Intelligent control system of intermediate frequency power supply

Abstract

The present application relates to the technical field of power electronic converter and induction heating intelligent control, and particularly relates to a kind of medium frequency power intelligent control system, comprising: data acquisition module, configured to real-time acquisition power conversion hardware end side voltage data and end side current data;Feature extraction module, configured to receive end side voltage data and the end side current data, and extract wideband waveform distortion feature;State evaluation module, configured to receive the wideband waveform distortion feature, and based on wideband waveform distortion feature solve out model trust entropy value;Adaptive control module, configured to receive the model trust entropy value, trigger second control mode to generate second switch instruction and send to the power conversion hardware;Model trust entropy value represents the degree of deviation of the prediction of the preset robust control model on the current real physical state of the power conversion hardware;The present application solves the problem of insufficient fine-grained service parameter perception in the background art.

Description

Technical Field

[0001] This invention relates to the field of power electronic converters and intelligent control technology for induction heating, specifically to an intelligent control system for medium-frequency power supplies. Background Technology

[0002] In industrial scenarios involving megawatt-level special metallurgy and superconducting material heating, the operational stability of medium-frequency power supply systems depends not only on the real-time response of electrical parameters, but also on a deep understanding of the underlying hardware physical state and business logic data. Existing technologies mostly focus on monitoring conventional electrical quantities, similar to link tracing in distributed systems. While they can locate macroscopic faults, they are difficult to capture more granular business parameters and underlying state evolution data.

[0003] Conventional filtering circuits and sampling mechanisms often filter out weak transient components in high-frequency switching environments, making it impossible to acquire fine-grained service data that reflects specific waveform distortions such as hardware aging and thermoelectric fatigue. Monitoring methods that rely on external physical sensors have significant risks of temperature rise delay and electromagnetic failure, making it difficult to accurately restore the parasitic parameter drift inside the power module without introducing invasive hardware. When faced with nonlinear impedance changes caused by material phase transitions, existing control logic blindly pursues dynamic response speed and lacks an adaptive adjustment mechanism based on the matching degree between the software model and the actual hardware state, which can easily lead to thermal breakdown of the power module under complex interference.

[0004] Therefore, how to achieve seamless, fine-grained data acquisition of underlying parasitic parameters and macroscopic model trust in the high-frequency, high-power operating environment of medium-frequency power supplies, and how to construct a multi-level adaptive control system from performance optimization to proactive self-healing based on quantified model trust indicators, has become a key issue in improving the survival probability of power conversion hardware under harsh operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent control system for medium-frequency power supplies, which solves the following technical problems:

[0006] To avoid the system from accelerating thermal collapse of the power module under complex phase change interference due to the pursuit of dynamic response speed, and to achieve zero-delay measurement of the internal micro-physical state of the inverter without relying on external physical sensors, thereby actively suppressing micro-thermal cycling and thermoelectric fatigue accumulation, and improving the survival probability and long-term process stability of the medium frequency power supply under harsh operating conditions.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A medium-frequency power supply intelligent control system, the system comprising:

[0009] The data acquisition module is configured to acquire the end-side voltage and end-side current data of the power conversion hardware in real time.

[0010] The feature extraction module is configured to receive end-side voltage data and end-side current data, and extract wideband waveform distortion features;

[0011] The state assessment module is configured to receive broadband waveform distortion features and calculate the model trust entropy value based on the broadband waveform distortion features.

[0012] The adaptive control module is configured to receive the model trust entropy value and compare it with a preset safety threshold: when the model trust entropy value is less than or equal to the preset safety threshold, a first control mode is triggered to generate a first switching command and send it to the power conversion hardware; when the model trust entropy value is greater than the preset safety threshold, a second control mode is triggered to generate a second switching command and send it to the power conversion hardware.

[0013] Among them, the model trust entropy value represents the degree of prediction deviation of the preset robust control model on the current real physical state of the power conversion hardware;

[0014] The control objective of the first control mode is to maximize active power output and dynamic response speed, while the control objective of the second control mode is to suppress microscopic thermal cycling and thermoelectric fatigue accumulation in the power conversion hardware.

[0015] Optionally, the feature extraction module includes:

[0016] The interference separation unit is configured to use frequency domain analysis algorithms to analyze the terminal voltage data and terminal current data, and separate the fundamental characteristics and high-frequency harmonic noise caused by the coupling between the grid side and the load side.

[0017] The distortion reconstruction unit is configured to reconstruct broadband waveform distortion features based on high-frequency harmonic noise, filter out pseudo-feature data with amplitudes lower than a preset noise threshold, and then send the cleaned broadband waveform distortion features to the state assessment module.

[0018] Optionally, the state assessment module includes:

[0019] The sensorless measurement unit is configured to input the broadband waveform distortion characteristics into a preset thermoelectric mapping matrix that reflects the relationship between waveform distortion and parasitic parameters without introducing an external temperature sensor, and calculate the internal parasitic parameter drift of the power conversion hardware.

[0020] The entropy quantization unit is configured to generate model trust entropy by calculating the information entropy quantization of the probability distribution of the difference between the internal parasitic parameter drift and the baseline parameter.

[0021] Optionally, the adaptive control module includes:

[0022] The reinforcement learning unit is configured to run a deep reinforcement learning algorithm to perform dynamic resonant frequency tracking based on end-side voltage data and end-side current data.

[0023] The extreme value optimization unit is configured to calculate the duty cycle parameters that satisfy the zero-voltage turn-on condition or the zero-current turn-on condition based on the results of dynamic resonant frequency tracking, and generate the first switching command accordingly.

[0024] Optionally, the adaptive control module includes:

[0025] The downgraded operation unit is configured to abandon dynamic resonant frequency tracking and actively reduce the switching frequency of the power conversion hardware;

[0026] The thermal distribution equalization unit is configured to calculate the phase parameters for forcing the system into a derating non-resonant operating mode by adjusting the phase shift angle of the drive signal, and introduce reactive circulating current inside the power conversion hardware according to the phase parameters to equalize the thermal distribution, thereby generating a second switching command.

[0027] Optionally, the adaptive control module is also configured to execute multi-level boundary defense logic:

[0028] Set a preset critical failure threshold, and the preset critical failure threshold is greater than a preset safety threshold;

[0029] When the model trust entropy value exceeds a preset safety threshold, the second control mode is triggered, which includes:

[0030] When the model trust entropy value is greater than the preset safety threshold and less than or equal to the preset critical failure threshold, the second control mode is executed.

[0031] When the model trust entropy value is greater than the preset critical failure threshold, the third control mode is triggered to generate a blocking pulse command and send it to the power conversion hardware.

[0032] The control objective of the third control mode is to cut off energy transmission within a preset hardware failure time boundary, which is set to two milliseconds.

[0033] Optionally, the system may also include:

[0034] The trust reconstruction module is configured to collect feedback voltage and feedback current data of the power conversion hardware under the second switching command during the second control mode operation.

[0035] The model update module is configured to calculate the prediction error using feedback voltage data and feedback current data, and adaptively correct the underlying parameters of the preset robust control model based on the prediction error, and send a recovery command to the adaptive control module after the correction is completed.

[0036] The adaptive control module is also configured to: in response to a recovery command, re-compare the model trust entropy value with a preset safety threshold.

[0037] Optionally, the power conversion hardware includes: an inverter bridge circuit, which consists of multiple insulated-gate bipolar transistor modules or silicon carbide power modules;

[0038] The drive protection circuit is configured to receive a first switching command or a second switching command and convert it into a physical level signal that drives the inverter bridge circuit to turn on and off.

[0039] Optionally, a special metallurgical smelting furnace or a superconducting material heating device is connected to the load side; the end-side voltage data and end-side current data include information on the nonlinear impedance mutation caused by the special metallurgical smelting furnace or the superconducting material heating device during the phase transition process.

[0040] Optionally, the data acquisition module and adaptive control module are deployed in the field edge controller to meet the microsecond-level real-time control requirements;

[0041] The feature extraction module and the status assessment module are deployed on a cloud server and interact with the field edge controller through an industrial network.

[0042] The beneficial effects of this invention are:

[0043] 1) This invention successfully captures weak transient components that are easily filtered out by conventional circuits in megawatt-level high-frequency switching environments by using a high-frequency Rogowski coil and a wide-band piezoresistive voltage divider; combined with a multi-dimensional signal stripping algorithm and distortion reconstruction mechanism, it can effectively filter out conventional electromagnetic interference and extract specific waveform distortion features that reflect hardware aging and thermoelectric fatigue, thus solving the problem of insufficient perception of fine-grained business parameters in the background technology.

[0044] 2) By introducing thermoelectric mapping matrix and information entropy quantization technology, this invention enables the system to accurately measure the drift of parasitic parameters inside the power module without relying on external physical sensors. This contactless calculation method avoids the temperature rise delay and electromagnetic failure risk of traditional temperature sensors, and uses model trust entropy quantification to quantify the risk of software and hardware trust collapse, providing a window for early prediction of hardware damage.

[0045] 3) This invention breaks away from the traditional logic of solely pursuing electrical response speed and constructs a multi-level defense system based on model trust entropy. In the early stage of the decline in software and hardware compatibility, the system can proactively switch from performance optimization mode to a more thermodynamically stable degraded operation mode by introducing reactive circulating current to balance heat distribution. This defense mechanism, which goes from performance optimization to proactive self-healing and then to physical isolation, significantly improves the survival probability of hardware under harsh operating conditions. Attached Figure Description

[0046] The invention will now be further described with reference to the accompanying drawings.

[0047] Figure 1 This is a module architecture diagram of a medium-frequency power supply intelligent control system according to an embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figure 1 As shown, a medium-frequency power supply intelligent control system includes: a data acquisition module configured to acquire end-side voltage data and end-side current data of power conversion hardware in real time;

[0050] The feature extraction module is configured to receive end-side voltage data and end-side current data, and extract wideband waveform distortion features;

[0051] The state assessment module is configured to receive broadband waveform distortion features and calculate the model trust entropy value based on the broadband waveform distortion features.

[0052] The adaptive control module is configured to receive the model trust entropy value and compare it with a preset safety threshold: when the model trust entropy value is less than or equal to the preset safety threshold, a first control mode is triggered to generate a first switching command and send it to the power conversion hardware; when the model trust entropy value is greater than the preset safety threshold, a second control mode is triggered to generate a second switching command and send it to the power conversion hardware.

[0053] Among them, the model trust entropy value represents the degree of prediction deviation of the preset robust control model on the current real physical state of the power conversion hardware; the control objective of the first control mode is to maximize active power output and dynamic response speed, and the control objective of the second control mode is to suppress the micro-thermal cycle and thermoelectric fatigue accumulation of the power conversion hardware.

[0054] The data acquisition module acquires terminal voltage and terminal current data through a high-frequency Rogowski coil and a wide-band piezoresistive voltage divider deployed at the output of the inverter topology. The aim is to capture weak transient components that are easily filtered out by conventional filter circuits in megawatt-level high-frequency switching environments.

[0055] The feature extraction module initiates a multi-dimensional signal stripping algorithm to extract broadband waveform distortion features from massive and mixed electrical signals. This action aims to filter out conventional electromagnetic interference and purify specific waveform distortions caused by hardware aging or thermoelectric fatigue. The state assessment module performs sensorless state mapping and calculates the core indicator model trust entropy value. Here, the specific term model trust entropy is introduced. In the specific technical environment of this invention, this term is different from the uncertainty measure in traditional information theory. It represents a dynamic indicator that quantifies the degree of accumulation of software and hardware trust collapse risk. It reflects the failure probability of the control system by measuring the model distortion caused by the drift of the underlying parasitic parameters.

[0056] Based on this, the adaptive control module acts as the decision-making center, executing dynamic hedging logic under extreme operating conditions. Specifically, regarding the setting of the preset safety threshold, the system uses an accelerated aging test bench to obtain empirical waveform data when the power hardware is in a critical thermal equilibrium state during the factory calibration stage and calculates the corresponding limit trust entropy value. This value is then multiplied by a preset safety reduction factor, which is set to a range of 0.75 to 0.85. Its significance lies in reserving a time margin for the junction temperature heat conduction delay of the underlying power module.

[0057] The specific method for obtaining the value is as follows: before the system leaves the factory, a destructive step load test is performed on the same batch of power conversion hardware to determine the average delay time from when the end-side signal shows an abnormal waveform to when the hardware experiences substantial thermal breakdown. The derating ratio is then calculated based on the ratio of this delay time to the system control cycle, thereby establishing the steady-state safety boundary.

[0058] When the model trust entropy value is less than or equal to a preset safety threshold, the system determines that the current hardware and software matching degree is high, triggering the first control mode to maximize the active power output to ensure the efficient advancement of the heating process. Conversely, when the model trust entropy value is greater than the preset safety threshold, the system determines that it is in a high-risk steady state, and the traditional optimization algorithm will become a catalyst for accelerating hardware damage. The system decisively triggers the second control mode to actively suppress microscopic thermal cycling and thermoelectric fatigue accumulation. This embodiment introduces the model trust entropy value as a hard criterion for switching multi-dimensional control modes. In megawatt-level special metallurgy and superconducting material heating scenarios, it successfully solves the long-standing problem of the system's pursuit of dynamic response speed leading to accelerated thermal collapse of power modules under complex phase transition interference. In the early stage of the decline in hardware and software trust degree, the system actively abandons the optimal solution in terms of electrical parameters and executes a more thermodynamically stable suboptimal solution, improving the survival probability of the intermediate frequency power supply under harsh operating conditions.

[0059] In a preferred embodiment of the present invention, the feature extraction module includes: an interference separation unit, configured to analyze the terminal voltage data and terminal current data using a frequency domain analysis algorithm to separate the fundamental wave characteristics and high-frequency harmonic noise caused by the coupling between the grid side and the load side;

[0060] The distortion reconstruction unit is configured to reconstruct broadband waveform distortion features based on high-frequency harmonic noise, filter out pseudo-feature data with amplitudes lower than a preset noise threshold, and send the cleaned broadband waveform distortion features to the state assessment module.

[0061] The interference separation unit receives raw electrical array data from the data acquisition module and uses multi-resolution wavelet transform to analyze the asymmetric coupled interference frequency band, separating the fundamental wave characteristics and high-frequency harmonic noise.

[0062] The nonlinear mutations caused by load phase changes can cause complex cross-modulation with the background harmonics of the power grid. Conventional Fourier transforms are difficult to capture transient mutations. Therefore, the introduction of frequency domain analysis algorithms can accurately locate the core frequency band of high-frequency harmonic storms. The distortion reconstruction unit reconstructs the wideband waveform distortion characteristics through integral operations and executes dynamic threshold cleaning logic.

[0063] If the extracted feature amplitude is lower than a preset noise threshold, the system determines it to be spurious feature data and removes it, aiming to prevent the underlying control algorithm from being misled by invalid noise and causing optimization oscillations; the system executes the following feature reconstruction model:

[0064]

[0065] During this feature reconstruction process, broadband waveform distortion features The integral calculation result of the distortion reconstruction unit characterizes the degree of distortion accumulation;

[0066] Specifically, in the formula The integral element representing the continuous time variable in the integration process, measured in seconds, is the integral of the system along this continuous time variable. The above describes the extraction of transient high-frequency voltage harmonic noise envelope amplitude from the terminal voltage data by the interference separation unit. A preset noise threshold with the same voltage dimension as calibrated by on-site electromagnetic compatibility benchmark testing. The net difference between them is accumulated over time.

[0067] It should be noted that, to prevent numerical shifts caused by low-amplitude noise, the system only applies the transient envelope amplitude. Strictly greater than the preset noise threshold The difference is only included in the integral when the condition is met; otherwise, the value of the infinitesimal element is zero, thus maintaining consistency with the linear rectification logic described later in the mathematical expression.

[0068] Meanwhile, for the end-side current data, the system uses parallel integral logic to calculate the current distortion characteristics in units of A·s. The two together constitute multi-dimensional broadband waveform distortion characteristics. In order to accurately execute the logic of filtering out false feature data at the code level and prevent the underlying algorithm from generating negative accumulation that leads to feature amplitude distortion, the system applies nonlinear truncation judgment processing to the net difference value before executing the integral.

[0069] Specifically, within each microsecond-level control clock step, the system performs real-time comparisons. and The size, only when Strictly greater than Only when the time is right will the difference be included in the cumulative calculation; if Less than or equal to If the difference under the current microelement is forced to be zero, this logical operation, which is equivalent to linear rectification, ensures from the physical calculation level that invalid high-frequency harmonics below the noise threshold will not produce any tiny numerical offset accumulation in the wideband waveform distortion feature reconstruction.

[0070] The integration interval of this operation is strictly defined within the start time of the control cycle provided by the underlying timer trigger signal. The control cycle end time provided by the reset signal between;

[0071] In this embodiment, the control cycle is the PWM switching cycle of the inverter, and its value range is set between 100μs and 500μs according to process requirements.

[0072] Specifically, considering that the feature extraction module is deployed on a cloud server in the subsequent architecture, when the field edge controller collects end-side voltage data and end-side current data, its internal high-frequency hardware timer adds a microsecond-level high-precision timestamp to each sampling point, and packages and uploads the synchronization flag bit calibrated with the timestamp of the PWM carrier zero point and carrier peak or the next zero point together with the electrical data. This allows the cloud to accurately restore the time boundary of the physical control cycle based on the timestamp when performing integral calculations, thereby ensuring the timing synchronization of the cleaning process.

[0073] By using interference separation and distortion reconstruction mechanism based on preset noise threshold, the pollution of the core control link caused by high-frequency harmonic storms caused by coupling between the network side and the load side is effectively isolated in special metallurgical smelting scenarios.

[0074] Filtering out spurious feature data cuts off dirty data input that induces high-frequency invalid frequency tuning in artificial intelligence models, thus stabilizing the working environment of the underlying hardware.

[0075] In a preferred embodiment of the present invention, the state assessment module includes: a sensorless measurement unit configured to input broadband waveform distortion characteristics into a preset thermoelectric mapping matrix reflecting the correlation between waveform distortion and parasitic parameters, and calculate the internal parasitic parameter drift of the power conversion hardware without introducing an external temperature sensor; and an entropy quantization unit configured to generate a model trust entropy value by calculating the information entropy quantization of the probability distribution of the difference between the internal parasitic parameter drift and the reference parameter.

[0076] The sensorless measurement unit acquires the broadband waveform distortion characteristics after cleaning and maps them to a preset thermoelectric mapping matrix. This aims to avoid the delay and failure risk of physical temperature sensors in the case of a small internal space and harsh electromagnetic environment of a high-frequency, high-power inverter.

[0077] Specifically, the thermoelectric mapping matrix The construction process is as follows: In the offline experiment phase, different ambient temperatures are simulated through a controllable environment test chamber and the power conversion hardware status at different aging stages is simulated through accelerated aging tests.

[0078] Specific excitation signals are injected under various states, and wideband waveform distortion characteristics are acquired synchronously. and the corresponding drift of the internal true parasitic parameters measured by a precision impedance analyzer. This forms a training sample set;

[0079] Minimize the sum of squared residuals using the least squares method Thus, a matrix is ​​fitted and generated. ;

[0080] In order to make the scalar features of the output of the previous stage meet the dimensionality requirements of matrix operations, the system performs sliding window sampling on the broadband waveform distortion feature scalars extracted under multiple consecutive control steps in the current diagnostic cycle and splices them according to the time sequence, thereby constructing a multi-dimensional input vector with time evolution information;

[0081] Specifically, if the monitored internal parasitic parameters are dimensional That is, parasitic resistance, stray inductance, and junction capacitance, and the control step size selected by the sliding window. Then the thermoelectric mapping matrix for A real matrix, where each row vector represents the sensitivity weight of the corresponding parasitic parameter to the waveform distortion characteristics of the past 10 cycles;

[0082] The system multiplies the multidimensional input vector with the thermoelectric mapping matrix to directly extract the thermal and aging effects from the electrical distortion and calculate the drift of internal parasitic parameters. Specifically, the system performs the following thermoelectric mapping matrix multiplication operation:

[0083]

[0084] in, The parsed internal parasitic parameter drift column vector has dimensions of , In this embodiment, the dimension of the monitored internal parasitic parameter is... Includes parasitic resistance drift, in units of Stray inductance drift, in units of and junction capacitance drift, in units of ; For continuous A multi-dimensional input column vector, composed of wideband waveform distortion feature scalars sampled and concatenated by a sliding window under each control step size, has a dimension of [missing information]. ; This is a thermoelectric mapping matrix, with dimensions set to 1. ;

[0085] This matrix maps the distortion features of the electrical dimension to the physical parameter dimension, thereby enabling a quantitative characterization of the degree of hardware aging.

[0086] Based on this, the information entropy of the probability distribution of the difference is calculated to generate the model trust entropy value; the system executes the following entropy value calculation model:

[0087]

[0088] in, The entropy value for model trust is calculated based on the entropy quantization unit and is dimensionless. The total number of parasitic parameter dimensions preset for the system, and the corresponding internal parasitic parameter drift vector. The dimension; For traversal indexes of the parameter dimensions; Indicates the first Probability distribution components in each dimension; For example, a preset minimum positive compensation factor, such as This is used to avoid the dead zone in logarithmic calculations; specifically, the initial proportional values ​​for calculating the drift of parameters in each dimension are:

[0089]

[0090] in, For the first The historical maximum deviation limit of the dimension parameters. This limit is determined during the system factory calibration stage. Specifically, it is obtained by conducting accelerated aging tests on the same type of power conversion hardware, and measuring the maximum drift values ​​of parasitic resistance, stray inductance, and junction capacitance when the hardware undergoes substantial performance degradation, which are then used as fixed reference values.

[0091] During system operation, this limit is stored as a constant in non-volatile memory and is not adjusted with real-time fluctuations, thus providing a stable normalized denominator; it is calculated by taking the absolute value of the real-time drift in the current dimension. Divide by the historical maximum deviation limit Obtain the initial ratio value ;

[0092] Its initial calculation method is to divide the difference of the current dimension by the historical maximum deviation limit; the ratio obtained by direct division does not necessarily meet the mathematical requirement that the sum of the probability distributions is 1, and when the parameters of a certain dimension have not drifted so that the difference is zero, substituting into the logarithmic function will cause the underlying microprocessor to trigger a numerical overflow exception.

[0093] Therefore, the system logically encapsulates this calculation during code execution: it performs a Softmax nonlinear mapping calculation on the initial scale values ​​of each dimension to ensure that all... The sum is strictly equal to 1, thus satisfying the prerequisite mathematical condition for information entropy calculation; to prevent extremely small probability values ​​from causing the logarithmic function calculation to collapse, the system... It internally embeds a preset, extremely small positive compensation factor, for example... Level, which is calculated when performing logarithmic operations. This smoothing process avoids division-by-zero faults in the microcontroller's calculation process, ensuring that the model trust entropy value can output a stable and convergent minimum safe value when the underlying hardware is completely healthy.

[0094] This embodiment abandons the traditional lagging monitoring method that relies on external physical sensors, and realizes the sensorless and zero-delay measurement of the internal microscopic physical state of the inverter in the superconducting material heating scenario; by using information entropy, the complex multidimensional drift of parasitic parameters is reduced to a single quantitative feedforward index, enabling the system to predict and intervene in advance before irreversible damage to the hardware occurs.

[0095] In a preferred embodiment of the present invention, in the first control mode, the adaptive control module includes: a reinforcement learning unit configured to run a deep reinforcement learning algorithm to perform dynamic resonant frequency tracking based on end-side voltage data and end-side current data; and an extreme value optimization unit configured to calculate the duty cycle parameter that satisfies the zero-voltage turn-on condition or the zero-current turn-on condition based on the result of the dynamic resonant frequency tracking, and generate a first switching command accordingly.

[0096] Upon receiving an instruction that the model trust entropy value is within a safe range, the reinforcement learning unit activates the deep reinforcement learning algorithm network. Specifically, the system employs a deep deterministic policy gradient algorithm, defining the time-series of continuously collected end-side voltage and current data as the state space, defining the fine-tuning compensation value of the inverter switching frequency as the action space, and using the weighted sum of maximizing active power output and minimizing switch conduction losses as the reward function. The system executes the following reward function model:

[0097]

[0098] in, Current control step size The instantaneous reward value is dimensionless. This refers to active power output. The total loss of the switching transistor is estimated in real time, including the sum of conduction loss and switching loss;

[0099] This is a preset active power reference value, which is usually set to the rated output power of the system and is used to normalize the real-time output power.

[0100] The system uses terminal voltage and current data, combined with the static current-voltage characteristic curves of power devices, to analyze these data. vs Calculate the conduction loss and use the instantaneous value at the moment of current turn-off. Estimating switching losses based on the voltage rise slope. This is the maximum allowable loss limit for the system;

[0101] and These are the active power weighting coefficient and the loss penalty weighting coefficient, respectively. Their specific values ​​are obtained through offline calibration at different heating stages of the special metallurgical process. For example, during the rapid melting period, in order to pursue the ultimate heating speed, [the following values ​​are used]. Set to 0.8. Set to 0.2, and adjust to [value] during the insulation period. Thus, by dynamically configuring weight coefficients, the deep deterministic strategy gradient algorithm is guided to find the optimal action compensation strategy under different process constraints.

[0102] Based on the reinforcement learning environment constructed above, the system continuously explores and executes dynamic resonant frequency tracking, aiming to fully unleash the exploratory capabilities of artificial intelligence and approximate the system's extreme efficiency point through high-frequency fine-tuning; the extreme value optimization unit extracts the convergence result of frequency tracking, that is, obtains the optimal switching frequency output by the deep reinforcement learning network.

[0103] The extreme value optimization unit substitutes the optimal switching frequency as a known constraint into the preset resonant tank circuit impedance model to calculate the dead time and conduction ratio required for the resonant tank circuit to exhibit slight inductive or slight capacitive behavior at this optimal frequency. Specifically, the extreme value optimization unit executes the following analytical model based on the impedance phase angle for dead time and duty cycle:

[0104]

[0105] in, For the resonant tank circuit at the optimal angular frequency That is, the optimal switching frequency multiplied by The complex impedance below, , , These are the equivalent resistance, equivalent inductance, and equivalent capacitance of the tank circuit, which are dynamically updated with the load phase change; to ensure zero-voltage turn-on, the system extracts the impedance phase angle. And calculate the required dead time based on the phase angle constraint conditions. With conduction duty cycle :

[0106]

[0107]

[0108] in, The equivalent output junction capacitance of the inverter module. This is the DC bus voltage. This represents the peak value of the resonant current at the turn-off moment. This is a preset safety dead zone margin based on the gate characteristics of the inverter module. This preset safety dead zone margin is determined by offline consultation of the power device datasheet combined with hardware propagation delay measurement, for example, by setting it to... nanosecond The switching period corresponds to the optimal switching frequency; the system uses this duty cycle parameter. Dead time Together with the optimal switching frequency, a first switching command is synthesized and sent to the physical execution layer;

[0109] Under the premise of ensuring sufficient safety margin of the underlying physical hardware, the first control mode accurately locks the zero-voltage turn-on or zero-current turn-on state through the synergy of deep reinforcement learning and extreme value optimization. In megawatt-level special metallurgical processes, this mechanism reduces switching losses, takes into account both active power output and dynamic response speed, and ensures the accurate execution of the heating process curve.

[0110] In a preferred embodiment of the present invention, in the second control mode, the adaptive control module includes: a degraded operation unit configured to abandon dynamic resonant frequency tracking and actively reduce the switching frequency of the power conversion hardware; and a thermal distribution equalization unit configured to calculate the phase parameters for forcing the system into a derated non-resonant operation mode by adjusting the phase shift angle of the drive signal, and to introduce reactive circulating current inside the power conversion hardware according to the phase parameters to equalize the thermal distribution and generate a second switching command.

[0111] In response to the warning signal that the model trust entropy value exceeds the limit, the downgraded operation unit immediately freezes all deep reinforcement learning exploration processes and actively abandons dynamic resonant frequency tracking; the system forcibly reduces the switching frequency of the power conversion hardware, aiming to directly reduce the main heat source of switching losses and curb thermodynamic divergence.

[0112] The thermal distribution equalization unit obtains the current operating status and calculates the phase parameters that force the system to enter the derating non-resonant operating mode by adjusting the phase shift angle of the drive signal. Based on this, the system introduces reactive circulating current inside the power conversion hardware according to the phase parameters, and uses the flow of reactive power to force equalize the thermal distribution of each power module, and generates a second switching command.

[0113] Specifically, in the phase parameter When converted into a specific second switching command, the system employs full-bridge phase-shift control logic for the inverter bridge circuit, setting the reference turn-on phase of the lead-bridge arm drive signal to be... The relative turn-on phase of the lagging bridge arm drive signal is set to... The system quantifies and controls the amplitude of the introduced reactive circulating current based on the formula:

[0114]

[0115] in, This is the DC bus voltage. To actively reduce the switching angle frequency, For the series stray inductance of the loop; by precisely applying the relative delay time for the hysteresis arm. This results in the generation of asymmetric reactive current circulation guidance at the moment of micro-commutation;

[0116] The heat distribution equalization unit calculates the phase parameters in radians. At that time, the state index integrates information theory dimensions and physical parasitic dimensions; specifically, the system executes the following phase adjustment model:

[0117]

[0118] Among them, the entropy sensitivity coefficient The method for obtaining the data is as follows: In an offline environment, a step-type thermal load is applied to an inverter of the same model, and the model trust entropy value is recorded. The relationship between the change and the local hot spot temperature rise was determined by linear regression to minimize the phase shift compensation gain of the hot spot temperature rise slope.

[0119] Drift compensation coefficient The method of obtaining the value is as follows: during the system factory benchmark calibration stage, by changing the phase shift angle of the bridge arm drive signal, the phase adjustment required to offset the reactive bias caused by a specific unit proportion of parasitic parameter drift is measured and used as the conversion gain.

[0120] To ensure that multi-dimensional hardware changes at the underlying level can be accurately quantified and to maintain the overall dimensional balance of the formula, this comprehensive modulus... The specific calculation method is as follows: extract the relative drift of parasitic parameters in each dimension output by the sensorless measurement unit, that is, the dimensionless percentage value obtained by dividing the actual physical drift by the corresponding reference parameter, calculate the L2 norm of the vector formed by these relative drifts, that is, the square root of the sum of squares of the relative drifts in each dimension, so as to reduce the multi-dimensional hardware aging characteristics into a single dimensionless scalar index.

[0121] This allows subsequent phase parameter calculations to comprehensively and evenly reflect the overall health status of the system's underlying structure.

[0122] The second control mode constructs an anti-fragile defense mechanism to cope with extreme harmonic interference. In the scenario where the phase transition of superconducting materials causes violent fluctuations, the system actively degrades its operation and cleverly uses phase parameters to introduce reactive circulating current, breaking the cumulative effect of local hot spots. This spatial thermal distribution balance curbs the deterioration trend of thermoelectric fatigue, provides a self-healing window for the underlying hardware, and avoids catastrophic physical collapse.

[0123] In a preferred embodiment of the present invention, the adaptive control module is further configured to execute multi-level boundary defense logic: setting a preset critical failure threshold, wherein the preset critical failure threshold is greater than a preset safety threshold;

[0124] When the model trust entropy value is greater than the preset safety threshold, the second control mode is triggered, which specifically includes: when the model trust entropy value is greater than the preset safety threshold and less than or equal to the preset critical failure threshold, the second control mode is executed.

[0125] When the model trust entropy value is greater than the preset critical failure threshold, the third control mode is triggered to generate a blocking pulse command and send it to the power conversion hardware; the control objective of the third control mode is to cut off energy transmission within the preset hardware failure time boundary, which is set to two milliseconds.

[0126] The adaptive control module initializes the parameter dictionary of the multi-level boundary defense logic in memory and sets a preset critical failure threshold. Specifically, the preset critical failure threshold is determined as follows: during the system development phase, extreme thermal stress cycle tests are applied to the same type of power conversion hardware, and the peak value of the model trust entropy value is recorded just before the chip solder layer undergoes substantial fatigue peeling. To ensure the reliability of the defense action, 0.9 times this peak value is taken as the preset critical failure threshold, for example, calibrated to 4.5, dimensionless.

[0127] The system monitors the change trajectory of the model trust entropy value in real time; when the model trust entropy value is greater than the preset safety threshold and less than or equal to the preset critical failure threshold, the system determines that it is currently in a controllable thermal fatigue period, smoothly transitions, and executes the second control mode to actively reduce derating and achieve thermal equilibrium; when extreme sudden interference causes the model trust entropy value to instantly break through the preset critical failure threshold, the system determines that the underlying hardware is on the verge of irreversible damage.

[0128] Based on this, the system bypasses all conventional algorithm links and directly triggers the third control mode; the system generates and issues a blocking pulse command within two milliseconds of the hardware damage time boundary, forcibly cutting off all energy transmission paths;

[0129] The two-millisecond hardware failure time boundary is derived from the short-circuit withstand time of the underlying insulated gate bipolar transistor or silicon carbide power module when a desaturation short-circuit fault occurs, combined with the overall system heat accumulation effect.

[0130] In the case of sudden phase change short circuit in special metallurgy, due to the limitation of current rise rate by the front-end filter inductor, the system has been confirmed by thermodynamic simulation that the time window from the occurrence of short circuit to the junction temperature exceeding the absolute damage red line of 175 degrees Celsius is about 2.5 milliseconds. Therefore, this boundary is strictly set to two milliseconds to reserve 0.5 milliseconds for hardware driver shutdown and physical arc extinguishing margin, so as to ensure that cutting off energy transmission within this boundary can absolutely avoid catastrophic physical explosion.

[0131] The multi-level boundary defense logic constructs three lines of defense: performance optimization, proactive degradation and self-healing, and absolute physical isolation. In extreme scenarios where a sudden phase transition short circuit occurs in a special metallurgical smelting furnace, the hardware damage time boundary is strictly limited to within two milliseconds, ensuring that pulse blocking can be executed with a microsecond-level response speed, thus preventing systemic disasters such as the explosion of expensive power modules and the scrapping of the entire furnace of special materials.

[0132] In a preferred embodiment of the present invention, the system further includes: a trust reconstruction module, configured to collect feedback voltage data and feedback current data of the power conversion hardware under the drive of a second switching command during the operation of the second control mode;

[0133] The model update module is configured to calculate the prediction error using feedback voltage data and feedback current data, and adaptively correct the underlying parameters of the preset robust control model based on the prediction error. After the correction is completed, it sends a recovery command to the adaptive control module. The adaptive control module is also configured to: in response to the recovery command, re-compare the model trust entropy value with the preset safety threshold.

[0134] During the stable thermodynamic window of the system in degraded operation, the trust reconstruction module acquires feedback voltage and current data from the power conversion hardware through sampling channels to extract the true baseline characteristics of the hardware after parameter drift. The model update module receives the above feedback data and calculates the forward prediction error by combining it with the expected output of the preset robust control model. Specifically, the preset robust control model is a physical mechanism model based on the prior circuit topology and is described by continuous-time state-space equations.

[0135]

[0136]

[0137] Wherein, the state vector Including inductor current and capacitor voltage, input vector The output vector is the duty cycle drive signal parsed from the second switching instruction. That is, by and The desired voltage and desired current vectors are formed; the system matrix. All are composed of multi-dimensional vectors of underlying parameters including equivalent parasitic resistance, stray inductance, and parasitic capacitance. Parameterized construction;

[0138] This model is used to predict the desired terminal voltage and current under a given second switching command. The system subtracts the desired voltage and current from the actual feedback voltage and current data. After obtaining the original error vector, directly calculating its partial derivative would cause gradient oscillation due to the cancellation of positive and negative errors. Therefore, the system further calculates the squared L2 norm of the voltage and current difference in the original error vector, thus constructing a scalar mean square error as the target loss function for the prediction error. The specific expression of the target loss function constructed by the system is as follows:

[0139]

[0140] in, The mean square error is the prediction error in scalar form. and The collected feedback voltage data and feedback current data vectors, and To pre-determine the desired voltage and desired current vectors output by the robust control model, It is the impedance matching dimension normalization coefficient. Its specific value is set offline based on the square of the ratio of the system's rated voltage to the rated current. It is used to eliminate the difference in absolute numerical magnitude between voltage and current and prevent subsequent gradient updates from being dominated by a single physical quantity.

[0141] Based on the prediction error in this scalar form, the system performs adaptive correction on the underlying parameters of the preset robust control model along the gradient inverse direction, aiming to realign the parameter benchmarks of the hardware and software.

[0142] Based on this, after the correction is completed, the system automatically sends a recovery command to the adaptive control module; the adaptive control module responds to the recovery command, recalculates and compares the model trust entropy value with the preset safety threshold;

[0143] In response to the entropy value falling back to a safe range, the system removes the downgrade lock and smoothly switches back to the first control mode; the system updates the model with the following underlying parameters:

[0144]

[0145] in, The modified multi-dimensional vector of underlying parameters of the preset robust control model is calculated by the model update module in the current iteration cycle.

[0146] The underlying parameter multidimensional vector of the pre-set robust control model before correction;

[0147] Let be the gradient vector of the prediction error, with dimensions of... ,in The physical dimensions representing the corresponding underlying parameters;

[0148] The learning rate step size parameter matrix is ​​preset; the system performs calculations by... and The product ensures dimensional homogeneity;

[0149] In terms of code implementation, the system relies on offline pre-compiled automatic differentiation logic to calculate the partial derivatives of the prediction error with respect to each underlying parameter based on the feedback signal collected at the current moment, in order to construct the gradient vector. Multiply by the learning rate step size parameter matrix mentioned above And by performing a subtraction operation, it is ensured that the multidimensional vector of underlying parameters evolves smoothly in the direction of minimizing the sum of squared residuals;

[0150] For the initial vector of the underlying parameters The initial values ​​are obtained through static calibration using an offline impedance analyzer before the system leaves the factory. For example, if the initial stray inductance of a certain batch of hardware is measured to be 15 nanohenries and the initial parasitic resistance to be 2.5 milliohms, the system uses these physical measurements as... The initial reference is stored in non-volatile memory and loaded into the memory cache during each power-on initialization.

[0151] The learning rate step size parameter matrix is ​​predefined, with its main diagonal elements corresponding to the learning rates of different underlying parameters. Specifically, since the multi-dimensional vector of underlying parameters contains parameters with vastly different physical quantities, such as equivalent parasitic resistance at the milliohm level and stray inductance at the Nahen level, using a single scalar learning rate would lead to extremely unbalanced gradient updates. Therefore, the system constructs... At that time, it is set as a diagonal matrix with the same dimensions as the underlying parameter multidimensional vector. The initial value of each diagonal element is proportional to the order of magnitude of the historical baseline value of the corresponding parameter. For example, for The resistance parameter is on the order of magnitude, and its corresponding learning rate is empirically tuned to the diagonal elements as follows: Level, and targeting The inductance parameters are on the order of magnitude, and the corresponding diagonal elements of the learning rate are tuned to... The system is designed to control the equalization rate of parameter iteration updates at different physical magnitudes and prevent the correction process from diverging.

[0152] In addition, to ensure the homogeneity of the dimensions on both sides of the formula, the learning rate step size parameter matrix... Each diagonal element has a built-in dimension transformation property at the physical level. Its specific dimension is set as the ratio of the square of the physical dimension of the corresponding underlying parameter to the square of the voltage dimension, thus strictly canceling out the prediction error gradient vector in mathematical multiplication operations. Its dimension is the square of the voltage divided by the dimension of the parameter. The dimension it carries ensures that the calculated updated compensation amount is consistent with the original multidimensional vector of the underlying parameters. Maintain absolute consistency in dimensions;

[0153] The gradient vector of the prediction error is actually the Jacobian matrix vector formed by taking the partial derivatives of the mean square error target loss function constructed above for each of the underlying parameters of the preset robust control model, such as equivalent parasitic resistance and stray inductance.

[0154] In terms of code implementation, the system relies on offline pre-compiled automatic differentiation logic to solve for partial derivative values ​​one by one based on the feedback signals collected at the current moment, and then multiply them by the aforementioned learning rate step size parameter matrix. The subtraction operation ensures that the multidimensional vector of underlying parameters evolves smoothly in the direction of minimizing the sum of squared residuals between the real feedback data and the output of the prediction model; the trust reconstruction and model update mechanism endows the medium frequency power supply system with closed-loop self-healing capability.

[0155] In the scenario of continuous heating of superconducting materials, the system utilizes the window period of degraded operation to realign the hardware and software parameters and dynamically reconstruct model trust, enabling the system to smoothly recover to a high-power output state after experiencing extreme disturbances and maintain long-term process stability.

[0156] In a preferred embodiment of the present invention, the power conversion hardware includes: an inverter bridge circuit, which is composed of multiple insulated gate bipolar transistor modules or silicon carbide power modules; and a drive protection circuit, configured to receive a first switching command or a second switching command and convert them into physical level signals that drive the inverter bridge circuit to turn on and off.

[0157] As the physical framework for megawatt-level energy conversion, the inverter bridge circuit uses multiple insulated gate bipolar transistor modules or silicon carbide power modules to construct the topology, aiming to provide extremely high upper limits of switching frequency and power density.

[0158] The drive protection circuit receives the first or second switching command from the adaptive control module through a high-speed optocoupler isolation channel; the system amplifies the weak digital command and converts it into a physical level signal sufficient to overcome the Miller capacitance effect, accurately driving the inverter bridge circuit to turn on and off.

[0159] Using wide-bandgap semiconductor power modules such as silicon carbide as the core of the inverter bridge, combined with high-frequency drive protection circuits, provides a solid physical execution foundation for upper-layer anti-fragile intelligent algorithms. In megawatt-level special metallurgical scenarios, it ensures that phase-shifting commands that introduce reactive circulating currents and blocking pulse commands within two milliseconds can be accurately dispatched into high-power electrical energy without distortion, guaranteeing the absolute responsiveness of the hardware layer.

[0160] In a preferred embodiment of the present invention, a special metallurgical smelting furnace or a superconducting material heating device is connected to the load side; the end-side voltage data and end-side current data include information on nonlinear impedance mutations caused by the special metallurgical smelting furnace or the superconducting material heating device during the phase transition process.

[0161] The system rigidly connects the load side to a special metallurgical smelting furnace or a superconducting material heating device at the physical interface layer; during the heating process, the material inside the furnace inevitably undergoes solid-liquid phase transition or lattice reconstruction; these microscopic physical morphological changes are directly mapped to drastic fluctuations in macroscopic electrical parameters.

[0162] During this period, the terminal voltage data and terminal current data acquired by the data acquisition module naturally carry information about the nonlinear impedance change caused by the phase transition process.

[0163] The connection to special metallurgical smelting furnaces or superconducting material heating devices reveals the physical root of harmonic storms coupled between the grid side and the load side. The nonlinear impedance mutation caused by material phase transition highlights the vulnerability of pursuing electrical response alone, and verifies the necessity and robustness of the technical solution to adopt multi-level defense and adaptive degradation control based on model trust entropy in extreme material processing scenarios.

[0164] In a preferred embodiment of the present invention, the data acquisition module and the adaptive control module are deployed in the field edge controller to meet the real-time control requirements at the microsecond level; the feature extraction module and the state evaluation module are deployed in the cloud server and interact with the field edge controller through the industrial network.

[0165] The system embeds the data acquisition module and adaptive control module into the field edge controller, which is closely connected to the power conversion hardware. This is intended to leverage the parallel processing advantages of the field programmable gate array (FPGA) to meet the microsecond-level hard real-time control requirements. The system deploys the feature extraction module and state evaluation module, which involve intensive matrix operations, in a cloud server with massive computing power. The field edge controller uploads the cleaned underlying electrical data to the cloud via a low-latency industrial Ethernet network.

[0166] Based on this, the cloud server completes broadband feature reconstruction and information entropy quantization calculation, and sends the calculated model trust entropy value to the field edge controller;

[0167] In addition, the field edge controller has a built-in communication disconnection protection mechanism. If it does not receive the updated entropy value from the cloud within the preset communication timeout threshold, the edge controller will automatically use the most recent valid model trust entropy value or force a switch to the second control mode to ensure that the hardware is in a thermodynamically safe state.

[0168] The field edge controller independently executes instructions and extreme protection actions such as pulse blocking based on this entropy value;

[0169] By adopting an asymmetric collaborative computing architecture between the edge and the cloud, the inherent contradiction between the huge computing power overhead of complex artificial intelligence algorithms and the microsecond-level hard real-time requirements of the underlying power electronic control is resolved in the industrial site of superconducting material heating. The edge safeguards the bottom line of physical security and the accuracy of high-frequency execution, while the cloud provides computing power support for macroscopic state assessment, ensuring the efficient operation of the entire antifragile control architecture.

[0170] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A medium-frequency power supply intelligent control system, characterized in that, The system includes: The data acquisition module is configured to acquire the end-side voltage and end-side current data of the power conversion hardware in real time. The feature extraction module is configured to receive the end-side voltage data and the end-side current data, and extract wideband waveform distortion features; The state assessment module is configured to receive the broadband waveform distortion features and calculate the model trust entropy value based on the broadband waveform distortion features. An adaptive control module is configured to receive the model trust entropy value and compare the model trust entropy value with a preset safety threshold: when the model trust entropy value is less than or equal to the preset safety threshold, a first control mode is triggered to generate a first switching command and send it to the power conversion hardware; when the model trust entropy value is greater than the preset safety threshold, a second control mode is triggered to generate a second switching command and send it to the power conversion hardware. The model trust entropy value represents the degree of prediction deviation of the preset robust control model on the current real physical state of the power conversion hardware. The control objective of the first control mode is to maximize active power output and dynamic response speed, while the control objective of the second control mode is to suppress microscopic thermal cycling and thermoelectric fatigue accumulation of the power conversion hardware. The status assessment module includes: The sensorless measurement unit is configured to input the broadband waveform distortion characteristics into a preset thermoelectric mapping matrix that reflects the correlation between waveform distortion and parasitic parameters without introducing an external temperature sensor, and calculate the internal parasitic parameter drift of the power conversion hardware. The entropy quantization unit is configured to generate the model trust entropy value by calculating the information entropy quantization of the probability distribution of the difference between the internal parasitic parameter drift and the reference parameter.

2. The intelligent control system for medium-frequency power supply according to claim 1, characterized in that, The feature extraction module includes: The interference separation unit is configured to use a frequency domain analysis algorithm to analyze the terminal voltage data and the terminal current data, and separate the fundamental characteristics and high-frequency harmonic noise caused by the coupling between the grid side and the load side. The distortion reconstruction unit is configured to reconstruct the broadband waveform distortion features based on the high-frequency harmonic noise, filter out pseudo-feature data with amplitudes lower than a preset noise threshold, and then send the cleaned broadband waveform distortion features to the state evaluation module.

3. The intelligent control system for medium-frequency power supply according to claim 2, characterized in that, In the first control mode, the adaptive control module includes: The reinforcement learning unit is configured to run a deep reinforcement learning algorithm to perform dynamic resonant frequency tracking based on the end-side voltage data and the end-side current data. The extreme value optimization unit is configured to calculate the duty cycle parameter that satisfies the zero voltage turn-on condition or the zero current turn-on condition based on the result of the dynamic resonant frequency tracking, and generate the first switching command accordingly.

4. The intelligent control system for medium-frequency power supply according to claim 3, characterized in that, In the second control mode, the adaptive control module includes: The downgraded operation unit is configured to abandon the execution of the dynamic resonant frequency tracking and actively reduce the switching frequency of the power conversion hardware; The thermal distribution equalization unit is configured to calculate the phase parameters for forcing the system into a derating non-resonant operating mode by adjusting the phase shift angle of the drive signal, and to introduce reactive circulating current inside the power conversion hardware according to the phase parameters to equalize the thermal distribution, thereby generating the second switching command.

5. The intelligent control system for medium-frequency power supply according to claim 1, characterized in that, The adaptive control module is also configured to execute multi-level boundary defense logic: A preset critical failure threshold is set, and the preset critical failure threshold is greater than the preset safety threshold; When the model trust entropy value is greater than the preset security threshold, the second control mode is triggered, specifically including: When the model trust entropy value is greater than the preset security threshold and less than or equal to the preset critical failure threshold, the second control mode is executed. When the model trust entropy value is greater than the preset critical failure threshold, the third control mode is triggered to generate a blocking pulse command and send it to the power conversion hardware. The control objective of the third control mode is to cut off energy transmission within a preset hardware failure time boundary, which is set to two milliseconds.

6. The intelligent control system for medium-frequency power supply according to any one of claims 1 to 5, characterized in that, The system also includes: The trust reconstruction module is configured to collect feedback voltage data and feedback current data of the power conversion hardware under the second switching command during the operation of the second control mode; The model update module is configured to calculate the prediction error using the feedback voltage data and the feedback current data, adaptively correct the underlying parameters of the preset robust control model based on the prediction error, and send a recovery command to the adaptive control module after the correction is completed. The adaptive control module is further configured to: in response to the recovery command, re-compare the model trust entropy value with the preset security threshold.

7. The intelligent control system for medium-frequency power supply according to claim 6, characterized in that, The power conversion hardware includes: The inverter bridge circuit is composed of multiple insulated gate bipolar transistor modules or silicon carbide power modules; The drive protection circuit is configured to receive the first switch command or the second switch command and convert them into physical level signals that drive the inverter bridge circuit to turn on and off.

8. The intelligent control system for medium-frequency power supply according to claim 3, characterized in that, The load side is connected to a special metallurgical smelting furnace or a superconducting material heating device. The end-side voltage data and the end-side current data contain information on nonlinear impedance abrupt changes caused by the special metallurgical smelting furnace or the superconducting material heating device during the phase transition process.

9. The intelligent control system for medium-frequency power supply according to claim 1, characterized in that, The data acquisition module and the adaptive control module are deployed in the field edge controller to meet the microsecond-level real-time control requirements. The feature extraction module and the state assessment module are deployed on a cloud server and interact with the field edge controller via an industrial network.

Images