Speed reducer self-adaptive torque control method and system based on load identification

Abstract

The invention discloses a speed reducer self-adaptive torque control method and system based on load identification, and relates to the technical field of speed reducer control. The method comprises the steps of collecting multi-source time sequence sensing data in real time when the speed reducer operates; performing time domain and frequency domain feature extraction on the data, and identifying current load features including load levels and types; predicting a future load change trend based on the load characteristics and historical data; a self-adaptive torque control instruction is generated according to the load characteristics, the change trend and the health state parameters, and the output torque is dynamically adjusted; and recording instruction execution efficiency data, and optimizing a load identification and torque control strategy by adopting a strategy gradient reinforcement learning algorithm. The system realizes the functions through cooperation of multiple modules. The problem that torque control of a traditional speed reducer cannot adapt to dynamic load changes is solved, precise self-adaptive adjustment of torque output is achieved, the operation efficiency and equipment safety are improved, and the service life is prolonged.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of speed reducer control, in particular to a speed reducer adaptive torque control method and system based on load identification. BACKGROUND

[0002] Speed reducer is a key power transmission component in industrial production, intelligent manufacturing, transportation and other fields. It realizes speed reduction and torque amplification through gear meshing, worm transmission or planetary gear train structure, and converts the high-speed small torque output of the prime mover into the low-speed large torque power required by the load, which is the core component to ensure the stable operation of mechanical equipment. In actual application, the load driven by the speed reducer is not constant in most scenarios. For example, the amount of material in the material sorting process of the logistics conveyor will suddenly increase or decrease, the load will change instantaneously when the robot grabs workpieces of different weights, and the load of the engineering machinery will fluctuate with the work intensity. These dynamic loads require real-time adaptation of the torque output of the speed reducer.

[0003] The existing torque control of the speed reducer mostly adopts a fixed control mode based on preset parameters, which cannot sense the dynamic changes of the load in real time. When the load suddenly changes or continuously fluctuates, the traditional control method is difficult to adjust the output torque in time, which may cause the equipment to stop due to insufficient torque, or cause gear wear, motor overheating and even structural damage due to excessive torque. Not only does this reduce the operating efficiency and energy economy of the transmission system, but it also shortens the service life of the speed reducer and associated components due to frequent overload shocks, and even causes production safety accidents. Therefore, it is necessary to improve the existing torque control technology of the speed reducer. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a speed reducer adaptive torque control method and system based on load identification, which aims to solve the problem that the existing speed reducer torque control method cannot respond to dynamic changes in load in real time, resulting in low operating efficiency, accelerated equipment wear and potential safety risks.

[0005] The technical solution of the present application is as follows: In a first aspect, the present application discloses a speed reducer adaptive torque control method based on load identification, comprising the following steps: Real-time acquisition of multi-source time sequence sensing data during speed reducer operation, wherein the multi-source time sequence sensing data includes motor input current, speed reducer output shaft vibration signal, output shaft speed and speed reducer housing temperature; Extracting and processing the time domain features and frequency domain features of the multi-source time sequence sensing data to identify the load characteristics of the current load, wherein the load characteristics include load level and load type; Based on the load characteristics and combined with historical load data, predict the load change trend in a future preset time window; Based on the load characteristics and load change trends, and combined with preset reducer health status parameters, an adaptive torque control command is generated to dynamically adjust the motor output torque and correspondingly adjust the transmission status of the reduction transmission module. Record the execution performance data of the adaptive torque control command, and based on the historical execution performance data, use the policy gradient reinforcement learning algorithm to iteratively optimize the load identification logic and torque control strategy.

[0006] This invention, through multi-source sensor data fusion analysis, achieves accurate identification of the reducer's load state and prediction of future trends, and generates forward-looking adaptive torque control commands based on this, effectively solving the problems of lag and rigidity in fixed control modes. Simultaneously, the introduction of health status assessment and reinforcement learning optimization mechanisms ensures that the control strategy improves response speed and energy efficiency while also considering equipment lifespan and operational safety, significantly enhancing the reducer's adaptability, reliability, and economy under dynamic operating conditions.

[0007] Furthermore, multi-source time-series sensor data is collected in real time during the operation of the reducer, including: The motor input current is collected by a current sensor, the vibration signal of the reducer output shaft is collected by a vibration sensor, the output shaft speed is collected by a speed sensor, and the reducer housing temperature is collected by a temperature sensor. Each sensor is electrically connected to the synchronous acquisition unit of the reducer control system, so that the acquired sensor data has a consistent timestamp. The acquired raw sensor data is sequentially filtered, denoised, and normalized to eliminate environmental electromagnetic interference signals and instantaneous pulse abnormal signals, thereby obtaining preprocessed multi-source time-series sensor data.

[0008] Through this technical solution, the present invention ensures the consistency of multi-source heterogeneous sensor data in time and the reliability of its quality by means of hardware synchronization and signal preprocessing, laying a solid data foundation for subsequent accurate feature extraction and state recognition, and effectively avoiding misjudgment caused by data asynchrony or noise interference.

[0009] Furthermore, time-domain and frequency-domain features are extracted from the multi-source time-series sensing data to identify the load characteristics of the current load, including: Perform a fast Fourier transform on the output shaft vibration signal to extract its main frequency amplitude features and harmonic distribution features, and construct a vibration feature vector; A fluctuation analysis of the motor input current is performed, and its short-time standard deviation and ripple coefficient are calculated to obtain the current fluctuation characteristics. The vibration feature vector, the current fluctuation feature, the real-time output shaft speed and the real-time housing temperature are fused to generate load feature input data. The load feature input data is input into a pre-trained load classification neural network model, which outputs the load type and the load level.

[0010] Through this technical solution, the present invention comprehensively characterizes the load state by combining time domain and frequency domain analysis and multi-feature fusion, and utilizes the powerful nonlinear classification capability of neural network models to achieve high-precision and automated identification of complex load types (such as constant, impact, periodic fluctuation, etc.) and fine load levels, overcoming the shortcomings of poor adaptability of traditional threshold judgment methods.

[0011] Furthermore, based on the load characteristics and combined with historical load data, predicting the load change trend within a preset time window includes: Historical load data are constructed into a load history sequence in chronological order, wherein each data point in the load history sequence includes load level, load type and corresponding collection timestamp; A time series decomposition algorithm is used to decompose the historical load sequence and extract periodic components, trend components, and residual components. The currently identified load features are concatenated with the periodic and trend components to generate trend prediction input data. The trend prediction input data is input into a long short-term memory network model for prediction, and the load change trend within the future preset time window is output.

[0012] Through this technical solution, the present invention decomposes the inherent patterns (cycles, trends) of historical load sequences and combines them with the current state, using the superior time series modeling capabilities of the LSTM model for prediction, thereby achieving advanced perception of future load changes. This enables torque control to shift from reactive response to proactive intervention, significantly improving the system's foresight and stability.

[0013] Furthermore, the preset speed reducer health status parameters are determined in the following way: Establish a temperature-torque-speed mapping model for the reducer under no-load conditions and at least three typical load conditions, as well as a baseline model for the output shaft vibration spectrum. The real-time collected multi-source time-series sensor data, including real-time housing temperature, motor input current, output shaft vibration signal and speed, are compared with the temperature-torque-speed mapping relationship model and the vibration spectrum baseline model, respectively, and the weighted Euclidean distance of each feature deviation is calculated. The weighting coefficient is preset according to the degree of influence of each feature on the health status of the reducer. The health status parameters of the reducer are determined based on the numerical results of the weighted Euclidean distance.

[0014] Through this technical solution, the present invention quantifies the health status by establishing a multi-condition benchmark model and calculating the comprehensive deviation between real-time data and the benchmark. This enables control decisions to fully consider the wear and performance degradation of the equipment itself, achieve dynamic matching between torque output and equipment load-bearing capacity, effectively prevent secondary damage caused by over-control or under-control when the health status is poor, and extend the equipment life.

[0015] Furthermore, based on the load characteristics and the load change trend, and combined with preset reducer health status parameters, an adaptive torque control command is generated, including: If the load type is identified as an impact load and the predicted trend indicates that the load will continue to increase, then an advance torque increase command is generated based on the predicted load change gradient and health status parameters. The command includes torque increase magnitude and increase rate parameters. If the load type is identified as a constant load and the predicted trend indicates that the load will decrease, an energy-saving torque reduction command is generated. The command is used to smoothly reduce the output torque to the optimal efficiency range that matches the current load level. If the health status parameter is lower than the first preset threshold, the upper limit of torque output is dynamically reduced; if the health status parameter is lower than the second preset threshold, a maintenance alarm signal is generated and a degraded operation suggestion is output. The first preset threshold is higher than the second preset threshold, and both are set based on the health benchmark value of the reducer under rated operating conditions.

[0016] Through this technical solution, the present invention generates differentiated and precise control commands based on multi-dimensional information such as load characteristics, changing trends and health status, thereby achieving early suppression of impact loads, energy-saving optimization of stable loads, and preventive protection against unhealthy conditions, comprehensively improving the intelligence level and robustness of the control system in the face of different scenarios.

[0017] Furthermore, based on historical performance data, a policy gradient reinforcement learning algorithm is used to iteratively optimize the load identification logic and torque control strategy, including: The execution performance data includes energy consumption change per unit time, output shaft vibration smoothness index, torque matching error, and housing temperature stability range. Construct a multi-objective optimization function with the goals of improving energy efficiency, ensuring stable operation, and extending equipment life; Based on the multi-objective optimization function, the weight parameters of the load classification neural network model and the decision threshold in the torque control command generation logic are adjusted online iteratively to update the load identification logic and torque control strategy.

[0018] Through this technical solution, the present invention introduces a reinforcement learning mechanism, enabling the system to continuously optimize itself based on actual operating performance (energy consumption, vibration, temperature, matching accuracy), and constantly adjust the parameters of the recognition model and the decision points of the control strategy, thereby adapting to long-term changes such as equipment aging, environmental changes, or load characteristic drift, so that the system always maintains optimal performance and has the ability to continuously evolve.

[0019] Secondly, the present invention also discloses a load-identification-based adaptive torque control system for a speed reducer used to implement the above method, comprising: The data acquisition module is used to collect multi-source time-series sensor data in real time during the operation of the speed reducer; The load identification and trend prediction module is used to extract time and frequency domain features from multi-source time-series sensor data to identify the load characteristics of the current load, and combine historical load data to predict the load change trend within a preset time window in the future. The health status assessment module is used to establish a temperature-torque-speed mapping model and an output shaft vibration spectrum baseline model for the reducer, and to obtain the health status parameters of the reducer by comparing real-time multi-source data with model data. The intelligent control module is used to generate adaptive torque control commands based on load characteristics, load change trends and health status parameters, so as to dynamically adjust the motor output torque and synchronously adapt to the transmission status of the reducer transmission module. The strategy optimization module is used to record the execution performance data of adaptive torque control commands and optimize the load identification logic and torque control strategy based on the strategy gradient reinforcement learning algorithm.

[0020] Through this technical solution, the present invention, through modular system design, materializes each step of the method into functionally defined software and hardware modules. These modules work together to achieve accurate perception, trend prediction, health assessment, intelligent control, and continuous optimization of the reducer load, providing reliable system support and implementation platform for the adaptive torque control method.

[0021] Furthermore, the data acquisition module includes: A current sensor is used to collect the input current of the motor; Vibration sensor, used to collect vibration signals from the output shaft of the reducer; A speed sensor is used to collect the output shaft speed of the reducer; Temperature sensor used to collect the temperature of the gearbox housing; Each sensor is equipped with a signal preprocessing unit, which is used to perform preliminary filtering on the raw sensing data.

[0022] Through this technical solution, the present invention integrates multiple dedicated sensors and is equipped with a local preprocessing unit to form a reliable data sensing layer, ensuring the accuracy, synchronization and initial cleanliness of the raw data acquisition, and providing high-quality data input for the effective operation of the upper-layer algorithm.

[0023] Beneficial effects This invention discloses a load-identification-based adaptive torque control method and system for speed reducers. By real-time acquisition of multi-source time-series data such as motor current, shaft vibration, speed, and temperature, and in-depth extraction of time and frequency domain features, the system can accurately identify the current load level and type. Furthermore, by combining historical load data and utilizing time-series decomposition and prediction models, it achieves advanced prediction of future load change trends. Based on this, it integrates current load characteristics, predicted trends, and speed reducer health status parameters obtained from benchmark model comparison to generate forward-looking and differentiated adaptive torque control commands. This dynamically adjusts the motor output torque and transmission state to accurately match real-time load requirements. Simultaneously, the system continuously records the execution performance data of the control commands and uses a policy gradient reinforcement learning algorithm to iteratively optimize the load identification model and control strategy, enabling continuous evolution.

[0024] The core advantage of this method and system lies in its breakthrough of the limitations of traditional fixed-parameter control modes, achieving a shift from passive response to active adaptation, and from single torque adjustment to multi-objective (efficiency, stability, and lifespan) collaborative optimization. It can effectively handle complex loads such as impacts and fluctuations, preventing overload and underload, and improving operating efficiency and equipment safety. Furthermore, by incorporating health status assessment and reinforcement learning optimization mechanisms, it deeply couples the control strategy with the actual performance state of the equipment, achieving preventative protection and long-term performance optimization, significantly extending the service life of the reducer, and providing an effective technical means to solve the problem of intelligent control of reducers under dynamic load conditions. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic flowchart of a speed reducer adaptive torque control method based on load identification provided in an embodiment of this application; Figure 2 This is a structural block diagram of the speed reducer adaptive torque control system based on load identification provided in the embodiments of this application; The meanings of the labels in the diagram are as follows: 100. Data Acquisition Module; 200. Load Identification and Trend Prediction Module; 300. Health Status Assessment Module; 400. Intelligent Control Module; 500. Strategy Optimization Module. Detailed Implementation

[0027] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0028] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0029] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0030] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0031] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0032] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0033] In related technologies, existing speed reducer torque control mostly adopts a fixed mode, which is difficult to adapt to dynamic load changes, resulting in slow response, low energy efficiency and shortened equipment life.

[0034] To address the aforementioned issues, this application provides a method and system for adaptive torque control of a speed reducer based on load identification. This method acquires multi-source sensor data in real time, identifies load characteristics and predicts changing trends, and generates adaptive control commands by combining these with health status parameters. This enables dynamic and precise adjustment of the output torque, while continuously optimizing the overall system performance.

[0035] The adaptive torque control method for speed reducers based on load identification provided in this application can be applied to speed reducer control systems. In this case, the speed reducer control system is the executing entity of the adaptive torque control method for speed reducers based on load identification provided in this application. This application does not impose any restrictions on the specific type of speed reducer control system.

[0036] For example, a speed reducer control system may include a data acquisition module, a load identification and trend prediction module, a health status assessment module, an intelligent control module, and a strategy optimization module. The data acquisition module includes multiple sensors (such as current sensors, vibration sensors, speed sensors, and temperature sensors) and a signal preprocessing unit; the load identification and trend prediction module includes a feature extraction unit and a prediction model unit; the health status assessment module includes a model library and a feature comparison unit; the intelligent control module includes an instruction generation unit and an execution unit; and the strategy optimization module includes a data recording unit and an optimization algorithm unit.

[0037] To better understand the adaptive torque control method for speed reducers based on load identification provided in the embodiments of this application, the specific implementation process of the adaptive torque control method for speed reducers based on load identification provided in the embodiments of this application will be described by way of example below.

[0038] Figure 1A schematic flowchart of a speed reducer adaptive torque control method based on load identification is shown in an embodiment of this application. The speed reducer adaptive torque control method based on load identification includes: S100 collects multi-source timing sensor data in real time during the operation of the reducer. The multi-source timing sensor data includes at least the motor input current, the reducer output shaft vibration signal, the output shaft speed, and the reducer housing temperature.

[0039] It can be understood that "multi-source time-series sensor data" refers to time-stamped sequential data synchronously collected from multiple sensors installed on the reducer and its drive motor, collectively reflecting the reducer's operating status. Real-time acquisition of multi-source time-series sensor data means synchronously collecting various physical quantity data of the reducer during operation through multiple sensors, ensuring consistent data timestamps.

[0040] For example, the data acquisition module of the reducer control system includes current sensors, vibration sensors, speed sensors, and temperature sensors. Current sensors (such as Hall effect current sensors) acquire the motor input current; vibration sensors (such as piezoelectric accelerometers) acquire the output shaft vibration signal; speed sensors (such as photoelectric encoders) acquire the output shaft speed; and temperature sensors (such as PT100 resistance thermometers) acquire the housing temperature. Each sensor ensures data synchronization through a synchronous acquisition unit (such as an FPGA-based synchronous acquisition card), and the sampling frequency can be dynamically adjusted according to operating conditions (e.g., the vibration signal sampling rate is set to 10kHz, and the temperature sampling rate is set to 1Hz). The acquired raw data undergoes filtering (such as low-pass filtering to remove high-frequency noise), noise reduction (such as wavelet thresholding), and normalization processing by a signal preprocessing unit to eliminate environmental interference and abnormal pulses, resulting in clean multi-source time-series sensor data.

[0041] S200 performs time-domain and frequency-domain feature extraction on multi-source time-series sensor data to identify the load characteristics of the current load. These load characteristics include load level and load type.

[0042] It can be understood that "load characteristics" refers to the feature information obtained by extracting features and recognizing patterns from sensor data, which can quantitatively describe the load size (level) and load nature (type).

[0043] For example, please refer to Figure 2The system performs a Fast Fourier Transform (FFT) on the output shaft vibration signal to extract the main frequency amplitude features (such as fundamental frequency amplitude and second harmonic amplitude) and harmonic distribution features (such as harmonic energy ratio) to construct a vibration feature vector. It then performs a fluctuation analysis on the motor input current, calculating the short-time standard deviation (such as the standard deviation per 1-second window) and ripple coefficient (the ratio of AC component to DC component) to obtain current fluctuation features. Finally, it fuses the vibration feature vector, current fluctuation features, real-time output shaft speed, and real-time casing temperature using multi-feature fusion (such as feature splicing or weighted fusion) to generate load feature input. This load feature is then input into a pre-trained load classification neural network model (such as a Convolutional Neural Network (CNN) or a Support Vector Machine (SVM)) to infer the output load type (such as impact load, constant load, periodic load, etc.) and load level (such as light load, medium load, heavy load).

[0044] S300 predicts load change trends within a preset time window based on load characteristics and historical load data.

[0045] It can be understood that "load change trend" refers to the possible direction and magnitude of load changes in the future, which are predicted based on historical patterns and the current state.

[0046] For example, please refer to Figure 2 The system constructs a historical load sequence by chronologically arranging historical load data. Each data point includes the load level, load type, and corresponding collection timestamp. A time series decomposition algorithm (such as STL seasonal decomposition) is used to decompose the historical load sequence, extracting periodic components (such as daily / weekly cycles), trend components (such as long-term upward or downward trends), and residual components (random fluctuations). The current load characteristics are concatenated with the periodic and trend component characteristics to generate trend prediction input data. The trend prediction input data is then input into a Long Short-Term Memory (LSTM) network model for prediction and inference, outputting the load change trend (such as the load level change curve and load type conversion probability) within a future preset time window (such as the next 5 minutes).

[0047] The S400 generates adaptive torque control commands based on load characteristics and load change trends, combined with preset reducer health status parameters, to dynamically adjust the motor output torque and synchronously adapt to the transmission status of the reducer's speed reduction transmission module.

[0048] As can be understood, "health status parameters" are quantitative indicators that reflect the degree of deviation of the reducer's current performance from its healthy baseline state, and are used to assess its load-bearing capacity. "Adaptive torque control commands" are a set of commands generated based on real-time status to dynamically adjust the motor's output torque, and may include parameters such as target torque value and rate of change.

[0049] For example, the preset health status parameters of the reducer are determined in the following way: A temperature-torque-speed mapping model (e.g., a polynomial model fitted with experimental data) is established for the reducer under no-load conditions and at least three typical load conditions (e.g., light load, medium load, heavy load), as well as an output shaft vibration spectrum baseline model (e.g., a vibration spectrum template under normal operating conditions). Real-time multi-source time-series sensor data (including real-time housing temperature, motor input current, output shaft vibration signal, and speed) are compared with the mapping model and the vibration spectrum baseline model, respectively, and the weighted Euclidean distance of each feature deviation is calculated (the weights are preset based on feature importance, such as a vibration feature weight of 0.4 and a temperature feature weight of 0.3). Based on the quantification results of the weighted Euclidean distance, the health status parameters of the reducer (e.g., health score, range 0-100) are determined.

[0050] Based on load characteristics and load change trends, combined with health status parameters, adaptive torque control commands are generated: If the load type is an impact load and the predicted trend is a continuous increase in load, a proactive torque increase command is generated based on the predicted load change gradient and health status parameters (e.g., torque increase of 20%, increase rate of 50 N·m / s); if the load type is a constant load and the predicted trend is a decrease in load, an energy-saving torque reduction command is generated (e.g., torque is smoothly reduced to the optimal efficiency range corresponding to the current load level); if the health status parameters are below the first threshold (e.g., health level < 80), the upper limit of torque output is dynamically attenuated (e.g., the maximum torque is limited to 90% of the rated value); if the health status parameters are below the second threshold (e.g., health level < 60), a maintenance alarm signal is generated and a degraded operation suggestion is output (e.g., maintenance is recommended and the torque is limited to 70% of the rated value).

[0051] The S500 records the execution performance data of adaptive torque control commands and optimizes the load identification logic and torque control strategy based on historical execution performance data. The "execution performance data" refers to multi-dimensional metrics used to evaluate the actual effectiveness of control commands.

[0052] For example, the performance data includes the change in energy consumption per unit time (e.g., kW·h / min), the smoothness index of output shaft vibration (e.g., the root mean square value of vibration acceleration), the torque matching error (the difference between the actual torque and the target torque), and the stable temperature range of the casing (e.g., the temperature fluctuation range). A policy gradient reinforcement learning algorithm (e.g., the PPO algorithm) is used to construct a multi-objective optimization function (e.g., a weighted summation form) with the objectives of "lowest energy consumption, minimum vibration, minimum matching error, and stable temperature". Based on the multi-objective optimization function, the weight parameters of the load classification neural network model and the decision threshold of the torque control command generation logic are iteratively optimized online to update the load identification logic and the torque control strategy.

[0053] In existing technologies, torque control of speed reducers mostly adopts a fixed mode, which cannot adapt to dynamic changes in load, resulting in lag response, low energy efficiency and equipment wear.

[0054] To address this, this application utilizes multi-source sensor data to identify load characteristics in real time, combines historical data to predict load change trends, and generates adaptive control commands based on health status parameters, thereby achieving dynamic and precise adjustment of torque output. Simultaneously, by recording execution performance data and continuously optimizing the control strategy, the system's response speed, energy efficiency, and equipment lifespan are improved.

[0055] In one possible implementation, S200 performs time-domain and frequency-domain feature extraction on multi-source time-series sensing data to identify the load characteristics of the current load, including: S210 performs a fast Fourier transform on the output shaft vibration signal to extract the main frequency amplitude features and harmonic distribution features, and constructs a vibration feature vector.

[0056] For example, the vibration signal is subjected to FFT transformation to obtain a spectrum diagram, and the amplitude of the fundamental frequency (such as the meshing frequency), the amplitude of the second harmonic, and the distribution of harmonic energy (such as the proportion of the energy of the 3rd to 5th harmonics) are extracted to construct the vibration feature vector [V1, V2, V3].

[0057] S220 performs fluctuation analysis on the motor input current, calculates the short-time standard deviation and ripple coefficient, and obtains the current fluctuation characteristics.

[0058] For example, the current signal is divided into frames (1 second per frame), the standard deviation of the data in each frame is calculated to obtain a short-time standard deviation sequence; the ratio of the effective value of the AC component to the DC component of the current is calculated to obtain the ripple coefficient; and the current fluctuation characteristics [C1, C2] are constructed.

[0059] S230 integrates vibration feature vectors, current fluctuation features, real-time output shaft speed, and real-time housing temperature to generate load feature input.

[0060] For example, the vibration feature vector [V1,V2,V3], current fluctuation feature [C1,C2], rotational speed R and temperature T are concatenated into a feature vector [V1,V2,V3,C1,C2,R,T] as the load feature input.

[0061] S240 inputs the load features into a pre-trained load classification neural network model and infers and outputs the load type and load level.

[0062] For example, the load classification neural network model is a CNN model. The input layer receives a 7-dimensional feature vector, which passes through 2 convolutional layers and 1 fully connected layer. The output layer uses the Softmax activation function to output the probability distribution of load type (e.g., 3 classes) and load level (e.g., 3 levels). The maximum probability value is taken as the recognition result.

[0063] Through the steps S210 to S240 above, multi-feature fusion fully utilizes the complementarity of vibration, current, speed and temperature information, improving the accuracy and robustness of load identification.

[0064] In one possible implementation, S300 predicts the load change trend within a preset time window based on load characteristics and historical load data, including: S310 constructs a load history sequence by chronologically arranging historical load data. Each data point in the load history sequence includes the load level, load type, and corresponding collection timestamp.

[0065] For example, historical load data includes load levels and load types recorded every 5 minutes over the past 7 days, and is constructed as a sequence [(L1,T1,TS1),(L2,T2,TS2),...] by time.

[0066] S320 uses a time series decomposition algorithm to decompose the historical load series and extract periodic components, trend components, and residual components.

[0067] For example, the STL decomposition algorithm is used to decompose the load history sequence into periodic components (such as daily cycles), trend components (long-term changes), and residual components (random fluctuations).

[0068] S330 concatenates the current load characteristics with the periodic and trend components to generate trend prediction input data.

[0069] For example, the current load characteristics (such as load level L_current, load type T_current) are concatenated with the periodic component value P_t and the trend component value Tr_t corresponding to the current time to form the input vector [L_current,T_current,P_t,Tr_t].

[0070] S340 inputs the trend prediction data into the long short-term memory network model for prediction and inference, and outputs the load change trend within a preset time window in the future.

[0071] For example, the LSTM model takes the above 4-dimensional vector as input and outputs the load level prediction and load type conversion probability for each minute within the next 5 minutes.

[0072] Through steps S310 to S340 above, combined with historical periodicity and trend characteristics, the LSTM model can effectively predict short-term load changes, providing a basis for proactive control.

[0073] In one possible implementation, the preset gearbox health status parameters in S400 are determined in the following way: S410, establish a temperature-torque-speed mapping model for the reducer under no-load conditions and at least three typical load conditions, as well as a baseline model for the output shaft vibration spectrum.

[0074] For example, through experimental testing, temperature (unit: °C), torque (unit: N·m), and speed (unit: rpm) data are collected under no-load, light-load, medium-load, and heavy-load conditions, and a temperature-torque-speed mapping relationship model (such as a multiple linear regression model) is fitted; vibration signals under normal operating conditions are collected, and a vibration spectrum baseline model (such as an average spectrum template) is established.

[0075] S420 compares the real-time acquired multi-source time-series sensor data with the mapping relationship model and the vibration spectrum baseline model to calculate the weighted Euclidean distance of each feature deviation. The weighting coefficients are preset based on the degree of influence of each feature on the health status of the reducer; all coefficients are positive and their sum is 1.

[0076] For example, the deviations between real-time temperature, torque, and speed and the predicted values ​​of the mapping model are calculated and denoted as ΔT (temperature deviation), ΔTorque (torque deviation), and ΔRPM (speed deviation), respectively. The difference between the real-time vibration spectrum and the baseline spectrum, ΔVib, is calculated (the vibration spectrum deviation can be obtained through methods such as the spectral correlation coefficient). These values ​​are then substituted into the formula to calculate the weighted Euclidean distance D = sqrt(w1 × ΔT² + w2 × ΔTorque² + w3 × ΔRPM² + w4 × ΔVib²), where the weighting coefficients are preset according to the importance of the features, for example, w1 = 0.3 (temperature), w2 = 0.3 (torque), w3 = 0.2 (speed), and w4 = 0.2 (vibration). The weighted Euclidean distance D is a comprehensive dimensionless index; a larger value indicates a greater overall deviation between the real-time state and the health baseline.

[0077] S430, based on the quantization results of the weighted Euclidean distance D, determine the health status parameters of the reducer.

[0078] For example, the distance D is mapped to a health score H using the formula H = 100 × exp(-0.1 × D). This health score H is a quantitative indicator ranging from 0 to 100: when D is 0 (no bias), H is 100 points, representing extremely good health; as the value of D increases, the value of H gradually approaches 0, representing a deterioration in health. The characteristics of the exponential function exp(-0.1D) make the score insensitive to small initial deviations, but produce a significant score decay for larger deviations, which aligns with the application logic in practical engineering.

[0079] Through the steps S410 to S430 above, the system can effectively quantify the real-time health status of the reducer by comparing multiple features and calculating weighted Euclidean distance, providing key equipment status information for subsequent precise adaptive torque control.

[0080] In one possible implementation, S400 generates adaptive torque control commands based on load characteristics and load change trends, combined with preset reducer health status parameters, including: S440: If the load type is identified as an impact load and the predicted trend indicates that the load will continue to increase, then a proactive torque increase command is generated based on the predicted load change gradient and the health status parameter H. The proactive torque increase command includes specific parameters for the torque increase magnitude and rate.

[0081] For example, if the health status parameter H=85 (good health status), then the command is generated: torque increase of 15%, increase rate of 40 N·m / s.

[0082] If the S450 identifies a constant load type and predicts a decrease in load, it generates an energy-saving torque reduction command. This command smoothly reduces the output torque to the optimal efficiency range that matches the current load level.

[0083] For example, if the current load level is medium load and its optimal efficiency range corresponds to 70%-80% of the rated torque, then the command generated is: smoothly reduce the torque to 75% of the rated torque at a rate of 20 N·m / s.

[0084] S460, if the health status parameter H is lower than the first preset threshold, the upper limit of torque output is dynamically reduced; if the health status parameter H is lower than the second preset threshold, a maintenance alarm signal is generated and a degraded operation suggestion is output. The first preset threshold is higher than the second preset threshold, and both are set based on the health benchmark value of the reducer under rated operating conditions.

[0085] For example, a first threshold H1 is set to 80 and a second threshold H2 to 60; if H < 80, the upper limit of torque output is reduced to 90% of the rated value; if H < 60, a maintenance alarm signal is generated and degraded operation is recommended (e.g., the upper limit of torque is limited to 70% of the rated value).

[0086] Through the above steps S440 to S460, the system can generate differentiated and precise control commands based on load type, changing trend and health status parameters, thereby realizing adaptive adjustment of torque output and preventive protection of equipment operating status.

[0087] In one possible implementation, the S500 optimizes the load identification logic and torque control strategy based on historical performance data, including: S510's performance data includes energy consumption change per unit time ΔE / Δt (the smaller the value, the higher the energy efficiency), output shaft vibration smoothness index Vibration (usually represented by the root mean square value of vibration acceleration; the smaller the value, the smoother the operation), torque matching error Error (the absolute error between the actual torque and the target torque; the smaller the value, the higher the control precision), and housing temperature stability range T (temperature fluctuation range; the smaller the value, the better the thermal stability).

[0088] S520 employs a policy gradient reinforcement learning algorithm to construct a multi-objective optimization function J with the goals of "improving energy efficiency, ensuring stable operation, and extending equipment lifespan".

[0089] For example, the function can be expressed as J=w1×ΔE+w2×Vibration+w3×Error+w4×ΔT, where w1 to w4 are weight coefficients, representing the degree of emphasis on different performance indicators during the optimization process, which can be dynamically adjusted through reinforcement learning algorithms (such as PPO).

[0090] S530, based on the multi-objective optimization function J, performs online iterative adjustments to the weight parameters of the load classification neural network model and the decision thresholds (such as load level classification thresholds, health status thresholds, etc.) in the torque control command generation logic, thereby realizing continuous updating and optimization of the load identification logic and torque control strategy.

[0091] Through the steps S510 to S530 above, the system can continuously optimize the load identification and torque control strategies based on multi-objective optimization functions and reinforcement learning algorithms, using actual operating performance data, thereby continuously improving the overall performance, adaptability, and economy of the system.

[0092] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0093] Corresponding to the load-identification-based adaptive torque control method for reducers described in the above embodiments, this application also provides a load-identification-based adaptive torque control system for reducers, wherein each module of the system can implement each step of the load-identification-based adaptive torque control method for reducers. Figure 2 The diagram shows a structural block diagram of a speed reducer adaptive torque control system based on load identification provided in an embodiment of this application. For ease of explanation, only the parts related to the embodiments of this application are shown.

[0094] Reference Figure 2 The system includes: The data acquisition module 100 is used to acquire multi-source time-series sensor data in real time during the operation of the speed reducer; The load identification and trend prediction module 200 is used to extract time-domain and frequency-domain features from multi-source time-series sensor data to identify the load characteristics of the current load, and combine historical load data to predict the load change trend within a preset time window in the future. The health status assessment module 300 is used to establish a temperature-torque-speed mapping model and an output shaft vibration spectrum baseline model for the reducer. By comparing real-time multi-source data with model data, the health status parameters of the reducer are obtained through quantification. The intelligent control module 400 is used to generate adaptive torque control commands based on load characteristics, load change trends and health status parameters, so as to dynamically adjust the motor output torque and synchronously adapt to the transmission status of the reducer transmission module. The strategy optimization module 500 is used to record the execution performance data of adaptive torque control commands and optimize the load identification logic and torque control strategy based on the strategy gradient reinforcement learning algorithm.

[0095] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0097] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0098] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for adaptive torque control of a speed reducer based on load identification, characterized in that, Includes the following steps: Real-time acquisition of multi-source time-series sensor data during the operation of the reducer, including motor input current, reducer output shaft vibration signal, output shaft speed and reducer housing temperature; The multi-source time-series sensing data is processed to extract time-domain and frequency-domain features to identify the load characteristics of the current load, which include load level and load type. Based on the load characteristics and combined with historical load data, predict the load change trend within a preset time window in the future. Based on the load characteristics and load change trends, and combined with preset reducer health status parameters, an adaptive torque control command is generated to dynamically adjust the motor output torque and correspondingly adjust the transmission status of the reduction transmission module. Record the execution performance data of the adaptive torque control command, and based on the historical execution performance data, use the policy gradient reinforcement learning algorithm to iteratively optimize the load identification logic and torque control strategy.

2. The method as described in claim 1, characterized in that, Real-time acquisition of multi-source time-series sensor data during the operation of the speed reducer, including: The motor input current is collected by a current sensor, the vibration signal of the reducer output shaft is collected by a vibration sensor, the output shaft speed is collected by a speed sensor, and the reducer housing temperature is collected by a temperature sensor. Each sensor is electrically connected to the synchronous acquisition unit of the reducer control system, so that the acquired sensor data has a consistent timestamp. The acquired raw sensor data is sequentially filtered, denoised, and normalized to eliminate environmental electromagnetic interference signals and instantaneous pulse abnormal signals, thereby obtaining preprocessed multi-source time-series sensor data.

3. The method as described in claim 1, characterized in that, The multi-source time-series sensing data is subjected to time-domain and frequency-domain feature extraction to identify the load characteristics of the current load, including: Perform a fast Fourier transform on the output shaft vibration signal to extract its main frequency amplitude features and harmonic distribution features, and construct a vibration feature vector; A fluctuation analysis of the motor input current is performed, and its short-time standard deviation and ripple coefficient are calculated to obtain the current fluctuation characteristics. The vibration feature vector, the current fluctuation feature, the real-time output shaft speed and the real-time housing temperature are fused to generate load feature input data. The load feature input data is input into a pre-trained load classification neural network model, which outputs the load type and the load level.

4. The method as described in claim 1, characterized in that, Based on the load characteristics and combined with historical load data, predict the load change trend within a preset time window in the future, including: Historical load data are constructed into a load history sequence in chronological order, wherein each data point in the load history sequence includes load level, load type and corresponding collection timestamp; A time series decomposition algorithm is used to decompose the historical load sequence and extract periodic components, trend components, and residual components. The currently identified load features are concatenated with the periodic and trend components to generate trend prediction input data. The trend prediction input data is input into a long short-term memory network model for prediction, and the load change trend within the future preset time window is output.

5. The method as described in claim 1, characterized in that, The preset speed reducer health status parameters are determined in the following way: Establish a temperature-torque-speed mapping model for the reducer under no-load conditions and at least three typical load conditions, as well as a baseline model for the output shaft vibration spectrum. The real-time collected multi-source time-series sensor data, including real-time housing temperature, motor input current, output shaft vibration signal and speed, are compared with the temperature-torque-speed mapping relationship model and the vibration spectrum baseline model, respectively, and the weighted Euclidean distance of each feature deviation is calculated. The weighting coefficient is preset according to the degree of influence of each feature on the health status of the reducer. The health status parameters of the reducer are determined based on the numerical results of the weighted Euclidean distance.

6. The method as described in claim 1, characterized in that, Based on the load characteristics and load change trends, and combined with preset reducer health status parameters, an adaptive torque control command is generated, including: If the load type is identified as an impact load and the predicted trend indicates that the load will continue to increase, then an advance torque increase command is generated based on the predicted load change gradient and health status parameters. The advance torque increase command includes torque increase magnitude and increase rate parameters. If the load type is identified as a constant load and the predicted trend indicates that the load will decrease, an energy-saving torque reduction command is generated. The energy-saving torque reduction command is used to smoothly reduce the output torque to the optimal efficiency range that matches the current load level. If the health status parameter is lower than the first preset threshold, the upper limit of torque output is dynamically reduced; if the health status parameter is lower than the second preset threshold, a maintenance alarm signal is generated and a degraded operation suggestion is output. The first preset threshold is higher than the second preset threshold, and both are set based on the health benchmark value of the reducer under rated operating conditions.

7. The method as described in claim 1, characterized in that, Based on historical performance data, a policy gradient reinforcement learning algorithm is used to iteratively optimize the load identification logic and torque control strategy, including: The execution performance data includes energy consumption change per unit time, output shaft vibration smoothness index, torque matching error, and housing temperature stability range. Construct a multi-objective optimization function with the goals of improving energy efficiency, ensuring stable operation, and extending equipment life; Based on the multi-objective optimization function, the weight parameters of the load classification neural network model and the decision threshold in the torque control command generation logic are adjusted online iteratively to update the load identification logic and torque control strategy.

8. A load-identification-based adaptive torque control system for a speed reducer used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source time-series sensor data in real time during the operation of the speed reducer; The load identification and trend prediction module is used to extract time and frequency domain features from multi-source time-series sensor data to identify the load characteristics of the current load, and combine historical load data to predict the load change trend within a preset time window in the future. The health status assessment module is used to establish a temperature-torque-speed mapping model and an output shaft vibration spectrum baseline model for the reducer, and to obtain the health status parameters of the reducer by comparing real-time multi-source data with model data. The intelligent control module is used to generate adaptive torque control commands based on load characteristics, load change trends and health status parameters, so as to dynamically adjust the motor output torque and synchronously adapt to the transmission status of the reducer transmission module. The strategy optimization module is used to record the execution performance data of adaptive torque control commands and optimize the load identification logic and torque control strategy based on the strategy gradient reinforcement learning algorithm.

9. The system as described in claim 8, characterized in that, The data acquisition module includes: A current sensor is used to collect the input current of the motor; Vibration sensor, used to collect vibration signals from the output shaft of the reducer; A speed sensor is used to collect the output shaft speed of the reducer; Temperature sensor used to collect the temperature of the gearbox housing; Each sensor is equipped with a signal preprocessing unit, which is used to perform preliminary filtering on the raw sensing data.

Images