action blacklist; 2. The model divergence degree D_i(k) exceeds the threshold D_th; 3. The thermal risk budget balance B(k) is less than the safety threshold B_safe; 4. The current temperature is close to or exceeds the upper limit of the target temperature; 5. The model error ε(k) continues to increase; 6. The uncertainty of the loading status q(k) exceeds the threshold.

[0106] Degradation protection output is: u_safe(k)= min(u(k-1), u_max(k) μ) in: u_safe(k): Degradation protection control output; u(k-1): Control output of the previous cycle; u_max(k): Current maximum allowed power; mu: Degradation factor, with a value ranging from 0 to 0.5.

[0107] If the current temperature is close to the target temperature, then force shutdown or low power hold is performed.

[0108] The determination of proximity to the target temperature is as follows: T(k)>= Ts(k)- T_near Where: T_near: close to the target temperature threshold.

[0109] If this condition is met, then: u_safe(k)=0 or: u_safe(k) = u_hold Where u_hold is the low power hold value, which can be 0 to 0.2.

[0110] Degradation protection may also include: 1. Limit maximum heating power; 2. Limit the rate of power change; 3. Power-up operation is prohibited; 4. Forcefully shut off the heating; 5. Extend the sampling confirmation time; 6. Enter soft recovery mode.

[0111] S13, Final Control Action Selection Step For candidate actions that are not included in the action blacklist and whose model divergence meets the requirements, the system calculates the comprehensive cost function.

[0112] Ji(k) = λ1 abs(T_i(k+H)-Ts(k)) + λ2 Ci(k) + λ3 D_i(k) + λ4 abs(u_i - u(k-1)) Where: Ji(k): comprehensive cost of candidate action u_i; T_i(k+H): temperature of candidate action u_i at the predicted endpoint; Ts(k): target temperature; Ci(k): risk consumption value of candidate action; D_i(k): model divergence degree of candidate action; abs(u_i-u(k-1)): change range of candidate action; λ1, λ2, λ3, λ4: weight coefficients.

[0113] The meanings of each item are as follows: lambdalabs(T_i(k+H)-Ts(k)): Target temperature tracking error; λ2Ci(k): Thermal risk consumption penalty; λ3D_i(k): Model uncertainty penalty; λ4abs(u_i-u(k-1)): Control action mutation penalty.

[0114] Ultimately, the candidate action with the lowest overall cost was selected as the control output: u_star(k)= action with minimum Ji(k) in U_allow(k) In other words, from the set of allowed candidate actions U_allow(k), the action with the smallest Ji(k) is selected as the final output.

[0115] If U_allow(k) is empty, then perform downgrade protection output: u_star(k) = u_safe(k) S14, Thermal Risk Budget Recovery Steps; When the system temperature stabilizes, the model error decreases, and the model divergence decreases, the system gradually replenishes the thermal risk budget account.

[0116] The conditions for determining a stable state in a system are: Abs(e(k)) <E_th Abs(v(k)) <V_th Abs(ε(k)) <M_th D(k) <D_th Where: E_th: temperature deviation threshold; V_th: temperature rise rate threshold; M_th: model error threshold; D_th: model divergence threshold.

[0117] If the above conditions are met for L consecutive sampling periods, then: B(k+1) = min(BO, B(k) + dB) Where: dB: Risk budget replenishment step size.

[0118] If any condition is not met, then: B(k+1)= max(0, B(k)-dB_penalty) Where: dB_penalty: risk penalty step size.

[0119] This mechanism ensures that the system will not immediately perform aggressive heating actions after it has just stabilized, but will gradually restore control.

[0120] S15, Steps for dynamically removing an action from the blacklist; The action blacklist is not a permanent ban, but is dynamically removed based on the current risk status.

[0121] Candidate action u_i can be removed from the action blacklist if it meets the following conditions: Ci(k) < release_ratio B(k) D_i(k) < D_th[[ID=~]] T_i_max(k) ≤ Ts(k) + T_allo No overshoot occurred in the last L consecutive sampling periods Wherein: release_ratio: Safety release coefficient, the value range can be from 0.3 to 0.8

[0122] After meeting the conditions: Remove the candidate action u_i from the action blacklist BlackList(k), so that it re - enters the set of allowed candidate actions U_allow(k) and the candidate action set again. This mechanism can prevent a certain action from being permanently disabled, and at the same time ensure that the action recovery must meet the safety conditions

[0123] S16, Control output execution step The control output module converts the final control action u_star(k) into the actual control signal of the heater

[0124] If pulse - width modulation is used for control, then within the control period Tc: t_on = u_star(k) Tc t_off = Tc - t_on Wherein: t_on: Heater conduction time; t_off: Heater off - time; Tc: Control period

[0125] If relay control is used, the minimum switching period T_min can be set to avoid frequent on - off of the relay

[0126] When: abs(u_star(k) - u(k - 1)) < u_min_change, keep the control action of the previous period unchanged. Wherein: u_min_change: Minimum control change threshold

[0127] If a solid - state relay, thyristor or power regulator is used, the control output module converts u_star(k) into the corresponding conduction angle, duty cycle or output voltage ratio

[0128] S17, System module structure and functions of each module As Figure 1 shown, the system of the present invention includes the following modules 1. Temperature acquisition module [[ID=5~]]For acquiring the furnace cavity temperature, ambient temperature and sensor status

[0129] 2. Target temperature input module Used to input or read heat treatment process curves, including heating curves, holding temperatures, holding times, and cooling requirements.

[0130] 3. Historical data storage module It is used to store historical temperature sequences, historical control actions, historical model errors, historical overshoot records, and historical batch response data.

[0131] 4. Thermal inertia dynamic identification module Used to update α(k), β(k), and γ(k) online based on T(k), u(k), Ta(k) and historical temperature sequences.

[0132] 5. Thermal Risk Budget Module Used to calculate B(k), Rh(K), Rv(k), Rl(k), Rm(k), Rd(k), and Rs(k).

[0133] 6. Uncertainty Calculation Module for Loading Status Used to calculate ρ(k), q(k)) and Rl(k).

[0134] 7. Candidate Action Generation Module Used to generate a candidate action set U(k), and to initially filter high-power actions based on u_max(k).

[0135] 8. Counterfact Prediction Module Used to predict the temperature trajectory for the next H sampling periods for each candidate action u_i.

[0136] 9. Action Risk Calculation Module This is used to calculate the risk cost value Ci(k) for each candidate action.

[0137] 10. Action Blacklist Module Used to determine whether an action should be disabled based on Ci(k), T_i_max(k), v_i_max(k), and u_max(k).

[0138] 11. Multi-shadow model prediction module It is used to output multiple sets of prediction results through fast response model, slow response model, conservative heat storage model, historical batch matching model and lightweight neural network model.

[0139] 12. Dispute Resolution Module Used to calculate D_i(k), which determines whether the predictions of multiple models are consistent.

[0140] 13. Degradation Protection Module This is used to output u_safe(k) when the risk exceeds the limit or the model divergence is too large.

[0141] 14. Final Action Selection Module It is used to select the candidate action with the minimum comprehensive cost value from the set of allowed candidate actions U_allow(k) based on the comprehensive cost function Ji(k), and determine it as the final control action u_star(k).

[0142] 15. Risk Recovery Module Used to gradually restore the thermal risk budget balance B(k) after the system stabilizes.

[0143] 16. Control Output Module Used to convert u_star(k) into a pulse width modulation signal, a relay control signal, a solid-state relay control signal, or a power regulation signal.

[0144] 17. Main Controller This is used to coordinate the operation of the above modules. The main controller can be a microcontroller, an embedded control board, a programmable logic controller, or an industrial computer.

[0145] The following existing technologies can be used in this invention: 1. Temperature sensor acquisition technology; 2. Thermocouple, resistance temperature detector (RTD), or infrared temperature measurement technology; 3. Pulse width modulation power control technology; 4. Relay, solid-state relay, thyristor, or power regulator control technology; 5. Recursive least squares parameter identification technique; 6. Sliding window parameter identification technology; 7. Multilayer perceptron neural network prediction technology; 8. Proportional-integral-derivative control technology; 9. Fuzzy control technology; 10. Industrial controller data storage and communication technology.

[0146] All of the above technologies can be implemented using existing mature technologies.

[0147] The innovation of this invention lies not in proposing temperature acquisition, pulse width modulation control, recursive least squares algorithm, or neural network algorithm separately, but in combining the above-mentioned existing technologies into a new temperature safety control system for heat treatment equipment, specifically including: 1. Map the risks of residual heat surge, temperature rise momentum, charging uncertainty, model error, sensor delay, and process stage into a unified thermal risk budget account; 2. Before the actual execution of the control action, counterfactual predictions are made for multiple candidate actions to predict in advance the temperature overshoot and risk consumption that each action may cause in the future; 3. Add candidate actions that exceed the thermal risk budget balance or may cause overshoot to the action blacklist and prohibit them from being executed in the current control cycle; 4. Multiple shadow models are used to predict the same candidate action in parallel, and the reliability of the current model prediction is judged by the model divergence degree; 5. When the model divergence is too large or the risk budget is insufficient, the degradation protection is triggered, limiting the maximum heating power or forcibly shutting down the heating; 6. Once the system stabilizes again, control authority is gradually restored through a risk recovery mechanism to prevent overshoot or oscillation from occurring again after the abnormal recovery.

[0148] Therefore, this invention is not a simple improvement of proportional-integral-derivative control, nor is it ordinary model predictive control. Instead, it is a linkage control method and system for heat treatment equipment, which combines "thermal risk budget account + counterfactual action blacklist + multi-shadow model divergence adjudication + degradation protection + risk compensation".

[0149] Examples are as follows: Taking a heat treatment furnace with a target temperature of 300 degrees Celsius as an example, the sampling period dt is 1 second, and the prediction step size H is 30, meaning the system predicts the temperature change over the next 30 seconds. Assume the current time is: T(k)=285; Ts(k)= 300; Ta(k)= 25; u(k-1)= 0.8; v(k)=1.2; v_safe=0.6 The system calculations show that the current temperature rise rate is relatively high, indicating that the furnace cavity is still in a state of sub-rapid heating.

[0150] First, the system predicts the maximum future temperature after the power outage based on the thermal inertia model: T_off_max(k) = 303. Therefore, the residual heat risk is: Rh(k) = wh max(o,303-300) means that even if the heating is currently turned off, the temperature may still continue to rise to 303 degrees Celsius, so the system will increase the residual heat risk deduction. The system then generates candidate actions: U(k)={0,0.2,0.4,0.6,0.8}. The system predicts the temperature for the next 30 seconds for each candidate action.

[0151] If the prediction result is: Action 0: Maximum temperature 301 Action 0.2: Maximum temperature 303 Action 0.4: Maximum temperature 306 Action 0.6: Maximum temperature 310°C Action 0.8: Maximum temperature 315°C If the allowed temperature limit is 305 degrees Celsius, actions 0.4, 0.6, and 0.8 will be blacklisted. The system then performs multi-shadow model predictions on the remaining actions 0 and 0.2.

[0152] If the fast-response model predicts a maximum temperature of 302°C, the slow-response model predicts a maximum temperature of 304°C, and the conservative thermal storage model predicts a maximum temperature of 308°C, then the model divergence is large, indicating that the current thermal state is uncertain. The system will then restrict or disable action 0.2. Ultimately, the system selects action 0, i.e., shutting down heating, as the control output for the current cycle.

[0153] When the temperature stabilizes, the rate of temperature rise decreases, the prediction error decreases, and the model divergence decreases over subsequent sampling periods, the system gradually replenishes B(k) and allows low-power heating actions to re-enter the candidate set. As can be seen from the above embodiments, this invention can prevent high-risk heating actions in advance based on residual heat risk and model divergence risk before the furnace temperature exceeds the target temperature, thereby reducing the probability of temperature overshoot.

[0154] In each control cycle, this invention first calculates the thermal risk budget that the current system can withstand, then performs counterfactual predictions on multiple candidate control actions to determine in advance whether each action might cause overshoot or oscillation, and prohibits high-risk actions through an action blacklist. Simultaneously, it uses the prediction discrepancies between multiple shadow models to determine the reliability of the current thermal state, and proactively initiates degradation protection when predictions are inconsistent. Therefore, this invention can improve the safety, stability, and disturbance rejection capability of temperature control in heat treatment equipment under the combined effects of thermal inertia, load changes, sensor delays, and model errors.

[0155] This invention can predict future risks associated with heating actions in advance, reducing temperature overshoot; it limits control actions through a thermal risk budget account to avoid continuing to output high-risk power when approaching the target temperature; it prohibits dangerous control actions in advance through an action blacklist mechanism, rather than remedying them after overshoot; it improves the safety of the control system under model uncertainty through multi-shadow model divergence judgment; it maintains stable control even with changes in charge, sensor delays, and environmental disturbances through a degradation protection mechanism; and it enables the equipment to smoothly recover from abnormal states and reduce oscillations through a risk recovery mechanism. It does not rely on high-cost, complex models and can be implemented on low-cost controllers. This invention is applicable to various heat treatment equipment such as box furnaces, tube furnaces, pit furnaces, ovens, muffle furnaces, experimental furnaces, sintering furnaces, and drying furnaces. Compared to simple proportional-integral-derivative control or ordinary model prediction, this invention emphasizes safety adjudication, action disabling, and model uncertainty protection, making it more practical and auditable.

[0156] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification, characterized in that, Includes the following steps: Collect operating data of the heat treatment equipment, and identify the thermal response parameters of the heat treatment equipment online based on the operating data; A thermal risk budget account is established based on the operating data, the thermal response parameters, and the current process stage, and the thermal risk budget balance for the current control cycle is determined. Multiple candidate control actions are generated, and counterfactual prediction is performed on each candidate control action to calculate its heat risk consumption value; Candidate control actions that meet any preset disabling conditions are added to the action blacklist. The preset disabling conditions include at least the thermal risk consumption value of the candidate control action being greater than the thermal risk budget balance. Candidate control actions not included in the action blacklist are input into multiple shadow models for future temperature prediction, and the model divergence is calculated based on the prediction results of each shadow model. If the model divergence exceeds the preset divergence threshold, then a degradation protection control is executed, and a degradation control action is output. If the model divergence does not exceed the preset divergence threshold, then from the candidate control actions that have never been included in the action blacklist, an optimal candidate control action is selected as the final control action based on the comprehensive cost function. The final control action or the degraded control action is output to the heater for execution, thereby realizing intelligent temperature control of the heat treatment equipment based on thermal inertia dynamic identification and shadow model verification.

2. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 1, characterized in that, The thermal response parameters are identified online using a recursive least squares algorithm, a sliding window least squares algorithm, or a Kalman filter algorithm. The thermal response parameters are updated in real time based on the prediction error at the current moment. The thermal response parameters include the thermal inertia coefficient, the heating effect coefficient, and the environmental effect coefficient.

3. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 1, characterized in that, Establishing the thermal risk budget account includes: setting a baseline safety limit for the thermal risk budget account; obtaining the current temperature, target temperature, temperature rise rate, model prediction error, and sensor delay time based on the operating data; determining the corresponding process stage weight based on the current process stage; calculating the residual heat surge risk, temperature rise momentum risk, loading status uncertainty risk, model error risk, sensor delay risk, and process stage sensitivity risk based on the thermal response parameters, the difference between the current temperature and the target temperature, the temperature rise rate, the model prediction error, the sensor delay time, and the process stage weight, respectively; and deducting the calculated risks from the baseline safety limit to obtain the thermal risk budget balance.

4. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 1, characterized in that, Generating the plurality of candidate control actions includes: Obtain a preset set of candidate actions, calculate the maximum power allowed in the current control cycle based on the thermal risk budget balance, and eliminate candidate control actions whose power values ​​are greater than the maximum power from the set of candidate actions.

5. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 4, characterized in that, When performing counterfactual prediction for each candidate control action, the thermal response parameters obtained through online identification are used as inputs, including the current temperature, the candidate control action, and the ambient temperature, to predict the temperature trajectory for multiple future sampling periods, and to extract the predicted maximum temperature and the predicted maximum temperature rise rate from it.

6. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 5, characterized in that, The calculation of the thermal risk consumption value of the candidate control action includes: The heat risk consumption value is obtained by weighted summing the overshoot of the predicted maximum temperature and target temperature corresponding to the candidate action, the excess of the predicted maximum temperature rise rate and the safe temperature rise rate threshold, the loading state uncertainty calculated based on the operating data, the model prediction error, and the change range of the candidate action and the control action of the previous cycle.

7. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 1, characterized in that, The preset disabling conditions also include: The predicted maximum temperature corresponding to this candidate action exceeds the upper limit of the target temperature. The predicted maximum temperature rise rate corresponding to this candidate action exceeds the absolute safe rate limit; The power value of the candidate action exceeds the maximum power allowed in the current control cycle.

8. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 1, characterized in that, When calculating the model divergence, for the same candidate control action, the temperature trajectory predicted by each shadow model for multiple future sampling periods is obtained; For each future sampling period, calculate the difference between the maximum and minimum values ​​of the predicted temperature in each shadow model, and take the maximum value of the difference in all future sampling periods as the model divergence degree; Alternatively, for each future sampling period, calculate the average value of the predicted temperature of each shadow model, then calculate the standard deviation of the predicted temperature of each shadow model relative to the average value, and take the average value of the standard deviations in all future sampling periods as the model divergence degree. The shadow model includes at least two of the following: fast response shadow model, slow response shadow model, conservative heat storage shadow model, historical batch matching shadow model, and lightweight neural network shadow model.

9. The intelligent temperature control method for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification according to claim 3, characterized in that, It also includes: when the absolute value of temperature deviation is less than the temperature deviation threshold, the absolute value of temperature rise rate is less than the temperature rise rate threshold, the absolute value of model prediction error is less than the model error threshold, and the model divergence degree is less than the preset divergence threshold for L consecutive sampling periods, the current thermal risk budget balance is increased by one replenishment step size until the thermal risk budget balance reaches the benchmark safety limit. When the thermal risk consumption value of a candidate control action is less than the product of the safety release coefficient and the current thermal risk budget balance, and the model divergence degree of the candidate control action is less than the preset divergence threshold, the candidate control action is removed from the action blacklist.

10. An intelligent temperature control system for heat treatment equipment based on thermal inertia dynamic identification and shadow model verification, used to execute the method according to any one of claims 1-9, characterized in that, include: The temperature acquisition module, the target temperature input module, and the historical data storage module are used to acquire operational data, input the target temperature, and store historical data, respectively. The main controller, and the following modules respectively connected to the main controller: A thermal inertia dynamic identification module is used to identify thermal response parameters online based on the operating data; The loading state uncertainty calculation module is used to calculate the loading state uncertainty. The thermal risk budget module is used to establish thermal risk budget accounts and determine thermal risk budget balances; The candidate action generation module is used to generate multiple candidate control actions; The counterfactual prediction module is used to predict future temperatures for candidate control actions; The action risk calculation module is used to calculate the thermal risk consumption value of candidate control actions; The action blacklist module is used to add candidate control actions that meet any preset disabling conditions to the action blacklist. The multi-shadow model prediction module is used to predict future temperatures using multiple shadow models. The disagreement resolution module is used to calculate the model disagreement degree; Degradation protection module, used to output degradation control actions; The final action selection module is used to select the final control action; The risk replenishment module is used to replenish the remaining balance of the hot risk budget. The control output module is used to convert control actions into heater control signals.

Images