An edge-computing-based motor adaptive prediction and optimization control system

Abstract

The application discloses a motor adaptive prediction and optimization control system based on edge computing and relates to the field of motor control. The system aims to solve the problem of control right loss in the physical blind area of motor commutation zero-crossing. The system comprises a physical feature anchoring unit, a dead zone window locking and a freewheeling clamp feature capturing unit, a zero point true value extraction unit for calibrating current, an entropy weight state perception unit for calculating information entropy and normalized entropy divergence and generating a dimensionless confidence factor representing observability, a probability prediction reconstruction unit for querying a prior probability model in the blind area and calculating a conditional mathematical expectation and outputting a reconstructed voltage feedforward vector, and a dynamic arbitration execution unit for weighting and fusing the basic instruction and the reconstructed vector according to the confidence factor. Through the closed loop of physical anchoring, entropy weight perception and probability reconstruction, the application effectively fills the information vacuum, realizes the soft switching of the control mode without jitter and the self-healing of the physical layer.

Description

Technical Field

[0001] This invention relates to the field of motor control technology, and in particular to a motor adaptive predictive and optimization control system based on edge computing. Background Technology

[0002] In high-performance control applications of permanent magnet synchronous motors (PMSMs), the commutation zero-crossing region of the inverter constitutes a nonlinear physical bottleneck in the control system. Because power semiconductor devices must have a physical dead time to prevent bridge arm shoot-through short circuits, coupled with the freewheeling effect of the anti-parallel diodes during the dead time, a physical interruption in voltage output occurs near the zero-crossing point of the motor stator windings. Simultaneously, within the small interval where the current approaches zero, the signal-to-noise ratio of low-cost current sensors (such as Hall elements) drops sharply, and they are highly susceptible to zero-point drift due to semiconductor temperature drift.

[0003] The superposition of the aforementioned physical breaks (dead zone) and information breaks (noise and drift) causes the control system to essentially lose the observability of the motor state within the commutation zero-crossing region, creating an information vacuum or control blind zone that cannot be effectively adjusted by conventional linear feedback mechanisms. Within this blind zone, traditional control strategies based on error feedback (such as PID or FOC) fail due to the lack of effective input, often resulting in divergent or oscillating output commands. Macroscopically, this manifests as increased motor torque pulsation, mechanical chatter at low speeds, and reduced system energy efficiency.

[0004] Existing technologies typically mitigate this problem by improving hardware precision (e.g., using high-resolution grating rulers or high-precision fluxgate sensors) or introducing deterministic model compensation algorithms (e.g., voltage compensation tables based on fixed dead time). However, hardware upgrades are costly and unsuitable for low-cost edge devices; while deterministic compensation algorithms struggle to adapt to the dynamic drift of physical parameters during motor operation and cannot fundamentally fill the information gaps caused by physical mechanisms. Therefore, how to reconstruct lost control under conditions of limited edge computing resources and physical blind spots is a pressing technical problem in the field of motor control. Summary of the Invention

[0005] This invention provides a motor adaptive predictive and optimization control system based on edge computing, which aims to solve the problem of loss of control in the commutation zero-crossing region of existing motor control systems due to physical dead zone effect and sensor noise coupling.

[0006] In view of the above problems, the present invention provides a motor adaptive predictive and optimization control system based on edge computing. The system runs in an edge computing controller and is used to reconfigure control rights in the physical blind zone where the motor commutation zero crosses. The system includes: The physical feature anchoring unit is configured to receive the original sampled current sequence output by the sensor, lock the interlock dead zone window in the inverter switching timing, extract the zero-point physical true value using the current clamping feature within the window, and output the pure current sequence after removing the zero-point drift. The entropy weight state sensing unit is configured to work in parallel with the physical feature anchoring unit, receive the original sampled current sequence, calculate the signal information entropy containing noise components within the sliding window, calculate the normalized entropy divergence based on the background noise benchmark, and generate a dimensionless confidence factor characterizing the observability of the system through nonlinear mapping. The probability prediction reconstructing unit is configured to store a voltage vector probability distribution model constructed based on prior data. When the dimensionless confidence factor indicates that the system has entered the blind zone, the unit queries the model based on the current motor state and calculates the conditional mathematical expectation of the ideal voltage vector, and outputs the reconstructed voltage feedforward vector. The dynamic arbitration execution unit is configured to receive the basic control command and the reconstructed voltage feedforward vector, perform weighted fusion according to the dimensionless confidence factor, and generate the final drive command to drive the motor.

[0007] Preferably, the physical feature anchoring unit is specifically configured to execute the following logic when extracting the zero-point physical truth value: Monitor the PWM gate drive signal to lock the interlock dead zone window, and calculate the rate of change of current within the window; The system is determined to be in a freewheeling clamp state only when the rate of change is lower than a preset clamping threshold, and the current sampling value in this state is determined as the sensor zero-point drift.

[0008] Preferably, the entropy weight state sensing unit is specifically configured to use Shannon entropy as the signal information entropy and perform the following normalization calculation: The thermal noise entropy of the current path in a static state is obtained as the background noise reference. Calculate the Shannon entropy of the real-time raw sampled current; Calculate the ratio of the real-time sampled Shannon entropy to the background noise baseline to generate the normalized entropy divergence; The normalized entropy divergence is used to eliminate the influence of sensor range and motor power level on the blind zone determination threshold.

[0009] Preferably, the entropy weight state sensing unit uses a hyperbolic tangent function model to generate the dimensionless confidence factor, and the calculation formula is as follows: in, The dimensionless confidence factor; It is the hyperbolic tangent function; The normalized entropy divergence; Sensitivity gain; This is the trigger threshold; The hyperbolic tangent function model is used to provide continuously differentiable weight changes during the switching process between the steady-state region and the blind zone, so as to achieve jitter-free soft switching of the control mode.

[0010] Preferably, the voltage vector probability distribution model stored in the probability prediction and reconstruction unit is a sparse lookup table structure; When calculating the conditional mathematical expectation, the probability prediction and reconstruction unit is configured to integrate or perform a weighted summation on the indexed probability density function, thereby reconstructing the lost physical quantity using prior probability in the zero-crossing region where the sensor signal-to-noise ratio is extremely low.

[0011] Preferably, the dynamic arbitration execution unit executes the following weighted fusion logic: in, This refers to the final driving instruction; These are the basic control commands; The reconstructed voltage feedforward vector; The dimensionless confidence factor; When the dimensionless confidence factor When the value approaches 1, the system automatically bypasses the basic control command, and the reconstructed voltage feedforward vector takes over the control to avoid amplification of feedback errors in the blind zone.

[0012] Preferably, the system runs in an edge computing controller and adopts a heterogeneous computing architecture: The physical feature anchoring unit is deployed in the FPGA logic unit to achieve microsecond-level dead zone feature capture using parallel hardware logic; The probability prediction and reconstruction unit is deployed in the MCU or NPU computing unit to perform probabilistic inference operations; The FPGA logic unit and the MCU or NPU computing unit interact with each other via an on-chip bus or high-speed interface.

[0013] The technical solution provided in this application has at least the following technical effects: This invention solves the problem of control loss caused by the loss of observability in the physical dead zone of conventional linear controllers. Unlike existing technologies that attempt to fit erroneous paths of random noise using deterministic algorithms, this invention calculates the conditional mathematical expectation of the ideal voltage vector through a probabilistic prediction reconstruction unit, constructing a high-confidence feedforward command in the information vacuum region. This logically reconstructs the broken control loop, realizes a smooth torque transition of the motor at the commutation zero-crossing point, and eliminates mechanical chatter caused by information loss.

[0014] This invention solves the problem of data reference failure caused by temperature drift and noise in the dead zone of low-cost sensors. It utilizes the inherent dead-zone freewheeling characteristic of the inverter as an absolute physical anchor point to achieve real-time online calibration of the sensor's zero point. This mechanism transforms the physical-level fracture (dead zone) into a calibration-level reference, endowing the system with physical-level self-healing capability in the absence of an external standard source, ensuring the objective accuracy of the control reference during long-term operation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the architecture of a motor adaptive prediction and optimization control system based on edge computing, provided for an embodiment of the present invention. Detailed Implementation

[0016] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0017] For examples, please refer to Figure 1 This invention provides a motor adaptive predictive and optimization control system based on edge computing. The system runs in an edge computing controller and is used to reconfigure control rights in the physical blind zone where the motor commutation zero crosses. The system includes: The physical feature anchoring unit is configured to receive the original sampled current sequence output by the sensor, lock the interlock dead zone window in the inverter switching timing, extract the zero-point physical true value using the current clamping feature within the window, and output the pure current sequence after removing the zero-point drift. The entropy weight state sensing unit is configured to work in parallel with the physical feature anchoring unit, receive the original sampled current sequence, calculate the signal information entropy containing noise components within the sliding window, calculate the normalized entropy divergence based on the background noise benchmark, and generate a dimensionless confidence factor characterizing the observability of the system through nonlinear mapping. The probability prediction reconstructing unit is configured to store a voltage vector probability distribution model constructed based on prior data. When the dimensionless confidence factor indicates that the system has entered the blind zone, the unit queries the model based on the current motor state and calculates the conditional mathematical expectation of the ideal voltage vector, and outputs the reconstructed voltage feedforward vector. The dynamic arbitration execution unit is configured to receive the basic control command and the reconstructed voltage feedforward vector, perform weighted fusion according to the dimensionless confidence factor, and generate the final drive command to drive the motor.

[0018] The operation of the edge computing-based adaptive predictive and optimization control system for motors begins with the reconfiguration of the underlying hardware logic and the establishment of heterogeneous computing links within the edge computing controller. The power management circuit integrated within the edge computing controller simultaneously triggers the parallel startup sequence of the field-programmable gate array (FPGA) logic unit and the microcontroller unit upon power-up. The FPGA logic unit instantiates the parallel processing circuitry of the physical feature anchoring unit at the gate circuit level by loading a pre-stored hardware description bitstream file, and simultaneously configures the on-chip high-speed bus interface logic. The microcontroller unit establishes a wideband data interaction channel with the FPGA logic unit via the on-chip high-speed bus interface through direct memory access or address mapping mechanisms, thereby completing the hardware link laying for the transmission of physical layer feature data to the application layer algorithm.

[0019] Once the hardware link is deployed, the physical feature anchoring unit, located within the field-programmable gate array (FPGA) logic unit, immediately takes over the real-time monitoring of the inverter's switching timing, performing microsecond-level locking of the interlock dead-time window. The physical feature anchoring unit synchronously acquires the upper and lower bridge arm gate drive signals of the same bridge arm of the inverter via a separate set of parallel input pins. The internal combinational logic circuitry of the physical feature anchoring unit performs NOR operations on the upper and lower bridge arm gate drive signals to identify overlapping time domains where the two switching transistors are simultaneously in a low-level off state. To ensure the accuracy of timing locking, the physical feature anchoring unit uses a high-frequency system clock to drive a hardware counter, starting counting on the falling edge of the upper bridge arm gate drive signal and continuously monitoring until the rising edge of the lower bridge arm gate drive signal arrives. A digital comparator compares the real-time count value of the hardware counter with a preset system dead-time parameter. Only when the duration represented by the count value falls within the preset dead-time range does the physical feature anchoring unit output a high-level active window lock signal.

[0020] During the time domain where the window lock signal remains at a valid level, the physical feature anchoring unit drives a high-speed analog-to-digital converter to continuously discretize and quantize the motor stator current at a preset high sampling frequency (e.g., 20 million times per second). The differential operation logic module integrated within the physical feature anchoring unit sequentially reads the original current sample value of the current clock cycle and the original current sample value of the previous clock cycle, which are temporarily stored in the input buffer register. The differential operation logic module performs a difference operation on these two values ​​and divides the difference result by the sampling time interval to calculate the rate of change of current over time, reflecting the slope of the transient current fluctuation.

[0021] A digital comparator deployed in a field-programmable gate array (FPGA) compares the absolute value of the calculated rate of change of current over time with a pre-configured clamping threshold (e.g., 0.1 amperes per microsecond) in real time. Only when the rate of change of current over time is below the clamping threshold for several consecutive clock cycles does the state machine logic inside the physical feature anchoring unit transition to a locked state and output a high-level active freewheeling clamping status confirmation flag. The generation of this freewheeling clamping status confirmation flag physically indicates that the inductive energy stored in the motor windings has been completely released through the anti-parallel diode circuit, and the stator current is objectively in a static zero-value state where it no longer changes significantly.

[0022] At the instant the freewheeling clamp status confirmation signal transitions to an active level, the physical feature anchoring unit triggers the zero-drift latch to perform a data hold operation. The zero-drift latch directly reads the raw current sample value output by the current high-speed analog-to-digital converter and stores this value as the sensor's zero-drift. This sensor zero-drift represents the static error component of the current sensor caused by temperature drift, aging, or electromagnetic interference under the specific boundary condition that the physical current is objectively zero.

[0023] The pipelined subtraction arithmetic logic unit built inside the physical feature anchoring unit then reads the sensor zero-point drift amount held in the zero-point drift latch. In the subsequent pulse width modulation control cycle after the freewheeling clamp status confirmation flag is reset, the subtraction arithmetic logic unit performs hardware subtraction debiasing on each raw current sample value arriving via the input interface, that is, subtracts the sensor zero-point drift amount from each raw current sample value in real time. The numerical stream generated after hardware subtraction correction is constructed as a pure current sequence. This pure current sequence is written to the microcontroller's direct memory access buffer via the on-chip high-speed bus interface, serving as the sole high-confidence data source for the subsequent entropy weight state sensing unit to define the blind zone boundary.

[0024] The high-confidence data stream constructed by the physical layer then enters the entropy-weighted state-aware unit running inside the microcontroller to perform a quantitative assessment of the system's observability.

[0025] During the static standby phase, when the edge computing controller has completed its power-on initialization sequence but has not yet output any drive voltage vector to the motor, the entropy weight state sensing unit automatically triggers the background noise calibration program. During this phase, although the inverter is off, the electronic components inside the current sensor and signal conditioning circuit are affected by thermal motion, still generating slight random thermal noise. The entropy weight state sensing unit continuously reads a preset length of clean current sequence through the on-chip high-speed bus interface and defines this sequence as a static noise sample set.

[0026] The arithmetic logic unit (ALU) within the microcontroller performs statistical analysis on the static noise sample set, mapping the values ​​within the sample set to preset discretized intervals to construct a static noise probability distribution histogram. Based on the probability values ​​of each interval calculated from this histogram, the ALU performs Shannon entropy algorithm calculations, and the calculated numerical result is defined as the background noise entropy benchmark. The background noise entropy benchmark characterizes the lowest inherent uncertainty level of the measurement system under the current hardware environment. The entropy weight state sensing unit writes the background noise entropy benchmark into a non-volatile memory register, which serves as a fixed denominator that must be referenced when performing dimensionless normalization calculations, thereby eliminating the numerical differences in hardware noise floor between sensors of different ranges.

[0027] Once the motor enters dynamic operation mode, the entropy weight state sensing unit allocates a fixed-depth first-in-first-out (FIFO) data buffer in the random access memory. This data buffer stores the most recently received clean current sequence data points, and its capacity is set to a preset window length, for example, containing 50 consecutive sampling points, thus forming a sliding observation window that moves over time. Whenever a new clean current sampling point is written to the end of the buffer, the oldest sampling point at the beginning of the buffer is removed, ensuring that the window always contains the most recent timing data.

[0028] The microcontroller unit uses its floating-point arithmetic unit to perform discretized Shannon entropy calculations on the data within the sliding observation window. The processor first iterates through all values ​​within the window, counting the frequency of data points falling within each pre-defined amplitude interval, and divides this by the window length to obtain the discrete probability value for each interval. For each interval with a probability value greater than 0, the processor calculates the product of that probability value and its logarithm. The processor then sums and inverts the products of all intervals; the result is the real-time Shannon entropy. This real-time Shannon entropy quantifies the degree of disorder and information richness of the current current waveform, serving as a dynamic indicator of whether the system is observable through linear feedback.

[0029] The arithmetic logic unit (ALU) within the microcontroller retrieves the previously calibrated baseline noise entropy from a register and performs a difference operation with the currently calculated real-time Shannon entropy to obtain the entropy difference characterizing the current dynamic uncertainty increment. To eliminate absolute numerical deviations caused by differences in sensor range and motor power levels, the ALU introduces a preset, extremely small positive number as a safety bias constant (e.g., one part per million), and adds this safety bias constant to the baseline noise entropy to form a non-zero denominator to prevent division by zero errors. The ALU then divides the entropy difference by this non-zero denominator to calculate the normalized entropy divergence.

[0030] Normalized entropy divergence, as a dimensionless ratio, physically represents the ratio of the disorder of the current waveform to the system's static thermal noise. Through this divisional normalization process, different motor control systems, whether driving micro-servo motors or industrial-grade servo motors, are unified in the same dimensionless coordinate system for determining the blind zone boundary. This approach eliminates the coupling effect of hardware parameters on the algorithm threshold, allowing the same set of control parameters to maintain consistent decision logic without requiring recalibration for specific motors.

[0031] The entropy weight state sensing unit further uses the calculated normalized entropy divergence as an input variable, substituting it into a pre-installed hyperbolic tangent nonlinear activation model in the firmware to generate the final dimensionless confidence factor. The microcontroller unit first calculates the algebraic difference between the normalized entropy divergence and a preset trigger threshold. The trigger threshold defines the minimum disorder multiplier required for the system to determine if it is entering the control blind zone (e.g., set to 1.0, meaning triggering begins when the disorder exceeds twice the noise floor). This algebraic difference is then multiplied by a preset sensitivity gain coefficient. The magnitude of the sensitivity gain coefficient determines the steepness of the control mode switching process in the time domain: a larger gain coefficient results in a faster response, while a smaller gain coefficient provides a wider transition range.

[0032] The entropy weighted state sensing unit calculates the normalized entropy divergence. As input variables, the hyperbolic tangent nonlinear activation model pre-installed in the firmware is substituted. The microcontroller unit executes the following calculation formula to generate the final dimensionless confidence factor. : In this formula, As the trigger threshold, For sensitivity gain, The hyperbolic tangent function is used. Through this mathematical model, the output of the hyperbolic tangent function is mapped to a dimensionless confidence factor whose value is strictly limited to [0,1]. The unique sigmoid saturation characteristic and global continuous differentiability of the hyperbolic tangent function ensure that the numerical change curve of the dimensionless confidence factor is smooth and continuous when the normalized entropy divergence changes.

[0033] When the value of the dimensionless confidence factor exceeds the preset zero-point threshold, the probability prediction reconstruction unit is activated by the task scheduler of the microcontroller to take over or assist the voltage vector generation task of the motor in the control vacuum zone.

[0034] The probabilistic prediction and reconfiguration unit pre-configures and maintains a sparse lookup table structure in the non-volatile flash memory region of the microcontroller. This sparse lookup table serves as the carrier of the probability distribution model, establishing a mapping from the motor's operating state space to the ideal voltage vector probability space. The index keys of the sparse lookup table consist of two physical quantities: the motor's real-time mechanical speed and the rotor's electrical angular position. The stored values ​​in the sparse lookup table are not single deterministic voltage values, but rather a set of discretized probability density function data corresponding to that operating state. This probability density function data includes a series of possible voltage vector amplitudes and their corresponding normalized probability weights.

[0035] It should be noted that the voltage vector probability distribution model is constructed as follows: During the offline calibration phase of the motor, a high-precision dynamometer test bench is used to synchronously record the voltage vector and actual current response under ideal sinusoidal drive conditions at full speed and full load using a 1MHz high-frequency data acquisition card. The kernel density estimation (KDE) algorithm is then used to statistically model the current distortion characteristics and voltage vector deviation during dead zone occurrence, generating a lookup table stored in the non-volatile memory of the edge controller as the prior data.

[0036] Due to the storage resource limitations of the edge computing controller, this sparse lookup table only fills specific nonlinear regions near the commutation zero-crossing point with high-resolution data, while leaving regions with good system linearity empty or filled with low-resolution data, thus forming a sparse structure. During the real-time control cycle, the probabilistic prediction and reconstruction unit obtains the current rotor electrical angle through the position sensor and calculates the real-time mechanical speed using position difference calculations. The probabilistic prediction and reconstruction unit uses these two physical quantities as address pointers to perform fast indexing and addressing operations in the sparse lookup table. If the current state point falls between two discrete nodes in the sparse lookup table, the probabilistic prediction and reconstruction unit uses a bilinear interpolation algorithm to synthesize the probability density function of the current state point based on the probability density data of adjacent nodes, ensuring a smooth transition of the model in the continuous state space.

[0037] After obtaining the discretized probability density function corresponding to the current state, the probability prediction and reconstruction unit starts the mathematical expectation calculation engine. This calculation engine performs weighted summation and integration operations. The calculation engine traverses each voltage vector sample point in the discretized probability density function, multiplies the magnitude and phase angle of the voltage vector sample point by its corresponding probability weight, and then performs vector accumulation of the products of all sample points.

[0038] The result of this vector summation is the conditional mathematical expectation of the ideal voltage vector. This conditional mathematical expectation is defined as the reconstructed voltage feedforward vector. From a signal processing perspective, this integral summation process is mathematically equivalent to performing a statistical average filter on a dead-zone signal with random noise. By utilizing the entire set of probability distribution data instead of a single predicted value in the calculation, the probabilistic prediction reconstruction unit effectively cancels transient measurement errors caused by random thermal noise or electromagnetic interference in the zero-crossing region. The final generated reconstructed voltage feedforward vector, in a physical sense, represents the ideal driving voltage that is statistically closest to the actual physical requirements under the current uncertain observation conditions, thus providing a high-confidence instruction source for subsequent fusion execution.

[0039] The dynamic arbitration execution unit performs a linear weighted summation operation on the aforementioned input physical quantities based on a pre-set weighted fusion logic. As the final aggregation node for control command outputs, the dynamic arbitration execution unit simultaneously receives basic control commands from the basic linear controller via its internal high-precision floating-point arithmetic unit. Reconstructed voltage feedforward vector from the probabilistic prediction reconstruction unit and the dimensionless confidence factor from the entropy weight state perception unit The dynamic arbitration execution unit performs a linear weighted summation operation on the above-mentioned input physical quantities according to a pre-set weighted fusion logic. This operation follows the mathematical formula below: in, This represents the final drive command. Under the logic defined by this formula, the final drive command is synthesized by linear interpolation of the basic control command and the reconstructed voltage feedforward vector based on a dimensionless confidence factor.

[0040] In this integrated logic's physical operation mechanism, the dimensionless confidence factor acts as a sliding rheostat between two independent control channels. When the motor operates in the steady-state region far from the zero-crossing point, the dimensionless confidence factor approaches 0, causing the weight coefficient of the basic control command to approach 1, while the weight coefficient of the reconstructed voltage feedforward vector approaches 0, thus maintaining the dominance of the basic linear controller to ensure steady-state accuracy. Conversely, when the dimensionless confidence factor approaches 1 as the system enters the blind zone, the weight coefficient of the basic control command rapidly decays to 0. This weight transformation process physically constructs an automatic bypass mechanism: in the blind zone where system observability is lost, the dynamic arbitration execution unit cuts off the signal path of the basic linear controller, and the reconstructed voltage feedforward vector completely takes over the inverter's drive authority. This automatic bypass mechanism effectively isolates the erroneous high-gain regulation caused by feedback signal distortion in the blind zone of the basic linear controller, avoiding the risk of voltage oscillation caused by feedback error amplification from the top level of the control architecture. The final generated drive command is sent to the space vector pulse width modulation module and converted into a switching signal to drive the inverter hardware.

[0041] To verify the objective effectiveness of the aforementioned control logic in engineering practice, a numerical simulation was conducted on an industrial robot joint motor control scenario operating under low-speed, heavy-load conditions.

[0042] During the system initialization phase, the entropy weight state sensing unit measures the background noise entropy of the current channel at a reference of 0.5 bits. When the motor enters the commutation zero-crossing region under load, the waveform distortion caused by the dead zone effect causes the real-time monitored Shannon entropy value to rise to 1.5 bits.

[0043] Based on the aforementioned normalization algorithm, the MCU calculates the normalized entropy divergence as the difference between 1.5 and 0.5, divided by 0.5, resulting in 2.0. The system sensitivity gain is set to 3.0, and the trigger threshold to 1.0. The MCU substitutes these values ​​into the hyperbolic tangent activation model for calculation: first, it calculates the difference between the normalized entropy divergence of 2.0 and the trigger threshold of 1.0, obtaining 1.0; then, it multiplies this difference by the sensitivity gain of 3.0, obtaining an intermediate variable of 3.0; subsequently, it calculates the hyperbolic tangent function value of 3.0, which is approximately 0.995; finally, it adds 1 to the hyperbolic tangent function value and multiplies it by 0.5. The final calculated dimensionless confidence factor is approximately 0.9975.

[0044] The calculation results show that the system has identified an extremely high confidence level in the dead zone. Based on this, the dynamic arbitration execution unit applies a dimensionless confidence factor of 0.9975 to the weighted fusion formula, resulting in the reconstructed voltage feedforward vector occupying 99.75% of the control weight, while the weight of the basic linear controller remains at only 0.25%, effectively bypassing the basic controller. Comparative test data shows that after applying this automatic bypass and reconstructing strategy, the total harmonic distortion rate of the motor phase current at the zero-crossing point decreases from 8.5% in the traditional hard switching method to 2.3%, enabling the system to eliminate physical dead zone interference.

[0045] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motor adaptive predictive and optimization control system based on edge computing, characterized in that, The system runs on an edge computing controller and is used for control reconfiguration in the physical blind zone where the motor commutation zero crosses. The system includes: The physical feature anchoring unit is configured to receive the original sampled current sequence output by the sensor, lock the interlock dead zone window in the inverter switching timing, extract the zero-point physical true value using the current clamping feature within the window, and output the pure current sequence after removing the zero-point drift. The entropy weight state sensing unit is configured to work in parallel with the physical feature anchoring unit, receive the original sampled current sequence, calculate the signal information entropy containing noise components within the sliding window, calculate the normalized entropy divergence based on the background noise benchmark, and generate a dimensionless confidence factor characterizing the observability of the system through nonlinear mapping. The probability prediction reconstructing unit is configured to store a voltage vector probability distribution model constructed based on prior data. When the dimensionless confidence factor indicates that the system has entered the blind zone, the unit queries the model based on the current motor state and calculates the conditional mathematical expectation of the ideal voltage vector, and outputs the reconstructed voltage feedforward vector. The dynamic arbitration execution unit is configured to receive the basic control command and the reconstructed voltage feedforward vector, perform weighted fusion according to the dimensionless confidence factor, and generate the final drive command to drive the motor.

2. The system according to claim 1, characterized in that, When extracting the zero-point physical truth value, the physical feature anchoring unit is specifically configured to execute the following logic: Monitor the PWM gate drive signal to lock the interlock dead zone window, and calculate the rate of change of current within the window; The system is determined to be in a freewheeling clamp state only when the rate of change is lower than a preset clamping threshold, and the current sampling value in this state is determined as the sensor zero-point drift.

3. The system according to claim 1, characterized in that, The entropy weight state sensing unit is specifically configured to use Shannon entropy as the signal information entropy and perform the following normalization calculation: The thermal noise entropy of the current path in a static state is obtained as the background noise reference. Calculate the Shannon entropy of the real-time raw sampled current; Calculate the ratio of the real-time sampled Shannon entropy to the background noise baseline to generate the normalized entropy divergence; The normalized entropy divergence is used to eliminate the influence of sensor range and motor power level on the blind zone determination threshold.

4. The system according to claim 3, characterized in that, The entropy weight state sensing unit uses a hyperbolic tangent function model to generate the dimensionless confidence factor, and the calculation formula is as follows: in, The dimensionless confidence factor; It is the hyperbolic tangent function; The normalized entropy divergence; Sensitivity gain; This is the trigger threshold; The hyperbolic tangent function model is used to provide continuously differentiable weight changes during the switching process between the steady-state region and the blind zone, so as to achieve jitter-free soft switching of the control mode.

5. The system according to claim 1, characterized in that, The voltage vector probability distribution model stored in the probability prediction and reconstruction unit is a sparse lookup table structure. When calculating the conditional mathematical expectation, the probability prediction and reconstruction unit is configured to integrate or perform a weighted summation on the indexed probability density function, thereby reconstructing the lost physical quantity using prior probability in the zero-crossing region where the sensor signal-to-noise ratio is extremely low.

6. The system according to claim 1, characterized in that, The dynamic arbitration execution unit executes the following weighted fusion logic: in, This refers to the final driving instruction; These are the basic control commands; The reconstructed voltage feedforward vector; The dimensionless confidence factor; When the dimensionless confidence factor When the value approaches 1, the system automatically bypasses the basic control command, and the reconstructed voltage feedforward vector takes over the control to avoid amplification of feedback errors in the blind zone.

7. The system according to claim 1, characterized in that, The system runs in an edge computing controller and employs a heterogeneous computing architecture: The physical feature anchoring unit is deployed in the FPGA logic unit to achieve microsecond-level dead zone feature capture using parallel hardware logic; The probability prediction and reconstruction unit is deployed in the MCU or NPU computing unit to perform probabilistic inference operations; The FPGA logic unit and the MCU or NPU computing unit interact with each other via an on-chip bus or high-speed interface.

8. A motor adaptive predictive and optimal control method based on edge computing, characterized in that, The method, executed by an edge computing controller, is used to reconfigure control in the physical blind zone where the motor commutation zero crosses. The method includes: Physical feature anchoring steps: Lock the interlock dead zone window in the inverter switching sequence, capture the current clamping characteristics of the freewheeling phase of the anti-parallel diode within the window, extract the zero-point physical true value, and output the pure current sequence after real-time calibration. Entropy weight state perception steps: Receive the pure current sequence, calculate the signal information entropy within the sliding window, calculate the normalized entropy divergence based on the background noise benchmark, and generate a dimensionless confidence factor characterizing the observability of the system through nonlinear mapping; Probabilistic prediction reconstruction steps: Store the voltage vector probability distribution model constructed based on prior data. When the dimensionless confidence factor indicates that the system has entered the blind zone, query the model according to the current motor state and calculate the conditional mathematical expectation of the ideal voltage vector, and output the reconstructed voltage feedforward vector. Dynamic arbitration execution steps: Receive the basic control command and the reconstructed voltage feedforward vector, perform weighted fusion according to the dimensionless confidence factor, and generate the final drive command to drive the motor.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method of claim 8.

Images