An intelligent electrical control system based on industrial equipment
By constructing an AI intelligent optimization channel and a deterministic safety fallback channel on a SoC heterogeneous computing platform, and combining deep reinforcement learning and device state-space equations, the problem of failing to effectively predict complex physical consequences and mechanical shocks in existing technologies has been solved, and safe and stable electrical control of industrial equipment has been achieved.
Patent Information
- Application Number
- CN202611457254.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-18
- Publication Date
- 2026-11-17
AI Technical Summary
The existing solution combining AI controllers with hardware upper and lower limit clamping protection modules fails to effectively consider the dynamic coupling relationship between various physical state quantities within the device, resulting in the inability to predict the complex physical consequences caused by control commands. Furthermore, the rigid cutoff strategy of the limiting circuit causes mechanical shocks. The lack of hardware-level real-time coordination and physical isolation mechanisms makes it difficult to achieve real-time closed-loop electrical control.
Employing a SoC heterogeneous computing platform, integrating an ARM processor, an NPU processor, and FPGA programmable logic, it constructs an AI intelligent optimization channel and a deterministic safety fallback channel. By collecting data through sensors and using deep reinforcement learning to generate candidate control instructions, the device state space equation is embedded in the FPGA for advance prediction and safety verification. The candidate instructions are then corrected to meet physical safety constraints, achieving smooth control instruction transitions.
It significantly improves the accuracy and completeness of physical safety prediction, avoids abrupt changes in control commands, extends equipment lifespan, and ensures the safety and stability of underlying electrical execution links through hardware-level coordination and isolation.
Smart Images

Figure CN123113580A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electrical automation control technology for industrial equipment, and in particular to an electrical intelligent control system based on industrial equipment, specifically a dual-channel electrical intelligent control system that integrates artificial intelligence optimization capabilities with deterministic physical safety constraints. Background Technology
[0002] With the development of Industry 4.0 and the deep integration of information technology and operational technology, the requirements for production efficiency and energy consumption optimization in industrial equipment such as high-end CNC machine tools, high-frequency servo drives, and precision industrial robots are becoming increasingly stringent. The industry hopes to introduce artificial intelligence technologies such as deep reinforcement learning to take over closed-loop electrical control, dynamically adjusting control quantities such as pulse width modulation duty cycle and current loop reference values based on real-time operating conditions, in order to reduce energy consumption and optimize trajectory tracking errors while ensuring operational accuracy. However, the underlying electrical execution links, such as the servo drive current loop and the inverter pulse width modulation control loop, have extremely stringent requirements for the physical safety of control commands. Safety verification typically needs to be completed within microsecond or even nanosecond cycles. Abnormal control commands will directly lead to electrical component breakdown, equipment malfunction, and mechanical structure damage. Currently, the industry's solution combines an artificial intelligence controller with hardware upper and lower limit clamping protection modules, with sensor acquisition units collecting equipment current and voltage. The system uses speed and position status data. The AI controller runs a deep reinforcement learning model to calculate control commands. The hardware limiting circuit compares the command with preset voltage and current thresholds. If the value exceeds the threshold, it is forcibly truncated to the boundary threshold; otherwise, it remains unchanged before being output to the power drive circuit. This solution has the following shortcomings: First, the hardware limiting circuit only performs independent threshold comparisons on the instantaneous values of a single control variable, without considering the dynamic coupling relationship between various physical state quantities within the equipment. It cannot predict the complex physical consequences such as resonance and magnetic saturation caused by the control command. Second, the limiting circuit adopts a rigid truncation strategy. After the protection is triggered, the command value undergoes a step jump, causing mechanical shock and affecting the service life of the equipment. Third, the AI controller and the limiting module are independent of each other and lack hardware-level real-time collaboration and physical isolation mechanisms. Safety judgment is only at the threshold comparison level, making it difficult for AI technology to be applied to the real-time closed-loop electrical control link. Summary of the Invention
[0003] The technical problem this invention aims to solve is that existing solutions combining artificial intelligence controllers with hardware upper and lower limit clamping protection modules rely on hardware limiting circuits that only perform independent threshold comparisons on the instantaneous values of single control variables, failing to consider the dynamic coupling relationships between various physical state variables within the equipment and thus unable to predict the complex physical consequences caused by control commands. Furthermore, the limiting circuits employ a rigid cutoff strategy, resulting in a step jump in control command values after protection is triggered, leading to mechanical shocks. The artificial intelligence controller and limiting module operate independently, lacking hardware-level real-time collaboration and physical isolation mechanisms, making it difficult to integrate artificial intelligence technology into real-time closed-loop electrical control. The purpose of this invention is to provide an intelligent electrical control system for industrial equipment that, while retaining the adaptive optimization capabilities of artificial intelligence algorithms for industrial equipment electrical control parameters, simultaneously ensures that the underlying electrical execution stages meet deterministic physical safety constraints, preventing destructive consequences such as electrical component breakdown, equipment malfunction, and mechanical structural damage.
[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an electrical intelligent control system based on industrial equipment, including a sensor acquisition unit, a SoC heterogeneous computing platform and a power drive circuit;
[0005] The SoC heterogeneous computing platform integrates an ARM processor, an NPU processor, and FPGA programmable logic, including an AI intelligent optimization channel and a deterministic security fallback channel;
[0006] The sensor acquisition unit is used to acquire current status data, voltage status data, speed status data, and temperature status data of industrial equipment.
[0007] The AI intelligent optimization channel runs on the ARM processor and the NPU processor, and includes a state feature extraction module and a deep reinforcement learning decision module, used to generate candidate control instructions based on the state data collected by the sensor acquisition unit.
[0008] The deterministic safety fallback channel runs on the FPGA programmable logic and includes a physical information constraint sandbox module. The physical information constraint sandbox module is connected to the AI intelligent optimization channel via a high-speed bus and is used to receive the candidate control commands and the state data collected by the sensor acquisition unit. Based on the device state space equation solidified in the FPGA programmable logic, it predicts the physical state of the device in the next control cycle in advance to obtain the predicted state vector.
[0009] The predicted state vector is compared with the pre-stored physical safety constraint parameters. If it is determined that the predicted state vector does not exceed the physical safety constraint parameters, the candidate control command is determined as a safety control command.
[0010] When it is determined that the predicted state vector exceeds the physical safety constraint parameters, the candidate control command is modified into a safe control command that satisfies all physical safety constraint parameters based on the safety manifold projection operation.
[0011] The power drive circuit is connected to the deterministic safety fallback channel and is used to receive the safety control command and drive the actuator of the industrial equipment to move.
[0012] In a preferred embodiment, the physical information constraint sandbox module includes a state observation register module, a state space prediction operation unit, a physical constraint boundary database module, a boundary violation decision logic unit, a safe manifold projection operation unit, and an instruction arbitration output module. These modules are connected sequentially to jointly perform advance prediction, boundary violation decision, and dynamic clamping compensation processing of candidate control instructions. The safe manifold projection operation unit incorporates a piecewise linear projection lookup table obtained from offline KKT condition solving. During online operation, projection can be completed through table lookup and multiplication-addition operations, avoiding the high computational latency caused by online solving.
[0013] The present invention has the following beneficial effects:
[0014] 1. By embedding the device state-space equation in the FPGA programmable logic, the device can perform advance prediction of the physical state of the next cycle caused by candidate control instructions in a pipeline manner. This can identify complex physical risks such as resonance and magnetic saturation caused by the coupling of multiple physical state variables. Compared with the scheme that only compares thresholds for a single control variable, this significantly improves the completeness and accuracy of physical safety prediction.
[0015] 2. By generating a safety manifold projection lookup table offline based on KKT conditions, when a boundary is crossed, the candidate control command is projected to the safety manifold boundary with the shortest distance. This ensures that the corrected safety control command and the original candidate control command maintain mathematical continuity, avoids the step jump in control command caused by hard truncation, achieves smooth safety degradation of control command, reduces mechanical shock, and extends the service life of equipment.
[0016] 3. The AI intelligent optimization channel and the deterministic safety fallback channel run on the ARM processor, NPU processor, and FPGA programmable logic of the SoC heterogeneous computing platform, respectively. They achieve data interaction through a high-speed bus and realize physical isolation and real-time collaboration at the hardware level. This allows candidate control commands to be pushed down to the real-time closed-loop electrical control link while the deterministic safety fallback channel performs cycle-by-cycle physical safety verification and correction. This ensures the physical and electrical safety of the underlying electrical execution link while retaining most of the optimization benefits of the artificial intelligence algorithm. Attached Figure Description
[0017] Figure 1This is a flowchart illustrating the overall system architecture of the present invention.
[0018] Figure 2 This is a flowchart illustrating the structure of the physical information constraint sandbox module of the present invention.
[0019] Figure 3 This is a flowchart of the physical semantic consequences prediction method of the present invention. Detailed Implementation
[0020] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0021] The overall architecture of the electrical intelligent control system based on industrial equipment provided in this embodiment is divided into three layers from bottom to top: the industrial equipment physical layer, the SoC heterogeneous computing layer, and the power drive execution layer. The industrial equipment physical layer includes servo motors, inverters, and sensor components. The sensor components include current sensors, voltage sensors, encoders, vibration sensors, and temperature sensors, which are used to collect real-time current status data, voltage status data, speed status data, position status data, vibration status data, and temperature status data of the equipment. This multi-dimensional physical status data is then input to the SoC heterogeneous computing layer. The SoC heterogeneous computing layer integrates an ARM processor, an NPU processor, and FPGA programmable logic. Internally, it forms two parallel control channels: one is an AI intelligent optimization channel, running on the ARM processor and NPU processor, using... The system consists of two channels: a first channel that calculates candidate control instructions based on real-time status data, and a second channel that provides a deterministic safety fallback. This second channel runs on FPGA programmable logic and is used to perform physical safety verification and correction on the candidate control instructions, outputting safety control instructions. The two channels interact with each other via a high-speed bus (such as an AXI bus or dual-port random access memory). The AI intelligent optimization channel operates on a millisecond-level cycle, while the deterministic safety fallback channel operates on a microsecond-level cycle. The two channels maintain physical isolation at the hardware level while maintaining logical coordination through the high-speed bus. The power drive execution layer receives the safety control instructions output by the deterministic safety fallback channel, converts them into drive signals via a pulse width modulation generator and power drive circuit, drives actuators such as servo motors and inverters, and feeds back the executed device status to the sensor components, forming a complete closed loop.
[0022] In the AI intelligent optimization channel, the state feature extraction module extracts features from the multi-dimensional physical state data collected by the sensor components to obtain state feature vectors. The deep reinforcement learning decision module takes maximizing the cumulative reward as the optimization objective. Based on the state feature vectors and preset optimization objectives (such as minimizing energy consumption or minimizing trajectory tracking error), it runs a deep reinforcement learning model and outputs candidate control commands. The components of the candidate control commands include the pulse width modulation duty cycle and the current loop reference value. The candidate control commands are sent to the deterministic safety fallback channel via a high-speed bus.
[0023] The core implementation carrier of the deterministic safety fallback channel is the physical information constraint sandbox module, which consists of the following functional units within the FPGA programmable logic: a state observation register module, used to receive and cache the device state vector of the current control cycle collected by the sensor components, denoted as the device state vector of the k-th control cycle; a state space prediction operation unit connected to the state observation register module, which internally stores the state transition matrix and control input matrix, and uses a pipelined structure to perform matrix multiplication and addition operations on the device state vector and candidate control instructions of the current control cycle in parallel to obtain the predicted state vector of the next control cycle; a physical constraint boundary database module, which stores the device's current constraint parameters, torque constraint parameters, vibration constraint parameters, and temperature constraint parameters in read-only memory, and also stores the safety manifold segmentation region information; and an out-of-bounds decision logic unit, which works in conjunction with the state space prediction operation unit and the physical constraint boundary database module. The constraint boundary database module is connected, comparing each component of the predicted state vector with the pre-stored constraint parameters in the physical constraint boundary database module to determine whether the predicted state vector exceeds the safety constraint boundaries such as current, voltage, torque, vibration frequency band, or temperature, and outputs the boundary violation judgment flag and boundary violation direction information. The safety manifold projection calculation unit is connected to the boundary violation judgment logic unit, and internally contains a piecewise linear projection lookup table obtained by offline solution of KKT conditions. Based on the boundary violation judgment flag, the piecewise linear projection operation is performed on the candidate control command to obtain the projected safety control command. The command arbitration output module is connected to the safety manifold projection calculation unit. When the boundary violation judgment flag indicates that the constraint boundary has not been exceeded, the candidate control command is directly determined as the safety control command and output. When the boundary violation judgment flag indicates that the constraint boundary has been exceeded, the projected safety control command is output to the power drive circuit.
[0024] The device state-space equation used by the state-space prediction computation unit represents the linear mapping relationship between the device state vector and the candidate control command in the k-th control cycle, and is used to obtain the predicted state vector in the k+1-th control cycle. Specifically, it is expressed as follows:
[0025] ;
[0026] In the above formula, This represents the physical state vector of the device in the k-th control cycle, whose components include current, rotational speed, position, temperature, and vibration amplitude. This is the system state transition matrix, which characterizes the dynamic coupling characteristics between various physical state variables within the equipment. This matrix is determined by identifying the mechanical and electrical parameters of the equipment or by offline modeling. The control input matrix represents the coefficients of the candidate control commands on the physical state. The candidate control command given by the AI intelligent optimization channel in the k-th control cycle includes the pulse width modulation duty cycle and the current loop reference value. The predicted state vector for the k+1th control cycle is derived by the state space prediction operation unit. The state transition matrix and the control input matrix are both pre-stamped in the digital signal processing slice of the FPGA, so that matrix multiplication and addition operations can be completed in a pipeline manner in microseconds or even shorter cycles, thereby realizing the advance prediction of the physical state of the device in the next control cycle.
[0027] The physical semantic consequences prediction method is executed cyclically in the following order: First, multi-dimensional physical state data of the equipment in the current cycle is collected by sensor components, and the current cycle equipment state vector is constructed in the state observation register module. Then, the physical information constraint sandbox module receives candidate control commands output by the AI intelligent optimization channel through the high-speed bus. Next, the state space prediction calculation unit uses the fixed state transition matrix and control input matrix to calculate the predicted state vector of the next control cycle in a pipelined parallel manner. Subsequently, the boundary judgment logic unit compares each component of the predicted state vector with the constraint parameters pre-stored in the physical constraint boundary database module to determine whether there is a risk of exceeding the current, voltage, torque, vibration frequency band, or temperature constraint boundary. If it is determined that the constraint boundary has not been exceeded, the candidate control command is marked as a safe command and released directly. If it is determined that the constraint boundary has been exceeded, the safety manifold projection calculation unit is triggered to perform correction processing. Finally, the determined safe control command is sent to the power drive circuit through the command arbitration output module to drive the actuators such as servo motors and inverters to act, and the above process is repeated in the next control cycle, forming a rolling prediction and closed-loop execution of the equipment physical state.
[0028] A safe manifold is composed of multiple physical safety constraint functions, each of which is expressed as:
[0029] ;
[0030] In the above formula, This represents the i-th physical safety constraint function, such as a current upper limit constraint function, a torque upper limit constraint function, or a vibration amplitude constraint function. This is the device state vector. To determine the total number of constraints, if the predicted state vector results in at least one constraint function having a value greater than zero, it is considered to be outside the constraint boundary. In this case, the following projection problem needs to be solved:
[0031] ;
[0032] In the above formula, These are safety control commands after projection correction. To satisfy the set of feasible control instructions that meet all physical safety constraints, i.e., the control domain corresponding to the safe manifold, For the candidate control commands output by the AI intelligent optimization channel, the square of the Euclidean distance represents the search for the feasible safe point with the shortest distance to the original candidate command in the multidimensional control command space.
[0033] To solve the above projection problem, the Lagrange function is introduced:
[0034] ;
[0035] In the above formula, For Lagrange functions, Let be the Lagrange multiplier corresponding to the i-th constraint, and satisfy the non-negativity condition. The optimal solution to this problem satisfies the following KKT conditions:
[0036] ;
[0037] In the above formula, This represents the gradient of the Lagrange function with respect to the control command. For the optimal safety control command solution of the projection problem, the complementary relaxation condition indicates that only the multipliers corresponding to the active constraints on the boundary are non-zero. Since the electrical and mechanical physical constraints of industrial equipment are mostly piecewise linear boundaries in the form of convex polyhedra, the above KKT conditions can be solved offline and transformed into linear projection expressions under each piecewise region. A safety manifold projection lookup table is pre-generated and stored in the FPGA storage unit. During online operation, the safety manifold projection operation unit reads the projection coefficient matrix and bias vector corresponding to the out-of-bounds piecewise region from the lookup table according to the out-of-bounds decision logic unit. It performs multiplication and addition operations on the candidate control command to obtain the safety control command located on or inside the safety manifold boundary and with the shortest Euclidean distance to the candidate control command. This avoids the high computational latency caused by online solving. Since the projection solution is mathematically continuous, it will not cause a jump in the control quantity when output to the power drive circuit, realizing a seamless and smooth electrical safety degradation transition.
[0038] The AI intelligent optimization channel updates candidate control commands once every millisecond-level control cycle (e.g., every five milliseconds), while the deterministic safety fallback channel continuously operates the state advance prediction and over-boundary judgment process every microsecond-level control cycle (e.g., every ten microseconds corresponds to one pulse width modulation carrier cycle). During the interval when the AI intelligent optimization channel does not update candidate control commands, the deterministic safety fallback channel repeatedly performs advance prediction and over-boundary judgment on the same candidate control command until the AI intelligent optimization channel updates the next candidate control command. This allows the two channels to asynchronously coordinate at different time scales and jointly form a closed-loop control.
[0039] The following describes the working process of this invention in a specific application scenario. In the servo driver of a precision industrial robot, the current constraint parameter is set to the safe range corresponding to the rated current of the equipment, and the vibration constraint parameter is the vibration amplitude range corresponding to the mechanical resonance frequency band. The current loop reference value in the candidate control command calculated by the AI intelligent optimization channel based on the current working condition is within the rated range. If only the existing hardware threshold comparison scheme is used, the candidate command will be judged as compliant and directly released. However, after the physical information constraint sandbox module uses the fixed equipment state space equation to predict the vibration amplitude of the next control cycle, it is found that the candidate command will cause the predicted vibration amplitude to exceed the preset vibration constraint parameter under the current working condition coupling, that is, trigger the boundary judgment. At this point, the safety manifold projection calculation unit, based on the out-of-bounds segment region number determined by the out-of-bounds decision logic unit, reads the corresponding projection coefficient matrix and bias vector from the pre-fixed safety manifold projection lookup table, performs multiplication and addition operations on the candidate control command, and obtains the projected safety control command. The current loop reference value of this safety control command undergoes continuous and smooth micro-correction compared to the original candidate command, causing the predicted vibration amplitude to fall back within the safety constraint range, while maintaining the shortest Euclidean distance from the original candidate command. The command arbitration output module outputs the safety control command to the power drive circuit, driving the servo motor to perform the action. The entire process of advance prediction, out-of-bounds decision, and projection correction is completed within a microsecond-level control cycle, without causing a step-like impact on the continuous operation of the equipment.
[0040] In an alternative implementation, the physical information constraint sandbox module may not be implemented using FPGA programmable logic, but instead uses a real-time embedded processor with hardware floating-point arithmetic units and digital signal processing instruction sets. Through software pipeline optimization, state space prediction operations and secure manifold projection operations are implemented to approximate microsecond-level response timing, which is suitable for application scenarios that are more sensitive to hardware costs.
[0041] In another alternative implementation, the safe manifold projection computation unit may not use the offline lookup table generation method, but instead deploy a lightweight online iterative solver in the FPGA programmable logic or dedicated coprocessor to iteratively solve the projection problem in real time for scenarios where the constraints change dynamically with the operating conditions, so as to obtain stronger constraint adaptability.
[0042] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.
Claims
1. An electrical intelligent control system based on industrial equipment, characterized in that, This includes sensor acquisition units, a SoC heterogeneous computing platform, and power drive circuitry; The SoC heterogeneous computing platform integrates an ARM processor, an NPU processor, and FPGA programmable logic, including an AI intelligent optimization channel and a deterministic security fallback channel; The sensor acquisition unit is used to acquire current status data, voltage status data, speed status data, and temperature status data of industrial equipment. The AI intelligent optimization channel runs on the ARM processor and the NPU processor, and includes a state feature extraction module and a deep reinforcement learning decision module, used to generate candidate control instructions based on the state data collected by the sensor acquisition unit. The deterministic safety fallback channel operates on the FPGA programmable logic and includes a physical information constraint sandbox module. The physical information constraint sandbox module is connected to the AI intelligent optimization channel via a high-speed bus and is used to receive the candidate control commands and the state data collected by the sensor acquisition unit. Based on the device state space equation embedded in the FPGA programmable logic, it performs advance prediction of the device physical state in the next control cycle to obtain a predicted state vector. The predicted state vector is compared with pre-stored physical safety constraint parameters. If it is determined that the predicted state vector does not exceed the physical safety constraint parameters, the candidate control command is determined as a safe control command. If it is determined that the predicted state vector exceeds the physical safety constraint parameters, the candidate control command is modified into a safe control command that satisfies all physical safety constraint parameters based on the safety manifold projection operation. The power drive circuit is connected to the deterministic safety fallback channel and is used to receive the safety control command and drive the actuator of the industrial equipment to move.
2. The system according to claim 1, characterized in that, The physical safety constraint parameters include current constraint parameters, torque constraint parameters, vibration constraint parameters, and temperature constraint parameters.
3. The system according to claim 1, characterized in that, The physical information constraint sandbox module includes: The state observation register module is used to cache the device state vector of the current control cycle acquired by the sensor acquisition unit; The state space prediction operation unit is connected to the state observation register module and has a built-in fixed state transition matrix and control input matrix. It is used to perform matrix multiplication and addition operations on the device state vector and the candidate control command in a pipeline manner to obtain the predicted state vector. The physical constraint boundary database module is used to store the physical safety constraint parameters; The boundary judgment logic unit is connected to the state space prediction operation unit and the physical constraint boundary database module, and is used to compare the predicted state vector with the physical safety constraint parameters item by item, and output the boundary judgment flag. The safe manifold projection operation unit is connected to the boundary judgment logic unit and has a built-in piecewise linear projection lookup table obtained offline, which is used to perform piecewise linear projection operation on the candidate control instruction according to the boundary judgment flag. The instruction arbitration output module is connected to the safety manifold projection calculation unit and is used to output the safety control instruction to the power drive circuit according to the boundary judgment flag.
4. The system according to claim 3, characterized in that, The device state space equation is used to characterize the linear mapping relationship between the device state vector of the k-th control cycle and the candidate control command of the k-th control cycle, so as to obtain the predicted state vector of the k+1-th control cycle. The components of the device state vector include current, rotational speed, position, temperature, and vibration amplitude; The linear mapping relationship is determined by the state transition matrix and the control input matrix. The state transition matrix represents the dynamic coupling relationship between the physical state variables inside the device, and the control input matrix represents the effect coefficient of the candidate control command on the physical state of the device.
5. The system according to claim 3, characterized in that, The piecewise linear projection lookup table is generated in the following way: A constraint set consisting of multiple physical safety constraint functions is established. The KKT conditions are used to solve the projection problem of the feasible control command with the shortest Euclidean distance to the candidate control command offline, so as to obtain the projection coefficient matrix and bias vector corresponding to each segment region. The projection coefficient matrix and the bias vector are then solidified into the piecewise linear projection lookup table.
6. The system according to claim 1, characterized in that, The AI intelligent optimization channel updates the candidate control instructions according to a millisecond-level control cycle, and the deterministic safety fallback channel repeatedly performs advance prediction and out-of-bounds judgment on the same candidate control instruction according to a microsecond-level control cycle until the AI intelligent optimization channel updates the next candidate control instruction.
7. The system according to claim 1, characterized in that, The state feature extraction module is used to extract features from the state data collected by the sensor acquisition unit to obtain a state feature vector. The deep reinforcement learning decision module is used to output the candidate control command based on the state feature vector. The components of the candidate control command include the pulse width modulation duty cycle and the current loop reference value.
8. A method for intelligent electrical control of industrial equipment, characterized in that, include: Collect current status data, voltage status data, speed status data, and temperature status data of industrial equipment; Based on the state data, candidate control instructions are generated by a deep reinforcement learning decision module running on an ARM processor and an NPU processor. The physical information constraint sandbox module running on the FPGA programmable logic receives the candidate control instructions and the state data, and predicts the physical state of the device in the next control cycle based on the device state space equation solidified in the FPGA programmable logic to obtain the predicted state vector. The predicted state vector is compared with the pre-stored physical safety constraint parameters. If the predicted state vector does not exceed the physical safety constraint parameters, the candidate control command is determined as a safety control command. If the predicted state vector exceeds the physical safety constraint parameters, the candidate control command is modified to a safety control command that satisfies all physical safety constraint parameters based on the safety manifold projection operation. The safety control command is output to the power drive circuit to drive the actuator of the industrial equipment.
9. The method according to claim 8, characterized in that, The steps of modifying the candidate control command into a safe control command based on safe manifold projection calculation include: Establish a constraint set consisting of multiple physical security constraint functions; Based on the out-of-bounds segment region number to which the predicted state vector belongs, the corresponding projection coefficient matrix and bias vector are read from the piecewise linear projection lookup table obtained offline. The candidate control command is multiplied and added using the projection coefficient matrix and the bias vector to obtain the safety control command with the shortest Euclidean distance to the candidate control command.
10. The method according to claim 8, characterized in that, The step of generating the candidate control instruction is executed according to a millisecond-level control cycle, and the step of advance prediction and comparison is executed according to a microsecond-level control cycle. During the period when the candidate control instruction is not updated, the step of advance prediction and comparison is repeated.
