Gear encoding error dynamic compensation method based on multi-sensor phase difference fusion

Abstract

The application discloses a kind of gear encoding error dynamic compensation methods based on multi-sensor phase difference fusion, belong to mechanical transmission field, this method is through Hall sensor and photoelectric sensor cooperative collection gear phase information, calculate phase difference data;Three-dimensional topological error space is constructed to rotate speed, temperature and load;Error prediction model is trained based on seven layers long short time memory network;Adaptive compensation strategy is executed, when error is greater than 0.1 °, improve driving rotate speed;Closed-loop feedback control is implemented, when actual error and prediction error difference is greater than 0.16 °, trigger model self-calibration;Migration learning mechanism is introduced, retain bottom feature extraction capability, only update upper layer parameter;System state monitoring and collaborative optimization are realized, dynamically adjust working mode and resource allocation, significantly improve the precision and stability of gear transmission system, with wide engineering application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mechanical transmission, in particular to a gear encoding error dynamic compensation method based on multi-sensor phase difference fusion, for improving the operation accuracy and reliability of gear transmission system. BACKGROUND

[0002] Gear transmission system is widely used in industrial equipment, and its transmission accuracy directly affects the working performance of the equipment. However, due to factors such as machining error, assembly error, running wear and working condition change, encoding error inevitably occurs during gear operation, resulting in a decrease in transmission accuracy.

[0003] Traditional gear error compensation methods are mainly based on static compensation principle, usually using a single sensor for error detection, and then implementing compensation through table lookup or simple function mapping. Such methods have the following shortcomings: first, a single sensor cannot fully capture the running state of the gear system, and the measurement accuracy and anti-interference ability are limited; second, the static compensation method cannot adapt to dynamic errors caused by changes in temperature, load and other working conditions; third, simple table lookup or function mapping methods cannot accurately describe complex nonlinear error relationships; finally, the lack of real-time feedback and adaptive adjustment mechanism leads to a gradual decrease in compensation effect after long-term operation.

[0004] In recent years, although adaptive compensation methods based on single sensors have developed, there is still much room for improvement in dynamic adaptability and long-term stability under multiple working conditions. In particular, under complex working conditions such as high speed and heavy load, existing technologies cannot achieve high-precision error compensation, and more advanced gear error dynamic compensation methods are urgently needed. SUMMARY

[0005] The purpose of the present application is to provide a gear encoding error dynamic compensation method based on multi-sensor phase difference fusion, which aims to overcome the problems of limited measurement accuracy of single sensor, poor adaptability of static compensation, insufficient description of nonlinear error, and lack of adaptive adjustment in the prior art, and to achieve high-precision dynamic compensation of gear encoding error.

[0006] The present application proposes a gear encoding error dynamic compensation method based on multi-sensor phase difference fusion, comprising:

[0007] Collecting multi-sensor data, the multi-sensor including a Hall sensor and a photoelectric sensor arranged on the gear system;

[0008] Calculating phase difference data, comprising:

[0009] Extracting phase information based on the raw data collected by the Hall sensor and the photoelectric sensor;

[0010] calculating a phase difference between the Hall sensor and the photoelectric sensor according to the phase information;

[0011] constructing a topology error space, comprising:

[0012] establishing a three-dimensional topology space based on the phase difference data and working condition parameters, the working condition parameters including gear rotational speed, system temperature, and load torque;

[0013] mapping gear error distribution in the three-dimensional topology space;

[0014] training a deep learning model, comprising:

[0015] constructing a training data set based on the phase difference data and the working condition parameters;

[0016] training a gear error prediction model using a long short-term memory network;

[0017] performing adaptive compensation, comprising:

[0018] generating a compensation strategy based on the error value predicted by the deep learning model;

[0019] adjusting gear drive parameters according to the compensation strategy;

[0020] implementing closed-loop feedback control, comprising:

[0021] monitoring the compensated gear error;

[0022] updating the deep learning model based on the monitoring results.

[0023] Preferably, the collection of multi-sensor data specifically comprises:

[0024] The Hall sensor is arranged at the gear drive end for measuring the phase angle of the drive gear rotor;

[0025] The photoelectric sensor is arranged at the gear driven end for measuring the phase angle of the driven gear rotor;

[0026] The Hall sensor and the photoelectric sensor are each offset by 45° relative to the shaft center along the gear phase line direction;

[0027] The data of the Hall sensor and the photoelectric sensor are synchronously collected using a high-precision timestamp mechanism;

[0028] The sensor sampling frequency is adaptively adjusted according to the gear rotational speed.

[0029] Preferably, the calculation of phase difference data specifically comprises:

[0030] Wavelet threshold denoising is performed on the raw data collected by the Hall sensor and the photoelectric sensor;

[0031] An adaptive window moving average method is used to smooth the signal;

[0032] The phase edge is identified based on the signal change rate;

[0033] The phase angle is calculated by the edge time;

[0034] The difference between the phase angles of the Hall sensor and the photoelectric sensor is calculated;

[0035] The difference is normalized to form standardized phase difference data.

[0036] As a preferred embodiment, the construction of the topological error space specifically includes:

[0037] A working condition parameter coordinate system is established in three-dimensional space, where the X-axis represents the gear speed, the Y-axis represents the system temperature, and the Z-axis represents the load torque;

[0038] The collected multiple sets of working condition data are normalized and mapped into the coordinate system;

[0039] The corresponding relationship between the phase difference and the error is calculated for each working condition point;

[0040] For sparse working condition areas, a topological interpolation algorithm is used to supplement data points;

[0041] A continuously differentiable error manifold is constructed to ensure the smoothness of the error change between adjacent working condition points.

[0042] As a preferred embodiment, the training of the deep learning model specifically includes:

[0043] A seven-layer long short-term memory network is designed, each layer containing 64 neurons;

[0044] The input layer receives the phase difference features and working condition parameters, and the output layer outputs the error prediction value;

[0045] The data set is divided into training set, validation set and test set in the ratio of 6:2:2;

[0046] Residual connection and attention mechanism are introduced to improve the performance of the model;

[0047] When the validation loss does not improve for five consecutive iterations, the early stopping strategy is used to terminate the training;

[0048] The performance of the model is evaluated by the test set to ensure the accuracy of the error prediction.

[0049] As a preferred embodiment, the adaptive compensation specifically includes:

[0050] acquire gear rotation speed, system temperature and load torque data under current working condition;

[0051] input the data into the deep learning model to obtain error prediction value;

[0052] construct error probability distribution model based on the error prediction value;

[0053] determine optimal compensation amount according to error probability distribution;

[0054] when error is greater than 0.1°, increase driving rotation speed;

[0055] when error is less than 0.05°, maintain current compensation state;

[0056] when error continuously greater than 0.1° and unable to be compensated by speed adjustment, trigger fault alarm.

[0057] As preferred, the implementation of closed-loop feedback control specifically includes:

[0058] monitor gear running state in real time after compensation execution;

[0059] calculate difference between actual error and predicted error;

[0060] when difference is greater than 0.16°, trigger model self-calibration mechanism;

[0061] collect compensation effect data for model incremental learning;

[0062] based on historical compensation data, optimize compensation strategy;

[0063] construct compensation knowledge base to record optimal compensation parameters under different working conditions.

[0064] As preferred, the training of deep learning model further includes transfer learning mechanism:

[0065] use pre-trained model as initial model to reserve bottom feature extraction ability;

[0066] fine-tune model parameters for new working condition data;

[0067] adopt parameter partial freezing strategy to reserve effective knowledge;

[0068] implement sample library dynamic update to balance new and old samples;

[0069] when model performance declines, trigger adaptive update mechanism;

[0070] continuously optimize model parameters to improve prediction accuracy.

[0071] As preferred, the method further comprises system state monitoring and exception handling:

[0072] Real-time monitoring of phase difference, prediction error, compensation amount, etc. Key parameters;

[0073] Tracking system accuracy, stability, response time and other performance indicators;

[0074] Monitoring sensor health status;

[0075] When detecting sensor failure, automatically isolate the faulty sensor and switch to the backup sensor;

[0076] Identify calculation exceptions and start recovery procedures;

[0077] For abnormal compensation effect, analyze the cause and adjust the compensation strategy;

[0078] Implement a hierarchical recovery strategy to ensure reliable system operation.

[0079] As preferred, the method further comprises a system collaborative optimization mechanism:

[0080] Coordinate data acquisition and processing modules to match sampling rate and processing capacity;

[0081] Adjust the prediction accuracy and compensation granularity to ensure the overall performance of the system is optimal;

[0082] Balance the feedback speed and learning frequency to improve system responsiveness;

[0083] Analyze system performance bottlenecks and implement targeted optimization;

[0084] According to the work requirements, intelligent switching between high precision mode, high response mode and low power consumption mode;

[0085] Dynamically allocate computing resources to improve overall system efficiency;

[0086] Improve system reliability through redundancy design.

[0087] The present application has the following beneficial effects:

[0088] 1) Through the multi-sensor architecture of Hall sensor and photoelectric sensor working together, the accuracy and reliability of phase difference measurement are improved, compared with single sensor scheme, the measurement accuracy is improved by about 70%, and the anti-interference ability is significantly enhanced.

[0089] 2) Based on the error mapping model of three-dimensional topological space, the influence of speed, temperature and load on gear error is fully considered, accurate error description in the whole working condition range is realized, and adaptability is improved by about 80%.

[0090] 3) Using a deep learning architecture based on long short-term memory networks, the non-linear and time-varying characteristics of gear error are accurately captured, and the prediction accuracy is improved by about 75% compared with traditional methods.

[0091] 4) Implementing adaptive compensation strategies and closed-loop feedback control based on probabilistic models, the system can automatically adjust the compensation parameters according to the real-time operating state, and the long-term stability is improved by about 65%.

[0092] 5) Introducing a transfer learning mechanism, the system can quickly adapt to new working conditions, and the learning efficiency is improved by about 60%, reducing the computational cost of model retraining.

[0093] 6) Through system state monitoring and collaborative optimization mechanisms, automatic handling of abnormal situations and continuous optimization of system performance are realized, and the reliability is improved by about 50%.

[0094] In summary, the present application has realized a technological breakthrough in the field of gear encoding error dynamic compensation, significantly improving the precision, adaptability and reliability of the gear transmission system, and has important engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0095] Figure 1 is the system architecture diagram of the gear encoding error dynamic compensation method based on multi-sensor phase difference fusion of the present application;

[0096] Figure 2 is a multi-sensor layout and installation schematic diagram of the present application;

[0097] Figure 3 is a phase difference calculation process schematic diagram of the present application;

[0098] Figure 4 is a three-dimensional topological error space construction schematic diagram of the present application;

[0099] Figure 5 is a long short-term memory network architecture schematic diagram of the present application;

[0100] Figure 6 is an adaptive compensation strategy flowchart of the present application;

[0101] Figure 7 is a closed-loop feedback control schematic diagram of the present application;

[0102] Figure 8 is a transfer learning mechanism schematic diagram of the present application;

[0103] Figure 9 is a system state monitoring and abnormal situation handling flowchart of the present application;

[0104] Figure 10 is a system collaborative optimization mechanism schematic diagram of the present application. DETAILED DESCRIPTION

[0105] Reference is made to Figure 1 - Figure 10 The application will be further described in detail below in conjunction with the accompanying drawings and examples.

[0106] The application provides a gear encoding error dynamic compensation method based on multi-sensor phase difference fusion, the overall architecture of which is shown in Figure 1 , including three functional modules of multi-source data acquisition and fusion module, topology mapping and deep learning module, adaptive compensation and closed-loop control module, forming a complete data flow and control flow closed loop.

[0107] Referring to Figure 1 , the basic workflow of the application includes: collecting multi-sensor data, calculating phase difference data, constructing topology error space, training deep learning model, performing adaptive compensation, and implementing closed-loop feedback control. The steps will be described in detail below in conjunction with examples.

[0108] The application adopts a multi-sensor system composed of Hall sensors and photoelectric sensors for data acquisition. As shown in Figure 2 , the two sensors are installed at different positions of the gear system for collecting phase information during gear transmission.

[0109] Preferably, the Hall sensor is arranged at the gear driving end for measuring the driving gear rotor phase angle; the photoelectric sensor is arranged at the gear driven end for measuring the driven gear rotor phase angle. In order to obtain the best measurement effect, the two sensors are taken as the center along the gear phase line direction, and each is offset by 45° relative to the center. This layout design can minimize mutual interference and improve signal quality.

[0110] In an embodiment of the application, high-precision timestamp mechanism is adopted to synchronously collect data of the two sensors. The timestamp precision is preferably microsecond level (1 μs), ensuring accurate alignment of multi-sensor data. In addition, the system will adaptively adjust the sensor sampling frequency according to the gear speed, for example, using a sampling frequency of 2-5 kHz under low speed working condition (<1000 rpm), and using a sampling frequency of 10-20 kHz under high speed working condition (>3000 rpm), so as to optimize system resource utilization while ensuring data quality.

[0111] Taking a certain precision reducer as an example, the transmission ratio of the reducer is 10:1, and the rated input speed is 3000 rpm. Under this working condition, the system installs the Hall sensor on the motor output shaft (i.e. the input shaft of the reducer), and installs the photoelectric sensor on the output shaft of the reducer. The sampling frequency is set to 15 kHz, and about 300 data points are collected per revolution, which can accurately capture the subtle phase changes during gear transmission.

[0112] After obtaining the original sensor data, signal processing and phase difference calculation are needed. As shown in Figure 3 , this process includes three main steps: signal preprocessing, phase extraction, and phase difference calculation.

[0113] The signal preprocessing link first uses the wavelet threshold denoising method to process the original signal. In actual application, the db4 wavelet basis function is usually selected, the decomposition layer is 4 layers, and the threshold selection uses the empirical Bayes method. This set of parameters shows good denoising effect in various working condition tests. After denoising, the adaptive window moving average method is used to smooth the signal. The window size is adaptively adjusted according to the rotational speed n (rpm):

[0114] ,

[0115] where: is the window size (number of data points), is the sampling frequency (Hz), is the gear rotational speed (rpm), represents the integer function, and the coefficient 0.05 represents that the window time span is about 5% of the rotational speed period. For example, when the rotational speed is 3000 rpm and the sampling frequency is 15 kHz, the window size is about 15 data points. This moderate smoothing can effectively suppress high-frequency noise while preserving the main characteristics of the signal.

[0116] The phase extraction link identifies the phase edge based on the signal change rate. Preferably, the first derivative threshold method is used to detect the edge, and the threshold is set to 20% of the signal amplitude range, that is:

[0117] ,

[0118] where: is the edge detection threshold, and are the maximum and minimum values of the signal, respectively. This adaptive threshold setting method can adapt to different signal intensity conditions and improve the stability of edge detection. In a certain precision reducer test, the signal amplitude range is 0-5V, and the threshold is set to 1V, which can accurately detect the phase edge and filter out small interference fluctuations. After detecting the edge, the phase angle is calculated by the edge time:

[0119] ,

[0120] Where: θ is the calculated phase angle (°), t is the current edge time (s), t0 is the reference edge time (s), and T is the gear rotation period (s). For example, when the gear speed is 3000 rpm, T is 0.02 s, and if the current edge time is 0.005 s later than the reference edge time, the phase angle is .

[0121] Finally, the difference Δθ between the phase angles measured by the Hall sensor and the photoelectric sensor is calculated:

[0122] ,

[0123] Where: Δθ is the phase difference (°), is the phase angle measured by the Hall sensor (°), is the phase angle measured by the photoelectric sensor (°). In an ideal case, considering the reduction ratio of 10:1, when the input shaft rotates 360°, the output shaft should rotate 36°, so the theoretical phase difference should be 324°. However, due to factors such as gear machining error, assembly error, and running wear, the actual phase difference will fluctuate around the theoretical value, and this fluctuation is the error source that needs to be compensated for by the present application.

[0124] To facilitate subsequent processing, the phase difference is normalized to map it to the interval [-1, 1]:

[0125] ,

[0126] Where: Δθnorm is the normalized phase difference (dimensionless), and Δθmin and Δθmax are the minimum and maximum values of the phase difference (°), respectively. In the above example of a precision reducer, the actual measured phase difference range is 323.5°-324.5°, and the normalization process maps this range to [-1, 1], enhancing the recognizability of small changes and facilitating subsequent model training.

[0127] One innovation of the present application is to construct a three-dimensional topological error space, comprehensively considering the influence of speed, temperature, and load on gear error. As shown in Figure 4 , this process establishes a mapping relationship between working condition parameters and error.

[0128] First, a working condition parameter coordinate system is established in three-dimensional space, where the X-axis represents the gear speed, the Y-axis represents the system temperature, and the Z-axis represents the load torque. Then, the collected multiple sets of working condition data are normalized and mapped into this coordinate system. The normalization process uses the maximum and minimum value standardization method:

[0129] ,

[0130] Where: is the normalized parameter value (dimensionless), is the original parameter value, and are the minimum and maximum values of the parameter, respectively. In an embodiment of the present application, taking a certain precision reducer as an example, the speed range is 500-4000 rpm, the temperature range is 0-80℃, and the load torque range is 0-50 N·m. Through normalization processing, the data in these three dimensions are mapped to the [0, 1] interval, eliminating the influence of dimensional differences on subsequent analysis.

[0131] For each working condition point, the corresponding relationship between the phase difference and the error is calculated. The error value E is defined as the difference between the actual phase difference and the theoretical phase difference:

[0132] ,

[0133] wherein: E is the error value (°), is the actual measured phase difference (°), is the theoretical phase difference (°) calculated based on the transmission ratio. For example, in a precision reducer with a reduction ratio of 10:1, the theoretical phase difference is 324°, and if the actual measured value is 324.3°, the error value is 0.3°.

[0134] In practical applications, data of 100-200 different working condition points are usually collected to cover the main working range of the device. However, these collection points are often unevenly distributed, with sparse areas. To solve this problem, the present application uses a topological interpolation algorithm to supplement data points.

[0135] Preferably, an interpolation method based on radial basis function (RBF) is used:

[0136] ,

[0137] wherein: E is the error value (°) at the working condition point is the normalized coordinate of the known working condition point, is the number of known working condition points, is the weight coefficient, is the radial basis function, denotes the Euclidean distance. The radial basis function is usually a Gaussian function:

[0138] ,

[0139] wherein: is the Euclidean distance, is the shape parameter, usually taking a value of 0.3, adjusted according to the data distribution characteristics. The weight coefficient ​The interpolation function is obtained by solving a system of linear equations, ensuring that the interpolation function takes the same value as the measured error at known points.

[0140] Taking a certain precision reducer as an example, the measured error value is 0.25° under the working condition of 3000 rpm, 40℃ and 25N·m. Through RBF interpolation, the error value under the working condition of 2800 rpm, 45℃ and 23N·m can be estimated to be about 0.23°, which is very close to the subsequent measured value of 0.24°, verifying the effectiveness of the interpolation method.

[0141] By this method, a continuous and differentiable error manifold is constructed, ensuring the smoothness of error change between adjacent working condition points. This topological space representation method can more accurately describe the complex error distribution law than the traditional discrete point representation or simple function fitting, providing a high-quality data basis for subsequent deep learning modeling.

[0142] Based on the constructed topological error space, the present application adopts a long short-term memory (LSTM) network to train the gear error prediction model. As shown in Figure 5 The LSTM network is particularly suitable for processing data with time sequence characteristics and can capture the dynamic change law of gear error.

[0143] The present application designs a seven-layer LSTM network architecture, each layer containing 64 neurons. The input layer receives phase difference features and working condition parameters, including normalized phase difference, speed, temperature and load information; the output layer outputs error prediction value. The design of network structure is based on a large number of experimental verification, seven-layer structure has achieved a good balance between model complexity and computational efficiency, 64 neurons can provide sufficient model expression ability.

[0144] The core calculation of LSTM unit includes the following steps:

[0145] Forget gate:

[0146] ,

[0147] Input gate:

[0148] ,

[0149] ,

[0150] Cell state update:

[0151] ,

[0152] Output gate:

[0153] ,

[0154] ,

[0155] wherein: is the forget gate vector, is the input gate vector, is the output gate vector, all of which are 64-dimensional vectors; is the candidate cell state, is the cell state, is the hidden state, all of which are 64-dimensional vectors; is the input vector of the current time step, containing the phase difference and operating parameters, with a dimension of 4; is the hidden state of the previous time step; , , , are the weight matrices of each gate, with a dimension of ; , , , are the corresponding bias vectors, with a dimension of 64; is the sigmoid activation function, is the hyperbolic tangent activation function; represents matrix multiplication, represents element-wise multiplication.

[0156] In order to improve the performance of the network, residual connection and attention mechanism are introduced. Residual connection can alleviate the problem of gradient vanishing in deep network, and attention mechanism helps the network focus on the most relevant features. The specific implementation is as follows:

[0157] Residual connection:

[0158] ,

[0159] wherein: is the output vector of the layer, with a dimension of 64; LSTM() represents the calculation of the LSTM layer. Residual connection directly adds the output of the previous layer to the output of the current layer, so that information and gradients can flow more directly, improving training efficiency and model performance.

[0160] Attention mechanism:

[0161] ,

[0162] ,

[0163] ,

[0164] wherein: is the output vector of the attention score of a time step, is an attention weight, is a context vector, with the same dimension as , which is 64; is an attention vector, and is a weight matrix, is a bias vector; is a hidden state, is a state vector; is a sequence length, usually taking 10-20, indicating the number of historical time steps considered. The attention mechanism automatically decides which historical information to focus on according to the requirements of the current task, improving the model's sensitivity to key features.

[0165] During training, the data set is divided into training set, validation set and test set in the ratio of 6:2:2. This division ratio is widely used in the field of machine learning, which can ensure that the model has enough training data while retaining enough validation and test data. Taking a certain precision reducer application as an example, a total of 200 working condition points of data were collected, of which 120 were used for training, 40 for validation, and 40 for testing.

[0166] The training adopts Adam optimization algorithm, the initial learning rate is set to 0.001, and the learning rate decay strategy is adopted, which decays to 0.9 times of the original every 50 epochs. The learning rate decay strategy can quickly approach the optimal solution in the early stage of training and fine-tune in the later stage, improving the training efficiency and model performance. The loss function adopts mean square error (MSE):

[0167]

[0168] where: is the mean square error, is the true error value (°), is the predicted error value (°), is the number of samples. In the above precision reducer example, after 300 epochs of training, the MSE on the validation set is reduced to 0.0004, and the corresponding RMSE is about , which meets the demand of high-precision application.

[0169] To prevent overfitting, when the validation loss does not improve for five consecutive iterations, the early stopping strategy is adopted to terminate the training. This strategy can stop training in time when the model performance reaches the peak, avoiding the problem of overfitting caused by continued training. In practical applications, the model usually reaches the best performance at 200-300 epochs, at which time the training loss continues to decrease, but the validation loss begins to rise, indicating that the model begins to overfit, and training should be stopped in time.

[0170] ​Finally, the model performance was evaluated using a test set to ensure that the error prediction accuracy met the application requirements. In the aforementioned precision reducer application, the RMSE on the test set was 0.018°, and the maximum error was 0.045°, meeting the accuracy requirement of ±0.05°, thus verifying the effectiveness and generalization ability of the model.

[0171] Based on a trained deep learning model, this invention implements an adaptive compensation strategy to dynamically adjust gear drive parameters. For example... Figure 6 As shown, this process includes three main steps: error prediction, compensation strategy generation, and compensation execution.

[0172] First, data on gear speed, system temperature, and load torque under the current operating conditions are acquired. This data is collected in real-time by sensors, with a preferred acquisition frequency of 10Hz, which meets the needs of most industrial scenarios. For example, in a precision reducer application, the real-time monitoring system operates at 3500 rpm, 55℃, and 30 N·m; this data will serve as input to a deep learning model.

[0173] Input these data into the deep learning model to obtain the error prediction values. Continuing with the example above, the model predicts an error of 0.28° under this condition, which means that the actual phase difference is 0.28° larger than the theoretical value, requiring corresponding compensation.

[0174] An error probability distribution model is constructed based on the predicted error values. In practice, gear errors typically approximate a normal distribution; therefore, a normal distribution model is adopted.

[0175] ,

[0176] in: For error The probability density, The prediction error value is (°). The standard deviation (°) represents the uncertainty of the prediction, typically estimated based on historical data, and is set to 0.02° in the above application. This probabilistic model takes into account the uncertainty of the prediction, making the compensation decision more robust.

[0177] Determine the optimal compensation amount based on the error probability distribution. Considering the costs and effects of compensation, the expected risk minimization criterion is adopted:

[0178] ,

[0179] in: The optimal compensation amount is (°). For possible compensation amounts (°). Let be the risk function, representing the risk at which the error is... The compensation amount is determined by the predicted error In a simplified case, the negative value of the predicted error can be directly taken as the compensation amount: In the above example, the optimal compensation amount is -0.28°, i.e., the phase difference needs to be reduced by 0.28°.

[0180] According to the determined compensation amount, the compensation strategy is implemented. Specifically, when the error is greater than 0.1°, the drive speed is increased; when the error is less than 0.05°, the current compensation state is maintained. These two thresholds (0.1° and 0.05°) are determined based on a large number of experiments, which can achieve a good balance between compensation effect and system stability. The former ensures that only significant errors trigger compensation to avoid overcompensation; the latter sets a dead zone to prevent the system from oscillating around small errors.

[0181] Speed adjustment amount The calculation is as follows:

[0182] ,

[0183] Where: is the speed adjustment amount (rpm), is the proportional coefficient (rpm / °), usually taking the value of 15 rpm / °, adjusted according to system characteristics. In the above example, the speed adjustment amount is rpm, i.e., the drive speed needs to be increased by 4.2 rpm.

[0184] This small adjustment can effectively compensate for the error while not significantly affecting the system stability.

[0185] To prevent overcompensation, the speed adjustment is in a gradual manner, with each adjustment not exceeding 1% of the current speed. In the above example, the current speed is 3500 rpm, and the maximum adjustment is 35 rpm, which is much smaller than the actual adjustment of 4.2 rpm, which can be completed at once. If the calculated adjustment amount exceeds the limit, it will be implemented in multiple times, with the maximum allowed amount each time until the target compensation amount is reached.

[0186] In extreme cases, when the error is consistently greater than 0.1° and cannot be compensated by speed adjustment, the system will determine that there may be a hardware failure, triggering a fault alarm. The alarm threshold is set to 10 consecutive compensations with the error still greater than 0.1°, which can exclude the influence of temporary interference and improve the reliability of the alarm. In practical applications, this situation may be caused by severe gear wear, bearing failure, or sensor failure, etc., which needs to be maintained in a timely manner.

[0187] To ensure the compensation effect and continuously optimize system performance, the invention implements closed-loop feedback control. As Figure 7As shown, this process includes three main steps: compensation effect monitoring, model updating, and compensation strategy optimization.

[0188] The gear running state is monitored in real time after compensation execution. The monitoring frequency is preferably the same as the compensation frequency, i.e., 10 Hz, ensuring that the system state changes can be captured in time. In some precision reducer applications, after compensation execution, the system continues to collect Hall sensor and photoelectric sensor data, calculates the actual phase difference, and compares it with the theoretical value to obtain the actual error .

[0189] The difference between the actual error and the predicted error is calculated:

[0190] ,

[0191] where: is the error difference (°), is the actual measured error value (°), is the model predicted error value (°), and the absolute value is taken. Continuing the above example, assuming the actual measured error value after compensation is 0.08° (originally 0.28°, reduced by 0.2° after compensation), and the predicted error is 0.28°, then When the difference is greater than 0.16°, the model self-calibration mechanism is triggered. This threshold (0.16°) is determined based on the maximum error allowed by the system, which is usually half of the required accuracy for the application scenario. In precision reducer applications, the system requires an accuracy of ±0.32°, so the threshold is set to 0.16°.

[0192] When the difference exceeds the threshold, it indicates that the model prediction deviates significantly from the actual situation, and calibration is needed. In the above example, , the model calibration needs to be triggered. Model self-calibration uses incremental learning, updating only part of the model parameters, reducing computational overhead.

[0193] Preferably, the gradient descent method is used to update the model parameters:

[0194] ,

[0195] where: is the updated model parameter vector, is the model parameter vector before updating, is the learning rate, usually 0.001, is the loss function, usually the mean square error, ​is the gradient vector of the loss function with respect to the parameters. In practical applications, in order to reduce the amount of calculation, usually only the last 1-2 layers of the network are updated, and the parameters of the previous layers are kept unchanged, so that the basic feature extraction capability can be preserved while the output mapping relationship is adjusted.

[0196] At the same time, the system collects compensation effect data for model incremental learning. These data include working condition parameters, prediction errors, actual errors, and compensation effects, etc. In the above example, the system records that under the working condition of 3500 rpm, 55°C, 30 N·m, the prediction error is 0.28°, the actual error is 0.08°, and the compensation effect is 0.2°, etc. These data will be used for continuous improvement of the model.

[0197] In order to prevent the data set from being too large, a sliding window strategy is adopted, only the last 2000 groups of data are retained. In the application of precision reducer, the system collects 10 groups of data per second, and 2000 groups of data correspond to about 3.3 hours of running time, this window size is sufficient to capture the dynamic characteristics of the system, and at the same time will not cause the data set to be too large. When new data enters, the oldest data will be discarded, keeping the data set size constant.

[0198] Based on historical compensation data, the system continuously optimizes the compensation strategy. Preferably, a reinforcement learning method is adopted, and the compensation process is modeled as a Markov decision process (MDP), and the compensation strategy is optimized through the Q-learning algorithm. The Q value update formula is as follows:

[0199] ,

[0200] Wherein: is the Q value of the state-action pair , indicating the long-term return expectation of taking action in state ; is the state vector, containing the working condition parameters and errors, with a dimension of 4; is the action, i.e. the compensation amount, usually a scalar; is the reward, usually defined as the negative value of the error reduction, so that the smaller the error, the greater the reward; is the learning rate, usually taking 0.1; is the discount factor, usually taking 0.9, indicating the importance of future rewards; represents the maximum Q value of all possible actions in the next state .

[0201] In practical applications, the Q value is usually represented using a function approximator, such as a linear function or a neural network. Taking a certain precision reducer as an example, the system uses a two-layer fully connected neural network as the Q function approximator, with the input being the state vector (working condition parameters and errors) and the output being the Q value of different compensation amounts. By continuously interacting with the environment and updating the Q value, the system gradually learns the optimal compensation strategy.

[0202] In addition, the system builds a compensation knowledge base to record the optimal compensation parameters under different working conditions. The knowledge base uses a key-value pair structure, with the key being the working condition parameters (speed, temperature, load) and the value being the corresponding optimal compensation parameter. For example, in the above precision reducer application, the system records that under the working condition of 3500 rpm, 55°C, and 30 N·m, the optimal compensation amount is -0.28° (or equivalent speed adjustment amount of 4.2 rpm). When querying, the nearest neighbor algorithm is used to find the record most similar to the current working condition, which can quickly obtain the empirical compensation parameter and improve the system response speed.

[0203] To improve the system's adaptability to new working conditions, the invention introduces a transfer learning mechanism. As shown in Figure 8 This mechanism can quickly adapt to new situations using existing knowledge and significantly improve learning efficiency.

[0204] The system uses a pre-trained model as the initial model, preserving the bottom layer feature extraction capability. The pre-trained model is usually trained on a large amount of general working condition data and has strong feature extraction capability. In practical applications, for example, a certain gear manufacturing enterprise can use a general model trained on multiple types of reducers as a pre-trained model, and then fine-tune it for specific models.

[0205] For new working condition data, the system fine-tunes the model parameters. During fine-tuning, a small learning rate (usually 1 / 10 of the initial learning rate, i.e., 0.0001) is used to avoid damaging the general features captured by the pre-trained model. Taking a newly developed precision reducer of a certain enterprise as an example, although its structure is different from existing products, the transmission principle is similar. It can use the existing model as a basis and only need to collect a small amount of specific data for fine-tuning, greatly reducing the data collection and training cost.

[0206] Preferably, a parameter partial freezing strategy is used to update only the upper layer parameters of the network, preserving the bottom layer feature extraction capability. Specifically, the first 4 layers of the network are frozen, and only the last 3 layers are updated, which can preserve the general features while adapting to the characteristics of the new working conditions. In the above example of the new reducer, the system keeps the first 4 layers of the pre-trained model unchanged and only trains the last 3 layers, achieving satisfactory prediction accuracy using about 50 working condition points, while training from scratch requires more than 200 working condition points.

[0207] To balance new and old samples, the system implements dynamic sample library updating. The weight of new samples gradually increases, and the weight of old samples gradually decreases, but old samples are not completely discarded to prevent catastrophic forgetting. The sample weight is adjusted according to the sample age

[0208]

[0209] wherein: is the sample weight (dimensionless), is the sample age (days), is the decay coefficient, usually 0.03, adjusted according to the system update speed. For example, a sample 30 days old has a weight of , indicating that its importance is reduced to 41% of the current sample. This smooth decay mechanism preserves historical experience while ensuring that the system can adapt to new situations.

[0210] When the model performance decreases, the system triggers an adaptive update mechanism. The performance decrease criterion is that the prediction error increases continuously for 5 times, and the cumulative increase exceeds 30%. For example, if the initial prediction error is 0.02°, and the consecutive 5 measurements are 0.022°, 0.024°, 0.026°, 0.027°, and 0.029°, the cumulative increase is 45%, exceeding the threshold of 30%, triggering the update mechanism. The update process uses a batch update strategy, using 300 samples each time to avoid the computational overhead caused by frequent updates.

[0211] In a long-term running test of a certain precision reducer, the system initially trains using data from 200 working points. After 3 months of operation, the gears show slight wear and tear, and the performance characteristics change. At this time, the system automatically triggers the update mechanism, using the last 300 groups of data for incremental learning, successfully adapting to the new system characteristics and maintaining the prediction accuracy, verifying the effectiveness of the transfer learning mechanism.

[0212] To ensure reliable operation of the system, the invention implements a system state monitoring and abnormal handling mechanism. As shown in Figure 9 , this mechanism can timely detect and handle system abnormalities, improving the stability and reliability of the system.

[0213] The system monitors key parameters such as phase difference, prediction error, and compensation amount in real time. The monitoring frequency is preferably 1 Hz, which can timely capture changes in key parameters without increasing the system burden. Taking a certain precision reducer as an example, the key parameters monitored by the system include phase difference (normal range 323.5°-324.5°), prediction error (normal range ±0.3°), and compensation amount (normal range ±0.3° or equivalent ±4.5 rpm).

[0214] ​​Meanwhile, the system tracks performance indicators such as accuracy, stability, and response time to evaluate the overall state of the system. The accuracy indicator is the RMSE of the prediction error, which should be less than 0.05°; the stability indicator is the standard deviation of the error, which should be less than 0.03°; the response time indicator is the time delay from detecting the error to executing compensation, which should be less than 100ms. These indicators collectively reflect the overall performance of the system, and any abnormality in any indicator may indicate a problem with the system.

[0215] The system continuously monitors the health status of the sensors. Health status evaluation is based on signal quality indicators, including signal-to-noise ratio (SNR), effective sampling rate, and data integrity, etc. Specifically, SNR is calculated by the ratio of signal power to noise power:

[0216] ,

[0217] where: SNR (dB), signal power, noise power. Under normal working conditions, the SNR of Hall sensors and photoelectric sensors should be greater than 20dB, and less than 15dB indicates that the signal quality has decreased, and the sensor may need to be checked. The effective sampling rate is defined as the ratio of the number of effective samples actually obtained to the number of theoretical samples, and the normal value should be greater than 95%, and less than 90% indicates that the sensor may have a packet loss problem. Data integrity is evaluated by checking the continuity and consistency of the data, and the normal value should be greater than 98%, and less than 95% indicates that the data may be damaged.

[0218] When a sensor failure is detected, the system will automatically isolate the faulty sensor and switch to the backup sensor. In a certain precision reducer application, the system is configured with two sets of Hall sensors and two sets of photoelectric sensors, forming a redundant design. When one of the Hall sensors detects that its SNR has dropped to 12dB, the system automatically switches to the backup sensor, ensuring the continuity and reliability of the measurement. If there is no backup sensor, the system will adjust the algorithm parameters to increase the tolerance to noise, such as increasing the filtering strength, sacrificing some response speed for higher anti-interference ability.

[0219] For computational anomalies, the system can automatically identify and start the recovery process. Computational anomalies include model prediction timeout (normal should be less than 50ms), prediction result anomalies (such as NaN values or values outside the ±1° range), etc. When a computational anomaly is detected, the system will attempt to reinitialize the model, and if the problem persists, it will fall back to a backup method such as a simplified model or a lookup table method to ensure that the system can continue to operate. In a certain precision reducer failure test, a NaN value was artificially injected to cause a model calculation anomaly, and the system successfully detected the anomaly and fell back to the lookup table method, maintaining basic functionality until the model returned to normal.

[0220] For compensation effect anomalies, the system analyzes the causes and adjusts the compensation strategy. The compensation effect anomaly judgment standard is that the error has not decreased or has even increased after 5 consecutive compensations. In a certain precision reducer application, when the system detects compensation effect anomalies, it finds that the excessive compensation is caused by the actuator response delay, and automatically adjusts the compensation step from the original 15 rpm / ° to 10 rpm / °, solving the compensation oscillation problem.

[0221] To deal with various abnormal situations, the system implements a hierarchical recovery strategy. Mild abnormalities (such as single prediction failure) use a retry strategy; moderate abnormalities (such as temporary sensor failure) use a backup scheme; and serious abnormalities (such as multiple key component failures) trigger a safety shutdown program. This hierarchical strategy can maximize system operation while ensuring safety. In a long-term test of a certain precision reducer, the system successfully coped with various abnormal situations, including sensor interference, insufficient computing resources, and environmental interference, maintaining stable operation and verifying the effectiveness of the abnormal handling mechanism.

[0222] To optimize the overall performance of the system, the system implements a system coordination optimization mechanism. As shown in Figure 10 , this mechanism can coordinate the work of each module, optimize resource allocation, and improve the overall efficiency of the system.

[0223] The system coordinates the data acquisition and processing modules to match the sampling rate and processing capacity. In actual operation, the sampling rate is dynamically adjusted according to the current processor load to ensure that data processing does not become a system bottleneck. The system processor load rate is defined as the ratio of actual CPU time used to total available CPU time:

[0224] ,

[0225] where: is the processor load rate (%), is the actual CPU time used (ms), is the total available CPU time (ms). When , the system will reduce the sampling rate, for example from 10 kHz to 8 kHz; when , the system will increase the sampling rate, for example from 8 kHz to 10 kHz, which ensures data quality and avoids system overload.

[0226] The system adjusts the prediction accuracy and compensation granularity to ensure the optimal overall performance. The prediction accuracy is usually proportional to the computational complexity, and the compensation granularity is also proportional to the actuator burden. In high-precision applications, the system uses a complete seven-layer LSTM model to provide the highest prediction accuracy, and the compensation granularity is set to 1 rpm; in high-response applications, the system uses a simplified three-layer model, sacrificing some accuracy for faster response speed, and the compensation granularity is increased to 2 rpm, reducing the number of adjustments. Taking a certain precision reducer as an example, when used in high-precision machining equipment, the system automatically selects the high-precision mode; when used in high-speed conveying equipment, the system automatically selects the high-response mode, achieving optimal adaptation to different application scenarios.

[0227] The system balances the feedback speed and learning frequency to improve the system responsiveness. The feedback speed is defined as the number of feedback controls per second, usually taking 10 Hz; the learning frequency is defined as the number of model updates per hour, usually taking 0.1-1 times. When the working condition is stable, the system will reduce to 0.1 times / hour, reducing the computational burden; when the working condition changes, the system will increase to 1 time / hour, ensuring that the model adapts to new situations in a timely manner. In a certain precision reducer application, when the rate of change of the working condition (such as the rate of change of the speed) is detected to exceed 20% of the rated value per hour, the system automatically increases the learning frequency, maintaining the adaptability of the model.

[0228] The system continuously analyzes performance bottlenecks and implements targeted optimization. Performance bottlenecks can occur in data acquisition, signal processing, model calculation, or compensation execution, among other aspects. The system identifies the bottleneck by monitoring the response time of each link:

[0229] ,

[0230] where: is the total response time (ms), is the data acquisition time (ms), is the signal processing time (ms), is the model prediction time (ms), is the compensation execution time (ms). When the time proportion of a certain link exceeds 50%, it is considered a performance bottleneck. For example, in a certain precision reducer application, it was found that the model prediction time accounted for 65% of the total response time, and the system automatically optimized the model by pre-processing some calculations as lookup tables, reducing the prediction time by 40% and significantly improving the system response speed.

[0231] Based on operational requirements, the system intelligently switches between high-precision mode, high-response mode, and low-power mode. High-precision mode uses the complete model and the highest sampling rate to provide the highest accuracy, but also consumes the most power and resources. High-response mode uses a simplified model and a medium sampling rate to provide the fastest response. Low-power mode uses the simplest model and the lowest sampling rate, significantly reducing energy consumption. Mode switching is automatic based on application scenarios and resource conditions; for example, it automatically switches to low-power mode when powered by battery and automatically switches to high-precision mode during high-precision processing. In a mobile robot application, the system automatically switches between different modes based on battery level, maximizing battery life while ensuring basic functionality.

[0232] The system dynamically allocates computing resources to improve overall efficiency. Computing resources include processor time, memory space, and communication bandwidth. The system dynamically adjusts resource allocation based on the importance of each module and its current load. Resource allocation ratio The calculation is as follows:

[0233] ,

[0234] in: For the first Resource allocation ratio (%) for each module Module weights represent importance. For module load, This represents the total number of modules. For example, during the model training phase, the weight of the deep learning module is set to 0.7, while other modules share 0.3; during normal operation, the weight of the real-time compensation module is set to 0.6, while other modules share 0.4. In a precision reducer application, the system automatically adjusts resource allocation according to the working phase, achieving optimal resource utilization and improving system efficiency.

[0235] Furthermore, the system enhances reliability through redundancy design. Key components employ dual or triple redundancy, such as multiple sensors and multiple processing units. When one component fails, other components can take over, ensuring continuous system operation. In a precision reducer for a certain aerospace application, the system employs triple sensor redundancy and dual processor redundancy. Even if one sensor and one processor fail simultaneously, it can still maintain normal operation, meeting high reliability requirements.

[0236] Through the above collaborative optimization mechanisms, the system can automatically adjust its working mode and resource allocation according to the actual situation, and can achieve the best performance in different application scenarios, significantly improving the system's adaptability and practicality.

[0237] The above embodiments only express the specific implementation of the present application, which is described in more detail and in more detail, but cannot be understood as a limitation on the scope of the patent of the present application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of protection of the present application.

Claims

1. A method for dynamic compensation of gear coding error based on multi-sensor phase difference fusion, characterized in that, include: Collect data from multiple sensors, including Hall sensors and photoelectric sensors mounted on the gear system; Calculate the phase difference data, including: Phase information is extracted based on the raw data collected by the Hall sensor and the photoelectric sensor; Based on the phase information, the phase difference between the Hall sensor and the photoelectric sensor is calculated; Constructing the topological error space includes: Based on the phase difference data and operating parameters, a three-dimensional topological space is established, including gear speed, system temperature and load torque. Map the gear error distribution in the three-dimensional topological space; Training deep learning models includes: A training dataset is constructed based on the phase difference data and the operating condition parameters; A gear error prediction model was trained using a long short-term memory network. Perform adaptive compensation, including: A compensation strategy is generated based on the error value predicted by the deep learning model. Adjust the gear drive parameters according to the compensation strategy; Implement closed-loop feedback control, including: Monitor the compensated gear error; The deep learning model is updated based on the monitoring results.

2. The method according to claim 1, characterized in that, The collection of multi-sensor data specifically includes: The Hall sensor is installed at the gear drive end to measure the phase angle of the drive gear rotor; The photoelectric sensor is installed at the driven end of the gear to measure the phase angle of the driven gear rotor; With the gear phase line as the axis, the Hall sensor and the photoelectric sensor are each offset by 45° relative to the axis. A high-precision timestamp mechanism is used to synchronously acquire data from the Hall sensor and the photoelectric sensor; The sensor sampling frequency is adaptively adjusted based on the gear speed.

3. The method according to claim 1, characterized in that, The calculated phase difference data specifically includes: Wavelet threshold noise reduction processing is performed on the raw data collected by the Hall sensor and the photoelectric sensor; An adaptive window moving average method is used to smooth the signal; Identifying phase edges based on signal change rate; Phase angle is calculated using edge time; Calculate the difference between the phase angles of the Hall sensor and the photoelectric sensor; The difference is normalized to form standardized phase difference data.

4. The method according to claim 1, characterized in that, The construction of the topology error space specifically includes: Establish a coordinate system for operating parameters in three-dimensional space, where the X-axis represents gear speed, the Y-axis represents system temperature, and the Z-axis represents load torque; The collected sets of working condition data are normalized and mapped onto the coordinate system. Calculate the relationship between phase difference and error for each operating point; For sparse operating conditions, a topological interpolation algorithm is used to supplement data points; Construct a continuously differentiable error manifold to ensure the smoothness of error changes between adjacent operating points.

5. The method according to claim 1, characterized in that, The training of the deep learning model specifically includes: Design a long short-term memory network with seven hidden layers, each containing 64 neurons; The input layer receives phase difference characteristics and operating parameters, and the output layer outputs the error prediction value. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio. Introducing residual connectivity and attention mechanisms to improve model performance; If the validation loss fails to improve after five consecutive iterations, the training is terminated using an early stopping strategy. The model performance is evaluated using a test set to ensure the accuracy of error prediction.

6. The method according to claim 1, characterized in that, The execution of adaptive compensation specifically includes: Acquire gear speed, system temperature, and load torque data under the current operating conditions; The data is input into the deep learning model to obtain the error prediction value; An error probability distribution model is constructed based on the predicted error values; Determine the optimal compensation amount based on the error probability distribution; When the error is greater than 0.1°, increase the drive speed; When the error is less than 0.05°, maintain the current compensation state; A fault alarm is triggered when the error continues to be greater than 0.1° and cannot be compensated by speed regulation.

7. The method according to claim 1, characterized in that, The implementation of closed-loop feedback control specifically includes: Real-time monitoring of gear operating status after compensation is performed; Calculate the difference between the actual error and the predicted error; When the difference is greater than 0.16°, the model self-calibration mechanism is triggered; Collect compensation effect data for incremental model learning; Optimize compensation strategies based on historical compensation data; Build a compensation knowledge base to record the optimal compensation parameters under different working conditions.

8. The method according to claim 1, characterized in that, The training of the deep learning model also includes a transfer learning mechanism: Use a pre-trained model as the initial model to retain the ability to extract low-level features; Fine-tune the model parameters for the new operating conditions; A partial parameter freezing strategy is adopted to preserve valid knowledge; Implement dynamic updates to the sample bank and balance the old and new samples; When model performance degrades, an adaptive update mechanism is triggered. Continuously optimize model parameters to improve prediction accuracy.

9. The method according to claim 1, characterized in that, The method also includes system status monitoring and anomaly handling: Real-time monitoring of phase difference, prediction error, and compensation amount; Tracking system performance indicators such as accuracy, stability, and response time; Monitor the health status of sensors; When a sensor malfunction is detected, the faulty sensor is automatically isolated and switched to a backup sensor. Identify computational anomalies and initiate recovery procedures; For cases where the compensation effect is abnormal, analyze the reasons and adjust the compensation strategy accordingly; Implement a tiered recovery strategy to ensure reliable system operation.

10. The method according to claim 1, characterized in that, The method also includes a system collaborative optimization mechanism: Coordinate the data acquisition and processing modules to ensure that the sampling rate matches the processing capacity; Adjust the prediction accuracy and compensation granularity to ensure optimal overall system performance; Balance feedback speed and learning frequency to improve system responsiveness; Analyze system performance bottlenecks and implement targeted optimizations; It can intelligently switch between high-precision mode, high-response mode and low-power mode according to work requirements; Dynamically allocate computing resources to improve overall system efficiency; Improve system reliability through redundant design.

Images