Speed reducer multi-field loss equalization method and system based on working condition self-adaptive feedback

Abstract

The invention relates to the field of mechanical engineering, discloses a speed reducer multi-field loss balancing method and system based on working condition self-adaptive feedback, and aims to solve the problems of local loss concentration, intensified temperature rise and low energy efficiency caused by neglect of a multi-physical field coupling relation in the prior art. The method comprises the following steps: acquiring mechanical, thermal, fluid and electromagnetic parameters in operation of the speed reducer through a distributed sensor, and generating a standardized working condition state vector; and inputting into a mechanical-thermal-fluid-electromagnetic four-field coupling model, and solving the space-time loss distribution of the key component by combining with a data driving correction term. According to the scheme, refined prediction and closed-loop regulation and control of multi-field loss distribution are achieved, the local temperature difference and energy consumption concentration are remarkably reduced, the energy efficiency of the whole machine is improved by 8%-15%, the service cycle is prolonged, and the method is suitable for a high-end equipment transmission system under the high-dynamic heavy-load working condition.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, specifically relating to a method and system for balancing multi-field losses in a reducer based on adaptive feedback of operating conditions. Background Technology

[0002] With the continuous advancement of industrial automation and intelligent manufacturing, reducers, as core components of transmission systems, are widely used in fields such as robotics, CNC machine tools, wind power equipment, and new energy vehicles. Their operating efficiency and energy consumption directly affect the performance and reliability of the entire system. Under complex and variable operating conditions, reducers involve the coupling of multiple physical fields, including mechanical, thermal, and lubrication fields. Energy losses between these fields influence each other, leading to increasingly prominent problems such as concentrated local losses, increased temperature rise, and reduced lifespan. Therefore, achieving synergistic optimization of loss distribution under multiple field environments has become crucial for improving the overall efficiency of transmission systems.

[0003] In terms of energy consumption control during reducer operation, existing technologies mainly focus on the optimization design of a single field. For example, they may reduce mechanical friction loss by modifying the tooth profile, improve lubrication characteristics by using high-efficiency lubricants, or add heat dissipation structures to alleviate heat accumulation effects. However, such methods often neglect the dynamic coupling relationship between different physical fields, making it difficult to adapt to non-steady-state conditions such as load fluctuations, frequent start-stops, and changes in ambient temperature during actual operation. This results in limited optimization effects and may even lead to a "cost-benefit" effect under certain conditions, where a reduction in one type of loss may trigger a significant increase in another type of loss.

[0004] Current technologies have not yet established a dynamic loss balancing mechanism based on feedback from actual operating conditions, lacking unified modeling and real-time control capabilities for multiple loss paths from mechanical, thermal, and lubrication sources. Furthermore, most solutions do not incorporate adaptive feedback strategies, failing to autonomously adjust control parameters based on operating conditions to achieve optimal global efficiency. In addition, they lack precise perception and prediction capabilities regarding the changing patterns of loss distribution characteristics under different operating conditions, causing the system to operate at suboptimal operating points for extended periods, resulting in energy waste and accelerated component aging. These problems are particularly prominent in high-dynamic, heavy-load application scenarios, severely hindering the development of high-end equipment transmission systems towards higher efficiency and intelligence. A novel control method and system architecture capable of achieving adaptive operation, multi-field coordination, and loss balancing is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies by providing a multi-field loss balancing method and system for reducers based on adaptive feedback under operating conditions, which can effectively solve the problems in the background technology. Currently, when reducers operate under unsteady conditions such as complex load changes, speed changes, and frequent start-stops, strong coupling exists between multiple physical fields within them, including mechanical, thermal, lubrication, and electromagnetic fields (such as integrated motor drives). This leads to problems such as localized stress concentration, uneven temperature rise, oil film rupture, and unbalanced energy loss distribution. Traditional optimization strategies typically only target a single performance index with static design or open-loop control, failing to achieve coordinated adjustment of multi-field dynamic responses and making it difficult to achieve a comprehensive balance between efficiency, lifespan, and reliability. The resulting localized over-loss phenomenon significantly reduces the overall system energy efficiency and service life, hindering the further development of transmission systems for high-end equipment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: On one hand, a multi-field loss equalization system for a reducer based on adaptive feedback of operating conditions. This system includes the following components: a real-time operating condition sensing module, used to collect multi-source sensor data such as input torque, output speed, housing vibration signal, lubricating oil temperature and pressure, ambient temperature and humidity, and drive motor current and voltage during the reducer's operation, and to perform time synchronization and preprocessing on the data to generate a standardized operating condition state vector; a multi-field coupling modeling module, which receives the operating condition state vector, and based on a preset physical mechanism model and data-driven correction mechanism, constructs a time-varying model of gear meshing stiffness, a bearing support stiffness model, a dynamic evolution model of lubricating oil film thickness, and an electromagnetic loss distribution model. Then, through inter-field interface variable transfer, it achieves the mechanical-thermal-fluid-electromagnetic four-field coupling solution, outputting a spatial loss density distribution cloud map and time-series changes of each key component. The system comprises a curve; a loss balancing decision module, which receives the multi-field loss distribution results and, combined with a preset global objective function—with "loss balancing" as the core guideline—comprehensively considers three sub-objectives: maximum single-point loss suppression, regional cumulative loss smoothness, and overall system efficiency improvement, and uses a multi-objective weighted optimization algorithm to generate the optimal control instruction set; an adaptive execution module, which receives the control instruction set and, by adjusting the lubricating oil pump flow / pressure, cooling fan speed, active lubrication nozzle switching timing, clutch engagement pressure, or motor output characteristic parameters, achieves closed-loop intervention and dynamic reconstruction of the multi-field distribution within the reducer; and a data feedback update module, which collects actual operating status data after control, compares the predicted loss distribution with measured indirect indicators (such as infrared thermal imaging zone temperature difference and vibration energy spectrum distribution), calculates the deviation residual, and uses it to correct the multi-field coupled model parameters online, forming a complete feedback link from perception to execution to model evolution. Furthermore, in the real-time operating condition sensing module, the input torque and output speed are acquired through a high-precision torque sensor and a photoelectric encoder, with a sampling frequency of no less than 10kHz to ensure the capture of transient impact loads; the shell vibration signal is collected by a triaxial accelerometer array arranged at key support positions of the housing, with a spatial resolution of no less than 4 measuring points per square meter, used to identify structural resonance modes and stress propagation paths; the lubricating oil temperature and pressure are respectively measured by distributed micro-miniature PT100 temperature probes and piezoresistive pressure transmitters at the oil inlet, oil outlet, and tooth side injection area, with a response time of less than 50ms, ensuring accurate characterization of thermal-fluid coupling boundary conditions; the ambient temperature and humidity are measured by an external meteorological sensor and participate in the heat exchange boundary condition compensation calculation; all sensor data are timestamped and outlier filtered by an industrial-grade edge computing node, and a stable and reliable operating condition state vector sequence is output by using a sliding window mean method and wavelet denoising. Preferably, in the multi-field coupling modeling module, the time-varying model of gear meshing stiffness is established based on finite element contact analysis and Hertzian theory, taking into account tooth surface modification parameters, manufacturing errors, and the effects of elastic deformation. Its stiffness curve changes continuously with the meshing position and is used as an excitation source input to the dynamic equation. The bearing support stiffness model distinguishes between the contact angle changes of the rolling elements and the inner and outer rings, dynamically updates the support point stiffness matrix under different load directions, and supports the solution of 6-DOF displacement-force mapping relationship. The dynamic evolution model of lubricating oil film thickness adopts the modified Reynolds equation, introduces the viscosity-density relationship that varies with temperature and pressure, and the boundary conditions are jointly determined by the relative speed of gear motion and surface roughness. This achieves precise delineation of the elastohydrodynamic (EHL) lubrication region. The electromagnetic loss distribution model is applicable to integrated electric drive reducers, including stator copper loss, core eddy current and hysteresis loss, and rotor stray loss components. Its spatial distribution is closely related to the motor magnetic field harmonics and is calibrated through magnetic field-circuit coupling simulation. Furthermore, the four-field coupling solution adopts a loosely coupled iterative strategy. The mechanical field dynamic equation outputs load distribution to drive the update of thermal and flow field boundaries. The thermal field temperature field reacts to the material properties and lubricating oil viscosity, and the flow field oil film pressure is fed back to the support stiffness and friction torque. The electromagnetic field loss is superimposed as a heat source in the temperature rise calculation. The convergence tolerance for each iteration is set to a relative change of less than 3% between field variables. Preferably, in the loss balancing decision module, the global objective function is defined as: minimizing ,in Indicates the first Instantaneous power loss of each grid cell, and The adjustable weighting coefficients have values ​​ranging from [0.8, 1.2] to [0.6, 1.0], automatically switching according to the operating condition. The regional cumulative loss smoothness is quantified by calculating the standard deviation of the loss integral in adjacent spatial regions, reflecting the risk of thermal stress gradient. The multi-objective weighted optimization algorithm uses an improved non-dominated sorting genetic algorithm (NSGA-II) with a population size of 100, a maximum number of generations of evolution of 50, a crossover probability of 0.9, a mutation probability of 0.1, and completes an optimization solution every 2 seconds, outputting the control combination closest to the ideal point in the Pareto front. The control instruction set includes discrete-continuous mixed variables such as lubricating oil distribution ratio, cooling intensity level, active lubrication trigger threshold, and motor derating coefficient, which are converted into executable commands by the equipment through rule mapping. Furthermore, in the adaptive execution module, the lubricating oil pump adopts frequency conversion control, with a flow rate adjustment range of 30% to 100% of the rated value, a step size of 5%, and a response delay of less than 200ms; the cooling fan is equipped with multi-level speed regulation gears, dynamically switching according to the hot spot temperature rise rate, and activating the highest gear forced air cooling when the local temperature rise slope exceeds 5°C / s; the active lubrication nozzles are independently controlled according to spatial zones, with the injection timing synchronized with the gear meshing phase, and the injection pulse width determined by the current meshing zone load size, ensuring priority oil supply to heavily loaded gear surfaces; the clutch engagement pressure adjustment is used for load transfer control during multi-speed reducer shifting, avoiding local over-damage caused by instantaneous impact; the motor output characteristic parameter adjustment includes torque slope limitation and field weakening control intervention timing adjustment, reducing additional iron losses caused by high-frequency harmonic currents; all actuators have a fail-safe mode, restoring to the default energy-saving operation state when communication is interrupted or commands exceed limits; Furthermore, in the data feedback update module, the measured indirect indicators are obtained through non-invasive means. An infrared thermal imager is deployed on a specific observation window on the outer wall of the reducer, with a spatial resolution of no less than 320×240 pixels and a frame rate of 25fps, used to extract the average temperature and maximum temperature difference of each functional area. The vibration energy spectrum distribution is obtained by performing a short-time Fourier transform on the acceleration signal, focusing on the degree of energy accumulation in the 1kHz to 8kHz frequency band as a criterion for friction state and oil film integrity. The deviation residual is defined as the normalized Euclidean distance between the predicted and measured indicators. When the mean residual value exceeds a preset threshold of 0.15 within 5 consecutive sampling periods, the online identification process of model parameters is triggered. The online identification uses recursive least squares (RLS) combined with a forgetting factor (λ=0.98) to estimate and update key uncertainty parameters such as lubricating oil viscosity-temperature coefficient, heat dissipation coefficient, and contact thermal resistance in real time, ensuring the long-term effectiveness of the model. On the other hand, a multi-field loss equalization method for reducers based on adaptive feedback of operating conditions is proposed. The specific steps of this method are as follows: Step S110: Real-time acquisition of multi-dimensional operating condition data during reducer operation via a distributed sensor network, including mechanical, thermal, fluid, and electromagnetic parameters, followed by time synchronization and noise suppression processing to generate a unified format operating condition state vector; Step S120: Inputting the operating condition state vector into a multi-field coupled physical model, and combining it with data-driven correction terms to solve for the spatial and temporal loss density distribution of key components such as gears, bearings, shafts, and motors; Step S130: Based on the loss distribution results, constructing a loss equalization method based on "loss..." The problem is a multi-objective optimization problem with the goal of "consumption balance". The optimal control instruction set is obtained by solving the problem, which can simultaneously reduce peak loss, smooth regional cumulative loss and improve the overall system efficiency. In step S140, the control instruction set is sent to the actuator to adjust the lubricating oil supply strategy, cooling intensity, active lubrication logic or motor operating point to realize the dynamic reconstruction of the multi-physics field distribution inside the reducer. In step S150, the actual operating state data after control is collected, indirect observable indicators are extracted and compared with the model prediction results, the deviation residual is calculated, and when the residual exceeds the tolerance range, the online update mechanism of model parameters is started to complete the self-correction and iterative evolution of the multi-field coupling model. Preferably, in step S110, the multi-dimensional operating condition data acquisition process uses a time-triggered communication protocol to ensure data consistency. All sensors mark the sampling time with a UTC timestamp, and the maximum allowable clock deviation is ±10μs. The noise suppression processing adopts an adaptive decomposition method based on empirical wavelet transform (EWT), selects the optimal mode segmentation method for different signal types, and removes random and periodic interference while retaining effective characteristic frequency bands. The operating condition state vector is encapsulated in the form of a fixed-length data frame, containing the original value, unit identifier, and quality label for subsequent module calls. Furthermore, in step S120, the multi-field coupled physical model has completed offline calibration across the entire operating condition range during the initialization phase, covering at least 100 sets of typical load spectra and temperature condition combinations; the data-driven correction term is composed of a deep residual network, with the input being the current operating condition state vector and the model prediction residual from the previous moment, and the output being the correction amount of key parameters of each sub-model, such as meshing stiffness offset, convective heat transfer coefficient gain factor, etc., thereby improving the model's generalization ability under unknown operating conditions; the solution process runs on an embedded high-performance computing unit, and the time consumed by a single solution is controlled within 80ms, meeting the real-time requirements; Furthermore, in step S130, the constraints of the multi-objective optimization problem include the physical limits of the actuator, the minimum lubrication guarantee requirements, and the temperature rise safety boundary. Any control command must not violate the above-mentioned hard restrictions. The selection process of the optimal control command set introduces a fuzzy decision-making mechanism, which transforms operator preferences (such as prioritizing lifespan or prioritizing efficiency) into weight coefficient adjustment rules, thereby enhancing the system's human-machine collaborative adaptability. Compared with the prior art, the present invention has the following beneficial effects: By constructing a coupled model of mechanical, thermal, fluid, and electromagnetic fields and implementing closed-loop feedback updates, a refined and dynamic prediction of the multi-physics field loss distribution inside the reducer was achieved for the first time, with the prediction accuracy improved by no less than 40% compared to the traditional single-field model. A multi-objective optimization strategy centered on "loss balance" is proposed, which breaks through the limitation of simply pursuing the maximization of total efficiency in the past. It effectively suppresses the formation of local hot spots, reduces the maximum temperature difference of key components by more than 25%, and significantly delays the initiation of fatigue cracks. The system adopts an adaptive feedback mechanism based on operating conditions, supports online model correction and dynamic evolution of control strategies. After 500 hours of continuous operation, the system still maintains a prediction deviation of less than 10%, and its long-term stability far exceeds that of an open-loop control system. It has achieved a technological leap from passively bearing losses to actively controlling losses, and the overall energy efficiency of the machine has been improved by 8%-15%, which is particularly significant under heavy-load variable speed conditions. The system is compatible with various types of reducer platforms and is especially suitable for high-dynamic demand scenarios such as wind power, construction machinery and new energy vehicles, and has broad engineering promotion value. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the overall technical architecture of the reducer multi-field loss equalization method and system based on working condition adaptive feedback proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the synergistic effect of multi-field coupling modeling and loss equalization decision-making in this invention. Detailed Implementation

[0008] Please refer to Figure 1 and Figure 2 To further illustrate the technical means and effects of the present invention in order to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0009] Example 1 This embodiment uses a condition-adaptive feedback multi-field loss balancing system applied to large wind turbine gearboxes as an example to describe in detail the complete technical implementation path of the system. Wind power equipment faces complex operating conditions such as sudden wind speed changes, frequent start-ups and shutdowns, and extreme ambient temperature fluctuations, leading to drastic changes in internal mechanical stress, easy rupture of the lubricating oil film, and significant local temperature rise. Traditional static optimization strategies are insufficient to cope with such highly unsteady dynamic processes. Therefore, this embodiment constructs a complete system architecture integrating real-time sensing, multi-physics coupled modeling, loss balancing decision-making, closed-loop execution, and online model evolution.

[0010] The system first collects multi-dimensional operating condition data through a distributed sensor network deployed at key parts of the reducer. A high-precision torque sensor is installed at the high-speed shaft inlet, employing a strain gauge full-bridge circuit design with a range covering -2000 Nm to +6000 Nm and a resolution of 0.1 Nm. The sampling frequency is set to 12 kHz to ensure accurate capture of instantaneous torque impacts caused by wind shear. At the output end, a photoelectric encoder is installed at the low-speed shaft end, using an incremental A / B / Z phase output mode with 4096 PPR lines. After quadrupling the frequency, the angular resolution reaches 0.022°, corresponding to a time resolution of 83.3 μs, ensuring sufficient signal density even under low wind speed conditions with a rotational speed below 15 rpm. The shell vibration monitoring uses a triaxial accelerometer array with a total of 16 measuring points, located in the front and rear bearing housings, the middle reinforcing rib of the housing, and the base connecting flange area. The spatial distribution meets the requirement of 3.8 measuring points per square meter. The sensor model is IEPE piezoelectric accelerometer with a sensitivity of 100 mV / g and a frequency response range of 0.5 Hz to 10 kHz. It is used to capture typical fault characteristic frequencies such as gear meshing frequency and its sideband component, rolling element passing frequency, etc.

[0011] The lubrication system is equipped with three miniature PT100 temperature probes, embedded in the inlet pipeline (20cm from the pump outlet), the oil collection groove at the outlet, and 5mm downstream of the active nozzle in the tooth surface projection area. The probe diameter does not exceed 4mm, the response time is controlled within 45ms, the measurement range is -40°C to 150°C, and the accuracy is ±0.3°C. Pressure detection uses a silicon piezoresistive transmitter, installed at the T-junction of the main oil supply pipeline, with a range of 0 to 1.6 MPa, an overload capacity of 2.5 MPa, an output of a 4-20mA standard signal, and a response time of less than 35ms. The ambient temperature and humidity sensors are placed inside the engine compartment near the gearbox housing but not affected by direct heat radiation. They employ a combination of a capacitive humidity sensing element and a platinum resistance thermometer, with a temperature measurement range of -40°C to 85°C and a relative humidity measurement range of 0% to 100%RH, with accuracies of ±0.5°C and ±2%RH respectively. These sensors participate in the dynamic compensation calculation of external natural convection heat transfer boundary conditions.

[0012] All sensor data is connected to an industrial-grade edge computing node, which is built on an ARM Cortex-A72 quad-core processor with a clock speed of 1.8GHz, equipped with 4GB DDR4 memory and 32GB solid-state storage, and runs a real-time Linux operating system (PREEMPT_RT patch) to ensure deterministic task scheduling. Data acquisition adopts the Time Triggered Communication Protocol (TTCAN), with all devices synchronized to a UTC time base. A GPS module provides nanosecond-level timing signals, and the maximum allowable clock deviation for each channel is strictly controlled within ±8μs. The raw data frame contains the following fields: timestamp (UTC microsecond level), sensor ID, raw value, unit code (predefined enumerated value), and quality label (0 for normal, 1 for abnormal and to be filtered, 2 for failure). The data preprocessing process sequentially performs sliding window mean filtering (window length 100 points) and adaptive denoising based on empirical wavelet transform (EWT). The EWT decomposition process automatically divides the scale intervals based on the signal spectrum characteristics. For vibration signals, it prioritizes retaining the IMF components in the 800Hz to 6kHz frequency band to suppress grid harmonic interference. For temperature and pressure signals, it focuses on low-frequency trend terms and removes high-frequency glitches. Finally, it generates a unified operating condition vector with a dimension of 1×28, including: input torque, output speed, RMS values ​​of 16 vibration channels, 3 temperature readings, 2 pressure values, ambient temperature, and ambient humidity. All values ​​are normalized to the [0,1] interval and a validity flag is added.

[0013] The state vector of this operating condition is transmitted in real time to the multi-field coupled modeling module, which is deployed on an embedded high-performance computing unit (equipped with an NVIDIA Jetson AGX Xavier platform) with 32 TOPS INT8 computing power and supports parallel model solving. The modeling process is divided into two levels: the bottom level is a four-field coupled model based on physical mechanisms, and the top level is a data-driven correction mechanism composed of a deep residual network. The time-varying model of gear meshing stiffness is pre-established based on three-dimensional finite element contact analysis, with meshing fined down to the tooth root fillet region, and the element type being quadratic tetrahedrons, totaling approximately 1.2 million degrees of freedom. The model takes into account the tooth surface bulging modification parameters (modification amount 35μm, modification length accounting for 60%), cumulative tooth pitch error (±18μm), tooth tilt error (±12μm), and the influence of shaft elastic deformation. Within each meshing cycle, 120 discrete points are defined along the meshing line. Contact pressure and deformation are calculated point-by-point using Hertzian contact theory, and then integrated to obtain the instantaneous meshing stiffness curve. The amplitude of this curve exhibits a nonlinear increasing trend with increasing load, and significant fluctuations occur in the alternating single and double tooth region. This stiffness sequence serves as the excitation source input to the multibody dynamics equations. in For the generalized mass matrix, It is the damping matrix (including material internal friction and lubrication shear loss). It is a nonlinear stiffness matrix. For generalized coordinate vectors, This is the external load vector.

[0014] The bearing support stiffness model adopts the Jones-Harris rolling bearing theory framework, treating the inner and outer rings, cage, and rolling elements as independent rigid bodies, and considering the Hertzian contact deformation between the balls and raceways. The model distinguishes between contact angle changes under radial and axial loads, dynamically updating the support point stiffness matrix across six degrees of freedom under different operating conditions. For example, when the axial force increases, the contact angle of the angular contact ball bearing increases from the initial 25° to 32°, leading to an increase in axial stiffness and a decrease in radial stiffness. The model outputs the displacement-force mapping relationship for each support point, serving as the boundary condition input for structural vibration analysis.

[0015] The dynamic evolution model of lubricating oil film thickness is solved based on the two-dimensional transient Reynolds equation, and the Dowson-Higginson viscosity-density relationship is introduced: in, The density is under reference conditions, typically the initial density at normal pressure and room temperature. For reference pressure, atmospheric pressure is generally used. The reference viscosity is the viscosity at normal pressure and reference temperature. For reference temperature, ambient temperature is usually taken. For the density of lubricating oil, For dynamic viscosity, For local oil film pressure, For temperature, The viscosity-pressure index is used. Boundary conditions are determined by the relative speed of the gears (up to 15 m / s) and the surface roughness (Ra = 0.4 μm). The solution domain is divided into a 200×150 grid, and the finite difference method is used for iterative solution to identify the spatial distribution of the elastohydrodynamic (EHL) region, the mixed lubrication region, and the boundary lubrication region. The predicted minimum oil film thickness is no less than 1.8 μm; if it is below this threshold, a risk of metal-to-metal contact is considered.

[0016] The electromagnetic loss distribution model is applicable to the electric drive reducer structure of integrated permanent magnet synchronous motors, and the stator copper loss is calculated according to... The calculations take into account the increase in AC resistance caused by the skin effect. The core loss is decomposed into three terms—eddy current loss, hysteresis loss, and residual stray loss—using the Bertotti separation model, which are proportional to the square, first, and 1.5 powers of the magnetic flux density, respectively. The rotor eddy current loss is obtained through three-dimensional transient magnetic field simulation, with a focus on the circulating heating caused by end leakage flux. All electromagnetic losses are superimposed on the right side of the heat conduction equation as spatially distributed heat sources.

[0017] The four-field coupled solution employs a loosely coupled iterative strategy, executing a complete solution cycle every 20ms. Initially, the current operating condition state vector is loaded, initiating the mechanical field solution to obtain the gear meshing force, bearing reaction force, and shaft deformation distribution. These results are used as boundary conditions for the thermal and flow fields, solving for the temperature field distribution based on Fourier's law of thermal conduction and the convective heat transfer equation. Simultaneously, the temperature field is fed back to the material property database, updating the gear steel's elastic modulus (which decreases linearly with increasing temperature), lubricating oil viscosity (exponential decay), and contact thermal resistance (related to surface roughness and pressure). The flow field oil film pressure distribution reacts to the support stiffness model, forming a lubrication-structure feedback loop. Electromagnetic losses participate in the temperature rise calculation as an additional heat source. Each iteration judges the relative rate of change of inter-field variables (such as maximum contact stress, hot spot temperature, and peak oil film pressure). Convergence is determined when the changes in all variables are less than 2.8% in three consecutive iterations, outputting the spatial loss density distribution cloud map and time-series curve for the current moment. The entire solution process takes an average of 76ms, meeting the 80ms cycle requirement of the control system.

[0018] The loss balancing decision module receives the loss distribution results from the modeling module and constructs a multi-objective optimization problem centered on "loss balancing". The objective function is defined as: in For finite element mesh element index, For the first Instantaneous power loss of each unit; The system is divided into nine regions for the preset functional areas (such as the first-stage gear pair meshing area, the intermediate shaft bearing area, and the final-stage planetary carrier support area). The standard deviation of the cumulative loss in the region reflects the risk of thermal stress gradient. The overall system efficiency is defined as the ratio of output power to input power; weighting coefficients. and Dynamically adjust based on operating conditions: Set to [specific settings] during the startup acceleration phase. =1.1, =0.7, focusing on suppressing peak losses; adjusted to during stable operation. =0.9, =0.9, balancing efficiency and uniformity; set to [value] during emergency braking. =1.2, =0.6, prioritize preventing local overheating.

[0019] The constraints include: the minimum flow rate of the lubricating oil pump shall not be less than 35% of the rated value to ensure minimum lubrication requirements; the maximum power consumption of the cooling fan shall not exceed 15% of the system auxiliary power; any control command shall not cause the predicted maximum temperature to exceed the allowable limit of the material (180°C for gear steel and 120°C for seals); the timing of the motor field weakening control intervention shall not be earlier than 85% of the rated speed to avoid sudden torque drop causing mechanical shock.

[0020] The optimization algorithm employs an improved NSGA-II, with a population size of 100. The encoding method is a mixed integer-real vector, containing five decision variables: lubricating oil pump frequency setpoint (real number, 30%-100%), cooling fan speed (integer, 1-5 levels), number of active lubrication nozzles opened (integer, 1-4 groups), clutch engagement pressure offset (real number, ±20% of baseline value), and motor derating factor (real number, 0.8-1.0). Crossover is performed using simulated binary crossover (SBX) with a distribution exponent of 15; mutation is performed using multinomial mutation with a distribution exponent of 20. An optimization solution is completed every 2 seconds, generating a Pareto front containing approximately 40 non-dominated solutions. The decision mechanism uses a fuzzy comprehensive evaluation method, transforming operator preferences into weight adjustment rules: if "lifetime priority" is selected, then... Increase by 10%, Reduce by 10%; if "efficiency first" is selected, then the opposite applies. The final selection is the point closest to the ideal (minimum). Minimum ,maximum The solution with the shortest Euclidean distance is taken as the optimal control instruction set.

[0021] After receiving the control instruction set, the adaptive execution module maps it to specific execution commands. The lubricating oil pump is driven by a frequency converter, adopting a V / F control strategy with a frequency adjustment step of 0.5Hz (corresponding to a flow rate step of approximately 5%). The actual measured response delay is 185ms, meeting the design requirement of less than 200ms. The cooling fan is equipped with five-level speed regulation, achieved through PWM duty cycle switching. When the infrared thermal imager detects that the temperature rise rate in any area exceeds 4.8°C / s, it immediately switches to the fifth level of forced air cooling, increasing the airflow to 140% of the rated value. The active lubrication nozzles are independently controlled according to spatial zones, with each group corresponding to a gear pair. The injection timing is triggered by an encoder signal, and the rising edge of the pulse width modulation signal is synchronized with the gear entering the meshing area 1.2ms in advance. The oil injection duration is linearly adjusted within the range of 2-8ms according to the current load to ensure sufficient lubrication of the heavily loaded gear surfaces. The clutch engagement pressure is regulated by an electro-hydraulic proportional valve, with a control signal of 0-10V analog quantity, corresponding to a pressure range of 0-3.5MPa. A ramp-type pressure loading is implemented during gear shifting, with a rise time set to 300ms to avoid sudden load changes. Motor output characteristic parameters are transmitted to the driver via the CAN bus, adjusting the torque slope limit (originally 500 Nm / s, can be reduced to 200 Nm / s) and the field weakening control starting speed (originally 4500 rpm, can be delayed to 4800 rpm) to reduce additional iron losses caused by high-frequency harmonic currents.

[0022] All actuators are equipped with safety monitoring logic. Upon detection of communication interruption, command overrun, or hardware failure, a fail-safe mode is immediately triggered: the lubricating oil pump resumes operation at 60% of its rated flow, the cooling fan maintains medium speed (level 3), the active lubrication system shuts down non-critical nozzles, retaining only basic spraying, and the motor switches to constant power output mode. While not optimal, this mode ensures the system continues to operate safely for at least 4 hours under abnormal conditions.

[0023] The data feedback update module continuously collects actual operating status data after adjustment. An infrared thermal imager is installed on the outer wall observation window of the reducer, using an uncooled vanadium oxide detector with a resolution of 384×288 pixels, a spectral response range of 7.5-14μm, and a frame rate of 25fps, acquiring one temperature field image every 40ms. After distortion correction and radiometric calibration, the average temperature and maximum temperature difference of nine functional zones are extracted. The vibration energy spectrum distribution is obtained by performing a short-time Fourier transform (STFT) on the triaxial acceleration signal, with a window length of 512 points, an overlap rate of 75%, and a spectral resolution of 4.88Hz. The focus is on analyzing the energy concentration in the 1kHz to 8kHz frequency band as a criterion for oil film integrity—if the energy in this frequency band accounts for more than 65% of the total energy, it indicates a risk of boundary lubrication.

[0024] The deviation residual is defined as the normalized Euclidean distance between the predicted index and the measured index: in, For the first Predicted values ​​for each variable, For the first The measured values ​​of each variable. For the first 10 observed variables (total) =12, including 9 regional temperatures and 3 vibration frequency band energy). Its standard deviation. Within 5 consecutive sampling periods (total 10 seconds) When the mean exceeds 0.145, the model is deemed inaccurate, and the online identification process is initiated. The identification algorithm employs recursive least squares (RLS) combined with a forgetting factor (λ=0.98). The parameters to be estimated include the lubricating oil viscosity-temperature coefficient (ISO VG320 baseline value corrected by ±15%), the shell convective heat transfer coefficient (baseline value ±30%), and the gear contact thermal resistance (baseline value ±25%). The parameter update cycle is 1 second, and the multi-field coupled model is reinitialized after each update to achieve self-correction and iterative evolution. Long-term operation tests show that after 520 hours of continuous variable load testing, the prediction deviation stabilizes within 9.7%, verifying the effectiveness of the model update mechanism.

[0025] Example 2 This embodiment focuses on the application scenario of a two-speed electric drive reducer in new energy vehicles, highlighting a loss balancing implementation method based on a different algorithm architecture. The core technical difference lies in the use of a surrogate model to replace the solution of some physical equations in the multi-field coupling modeling module, thereby significantly reducing the computational load and adapting to the resource-constrained characteristics of the vehicle controller. Compared to Embodiment 1, this embodiment does not simply change the application scenario, but rather reconstructs the core modeling logic while ensuring functional consistency, resulting in a substantial difference in technical approach.

[0026] In new energy vehicle operating conditions, the reducer needs to frequently respond to the driver's accelerator pedal commands, experiencing various transient processes such as rapid acceleration, regenerative braking, and idle start-stop. The system has extremely high real-time control requirements, and must complete the entire link response from sensing to execution within 20ms. Therefore, the high-fidelity but computationally intensive finite element-Reynolds equation joint solution scheme in Example 1 cannot be directly used. To address this, this example proposes a hybrid modeling paradigm of "physical guidance + data fitting".

[0027] The sensor configuration and data preprocessing process of the real-time operating condition perception module are basically the same as in Implementation Example 1, but the sampling strategy has been adjusted: to reduce the load on the vehicle CAN FD bus, vibration signals are sampled using event triggering. Full-speed acquisition at 10kHz is initiated when the detected acceleration peak exceeds 0.8g, and reduced to 1kHz for the remaining time. The sampling interval for temperature and pressure sensors is extended from 100ms to 200ms, and Kalman filtering is used to predict intermediate values. The operating condition state vector is compressed to a 1×20 dimension, removing some redundant vibration channels while retaining key position information.

[0028] The core innovation of the multi-field coupling modeling module lies in introducing three types of surrogate models to replace the traditional numerical solver. The first type is the gear meshing stiffness surrogate model, whose training data comes from an offline high-precision finite element simulation database. It covers a three-dimensional working condition mesh with input torque of 0-3000 Nm, speed of 0-8000 rpm, and oil temperature of -20°C to 100°C, generating a total of 120,000 samples. A deep neural network (DNN) is used for fitting, with a network structure of 6 fully connected layers (28-dimensional input layer, 512, 256, 128, 64, and 32 hidden layer nodes respectively, and a 120-dimensional output layer corresponding to the meshing stiffness sequence). The activation function is Swish, and the loss function is MAE. After training, the prediction error is less than 3.5%. This model directly receives the working condition state vector as input and outputs a complete time-varying meshing stiffness curve, with a computation time of only 8 ms.

[0029] The second type is the thermal field distribution surrogate model, constructed based on Gaussian process regression (GPR). Training data is obtained from three-dimensional transient heat conduction simulations, considering the temperature field response under different combinations of cooling airflow, oil flow distribution, and electromagnetic loss inputs. The first 20 modes are selected as basis functions, and the kernel function parameters (RBF kernel) are automatically selected using Bayesian optimization to achieve efficient approximation of complex nonlinear heat conduction behavior. The model input consists of historical temperature sequences (10 steps long) from six representative temperature measurement points, the current cooling intensity level, and lubricating oil flow rate. The output is a predicted temperature distribution across the entire enclosure surface, maintaining spatial resolution consistent with the infrared camera. A single inference takes 12 ms.

[0030] The third type is the lubrication state classification model, implemented using a lightweight convolutional neural network (CNN-Lite). The input is the energy spectrum matrix (64×64) of the vibration signal in the 1kHz-8kHz frequency band, and the output is the probability distribution of four lubrication states: complete elastohydrodynamic lubrication, mixed lubrication, boundary lubrication, and dry friction. The network contains only 3 convolutional layers (kernel size 3×3, channel numbers 16 / 32 / 64), 2 pooling layers, and 1 fully connected layer, with a total of less than 50,000 parameters, allowing it to run on MCU-level processors. This model replaces the traditional Reynolds equation solution, achieving rapid discrimination of lubrication states.

[0031] The four-field coupling relationship is realized through a state transition diagram: mechanical field output load distribution → lookup table to obtain bearing stiffness change → input to dynamic model → generate vibration spectrum → input to lubrication state classification model → output lubrication rating → feedback to friction coefficient correction term → affect heat source input → input to thermal field proxy model → output temperature distribution → react on material properties and oil parameters → update to mechanical field and lubrication models. The entire coupling solution cycle is controlled within 18ms, meeting the real-time requirements of vehicle operation.

[0032] The objective function of the loss equalization decision module remains unchanged, but the optimization algorithm is replaced by the Fast Approximate Pareto Search (FAPS) algorithm, designed specifically for embedded platforms, instead of NSGA-II. This algorithm uses a reference-point-guided local search strategy, directly locating feasible solutions near ideal points instead of generating a complete Pareto front. The population size is reduced to 20, a greedy initialization strategy is adopted, and convergence occurs in only 10 iterations. Decision variables are simplified to discrete types: lubricating oil pump speed (levels 1-5), cooling fan speed (levels 1-3), active lubrication enable switch (0 / 1), and motor operating mode (standard / energy-saving / sport). An optimization is performed every 1.5 seconds, with power consumption only 23% of the original algorithm.

[0033] The adaptive execution module has been modified for vehicle environment adaptation. The lubricating oil pump has been changed to a combination of a constant displacement pump and a proportional relief valve, achieving flow control by adjusting the valve opening, improving the response time to 120ms. The cooling fan is integrated into the motor cooling system, sharing the same PWM drive circuit, and gear switching is coordinated by the vehicle controller. The active lubrication nozzle is only activated when running at high speed, controlled by a solenoid valve, with the injection timing linked to the clutch disengagement action, opening 150ms in advance to pre-lubricate the meshing gear surfaces. Motor output characteristic adjustment is achieved by modifying the torque mapping curve in the vehicle VCU, with the derating factor pre-stored in MAP form and dynamically called according to battery SOC and motor temperature.

[0034] The data feedback update module employs an incremental model calibration mechanism. Due to the inability to mount an infrared thermal imager, a distributed NTC temperature sensor array (12 measurement points) was used instead to estimate the surface temperature gradient. The deviation residual calculation is based on the weighted mean square error between the surrogate model's predicted temperature and the measured temperature, with weights allocated according to the importance of the measurement points (bearing housing weight 1.0, middle of the housing 0.6, and areas far from the heat source 0.3). When the residual continuously exceeds the limit, a surrogate model fine-tuning procedure is initiated: the weights of deep networks are frozen, online gradient descent updates are performed only on the last fully connected layer, the learning rate is set to 0.001, and each update uses only the latest 5 sets of data to avoid catastrophic forgetting. Experiments show that this mechanism can reduce the thermal model prediction error from the initial 18% to 6.2% after the vehicle has traveled 1000 kilometers.

[0035] This embodiment, by introducing a proxy model system, successfully reduces the system's computational requirements to a level that the vehicle ECU can handle without sacrificing core functionality. This demonstrates the engineering applicability of the invention on resource-constrained platforms and forms a complementary technical solution with Embodiment 1.

[0036] Example 3 This embodiment addresses the special needs of planetary reducers used in engineering machinery under extreme heavy loads and dust pollution environments. It proposes a loss balancing architecture based on multimodal sensor fusion and reinforcement learning decision-making. The essential difference lies in that the decision-making module abandons traditional optimization algorithms and instead adopts a deep Q-network (DQN) to achieve end-to-end policy learning, solving the problems of difficulty in manually setting multi-objective weights and complex and variable constraints.

[0037] Engineering machinery speed reducers often operate in harsh environments such as mines and tunnels, with load spectra exhibiting strong randomness and suddenness, and long maintenance cycles, requiring control systems with a high degree of autonomous decision-making capability. Traditional model-based optimization methods rely on precise parameter calibration, and their performance deteriorates sharply under the influence of uncertainties such as oil deterioration and accelerated wear. Therefore, this embodiment constructs an intelligent decision-making framework based on experience accumulation.

[0038] The sensing module incorporates two new types of sensors: first, a fiber Bragg grating (FBG) strain sensor embedded in the gear root to directly measure local stress changes, with a sampling frequency of 5kHz and an accuracy of ±5με; second, an oil particle counter installed in the return oil line to monitor the ISO 4406 contamination level in real time, outputting data once per minute. These two types of signals are fused with existing data to form an expanded operating condition state vector (1×32-dimensional), which better reflects the internal damage evolution trend.

[0039] The multi-field coupled modeling module still uses the high-fidelity physical model from Example 1, but its role has changed: it no longer serves as a real-time decision-making basis, but rather as a digital twin environment for training reinforcement learning agents. Before leaving the factory, the system uses historical operating data and virtual simulation to generate millions of operating condition segments, covering dozens of degradation modes such as normal wear, poor lubrication, and localized spalling. It runs a complete four-field coupled solution on a high-performance server, generating a "state-action-reward" triplet database. The state is the operating condition vector, the action is the combination of control commands, and the reward... Defined as: in, For time steps The reward value, To allow for extreme temperatures, A preset upper limit for safe temperature difference is used as a reference temperature difference for normalization. The rated efficiency is the design efficiency under ideal operating conditions. The maximum unit temperature, The temperature difference between adjacent areas For efficiency, For execution costs (such as energy consumption, mechanical wear increments). to This represents the task weight. The reward function inherently encodes the design philosophy of "loss leveling".

[0040] The Deep Q-Network employs a dual-stream convolutional structure: one branch processes numerical operating condition vectors (normalized and input into 4 fully connected layers), while the other branch processes vibration spectrum maps (input into 3 convolutional layers). Finally, the features are concatenated and fed into the value evaluation layer. The action space is discretized into 36 combinations (6 speeds for the lubricating oil pump × 3 speeds for the cooling fan × 2 types of lubrication logic × 3 types of motor modes). The training process utilizes experience replay and target network freezing techniques to ensure stable convergence. After training, the network parameters are embedded into the onboard AI chip (Horizon Robotics Journey 5), achieving an inference latency of less than 5ms.

[0041] When running online, the system no longer solves an optimization problem; instead, DQN directly outputs the optimal action based on the current state. The system collects the state input to the network every 5 seconds, obtains the action instruction with the highest Q-value, and executes it. Simultaneously, it records newly generated (s, a, r, s') sequences and periodically uploads them to the cloud for model retraining, achieving knowledge sharing within the group.

[0042] The data feedback update module now includes a health status assessment function. Based on the peak-to-peak value and skewness indices of the FBG stress signal, combined with the particle count growth rate, a fuzzy evaluation model for wear degree is established, outputting a health index HI∈[0,1]. When HI<0.6, a model retraining request is automatically triggered, and preventative maintenance suggestions are pushed to maintenance personnel. This mechanism enables the system not only to regulate wear but also to predict the end of its lifespan, achieving a leap from "passive balancing" to "proactive health management."

[0043] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for multi-field loss equalization of a reducer based on adaptive feedback under operating conditions, characterized in that, include: Multi-dimensional operating condition data during the operation of the reducer are collected in real time through a distributed sensor network. The multi-dimensional operating condition data includes mechanical parameters, thermal parameters, fluid parameters, and electromagnetic parameters. The multidimensional operating condition data is subjected to time synchronization and noise suppression processing to generate a standardized operating condition state vector. The operating condition vector is input into the multi-field coupled physical model, and combined with the data-driven correction term, the multi-field loss density distribution of the key components of the reducer in the spatial and temporal dimensions is obtained. Based on the multi-field loss density distribution, a multi-objective optimization problem with loss balance as the core orientation is constructed, and the optimal control instruction set is obtained by solving the problem. The objective function of the multi-objective optimization problem comprehensively considers the suppression of maximum single-point loss, the smoothness of regional cumulative loss, and the improvement of the overall system efficiency. The optimal control instruction set is sent to the actuator to adjust the lubricating oil supply strategy, cooling intensity, active lubrication logic or motor operating point, so as to realize the dynamic reconstruction of the multi-physics field distribution inside the reducer. The actual operating status data after the control is collected, indirect observable indicators are extracted and compared with the multi-field loss density distribution results predicted by the model, and the deviation residual is calculated. When the deviation residual exceeds the preset tolerance range, the online update mechanism of the model parameters is activated to correct the multi-field coupled physical model, and the model self-correction and iterative evolution are completed.

2. The method for multi-field loss equalization of a reducer based on adaptive feedback under operating conditions as described in claim 1, characterized in that, The multi-dimensional operating condition data includes input torque, output speed, housing vibration signal, lubricating oil temperature, lubricating oil pressure, ambient temperature and humidity, and drive motor current and voltage.

3. The method for multi-field loss equalization of a reducer based on adaptive feedback according to claim 1, characterized in that, The multi-dimensional operating condition data undergoes time synchronization and noise suppression processing to generate a standardized operating condition state vector, including: A time-triggered communication protocol is used to timestamp-align all sensor data, with the maximum allowed clock deviation being a preset threshold. An adaptive decomposition method based on empirical wavelet transform is used to suppress noise in different types of signals, preserving effective characteristic frequency bands while removing random and periodic interference. The processed data is encapsulated into a fixed-length data frame containing the original numerical value, unit identifier, and quality label, which serves as the operating condition state vector.

4. The method for multi-field loss equalization of a reducer based on adaptive feedback under operating conditions according to claim 1, characterized in that, The multi-field coupled physical model includes a time-varying model of gear meshing stiffness, a bearing support stiffness model, a dynamic evolution model of lubricating oil film thickness, and an electromagnetic loss distribution model. The multi-field coupled physical model achieves loosely coupled iterative solution of mechanical field, thermal field, fluid field, and electromagnetic field through the transfer of variables at the inter-field interface.

5. The method for multi-field loss equalization of a reducer based on adaptive feedback according to claim 4, characterized in that, The data-driven correction term is composed of a deep neural network. Its input is the current operating condition state vector and the model prediction residual at the previous moment. Its output is the correction amount of the key parameters of each sub-model in the multi-field coupled physical model.

6. The method for multi-field loss equalization of a reducer based on adaptive feedback according to claim 1, characterized in that, The constraints of the multi-objective optimization problem include the physical limits of the actuator, the minimum lubrication guarantee requirements, and the temperature rise safety boundary.

7. The method for multi-field loss equalization of a reducer based on adaptive feedback according to claim 1, characterized in that, The indirect observable indicators include the temperature difference between different sections of the outer wall of the reducer obtained by an infrared thermal imager, and the vibration energy spectrum distribution of a specific frequency band obtained by performing a short-time Fourier transform on the vibration signal of the casing.

8. The method for multi-field loss equalization of a reducer based on adaptive feedback according to claim 1, characterized in that, The online parameter update mechanism of the model adopts the recursive least squares method combined with the forgetting factor to estimate and update the lubricating oil viscosity-temperature coefficient, heat dissipation coefficient and contact thermal resistance parameters in real time.

9. A multi-field loss equalization system for a reducer based on adaptive feedback under operating conditions, characterized in that, include: The real-time operating condition perception module is used to collect multi-dimensional operating condition data during the operation of the reducer, and to perform time synchronization and preprocessing on the multi-dimensional operating condition data to generate a standardized operating condition state vector. The multi-field coupling modeling module is used to receive the working condition state vector, construct a multi-field coupling physical model based on a preset physical mechanism model and a data-driven correction mechanism, and solve the multi-field loss density distribution of the key components of the reducer. The loss equalization decision module is used to receive the multi-field loss density distribution, combine it with a preset global objective function oriented towards loss equalization, and use a multi-objective weighted optimization algorithm to generate the optimal control instruction set. An adaptive execution module is used to receive the optimal control instruction set and adjust at least one of the lubricating oil pump, cooling fan, active lubrication nozzle, clutch or motor to achieve closed-loop intervention on the multi-physical field distribution inside the reducer. The data feedback update module is used to collect actual operating status data after regulation, compare the predicted multi-field loss density distribution with the measured indirect observable indicators, calculate the deviation residual and use it to correct the parameters of the multi-field coupled physical model online, forming a complete feedback link from perception to execution to model evolution.

10. The multi-field loss equalization system for a reducer based on adaptive feedback according to claim 9, characterized in that, The real-time operating condition sensing module includes a high-precision torque sensor, an optical encoder, a triaxial accelerometer array, a distributed temperature probe, a piezoresistive pressure transmitter, and a meteorological environmental sensor; the data feedback and update module includes an infrared thermal imager deployed on the observation window of the reducer's outer wall and a spectrum analysis unit for processing vibration signals.

Images