Energy efficiency optimization control method for motor system driven by industrial Internet of Things

Abstract

The invention discloses an energy efficiency optimization control method for a motor system driven by an industrial internet of things, and provides a three-stage progressive energy efficiency optimization algorithm: in the first stage, a collaborative optimization mechanism of a multi-physics coupling loss model and fuzzy logic reasoning is established; carrying out nonlinear mapping on iron loss, copper loss and mechanical loss by adopting a variable universe adaptive fuzzy controller; in the second stage, a self-adaptive parameter identification engine based on improved particle swarm optimization (PSO) is designed, and motor parameter online drift compensation and loss model coefficient dynamic correction are achieved; and in the third stage, a real-time execution strategy combining model predictive control (MPC) and dynamic flux linkage trajectory planning is implemented, and the amplitude and phase angle of the stator flux linkage are actively adjusted according to a load torque prediction result. The method overcomes the defects that traditional loss model control (LMC) is sensitive to parameters and search control (SC) is slow in convergence.

Description

Technical Field

[0001] This invention belongs to the field of industrial automation control technology, specifically relating to an energy efficiency optimization control method for an industrial Internet of Things driven motor system. Background Technology

[0002] With the deepening of intelligent manufacturing, motor systems, as the primary energy consumer in the industrial sector, directly impact enterprise operating costs and carbon emission indicators through their energy efficiency. Existing technologies mainly suffer from the following three problems: First, traditional motor energy efficiency optimization control methods are mainly divided into two categories: Loss Model Control (LMC) and Search Control (SC). The LMC method relies on precise motor parameters to establish a mathematical model of losses and analytically solves for the optimal excitation current to maximize efficiency. However, it suffers from sensitivity to motor parameter drift and poor robustness. The SC method finds the minimum input power point by addressing small disturbances near the operating point, requiring no precise model, but it suffers from slow convergence, poor dynamic response, and susceptibility to oscillations. Both methods are ill-suited to the complex and ever-changing operating conditions in industrial settings.

[0003] Second, existing motor control systems suffer from severe information silos and lack a unified industrial internet communication architecture. The field layer primarily uses traditional fieldbus protocols such as MODBUS and PROFIBUS, which have low data sampling rates and high transmission delays, making it difficult to support the high-frequency, real-time data requirements of energy efficiency optimization algorithms. The lack of interoperability between the management and execution layers prevents energy efficiency data from being effectively uploaded to the cloud platform for big data analysis, and also hinders the rapid dissemination of optimization decisions from the cloud to the execution layer, creating a "data can't get in, instructions can't get out" dilemma.

[0004] Third, the accuracy of motor loss modeling is insufficient. Existing technologies mostly use simplified iron loss equivalent resistance models, neglecting multi-physical field effects such as frequency nonlinearity, magnetic flux saturation, and temperature coupling, resulting in model prediction errors exceeding 15%. Copper loss models do not fully consider the impact of dynamic changes in winding temperature on resistance values, and mechanical loss models do not effectively integrate the nonlinear characteristics of bearing friction and wind resistance, significantly reducing the actual effectiveness of model-based optimization control strategies.

[0005] To address the aforementioned technical issues, this invention proposes an energy efficiency optimization and control method for motor systems driven by the Industrial Internet of Things (IIoT). By constructing an "edge-cloud" collaborative architecture integrating OPC UA-MQTT protocols and designing a three-stage progressive optimization algorithm driven by both "model" and "data", the method achieves online precise optimization and intelligent operation and maintenance of motor system energy efficiency. Summary of the Invention

[0006] The purpose of this invention is to provide an energy efficiency optimization control method for an industrial Internet of Things driven motor system, the core of which is reflected in the following five aspects: (I) Heterogeneous Protocol Fusion Architecture

[0007] A two-layer communication architecture integrating OPC UA and MQTT protocols is constructed. At the device layer, a lightweight OPC UA server is deployed within the embedded controller of the motor driver, utilizing OPC UA's deterministic communication mechanism to collect motor status data (stator current). Rotor speed DC bus voltage Millisecond-level data acquisition and command issuance, sampling period The 5-10ms timeout setting meets the high-speed real-time requirements of the motor vector control loop. Between the edge layer and the cloud platform, an asynchronous transmission of energy efficiency optimization parameters is achieved using the MQTT publish / subscribe protocol. Topic naming follows the IEC 62541 standard naming convention, and the Message Quality of Service (QoS) level is set to Level 1 to ensure reliable "at least once" delivery of critical control commands. The edge gateway incorporates a Protocol Conversion Engine (PCE) to dynamically map the OPC UA address space to the MQTT topic tree. Mapping rules are stored in an embedded SQLite database, supporting hot updates. This architecture ensures both real-time determinism in device-level control and flexibility and scalability in cloud-edge collaboration.

[0008] (II) Multiphysics Coupling Loss Model

[0009] A refined loss model incorporating iron loss, copper loss, and mechanical loss is established. The iron loss model employs a variable-coefficient Bertotti separation model, which incorporates the eddy current loss coefficients. With hysteresis loss coefficient Designed for frequency With magnetic flux density Bivariate functions: ; ; in, and The reference loss coefficient under rated operating conditions is calibrated from the motor's factory test data, and its value range is as follows: and ; , This is the frequency correction factor; , This is the saturation correction factor; , It is a non-linear exponent; The saturation magnetic flux density ranges from 1.2T to 1.6T. The reference frequency is set to 50Hz.

[0010] The hysteresis loss coefficient is the product of the reference hysteresis loss coefficient, the frequency correction term, and the flux saturation correction term. The frequency correction term is 1 plus the product of the frequency correction coefficient and the difference between the actual frequency and the reference frequency. The flux saturation correction term is 1 plus the product of the saturation correction coefficient and the nonlinear exponent of the ratio of the actual flux density to the saturation flux density.

[0011] The eddy current loss coefficient is the product of the reference eddy current loss coefficient, the frequency square correction term, and the magnetic flux saturation correction term. The frequency square correction term is 1 plus the product of the frequency correction coefficient and the square of the difference between the actual frequency and the reference frequency. The magnetic flux saturation correction term is 1 plus the product of the saturation magnetic flux correction coefficient and the nonlinear exponent of the ratio of the actual magnetic flux density to the saturation magnetic flux density.

[0012] The copper loss model incorporates a dynamic temperature compensation mechanism: ; ; in, , for - Shaft stator current components, The winding resistance at 20℃ The temperature coefficient of resistance of copper. The real-time temperature of the windings is measured using a PT100 resistance temperature detector (RTD) sensor. This is a reference temperature.

[0013] The instantaneous power loss of copper is the product of the coefficient and the sum of the squares of the direct and quadrature axis stator current components, and then multiplied by the winding resistance considering the real-time temperature.

[0014] The real-time winding resistance is the winding resistance at the reference temperature, multiplied by 1 plus the product of the copper resistance temperature coefficient and the difference between the real-time winding temperature and the reference temperature.

[0015] The mechanical loss model uses a secondary wind speed model: ; in, The bearing friction coefficient, with a range of values. N·m·s / rad, The drag coefficient has a range of values. N·m·s² / rad².

[0016] Mechanical loss power is the sum of bearing friction loss and wind resistance loss; where bearing friction loss is the product of bearing friction coefficient and rotor speed, and wind resistance loss is the product of wind resistance coefficient and the square of rotor speed.

[0017] (III) Adaptive Fuzzy Controller with Variable Universe

[0018] Design a dual-input, single-output variable universe-of-discourse fuzzy controller, with the input variable being the real-time efficiency deviation of the motor. and its rate of change The output variable is the excitation current adjustment coefficient. Traditional fuzzy controllers suffer from a contradiction between control accuracy and the number of rules due to a fixed universe of discourse. This invention introduces a scaling factor to dynamically adjust the universe of discourse range. Input universe scaling factor: ; ; The scaling factor of the first input domain at the next moment is equal to the scaling factor at the current moment, plus the product of the first learning rate, the current efficiency deviation, and the rate of change of efficiency deviation.

[0019] The scaling factor of the second input domain at the next moment is equal to the scaling factor at the current moment, plus the product of the second learning rate and the absolute value of the rate of change of efficiency deviation.

[0020] Output universe scaling factor: ; The output universe scaling factor at the current moment is equal to the product of the reference scaling factor, 1, the efficiency deviation sensitivity coefficient, and the absolute value of the efficiency deviation.

[0021] in, , For learning rate, As the baseline scaling factor, The efficiency deviation sensitivity coefficient is used. The fuzzy rule base adopts Mamdani-type inference, with a total of 25 rules. The antecedent of each rule is... and The level 7 language variables {NB, NM, NS, ZO, PS, PM, PB} have the consequent as... The five-level variables {NM, NS, ZO, PS, PM} are defuzzified using the centroid method.

[0022] (iv) Improve the particle swarm adaptive parameter identification engine

[0023] To address the motor parameter drift problem, a dual-loop nested particle swarm optimization (DN-PSO) algorithm is designed. The outer PSO layer is responsible for identifying slowly time-varying parameters (stator resistance). Rotor resistance ), particle dimension Population size Inertial weight Adopt a linear decreasing strategy: Learning factor The inner PSO layer is responsible for identifying fast time-varying parameters (iron loss equivalent resistance). Mutual induction ), particle dimension Population size Inertial weight Learning factor The fitness function is designed as follows: ; in, The sampling window length, The regularization coefficient is . These are the parameter ratings. When the parameter identification result changes by more than 5% compared to the last update, or the cumulative running time exceeds 300 seconds, the model prediction controller parameters are refreshed.

[0024] The fitness function is the sum of the squares of the differences between the measured loss and the model loss within the sampling window, plus the product of the regularization coefficient and the sum of the absolute values ​​of the differences between the parameter to be identified and its nominal value.

[0025] (v) Model Prediction of Dynamic Flux Linkage Trajectory Planning

[0026] To address the energy efficiency limitations of traditional vector control fixed flux commands, this invention designs a finite-time-domain rolling optimization strategy based on load torque prediction.

[0027] The objective function is designed as a multi-objective weighted form: ; in, To predict the time domain, As energy consumption weight, For torque tracking weights, Current smoothing weights are automatically adjusted based on the load factor: when the load factor... hour, Take the upper limit. Take the lower limit; when At the same time, the weighting is reversed.

[0028] The objective function value is the sum of the product of energy consumption weight and total loss, torque tracking weight and the square of torque tracking error, and current smoothing weight and the square of stator current deviation in the prediction time domain.

[0029] The magnetic flux trajectory constraint boundary adopts an elliptical dynamic adjustment mechanism: ; in, For load rate Functions: , This is the baseline ratio under no-load conditions. and These represent the maximum allowable offsets of the d-axis and q-axis currents, respectively. This constraint ensures increased freedom of flux linkage adjustment under light loads to maximize energy savings, while tightening the boundary under heavy loads to guarantee torque response capability.

[0030] The square of the ratio of the difference between the quadrature axis current and the quadrature axis reference current to the maximum permissible offset of the quadrature axis, plus the square of the ratio of the difference between the direct axis current and the direct axis reference current to the maximum permissible offset of the direct axis, results in a value less than or equal to the square of the load factor correlation coefficient.

[0031] This invention overcomes the shortcomings of traditional loss model control (LMC) in terms of parameter sensitivity and search control (SC) in terms of slow convergence. Attached Figure Description

[0032] Figure 1 This is a flowchart of an energy efficiency optimization control method for an industrial IoT-driven motor system according to the present invention. Detailed Implementation

[0033] like Figure 1 As shown, the energy efficiency optimization execution algorithm of this invention adopts a three-stage progressive architecture, with each stage containing several sub-steps. These steps are tightly coupled through data flow and control flow, forming a closed-loop optimization. The algorithm execution cycle is synchronized with the motor control cycle, with a base cycle... Running on an STM32H743 microcontroller, the CPU utilization is kept below 65% to ensure real-time performance.

[0034] Phase 1: Data Fusion and Co-optimization of Fuzzy-Loss Model

[0035] The core of the first stage is to collect raw data from multi-source heterogeneous sensors, perform spatiotemporal alignment and quality verification, and then input the data into a multiphysics loss model and a variable-domain fuzzy logic controller to generate a preliminary optimal setpoint for the excitation current. This stage includes three sub-steps: multi-sensor data fusion, online calculation of the loss model, and fuzzy logic inference decision-making.

[0036] Sub-step 1.1: Multi-sensor data fusion

[0037] Multi-sensor data fusion involves spatiotemporal registration and information complementation of sensor data from different sampling frequencies, physical dimensions, and time delays, providing a unified, accurate, and reliable input vector for subsequent loss models. This sub-step is implemented in the driver microcontroller through DMA (Direct Memory Access) and interrupt service routines, automatically completing data acquisition without CPU intervention.

[0038] The primary challenge in data fusion is the asynchronous sampling problem of sensors. Stator current sensors operate at switching frequencies... Synchronous sampling (e.g., 10kHz), speed encoders trigger sampling based on position change events (e.g., every 1 / 1024 revolutions), temperature sensors poll via an SPI interface at a fixed frequency of 1kHz, and flux sensors output analog signals at a high frequency of 50kHz. Directly using the latest data for splicing would introduce model errors due to time deviations. This invention employs a unified timestamp alignment mechanism. All sensor data is immediately marked with a high-precision timestamp based on the CYCCNT register value of the ARM Cortex-M7 core DWT (DataWatchpointandTrace) module at the moment of acquisition. This register counts at the CPU clock frequency (400MHz) with a resolution of 2.5ns.

[0039] The data fusion buffer mechanism employs a triple buffering structure, defining three circular buffers: RawBuffer (raw data area), FusedBuffer (fused data area), and ValidatedBuffer (validation data area), each with a capacity of 128 units. The DMA controller directly writes the ADC conversion results into the RawBuffer. When the RawBuffer is half full (64 units), a DMAHalfTransfer interrupt is triggered, initiating timestamp marking and preliminary filtering within the interrupt service routine. The preliminary filtering uses a five-point cubic smoothing algorithm with a window width of 5 sampling points. The smoothing coefficients are determined through offline least-squares fitting, effectively suppressing ADC quantization noise and electromagnetic interference.

[0040] In each control cycle At the start, the CPU reads all sensor data from the ValidatedBuffer within the most recent 10ms window and calculates the median of the valid timestamps for each sensor data point. Select timestamp and Data with a deviation of less than 2ms is considered valid data. If a single sensor has no valid data within this window (e.g., a temperature sensor with a sampling period of 1ms can always acquire data), then linear extrapolation is used for estimation. ,in This represents the rate of change of the sensor's historical data.

[0041] To address sensor malfunctions or data anomalies, this invention implements a data quality check. For the stator current, the sum of the instantaneous values ​​of the three-phase currents is calculated. Ideally, it should be close to zero; if its absolute value exceeds the threshold for five consecutive cycles... ( If the current is not within the rated current range, it is determined that the current sensor is malfunctioning, triggering an alarm and replacing it with redundant current sensor data (if available) or observations based on a voltage model. For temperature sensors, a rationality check based on physical constraints is performed; the winding temperature must meet the following requirements. and ( (The maximum allowable temperature for the insulation class, e.g., 155℃ for Class F); otherwise, it is considered a sensor fault, and the estimated value is obtained using a thermal model. : ; in, For heat capacity, Thermal resistance is calculated from the motor's structural parameters.

[0042] The rate of change of the estimated winding temperature is equal to the ratio of the sum of copper losses and iron losses to the heat capacity, minus the ratio of the difference between the estimated temperature and the ambient temperature to the thermal resistance.

[0043] Sub-step 1.2: Online calculation of the loss model

[0044] After data fusion is completed, the input vector with a unified time scale will be used. Input a multi-physics coupling loss model to calculate iron loss under the current operating conditions in real time. Copper loss Mechanical damage and total loss This provides a basis for decision-making in fuzzy logic controllers.

[0045] The core of online loss model calculation lies in the real-time updating of the variable coefficient mechanism. Traditional models use fixed coefficients, while this invention uses the iron loss coefficient... and Designed for frequency With magnetic flux density The bivariate function reflects frequency nonlinearity and magnetic saturation effect. Frequency From rotational speed With extreme logarithms calculate: Magnetic flux density Through stator magnetic flux Estimate: ,in ( For stator inductance, (for mutual induction) ( (Effective air gap cross-sectional area). This dynamic coefficient calculation avoids the large memory consumption problem of traditional table lookup methods and can capture the loss characteristics of the motor under all operating conditions from no-load to overload.

[0046] Iron loss calculation employs time-domain integration rather than frequency-domain analysis to meet the real-time requirements of the control cycle. Eddy current loss components are calculated as follows: ; The instantaneous value of eddy current loss is the product of the eddy current loss coefficient and the square of the stator flux change rate, divided by the equivalent resistance of iron loss.

[0047] in, Approximation using the flux difference between the current period and the previous period: , The iron loss equivalent resistance is initially fitted using no-load test data. The hysteresis loss component is calculated as follows: ; in, The shape factor of the hysteresis loop, with a range of values. The coefficient is determined by the silicon steel sheet material type (such as 50W470, 35WW300), and it corrects the approximate calculation error of hysteresis loss.

[0048] The instantaneous value of hysteresis loss is the product of the hysteresis loss coefficient, the motor frequency, the square of the stator flux linkage, and the hysteresis loop shape coefficient.

[0049] In copper loss calculation, winding resistance Temperature compensation is achieved through real-time winding temperature. accomplish: ; in, By DC injection method in the cold state of the motor ( The measurement is performed and stored in non-volatile memory (Flash). The measurement current is 10% of the rated current, and the measurement time is 200ms. Motor rotation should be avoided.

[0050] The real-time winding resistance is the winding resistance at the reference temperature, multiplied by 1 plus the product of the copper resistance temperature coefficient and the difference between the real-time winding temperature and the reference temperature.

[0051] After calculating the instantaneous power of copper loss, a 1-second moving average filter is required to eliminate PWM switching noise. ; The average power of copper loss is the sum of the instantaneous values ​​of copper loss at each moment within the moving average window, and the ratio of the window length.

[0052] In mechanical loss calculation, bearing friction coefficient With drag coefficient It is not a constant, but a function of rotational speed and temperature: ; ; in, For bearing temperature, For reference temperature, This refers to the temperature coefficient of bearing grease. The nonlinear coefficient of wind resistance. This is the rated speed.

[0053] The real-time bearing friction coefficient is the product of the reference bearing friction coefficient and 1, plus the product of the bearing grease temperature coefficient and the difference between the real-time bearing temperature and the reference temperature.

[0054] The real-time drag coefficient is the product of the reference drag coefficient and 1, plus the drag nonlinearity coefficient and the square of the ratio of the actual rotor speed to the rated speed.

[0055] In each The cycle begins by reading the fused input vector. Calculate in sequence , , Then update , , , , The time-varying coefficients are calculated last. , , and total loss The calculation results are stored in a structure called LossStruct, which contains the loss values ​​of each component and the total loss value, for subsequent use by the fuzzy controller.

[0056] To verify the accuracy of the loss model calculation, this invention implements a model output cross-validation mechanism. The total loss calculated by the model is then... The measured loss is calculated by the difference between electrical power input and mechanical power output. Compare: ; in, This is the DC bus current, measured through a DC-side shunt (such as Tyco's 0.5mΩ / 50W). The electromagnetic torque is measured or estimated directly using a torque sensor (such as HBMT40B). If the deviation between the two exceeds 15% for more than 10 cycles, the model self-learning flag is triggered, and the parameter identification engine (second-stage algorithm) is started for online correction.

[0057] The measured total loss is the difference between the input electrical power and the output mechanical power; where the input electrical power is the product of the DC bus voltage and the DC bus current, and the output mechanical power is the product of the electromagnetic torque and the rotor speed.

[0058] Sub-step 1.3: Variable universe fuzzy logic reasoning decision

[0059] Total loss output based on loss model Calculate the real-time efficiency of the motor This leads to efficiency deviation. Efficiency deviation is the difference between the efficiency reference value and the real-time efficiency. Efficiency reference value It is not a fixed value, but depends on the load rate. Dynamic tuning: ,in This is for the motor's maximum efficiency (e.g., 94%). This nonlinear tuning ensures that the efficiency target is not overly stringent under light loads, while approaching the maximum efficiency under heavy loads. The fuzzy controller uses... and its rate of change The input is scaled by varying the universe of discourse, then fuzzified, rule-based inference is performed, and defuzzification is applied to output the excitation current adjustment coefficient. The rate of change of efficiency deviation is the ratio of the difference between the efficiency deviation at the current moment and that of the previous control cycle to the duration of the control cycle.

[0060] The core of variable universe fuzzy control lies in the dynamic adjustment of the scaling factor, which solves the "rule explosion" problem in traditional fuzzy control. When efficiency deviation... When it is large, the scaling factor Automatically increasing and expanding the input domain enables the fuzzy controller to respond quickly; when the deviation decreases, Scaling down improves control precision. This adaptive mechanism keeps the number of fuzzy rules constant at 25, but achieves equivalent control precision to a traditional controller with over 100 rules. The scaling factor update employs an incremental learning strategy, avoiding complex gradient calculations and making it suitable for real-time implementation in microcontrollers.

[0061] Input variables The basic domain of discourse is defined as (i.e., ±10% efficiency deviation) The domain is (That is, changing by 50% per second). After scaling factor. , After scaling, the actual universe of discourse is and Fuzzy subsets are assigned Gaussian membership functions: ; in, As the membership function center, the level 7 language variables {NB, NM, NS, ZO, PS, PM, PB} are respectively set as , The width is uniformly set to 0.8. Output variable. The domain is After scaling factor The actual output range after scaling is .

[0062] The membership degree of a fuzzy subset is the value of an exponential function with the natural constant as the base, which is the ratio of the squared difference between the input variable with a negative exponent and the center of the membership function to twice the squared width of the membership function.

[0063] The fuzzy rule base uses Mamdani-type inference and has 25 rules. The antecedent of each rule is... and The 7×7 combination, the latter being The 5-level variables. The rule design follows the principle of "if the efficiency is low and continues to decrease, the excitation will be increased significantly". The rules are stored as a two-dimensional array `rule_base[7][7], and the element value is the output fuzzy subset index, which occupies only 49 bytes of memory. The inference process adopts MIN-MAX synthesis: first calculate the activation intensity of each rule. Then, for rules with the same consequent, the maximum activation strength is taken as the final output membership function. .

[0064] In each Period, first calculate and Then update the scaling factor. , , Next, the input variables are scaled to the fuzzy universe, and the activation degree of each membership function is calculated. The 25 rules are iterated through, and the activation strength of each rule is calculated to synthesize the output membership function. Finally, the centroid method is used to defuzzify the system. ; in, To output the center point of the fuzzy subset, the value is... .

[0065] The excitation current adjustment coefficient is the product of the output universe of discourse scaling factor, the sum of the products of the center points and corresponding membership degrees of the output fuzzy subsets, and the ratio of the sum of the membership degrees of the output fuzzy subsets.

[0066] The adjustment coefficient output by the fuzzy controller is applied to the excitation current reference value: ; in, The base excitation current is taken as the no-load excitation current for induction motors. For permanent magnet synchronous motors, zero is used. (Adjusted) The data is fed into the second-stage adaptive parameter identification engine for correction, and finally into the third-stage model prediction controller.

[0067] The excitation current reference value is the sum of the basic excitation current and the excitation current adjustment coefficient.

[0068] Learning rate of a variable universe fuzzy controller and The tuning of this setting is crucial for system stability. If Too large, scaling factor Severe fluctuations will lead to control oscillations; if the value is too small, the adaptive capability will be insufficient. This invention provides a tuning method based on Lyapunov stability theory: selecting a Lyapunov function... , require its difference The derivation yields... The upper limit is In actual engineering projects, conservative values ​​are adopted. , Experiments have verified that the system can achieve stable convergence when the load torque changes by a step (±50%), with an overshoot of less than 3%.

[0069] Phase 2: Adaptive Parameter Optimization and Online Identification

[0070] The core of the second stage is to perform online identification and adaptive optimization of the time-varying parameters in the loss model of the first stage, ensuring that the model accuracy is dynamically maintained despite factors such as motor aging, temperature changes, and magnetic circuit saturation. This stage includes two sub-steps: improved particle swarm optimization parameter identification and model predictive controller parameter refresh.

[0071] Sub-step 2.1: Improved Particle Swarm Optimization (DN-PSO) Parameter Identification

[0072] The DN-PSO algorithm addresses the problem of motor parameter identification by employing a double-ring nested structure to handle slow time-varying parameters and fast time-varying parameters respectively. The outer PSO identifies stator and rotor resistance (time constant in minutes), while the inner PSO identifies iron loss equivalent resistance and mutual inductance (time constant in seconds). Each layer independently sets the population size, inertia weight, and learning factor to improve identification efficiency and accuracy.

[0073] Traditional PSO (Programmable Optimization Search) is prone to premature convergence in parameter identification, especially in multi-extremum parameter spaces where it gets trapped in local optima. This invention's double-nested strategy decouples the high-dimensional parameter space: the outer layer searches in a low-dimensional, slowly changing space, providing initial values ​​for the inner layer; the inner layer performs a fine-grained local search, rapidly tracking parameter changes. Furthermore, the dynamic adjustment strategy of inertial weights balances global exploration and local exploitation capabilities: large weights initially ensure extensive particle exploration, while small weights later promote precise convergence. The regularization term of the fitness function prevents parameters from deviating excessively from their nominal values, improving the physical plausibility of the identification results.

[0074] The outer PSO algorithm parameters are configured as follows: particle dimension Corresponding optimization variables ,in This refers to the rotor resistance (induction motor) or the equivalent damping winding resistance (permanent magnet synchronous motor). Population size. Particle positions are initialized within ±20% of the nominal value: , Particle velocity limit To prevent sudden changes in position, the inertia weight decreases linearly. ; in, , , Maximum number of iterations. Learning factor. The individual optimal weight is equal to the global optimal weight.

[0075] The inertial weight of the outer particle swarm at the current iteration is equal to the maximum inertial weight minus the product of the difference between the maximum and minimum inertial weights and the ratio of the current iteration number to the maximum iteration number of the outer layer.

[0076] Inner PSO algorithm parameter configuration: Particle dimension Corresponding variables , This is the iron loss equivalent resistance. Mutual intuition. Population size. Position initialization range: , Inertia weight , The inertia weight of the inner particle swarm at the current iteration is equal to the initial inertia weight minus the product of the change in inertia weight and the ratio of the current iteration number to the maximum number of iterations in the inner layer. Learning factor. Enhance local development capabilities.

[0077] The fitness function design reflects the relationship between model error and parameter constraints: ; in, The parameter vector to be identified, This is the number of samples taken in the sliding window, covering 500ms of runtime data. The model loss is calculated using the current parameter vector. This represents the measured loss. Regularization coefficient. Balancing model accuracy with parameter rationality to prevent Parameters fluctuate abnormally due to measurement noise.

[0078] The fitness function is the sum of the squares of the differences between the measured loss and the model loss within the sampling window, plus the product of the regularization coefficient, the sum of the squares of the differences between the parameter to be identified and its nominal value, and the ratio of the squares of the parameter nominal values.

[0079] DN-PSO triggers a full identification every 300 seconds, or an emergency trigger when data quality verification reveals a model error exceeding 15%. Each iteration of the outer PSO calls the inner PSO for 20 rapid optimizations, using the optimal fitness value of the inner layer as the evaluation criterion for the outer particles. The specific process is as follows: 1. Initialization: Randomly generate 20 particles in the outer parameter space and record their optimal positions. with the global optimal position .

[0080] 2. Outer layer iteration: For each particle Fix its value and call the inner PSO optimization. Inner PSO (Previous identification result) is used as the center for initialization.

[0081] 3. Inner layer optimization: The inner PSO is executed for 80 iterations and returns the optimal fitness. Corresponding .

[0082] 4. Outer layer evaluation: As particles Adaptability, update and .

[0083] 5. Velocity Update: Update particle velocity and position according to the standard PSO formula: ; ; in, It is a random number in the range [0,1].

[0084] The particle velocity at the next moment is equal to the product of the current inertial weight and the current particle velocity, plus the product of the individual learning factor and the random number, the difference between the individual particle's optimal position and the current position, plus the product of the global learning factor and the random number, the difference between the population's global optimal position and the current position.

[0085] The particle's position at the next moment is equal to the sum of the particle's current position and the particle's velocity at the next moment.

[0086] 6. Convergence judgment: When the global optimal fitness improves by less than 0.1% for 5 consecutive iterations or reaches the maximum number of iterations of 100, the outer loop is terminated and the final identification parameters are output.

[0087] DN-PSO involves significant computation, requiring approximately 5 seconds of CPU time for a single complete identification, far exceeding the 10ms control cycle. Therefore, this invention employs a background task mechanism, running DN-PSO in a low-priority thread, parallel to the real-time control thread. The control thread uses the previous identification result each cycle, and the background thread atomically updates the parameters upon completion, using double buffering to avoid data races. Furthermore, to accelerate convergence, prior knowledge is introduced: a database of initial parameter values ​​is established using the motor's factory test data. Initial values ​​are queried based on the current load rate and temperature, bringing the initial PSO population closer to the true values, reducing the number of iterations by approximately 40%.

[0088] Sub-step 2.2: Refreshing Model Predictive Controller Parameters

[0089] After the DN-PSO completes parameter identification, the updated parameters need to be passed to the Model Predictive Controller (MPC) in the third stage to refresh the prediction model matrix, constraint boundaries and objective function weights, so as to ensure that the MPC makes optimization decisions based on the latest motor characteristics.

[0090] MPC performance is highly dependent on model accuracy; parameter mismatch can lead to accumulated prediction errors and even instability. This invention employs a dual mechanism: parameter change rate triggering and timed triggering. A refresh is immediately triggered when the change rate of any parameter exceeds 5%; a forced refresh is also applied even if the change rate is small but the accumulated runtime exceeds 300 seconds, preventing performance degradation due to slow parameter drift. The refresh process uses atomic operations to ensure that MPC does not experience instantaneous anomalies during parameter switching.

[0091] Parameter refresh involves updating multiple data structures within the MPC, specifically including: 1. Prediction model matrix update: update the identified matrix. Substitute into the discrete state space matrix Recalculate the prediction time domain The state transition matrix inside With input matrix : ; 2. Loss model coefficient update: update the identified coefficients. The coefficients are written into the LossModelParams structure, which is defined as follows: typedef struct { floatRs; / / Stator resistance floatRr; / / Rotor resistance floatRfe; / / Iron loss equivalent resistance floatLm; / / Mutual inductance floatkh0; / / Reference hysteresis loss coefficient floatke0; / / Reference eddy current loss coefficient floatalpha_h; / / Frequency correction coefficient floatalpha_e; / / Frequency correction coefficient / / ...other coefficients LossModelParams; The parameter update uses the volatile modifier to ensure that the compiler does not perform optimizations and to guarantee data consistency.

[0092] 3. MPC constraint boundary update: stator current amplitude limit It is determined by the motor's heat capacity, and the formula is: ; in, The thermal resistance from the winding to the environment, with a range of values. K / W is determined by the motor cooling method (self-cooling, air cooling, water cooling). This dynamic constraint ensures that the motor winding temperature does not exceed the insulation class limit under any operating conditions.

[0093] The maximum allowable amplitude of the stator current is the square root of the difference between the rated temperature and the real-time winding temperature, and the product of the real-time winding resistance and the copper thermal resistance.

[0094] 4. Objective Function Weight Update: Weight coefficients in the MPC objective function Based on load rate Automatic adjustment. The load factor is obtained from the torque estimate after low-pass filtering: ; in, These are the filter coefficients.

[0095] The filtered load rate is the product of the filter coefficient and the ratio of the actual torque to the rated torque, plus 1 and minus the filter coefficient, and the load rate of the previous control cycle.

[0096] The weighting adjustment rules are as follows: Light load area ( ): Prioritize energy conservation Medium load area ( ): Balancing efficiency and dynamism Heavy load area ( ): Priority torque response Parameter updates are performed in a background thread. When DN-PSO generates new parameters, the atomic flag `params_updated_flag` is set to 1. The MPC main thread checks this flag in each control cycle. If it is 1, a refresh sequence is initiated: first, one row of state matrix A is updated, and the cycle counter is incremented by 1. When the counter reaches 20, all updates are complete, and the flag is cleared. This step-by-step refresh mechanism avoids excessively long single refresh times that could affect real-time performance.

[0097] To ensure a smooth refresh process, this invention introduces a parameter transition mechanism. New parameters do not directly replace old parameters, but are transitioned through a first-order low-pass filter: ; Among them, the transition coefficient The parameter switching is completed smoothly over approximately 30 cycles (300ms), avoiding current surges caused by sudden changes in control values.

[0098] The smoothed new parameters are the product of the transition coefficient and the new identification parameter, plus 1 minus the transition coefficient, and the product of the old parameters.

[0099] Phase 3: Real-time control execution and energy efficiency closed loop

[0100] The third stage is the core execution layer. Based on the optimization goals and accurate models provided in the first two stages, it implements model predictive control (MPC) and dynamic flux linkage trajectory planning to generate the final voltage command and drive the inverter. At the same time, it establishes an energy efficiency closed-loop feedback mechanism to feed back the actual energy efficiency indicators to the first stage, forming a complete control closed loop.

[0101] Sub-step 3.1: Model Predictive Control (MPC) Rolling Optimization

[0102] MPC solves a finite-time open-loop optimization problem in each control cycle and uses a predictive model to estimate the future. The system response is evaluated step by step, and the control sequence that optimizes the objective function is selected. Only the first step is executed, and the solution is recalculated in the next cycle to achieve rolling optimization.

[0103] MPC's advantage lies in its explicit handling of constraints and multi-objective optimization, but it involves enormous computational costs. This invention, tailored to the characteristics of motor control, employs a strategy combining offline explicit MPC computation with online table lookup. In the offline phase, parametric polytope partitioning is performed in the cloud, dividing the state space into several regions, each corresponding to a set of optimal control gains. In the online phase, only the region to which the current state belongs needs to be determined; the optimal control quantity can then be obtained through matrix multiplication, reducing the time to solve the quadratic programming (QP) problem from milliseconds to microseconds.

[0104] Offline computation is performed in the cloud, involving multi-parameter planning. (State vector) With control vector The constraint set is: ; ; The quadratic matrix of the objective function , ( (For control weights). Offline computation took approximately 2 hours (Intel Xeon Gold CPU), generating a lookup table file empc_table.bin, approximately 5MB in size, containing 5000 region divisions and corresponding gain matrices. This file was sent to the edge gateway via MQTT, and the gateway pushed it to the drive's Flash storage via FTP.

[0105] The state vector is constrained to ensure that the stator current of the quadrature and direct axes does not exceed the maximum allowable current and the rotor speed is within the range of minimum and maximum allowable speed.

[0106] The control vector is constrained so that the AC and DC axis voltage commands do not exceed the maximum voltage value corresponding to the DC bus voltage.

[0107] During online execution, the state partition identification uses a binary tree search algorithm, with an average number of comparisons. This is significantly lower than the 5000 searches required for a brute-force search. Each region is defined by a set of linear inequalities: ; in, For a matrix, It is a vector. The recognition process starts from the root node and calculates... If all elements are less than or equal to 0, proceed to the left subtree; otherwise, proceed to the right subtree, until a leaf node is reached.

[0108] In each During the cycle, MPC performs the following steps: 1. State prediction: based on the current measured state Compared with the control amount of the previous cycle Predicting the future using discrete models Step state sequence .

[0109] 2. Region identification: Given a binary tree, quickly locate the index of the region it belongs to. .

[0110] 3. Control Calculation: Reading Area The corresponding gain matrix Calculate the optimal control increment: ; in, , Output from the first-stage fuzzy controller Generated by the torque controller.

[0111] The optimal control increment is the product of the explicit gain matrix of the region to which the current state belongs and the difference between the state reference value and the measured state value.

[0112] 4. Constrained Projection: Projected onto the voltage constraint set: ; in, is the maximum allowable voltage increment, and sat is a saturation function.

[0113] The current control quantity is the sum of the control quantity of the previous cycle and the saturation value of the optimal control increment; where the saturation value is the result of limiting the optimal control increment to the maximum allowable control increment range.

[0114] 5. Command output: Converted into a three-phase voltage command via inverse Park transform The signal is fed into the SVPWM (Space Vector Pulse Width Modulation) module to generate the inverter switching signal.

[0115] To handle strong disturbances such as sudden load changes, MPC introduces a Disturbance Observer (DOB). This observeer measures the total disturbance... Expand to state: ; in, The rate of change of the disturbance is considered as white noise. The disturbance is observed using a Kalman filter and compensated for in the prediction model to improve robustness. Observer gain. The Riccati equation can be solved offline, requiring only matrix multiplication online, with an additional time consumption of less than 10μs.

[0116] Sub-step 3.2: Dynamic flux linkage trajectory planning and execution

[0117] The voltage command output by the MPC is based on a predictive model and does not fully consider magnetic circuit saturation and loss nonlinearity. This invention adds a dynamic flux linkage trajectory planning layer to the MPC, which actively adjusts the stator flux linkage amplitude and phase angle according to the load prediction results, thereby further improving energy efficiency.

[0118] Traditional vector control uses constant flux linkage control, meaning the flux linkage amplitude is the same under both light and heavy loads. However, iron loss dominates under light loads, and appropriately reducing the flux linkage can significantly reduce iron loss (iron loss ∝ While the output torque still meets the requirements, dynamic flux linkage trajectory planning is performed in... - An elliptical constraint boundary is constructed in the current coordinate system, and the boundary shape is adjusted in real time to guide the current operating point to move along the optimal efficiency curve.

[0119] The center of the elliptical boundary is ,in This is the current excitation current. For torque current. Along the major axis of the ellipse. axial direction, minor axis along Axial direction, ratio of major to minor axis By load rate Decide: ; This design means that under light load ( ), The ellipse approaches a circle, allowing for a larger size. Shaft adjustment range; under heavy load ( ), Ellipse flattened, restrictive shaft current to ensure The shaft current is sufficient to output torque.

[0120] The equations for the elliptic constraint are: ; in, , , , This is the maximum allowed offset.

[0121] The square of the ratio of the difference between the direct-axis current and the current at the center of the direct axis to the length of the semi-axis of the ellipse, plus the square of the ratio of the difference between the current at the cross-axis current and the current at the center of the cross-axis to the length of the semi-axis of the cross-axis of the ellipse, results in a value less than or equal to 1.

[0122] To achieve optimal efficiency trajectory tracking, the Lagrange multiplier method is introduced to transform the efficiency optimization problem into a constrained extremum problem. The Lagrange function is constructed as follows: ; The Lagrange function is the total power loss plus the product of the Lagrange multiplier and the difference between the torque reference value and the calculated electromagnetic torque value.

[0123] right Taking the partial derivative and setting it to zero, we obtain the optimal efficiency condition: ; The partial derivative of total losses with respect to the direct-axis current is zero, and the product of the partial derivative of total losses with respect to the quadrature-axis current and the correlation coefficient of the Lagrange multiplier and the electromagnetic torque is zero.

[0124] The optimal excitation current can be obtained by solving the problem. With optimal torque current The implicit relationship can be solved online using numerical methods (such as Newton's iteration). The initial value of the iteration is taken from the MPC output, and it usually converges in 2-3 iterations.

[0125] In each After the MPC outputs the voltage command, the dynamic flux linkage planning layer executes the following: 1. Load Forecasting: Read the torque forecast value from the edge gateway. (Through MQTT subscription) Calculate load rate .

[0126] 2. Ellipse boundary update: based on calculate , , Update the MPC constraint set.

[0127] 3. Optimal Trajectory Solution: Solve the Lagrange extremum problem to obtain the optimal current command. .

[0128] 4. Instruction Correction: Modify MPC output Projecting the image into the ellipse boundary; if it extends beyond the boundary, then shrinking it along the gradient direction to the boundary point: ; The final current command is the result of projecting the model's predicted control output current command onto the elliptical constraint boundary. If it exceeds the boundary, it shrinks along the gradient direction to the boundary point.

[0129] 5. Feedforward compensation: The optimal excitation current is fed forward to the voltage command. ; Among them, feedforward gain , It is a direct-axis inductor.

[0130] The direct-axis voltage command is the direct-axis voltage command output by the model predictive control, plus the product of the feedforward gain and the difference between the optimal direct-axis current and the direct-axis current output by the model predictive control.

[0131] Dynamic flux linkage planning significantly improves energy efficiency under light loads. Experimental data shows that when the load rate... At that time, the efficiency of traditional constant magnet linkage control was 78%, while the efficiency of this invention, after dynamic programming, was increased to 85.5%, achieving an energy saving effect of 9.6%. Under heavy load ( This results in an efficiency improvement of approximately 1.2%, proving the effectiveness of the strategy under both light and heavy loads. Elliptical boundary parameters. and The value needs to be tuned offline based on the motor torque-current characteristics. Too high a value will cause torque pulsation, while too low a value will limit energy-saving potential. It is recommended to determine the optimal value through finite element simulation.

[0132] Sub-step 3.3: Energy efficiency closed-loop feedback and self-learning

[0133] To achieve optimal long-term energy efficiency, this invention establishes an energy efficiency closed-loop feedback mechanism, feeding back actual operating energy efficiency indicators to the first stage to correct the fuzzy controller efficiency reference value. Together with the loss model coefficients, they form a self-learning closed loop.

[0134] During long-term operation of a motor, bearing wear leads to increased friction coefficient, insulation aging leads to increased leakage current, and permanent magnet demagnetization leads to decreased flux linkage. These slow changes gradually cause the initial model to become inaccurate. The energy efficiency closed-loop system updates fuzzy rules and model parameters online by statistically analyzing historical data of the optimal energy efficiency point, enabling the control system to be adaptive. This mechanism borrows from reinforcement learning, using energy efficiency improvement as a reward signal to guide the controller parameters towards the optimal direction.

[0135] Energy efficiency statistics are collected at the edge gateway, and the average efficiency is calculated every hour. Proportion to energy loss. Define the energy efficiency reward function: ; The energy efficiency bonus is the difference between the average efficiency at the current moment and the average efficiency one hour ago, and the ratio of the current average efficiency to the average efficiency one hour ago, which is then converted into a percentage.

[0136] like If the current control strategy is considered effective after 3 hours, positive learning is triggered: the optimal operating point in the past 3 hours ( The combined records are stored in the optimal working point database. The database uses a KD-tree (K-Dimensional Tree) structure for storage, supporting fast nearest neighbor queries. When the number of records in the database exceeds 1000, an incremental fuzzy rule extraction algorithm is activated. Based on FCM (FuzzyC-Means) clustering, the working points are clustered into 7 classes. A new fuzzy rule is generated for each class center, replacing the original poorly performing rules (rules with activation below the threshold).

[0137] Model parameter self-learning is only applicable to slowly changing parameters, such as bearing friction coefficient. By monitoring no-load loss (make The changing trend of ) can be corrected online. : ; Among them, learning rate The correction cycle is 1 hour to ensure smooth parameter evolution.

[0138] The bearing friction coefficient at the next moment is equal to the coefficient at the current moment, plus the product of the learning rate, the difference between the measured loss under no-load and the loss under no-load model, and the ratio of the rotor speed.

[0139] The energy efficiency closed loop performs the following operations per cycle: 1. Energy efficiency calculation: Instantaneous efficiency is calculated based on measured data. After a 1-second sliding average, the result is... .

[0140] 2. Reward Assessment: Rewards are calculated hourly. Determine the direction of learning.

[0141] 3. Data Recording: If , will the current working point Insert into the KD-tree database.

[0142] 4. Rule Update: Fuzzy rule extraction is triggered once a month. When the database size exceeds the threshold, the fcm_cluster() function is called to cluster and generate a new rule library. The new rule library is uploaded to the cloud via MQTT and then (optionally) issued after manual confirmation.

[0143] 5. Parameter correction: Under no-load conditions ( ), executed every 10 minutes Correction to prevent model errors caused by bearing aging.

[0144] Energy efficiency closed-loop self-learning enables the system to have long-term adaptability. In experiments simulating bearing aging, the friction coefficient is... The efficiency decreased by 2.5% when the self-learning mechanism was not present, but decreased by 0.02 when the self-learning mechanism was present. With the self-learning mechanism, the system adjusted... Reducing the flux linkage partially offsets frictional losses, resulting in an efficiency decrease of only 0.8%, and the learning process converges within 2 hours. The KD-tree database can be stored in SQLite on the edge gateway, with low query complexity. It has minimal impact on real-time performance.

[0145] This invention discloses an energy efficiency optimization control method for an industrial Internet of Things-driven motor system, proposing a three-stage progressive energy efficiency optimization algorithm: The first stage establishes a collaborative optimization mechanism combining a multi-physics coupled loss model and fuzzy logic reasoning, employing a variable universe adaptive fuzzy controller to perform nonlinear mapping of iron losses, copper losses, and mechanical losses; the second stage designs an adaptive parameter identification engine based on improved particle swarm optimization (PSO) to achieve online drift compensation of motor parameters and dynamic correction of loss model coefficients; the third stage implements a real-time execution strategy combining model predictive control (MPC) and dynamic flux linkage trajectory planning, actively adjusting the stator flux linkage amplitude and phase angle based on load torque prediction results. This invention overcomes the shortcomings of traditional loss model control (LMC) in terms of parameter sensitivity and slow convergence of search control (SC).

[0146] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art will be able to make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the preceding claims.

Claims

1. A method for optimizing the energy efficiency control of a motor system driven by the Industrial Internet of Things (IIoT), characterized in that, Includes the following steps: A two-layer heterogeneous communication architecture based on OPC UA and MQTT protocols is constructed. An OPC UA server is deployed at the device layer to realize real-time control data interaction of motors, and the MQTT protocol is used between the edge layer and the cloud platform for the publish / subscribe transmission of energy efficiency optimization commands and status monitoring data. A multi-physics coupled motor loss model is established, which includes iron loss components, copper loss components and mechanical loss components. The iron loss component adopts a variable coefficient model that considers frequency nonlinearity and magnetic flux saturation effect, the copper loss component adopts a winding resistance model that takes temperature correction into account, and the mechanical loss component adopts a wind friction loss model that is related to the square of the rotational speed. The variable universe adaptive fuzzy logic controller uses the real-time efficiency deviation and torque ripple rate of the motor as input variables and the excitation current adjustment coefficient as output variable. It dynamically adjusts the input and output universe ranges through a scaling factor to achieve online efficiency optimization of the motor under light load conditions.

2. The method according to claim 1, characterized in that, In the aforementioned two-layer heterogeneous communication architecture, the OPC UA server is deployed within the embedded controller of the motor driver, providing real-time data access services for motor current, voltage, and speed through the OPC UA client-server mode, with a sampling period of no more than 10ms; the MQTT message broker is deployed on the edge gateway, using QoS level 1 to ensure the transmission of energy efficiency control commands, with the release period dynamically adjusted according to the load, ranging from 50ms to 1000ms.

3. The method according to claim 1, characterized in that, The formula for calculating the iron loss component in the multiphysics coupling loss model includes the eddy current loss coefficient. and hysteresis loss coefficient These two coefficients are frequency and magnetic flux density The bivariate function was obtained by combining offline finite element simulation and online recursive least squares method, and the initial value was calibrated using the motor factory test data.

4. The method according to claim 1, characterized in that, The scaling factor of the variable universe adaptive fuzzy logic controller is based on... The pattern adapts and adjusts itself, among which The learning rate, with a value range of [0.01, 0.05]. For efficiency deviation, The efficiency deviation rate is the scaling factor at the next time step. The scaling factor at the current time step is equal to the scaling factor at the current time step plus the product of the learning rate, the current efficiency deviation, and the efficiency deviation rate.

5. The method according to any one of claims 1 to 4, characterized in that, The adaptive parameter identification engine employs an improved particle swarm optimization algorithm, where the particle position vector is defined as the stator resistance of the motor. Rotor resistance Iron loss equivalent resistance The set of values, the fitness function is designed as the mean square error between the model's predicted loss value and the actual measured loss value, and a dynamic inertial weight is introduced into the particle velocity update formula. ,in This represents the maximum number of iterations, ranging from 100 to 150. This represents the current iteration number; The inertia weight at the current iteration is equal to the maximum inertia weight, minus the product of the difference between the maximum and minimum inertia weights and the ratio of the current iteration number to the maximum iteration number.

6. The method according to claim 5, characterized in that, After each identification cycle, the adaptive parameter identification engine updates the optimized motor parameters into the loss model and simultaneously activates the parameter refresh mechanism of the model prediction controller. The refresh trigger condition is that the parameter change rate exceeds 5% or the cumulative running time reaches 300 seconds.

7. The method according to claim 1, characterized in that, The real-time control execution phase employs a finite-time-domain rolling optimization strategy, with the objective function being: ,in This represents the total power loss. For torque tracking error, For prediction in the time domain, the values ​​range from 0.5 seconds to 2 seconds, and the weighting coefficients are... and Adjust online based on load type, especially during light loads. The ratio is greater than 10, and the ratio is less than 2 when fully loaded; The objective function is the product of energy consumption weight and total power loss, and the product of torque tracking weight and the square of torque tracking error, integrated over the prediction time domain.

8. The method according to claim 7, characterized in that, The dynamic flux linkage trajectory planning is based on the load torque prediction results. - Construct an elliptical magnetic flux constraint boundary in a coordinate system, where the ratio of the major axis to the minor axis of this boundary is... Determined by the torque-flux flux optimal efficiency curve, when the load torque is predicted... hour, Take 1.2, when hour, We set the value to 0.8 and used linear interpolation in the middle region.

9. The method according to claim 1, characterized in that, The industrial IoT platform deploys a digital twin, which subscribes to the operating data of the physical motor entity via the OPC UA protocol, receives optimization instructions from the cloud platform via the MQTT protocol, and uses Unreal Engine or Unity3D engine to build a three-dimensional visualization interface to display the internal magnetic field distribution, loss hotspots and energy efficiency indicators of the motor in real time. The data synchronization delay between the twin and the physical entity is less than 200ms.

Images