A high-temperature robot dog driving correction control method

Abstract

The application discloses a high-temperature robot dog driving correction control method, and belongs to the technical field of robot motion control and motor driving. Specifically includes the following steps: S1, historical data acquisition and preprocessing: collecting multi-source historical data of the robot dog and preprocessing to obtain a standardized data set; S2, deep feature extraction: based on the standardized data set, first-level features, second-level features and third-level features are extracted in sequence, and are combined into deep features; S3, prediction model training and optimization: training a prediction model based on the deep features, and deploying after optimization. A three-level completely vertical dependence feature extraction architecture of "intrinsic representation-cause decoupling-evolution modeling" is constructed, the second level is only based on the output of the first level, the third level is only based on the output of the second level, and there is no cross-level data input, so that the defects of the existing scheme, such as feature extraction level confusion and unclear dependence relationship, are completely solved, and the rigor and scalability of the architecture are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of robot motion control and motor drive technology, and in particular to a method for correcting and controlling the drive of a robot dog under high temperature conditions. Background Technology

[0002] In heavy industrial settings such as metallurgy and chemical engineering, quadruped robot dogs are widely used for inspection and operation in high-temperature areas due to their excellent obstacle-crossing and terrain adaptability. However, ambient temperatures consistently above 80°C can cause thermal expansion, magnet demagnetization, or electronic component drift in the motor's built-in encoder (such as a magnetic encoder or photoelectric encoder), resulting in inconsistent zero-point offsets (i.e., different encoder drift amounts for different legs). This drift disrupts the originally precise multi-joint coordination. Taking the common trot gait as an example, it relies on the left front leg and right hind leg maintaining a strict 180° phase difference. If, due to high temperatures, the left front leg encoder drifts +3° in the positive direction, while the right hind leg drifts -2°, the actual phase difference becomes 185°, causing the two legs to exert force asynchronously in the support phase, leading to body vibration, increased energy consumption, and in severe cases, mechanical resonance or falls.

[0003] In the prior art, the solutions to the above-mentioned synchronization problem mainly include: Hardware redundancy: Dual encoders or high-precision absolute encoders are used, but the cost is high and high-temperature drift cannot be completely avoided. IMU-assisted correction: The aircraft detects attitude abnormalities through the in-flight IMU and infers the asynchrony between the legs. However, the IMU also exhibits zero-bias drift at high temperatures and has a lag in response. Periodic calibration: Manual calibration is performed after shutdown, which cannot meet the needs of continuous operation.

[0004] Furthermore, existing correction schemes for encoder thermal drift are mostly based on fixed mathematical models or simple statistical feature extraction methods, which have three major drawbacks: First, the feature extraction does not embed the physical priors of motor drive and encoder thermal drift, and is mostly based on general time-domain statistical features, which cannot accurately characterize the core causes of drift. Second, it cannot distinguish between "synchronization error caused by encoder thermal drift" and "non-drift error caused by load fluctuation, terrain disturbance, and gait adjustment," which is prone to miscorrection. Third, it does not consider the hysteresis nonlinear characteristics of encoder drift such as thermal creep and hysteresis loop at high temperatures, and cannot capture the irreversible correlation between drift and temperature history path, resulting in severely insufficient drift estimation accuracy and poor generalization in rapid temperature change scenarios. At the same time, the few correction schemes based on machine learning lack a strict vertical progressive architecture for feature extraction, have cross-level data dependencies, have loose hierarchical logic, and have high model complexity, making them unsuitable for real-time deployment on embedded controllers.

[0005] Therefore, there is an urgent need for a pure software, peripheral-free, real-time drive correction control method with a strict hierarchical feature extraction architecture, strong physical interpretability, ability to decouple error causes, and adaptability to nonlinear drift characteristics, so as to improve the motion robustness of robot dogs in high-temperature environments at low cost. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a method for correcting and controlling the drive of a robot dog under high temperature conditions; it can solve the problems of multi-motor synchronization phase inaccuracy and poor gait stability caused by encoder thermal drift of the robot dog motor under high temperature conditions in the prior art.

[0007] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, a method for drive correction control of a robot dog under high temperature, specifically including the following steps: S1. Historical Data Acquisition and Preprocessing: Collect multi-source historical data of the robot dog and preprocess it to obtain a standardized dataset; S2. Deep Feature Extraction: Based on the standardized dataset, first-level features, second-level features, and third-level features are extracted sequentially and merged into deep features; S3. Predictive Model Training and Optimization: Train the predictive model with deep features, optimize it, and then deploy it. S4. Pre-operation room temperature calibration: The symmetrical leg walks at a low speed, and records the synthetic current amplitude and rotor angle under the same synchronous phase to establish an ideal symmetrical reference and correct the trigger threshold. S5. Real-time data acquisition: During high-temperature operation, the main motion controller acquires the rotor angle, three-phase current, dq axis current, encoder and winding temperature, and gait phase synchronization stamp of each joint motor driver in real time. S6. Error observation window trigger judgment: Select the symmetrical leg pair and determine whether they are simultaneously in the middle of the support phase; if so, enter the error observation window; otherwise, maintain normal synchronous control. S7. Initial judgment of synchronization error: Calculate the amplitude of the three-phase combined current and the degree of asymmetry of the symmetrical leg pair. If it exceeds the correction trigger threshold, it is determined that there is a synchronization phase error and the dual-path drift estimation is initiated. S8. Dual-path drift estimation: One path inputs real-time operating data into a pre-trained drift prediction model and outputs the encoder zero-point drift prediction value; the other path calculates the relative phase deviation of the symmetrical leg pair and estimates the drift observation value online using the least squares method based on multi-cycle deviation data. S9. Drift fusion and correction compensation: The predicted value and the observed value are adaptively weighted and fused to obtain the final zero-point drift, and the position correction value is generated as the motor position loop feedback input to realize closed-loop compensation. S10, Stable control of correction parameters: Exponential smoothing filter is applied to the position correction amount. When the current asymmetry is lower than the threshold for multiple consecutive cycles, the compensation parameters are frozen to prevent overcorrection.

[0008] Furthermore, in step S1, the historical operating data includes: three-phase current data, rotor electrical angle data, dq axis current data, encoder and winding temperature data, gait phase synchronization stamp data, gait condition data, and encoder zero-point drift amount annotation data under the corresponding operating condition obtained through high-precision calibration equipment; preprocessing includes: outlier removal, timing synchronization alignment, dimensional normalization, and hierarchical division of the dataset; Outlier removal: Outliers in current, angle, and temperature data are removed using the 3σ criterion; missing values ​​in the samples are filled using linear interpolation. Timing synchronization alignment: Based on the EtherCAT synchronization clock, all data are aligned to the same sampling interval to ensure that the timestamps of the data of each joint are completely consistent at the same time. Dimensional normalization: z-score normalization is used to normalize all feature dimensions to eliminate the influence of dimensions; Dataset hierarchical partitioning: The dataset is partitioned according to temperature gradient, and then divided into training set, validation set, and test set according to a preset ratio.

[0009] Furthermore, in step S2, the extraction process for the first-level features is specifically as follows: S211. Using a single complete gait cycle as the smallest analysis unit, based on the torque generation mechanism of field-oriented control (FOC), extract time-domain statistical features with physical constraints from the d-axis excitation current, q-axis torque current, composite current amplitude, and rotor angle deviation sequence, including torque domain features, excitation domain features, and angle domain features, and output time-domain intrinsic features. S212. Using the gait period as the fundamental frequency, perform fast Fourier transform and wavelet packet decomposition on the synthetic current and rotor angle deviation sequence within a single period to extract frequency domain features synchronized with the gait, including: fundamental frequency and harmonic features, wavelet packet energy entropy features, and output frequency domain eigenvalues. S213. Based on the thermal expansion and magnet demagnetization mechanism of encoder thermal drift, construct thermal-electric coupling characteristics, including: thermal steady-state characteristics, thermal-electric coupling correlation characteristics, and output thermal-electric coupling intrinsic characteristics; S214. The time-domain intrinsic features, frequency-domain intrinsic features, and thermo-electric coupling intrinsic features are standardized to obtain the first-level features.

[0010] Furthermore, in step S2, the extraction process for the second-level features is specifically as follows: S221. Based on the ideal symmetry relationship of the target gait, perform gait phase alignment feature mapping on the eigenvalue sets of two single joints of the symmetrical leg pair, construct cross-joint coupling features, including: symmetry feature deviation matrix, coupling correlation matrix, cross-cycle temporal consistency features, and output spatiotemporal coupling features. S222. Based on the essential difference between thermal drift error and non-drift error, through orthogonal projection decomposition, the "synchronization error characteristics caused by encoder thermal drift" and "error characteristics caused by non-drift factors" are accurately decoupled, and the drift-specific features and interference features after decoupling are extracted and the decoupled features are output. S2221. Based on the thermal-electric coupling characteristics in the first-level feature set, construct a drift-related projection basis; based on the load fluctuation characteristics in the torque domain features and the transient impact characteristics in the frequency domain features, construct a non-drift projection basis. S2222. Project the spatiotemporal coupling feature set orthogonally onto the drift-related projection base and the non-drift projection base respectively to obtain the drift-specific projection feature set and the non-drift interference projection feature set. S2223. Extract the projection energy ratio, principal component eigenvalues, temporal accumulation of projection features, and correlation with temperature features from the drift-specific projection feature set to form decoupled drift-specific features; extract the projection energy ratio and transient impact amplitude from the non-drift interference projection feature set to form interference degree features. S223. Combine the spatiotemporal coupling features with the decoupling features to form the second-level features.

[0011] Furthermore, in step S2, the extraction process for the third-level features is specifically as follows: S231. To address the thermal creep hysteresis nonlinearity and hysteresis loop characteristics of encoder zero-point drift at high temperatures, a Preisach operator grid is constructed based on the time sequence of the second-level decoupled feature set. Principal component features and hysteresis loop features of the weight matrix are extracted, and hysteresis nonlinearity features are output. S232. Based on the continuous periodic time series features in the second-level feature set, construct drift evolution features with long-term memory, including: long-term and short-term trend features, temperature gradient mapping features, and cross-gait generalization features, and output drift evolution time series memory features. S233. Based on hysteresis and evolutionary features, nonlinear mapping is performed through Gaussian kernel principal component analysis to extract kernel principal component features and form third-level features.

[0012] Furthermore, in step S3, the drift prediction model adopts a lightweight temporal convolutional network (TCN) or a lightweight long short-term memory (LSTM) architecture. The model input is deep features, and the output is the zero-point drift of the corresponding joint encoder. After the model training is completed, INT8 quantization compression is performed to meet the real-time inference requirements of the embedded controller within the control cycle.

[0013] Furthermore, in step S6, the criteria for determining the mid-support phase are: the reading of the leg Z-axis force sensor is greater than the preset load threshold, and the linear velocity of the leg joint movement is close to zero; symmetrical leg pairs include diagonal leg pairs in trot gait, same-side leg pairs in pace gait, and front and rear leg pairs in bound gait.

[0014] Furthermore, in step S7, the formula for calculating the magnitude of the synthesized current is: in, For the first i The combined current amplitude of the joint motor 、 、 These are the three-phase current values ​​of the motor; The formula for calculating current asymmetry is: in, For current asymmetry, , The sum of the current amplitudes of the two joint motors in the symmetrical leg pair.

[0015] Furthermore, in step S8, the formula for calculating the relative phase deviation is: in, The relative phase deviation of the symmetrical leg pairs. , These are the measured rotor angles of the two motors aligned with the symmetrical legs; The model for the online estimation using the least squares method is as follows: in, The zero-point drift of each leg encoder. To observe noise.

[0016] Furthermore, in step S9, the weights of the adaptive weighted fusion are adaptively adjusted based on the interference degree features and joint temperature change rate output from the second-level feature extraction: when the interference degree features exceed a threshold or the temperature change rate is greater than a preset threshold, the weight of the predicted drift value is increased; when the interference degree features are lower than a threshold and the temperature change rate is less than or equal to a preset threshold, the weight of the observed drift value is increased; the calculation formula for the position correction amount is: in, Let be the position correction amount for the i-th motor. This represents the measured rotor angle of the motor. This represents the final zero-point drift of the motor.

[0017] Beneficial effects: A three-level feature extraction architecture with complete vertical dependence of "intrinsic representation - cause decoupling - evolutionary modeling" was constructed. The second level is based only on the output of the first level, and the third level is based only on the output of the second level, without any cross-level data input. This completely solves the defects of the existing schemes in feature extraction hierarchy disorder and unclear dependency relationship, and greatly improves the rigor and scalability of the architecture.

[0018] Each level of feature extraction incorporates the physical mechanisms of motor FOC drive, encoder thermal drift, and quadrupedal gait symmetry. It extracts intrinsic features with physical constraints from multiple domains, rather than general, undifferentiated statistical features. This approach fundamentally addresses the core causes of encoder drift at high temperatures, resulting in highly robust and targeted features that avoid the black-box defects of data-driven models.

[0019] By using orthogonal projection decomposition, the "synchronization error caused by encoder thermal drift" and the "non-drift error caused by load fluctuation, terrain disturbance and gait adjustment" are accurately distinguished. This solves the industry pain point that existing solutions cannot identify interference and are prone to miscalibration, and greatly improves the reliability of calibration and the adaptability to industrial scenarios.

[0020] By introducing the Preisach hysteresis model into encoder thermal drift feature extraction, the thermal creep, hysteresis loop, and irreversible characteristics related to temperature history path of magnetic encoder drift under high temperature are accurately captured. This completely solves the problem of insufficient prediction accuracy of existing solutions for rapid temperature change and nonlinear drift scenarios. The average absolute error of drift prediction is ≤0.3°, and the generalization is greatly improved.

[0021] Utilizing feedback data from the motor driver itself, it can be directly adapted to the embedded control system of existing quadruped robot dogs without the need for additional sensors or hardware modifications. The lightweight model and quantization compression design can complete real-time inference and correction within a 1kHz control cycle without affecting the real-time performance of the original gait control, supporting the robot dog to operate continuously and uninterruptedly in high-temperature industrial environments.

[0022] A dual closed-loop correction architecture of "model prediction + online observation" is constructed. It achieves accurate closed-loop correction in steady-state scenarios through online observation, and enables early prediction of drift in scenarios with rapid temperature changes and strong interference through deep feature models. It is perfectly adapted to the full-condition operation requirements of complex high-temperature industrial scenarios such as metallurgy and chemical industry. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the structure of a multi-motor synchronization error self-correction system for a robot dog under high temperature conditions according to the present invention; Figure 2 This invention provides a timing interaction diagram for multi-motor synchronization error correction in a robot dog under high temperature conditions. Figure 3 This is a state transition diagram for synchronous correction of a robot dog under high temperature conditions according to the present invention. Detailed Implementation

[0024] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 S1. Historical Data Acquisition and Preprocessing Data Acquisition: A high and low temperature environment test chamber was built to simulate multiple temperature gradients. Multi-source data of 1000 complete gait cycles were collected under each temperature gradient within the trot gait speed range of 0.3~1.2m / s. This included: three-phase current of each joint motor, rotor electrical angle, dq axis current, encoder and winding temperature, gait phase synchronization stamp, and gait operating condition data. Simultaneously, a high-precision photoelectric angle calibrator with an accuracy of ±0.05° was used to collect the labeled data of encoder zero-point drift under the corresponding operating conditions.

[0026] Data preprocessing: Outlier removal: Outliers in current, angle, and temperature data are removed using the 3σ criterion, and missing values ​​are filled in using linear interpolation, improving data integrity to 99.8%. Timing synchronization alignment: Based on the EtherCAT synchronization clock, all data are aligned to a 1ms sampling interval to ensure that the timestamp deviation of data from each joint at the same time is ≤10μs; Dimensional normalization: z-score normalization is used to normalize all feature dimensions to eliminate the influence of dimensions; Dataset stratification: The dataset is stratified according to the temperature gradient, and divided into training set, validation set and test set in a ratio of 7:2:1 to ensure that the temperature gradient distribution of each dataset is consistent.

[0027] S2 Deep Feature Extraction Based on a standardized dataset, three levels of vertically dependent features are extracted sequentially without cross-level data input. These features are then merged to obtain deep features. The specific process is as follows: First-level feature extraction: using a complete gait cycle of 0.5 seconds as the smallest unit of analysis. Time-domain intrinsic feature extraction: Based on the FOC torque generation mechanism, extract torque domain features (mean, peak value, standard deviation, and fluctuation coefficient of q-axis current), excitation domain features (steady-state deviation and harmonic content of d-axis current), and angle domain features (range and cumulative deviation of rotor angle deviation). Frequency domain intrinsic feature extraction: FFT is performed on the single-cycle synthetic current and rotor angle deviation sequence using the 2Hz gait fundamental frequency to extract the fundamental frequency and the amplitude of the 2nd / 3rd harmonics; db4 wavelet is used to perform 3-level wavelet packet decomposition to extract the wavelet packet energy entropy of each frequency band; Thermo-electric coupling intrinsic feature extraction: Based on the encoder thermal expansion and magnet demagnetization mechanism, extract thermal steady-state features (encoder temperature mean, temperature change rate) and thermo-electric coupling correlation features (Pearson correlation coefficient between winding temperature and d-axis current deviation, cross-correlation coefficient between encoder temperature and rotor angle deviation). The above three types of intrinsic features are standardized to obtain the first-level features.

[0028] Second-level feature extraction: Spatiotemporal coupling feature construction: For the diagonally symmetrical leg pairs (left front-right rear leg, right front-left rear leg) of the trot gait, the gait phase of the first-level feature set of the two single joints in the leg pair is aligned from 0° to 360°. The symmetrical feature deviation matrix, coupling correlation matrix, and temporal consistency features across 3 cycles are constructed, and the spatiotemporal coupling features are output. Drift and interference decoupling feature extraction: A drift-related projection basis is constructed based on the first-level thermo-electric coupling features, and a non-drift projection basis is constructed based on the torque domain load fluctuation features and the frequency domain transient impact features. The spatiotemporal coupling feature set is orthogonally projected onto the two projection basis sets to obtain the drift-specific projection feature set and the non-drift interference projection feature set. The drift-specific projection feature set is extracted by extracting the projection energy ratio, the eigenvalues ​​of the first three principal components, the time-series accumulation of the projection features, and the correlation with the temperature features to form the drift-specific features. The non-drift interference projection feature set is extracted by extracting the projection energy ratio and the transient impact amplitude to form the interference degree features. By merging the spatiotemporal coupling features and the decoupling features, the second-level features are obtained.

[0029] Third-level feature extraction: Performed only based on the output of the second-level features. Hysteresis nonlinear feature extraction: For the thermal creep and hysteresis loop characteristics of encoder drift at high temperature, a 10×10 Preisach operator grid is constructed based on the second-level decoupling feature time sequence to extract the first 5 principal component features of the weight matrix, the coercivity and remanence of the hysteresis loop, and output the hysteresis nonlinear features. Drift evolution temporal memory feature extraction: Based on the second-level continuous periodic temporal features, extract long-term and short-term trend features (moving average and first-order difference of drift amount over 5 consecutive periods), temperature gradient mapping features (mapping relationship between temperature change rate and drift change rate), and cross-gait generalization features (consistency coefficient of drift features under different walking speeds). Gaussian kernel principal component analysis (kernel function width σ=1.2) was used to perform nonlinear mapping on the above features, and the first 8 kernel principal component features were extracted to form the third-level features.

[0030] S3 Prediction Model Training and Optimization The drift prediction model adopts a lightweight TCN architecture with the following core parameters: kernel size 3, number of network layers 4, dilation coefficient [1, 2, 4, 8], and dropout rate 0.2. The model input is the three-level merged deep features, and the output is the zero-point drift amount of the corresponding joint encoder.

[0031] Model training: The loss function is mean squared error (MSE), the optimizer is Adam, the initial learning rate is 1e-4, the number of training epochs is 200, and an early stopping strategy with patience=10 is set to prevent overfitting. Model optimization and deployment: After training, INT8 quantization compression is performed, and the model size is compressed to 1 / 4 of the original model. The single-frame inference time is ≤0.5ms, which meets the real-time requirement of 1kHz control cycle. Finally, it is deployed to the robot dog's main controller.

[0032] S4 pre-operation room temperature calibration In a normal temperature environment in the workshop, the robot dog was controlled to walk stably at a low speed of 0.3 m / s with a trot gait. Data was continuously collected for 10 complete gait cycles. The amplitude of the three-phase combined current and the rotor angle of the diagonally symmetrical legs at the same joint were recorded under the same gait phase to establish an ideal symmetry reference and correct the trigger threshold. Ideal symmetrical reference: Under the trot gait, the ideal phase difference between the diagonal legs and the rotor angles of the same joint is 180°, and the ideal deviation of the combined current amplitude is ≤0.2A; Correction trigger threshold: current asymmetry ≥ 0.8A, or relative phase deviation ≥ 2°.

[0033] S5 Real-time Data Acquisition The robot dog enters the high-temperature inspection area and walks continuously at a trot gait of 0.8 m / s. The main motion controller collects the rotor angle, three-phase current, dq axis current, encoder and winding temperature of the 12 joint motor drivers in real time at a frequency of 1kHz via the EtherCAT bus. It also collects the leg six-dimensional force sensor data and gait phase synchronization stamp simultaneously. All data are aligned to the same timestamp with a time deviation of ≤10μs.

[0034] S6 Error Observation Window Trigger Judgment For the diagonally symmetrical leg pairs (left front-right rear, right front-left rear) of the trot gait, it is determined in real time whether the leg pairs are simultaneously in the middle of the support phase. The determination criteria are: the reading of the leg Z-axis force sensor is ≥50N (preset load threshold, corresponding to 1 / 4 of the body weight), and the linear velocity of the leg joint movement is ≤5mm / s.

[0035] If the symmetrical leg pairs simultaneously meet the above conditions, the error observation window will be entered; otherwise, normal gait synchronization control will be maintained and the correction process will not be initiated.

[0036] S7 Synchronization Error Preliminary Assessment Within the error observation window, the three-phase combined current amplitude and current asymmetry of the symmetrical leg motor with the same joint are calculated according to the following formula: Synthetic current amplitude ： Current asymmetry: In this embodiment, the measured amplitude of the synthesized current of the left foreleg hip joint motor was determined. The amplitude of the synthesized current of the motor in the right hind leg hip joint The current asymmetry was calculated. If the error exceeds the correction trigger threshold of 0.8A, a synchronization phase error is determined to exist, and dual-path drift estimation is initiated.

[0037] S8 Dual-Path Drift Estimation A dual-path parallel drift estimation method combining "model prediction + online observation" is adopted: Model prediction branch: The real-time collected working condition data, after preprocessing and feature extraction in the same manner as S1 and S2, is input into the pre-trained TCN drift prediction model, and the output is the predicted value of the zero-point drift of the left foreleg hip joint encoder. Predicted zero-point drift value of the right hind leg hip joint encoder ; Online observation branch: according to the formula Calculate the relative phase deviation of the symmetrical leg pair, and obtain the relative phase deviation. Based on the relative phase deviation data of five consecutive gait cycles, the observed value of the left foreleg drift was obtained through online estimation using the least squares method. Observed value of right hind leg drift Observation noise .

[0038] S9 Drift Fusion and Correction Compensation Adaptive weighted fusion: Obtain the interference feature output of the second-level feature extraction. Under this working condition, the interference feature value is 0.12, which is lower than the interference threshold of 0.3. At the same time, the joint temperature change rate is 0.5℃ / s, which is less than the preset threshold of 2℃ / s. Therefore, the weight of the observed value is adaptively increased, and the weight of the predicted value is set to 0.3 and the weight of the observed value is 0.7. Calculate the final zero-point drift: Closed-loop correction compensation: according to the formula Generate position correction amount, left foreleg hip joint position correction amount Correction amount of right hind leg hip joint position The position correction value is used as the feedback input of the motor FOC control position loop to achieve real-time closed-loop compensation.

[0039] S10 correction parameter stability control The position correction amount is subjected to exponential smoothing filtering with a smoothing coefficient α=0.3 to suppress the body vibration caused by sudden changes in the correction amount. After continuous monitoring for 8 complete gait cycles, the current asymmetry of the symmetrical leg pair decreases from 1.1A to 0.15A, which remains below the threshold of 0.8A. The relative phase deviation decreases from 1.9° to 0.25°, which is below the threshold of 2°. At this point, the compensation parameters are frozen and the correction iteration is stopped to prevent overcorrection. When the subsequent operating conditions change and the current asymmetry exceeds the threshold again, the correction process is restarted.

[0040] Example 2 S1 Historical Data Acquisition and Preprocessing Data acquisition: A high and low temperature alternating test chamber was built to simulate multiple temperature gradients, multiple load gradients, and rapid temperature change conditions of 1℃ / s and 5℃ / s. Multi-source data of 2000 complete gait cycles within the pace gait speed range of 0.2~0.8m / s under each condition were collected. Simultaneously, a high-precision photoelectric angle calibrator was used to collect encoder zero-point drift annotation data.

[0041] Data preprocessing: Outlier removal and completion: Outliers are removed using the 3σ criterion, and missing values ​​are completed using linear interpolation; Timing synchronization alignment: Based on the EtherCAT synchronization clock, all data is aligned to a 0.5ms sampling interval, with a timestamp deviation ≤5μs; Dimensional normalization: z-score standardization is used to process all feature dimensions; Dataset stratification: The dataset is stratified according to both temperature gradient and load gradient, and divided into training set, validation set and test set in a 7:2:1 ratio to ensure uniform distribution of working conditions in the dataset.

[0042] S2 Deep Feature Extraction Based on a standardized dataset, three levels of vertical dependency features are extracted sequentially, and the feature extraction parameters are optimized to adapt to heavy-load and rapid temperature change conditions. First-level feature extraction: Taking a complete gait cycle of 0.6s as the smallest analysis unit, torque fluctuation features and load-current correlation features for heavy load adaptation are added to the time domain features; low-frequency load disturbance features are added to the frequency domain features; and temperature-drift dynamic correlation features under rapid temperature change are added to the thermal-electric coupling features. After standardization, the first-level features are obtained.

[0043] Second-level feature extraction: Performed only based on the output of the first-level features. Spatiotemporal coupling feature construction: For symmetrical leg pairs on the same side of the pace gait (left front-left rear leg, right front-right rear leg), perform gait phase alignment on the first-level feature set of single joints of the leg pair, construct the symmetrical feature deviation matrix, load-torque coupling correlation matrix, and temporal consistency features across 5 cycles under heavy load conditions, and output the spatiotemporal coupling features. Drift and disturbance decoupling feature extraction: A drift-related projection basis is constructed based on the first-level thermo-electric coupling features, and a non-drift projection basis is constructed based on the torque domain heavy load fluctuation features, frequency domain transient impact features, and terrain disturbance features; the drift features and disturbance features are accurately decoupled through orthogonal projection decomposition, and drift-specific features and disturbance degree features are extracted. By merging the spatiotemporal coupling features and the decoupling features, the second-level features are obtained.

[0044] Third-level feature extraction: Performed only based on the output of the second-level features. Hysteresis nonlinearity feature extraction: To address the strong hysteresis nonlinearity of encoder drift under rapid temperature changes, a 15×15 Preisach operator grid is constructed to extract the principal component features of the weight matrix and the full-range features of the hysteresis loop, and output the hysteresis nonlinearity features. Drift evolution temporal memory feature extraction: extracting long-term and short-term trend features under rapid temperature changes, temperature gradient-drift rate mapping features, and cross-gait generalization features under heavy load; Gaussian kernel principal component analysis is used to perform nonlinear mapping, extract kernel principal component features to form third-level features, and merge the third-level features to obtain deep features.

[0045] S3 Prediction Model Training and Optimization The drift prediction model adopts a lightweight LSTM architecture to adapt to the computing power limitations of heterogeneous controllers. The core parameters are: 2-layer LSTM network, 32 hidden layer units, and dropout rate of 0.1. The input is deep features, and the output is the zero-point drift amount of the corresponding joint encoder.

[0046] Model training: The loss function used is MSE, the optimizer is Adam, the initial learning rate is 5e-5, the number of training epochs is 300, and the early stopping strategy is patience=15. Model optimization and deployment: After training, INT8 quantization compression is performed. After quantization, the inference time per frame is ≤0.3ms, which meets the real-time requirement of 2kHz control cycle and is deployed to heterogeneous controllers.

[0047] S4 pre-operation room temperature calibration Under normal temperature (25℃) and rated load of 15kg, the robot dog was controlled to walk stably at a low speed of 0.2m / s with a pace gait, and data from 15 complete gait cycles were continuously collected to establish an ideal symmetrical benchmark and correct the trigger threshold. Ideal symmetry benchmark: Under pace gait, the ideal phase difference of the same-side leg with respect to the same joint rotor angle is 0°, and the ideal deviation of the synthesized current amplitude is ≤0.3A; Correction trigger threshold: current asymmetry ≥1.0A, or relative phase deviation ≥2.5°.

[0048] S5 Real-time Data Acquisition The robot dog enters a high-temperature working area, carrying a 15kg detection device and walks continuously at a pace of 0.5m / s. The main controller collects the rotor angle, three-phase current, dq axis current, encoder and winding temperature of each joint motor in real time at a frequency of 2kHz. It also collects leg force sensor data, gait phase synchronization stamp, and real-time temperature change rate data simultaneously. All data are strictly time-aligned.

[0049] S6 Error Observation Window Trigger Judgment For symmetrical leg pairs on the same side of the pace gait (left front-left back, right front-right back), it is determined in real time whether the leg pairs are simultaneously in the middle of the support phase. The determination criteria are: the Z-axis force sensor reading of the leg is ≥80N (1 / 4 of the total weight of the machine body and load), and the linear velocity of the leg joint movement is ≤3mm / s.

[0050] If both symmetrical legs meet the conditions simultaneously, the error observation window is entered; otherwise, normal synchronization control is maintained.

[0051] S7 Synchronization Error Preliminary Assessment Within the error observation window, the combined current amplitude and current asymmetry of the motors with the same joint for the symmetrical leg are calculated, and the combined current amplitude of the motor of the right foreleg thigh joint is measured. The amplitude of the synthesized current of the motor at the right hind thigh joint The current asymmetry was calculated. If the error exceeds the correction trigger threshold of 1.0A, a synchronization phase error is determined to exist, and dual-path drift estimation is initiated.

[0052] S8 Dual-Path Drift Estimation Model prediction branch: After preprocessing and feature extraction, real-time operating data is input into a pre-trained LSTM drift prediction model, which outputs the predicted zero-point drift value of the right foreleg thigh joint encoder. Predicted zero-point drift value of the encoder at the right hind leg thigh joint ; Online observation branch: according to the formula Calculate the relative phase deviation of the same-side leg pair, and obtain the relative phase deviation. Based on the relative phase deviation data of eight consecutive gait cycles, the observed value of the right foreleg drift was obtained through online estimation using the least squares method. Observed value of right hind leg drift .

[0053] S9 Drift Fusion and Correction Compensation Adaptive weighted fusion: Under this operating condition, the joint temperature change rate is 4.2℃ / s, which exceeds the preset threshold of 2℃ / s. At the same time, the interference characteristic value is 0.38, which exceeds the interference threshold of 0.3. Therefore, the weight of the predicted value is adaptively increased, and the weight of the predicted value is set to 0.7 and the weight of the observed value is 0.3. Calculate the final zero-point drift: Closed-loop correction compensation: Position correction values ​​are generated according to the formula, including the right foreleg thigh joint position correction value. Correction amount of right hind leg thigh joint position The correction value is input into the motor position loop to achieve real-time closed-loop compensation.

[0054] S10 correction parameter stability control The position correction is subjected to exponential smoothing filtering with a smoothing coefficient α=0.2 to adapt to the dynamic adjustment requirements under rapid temperature change conditions. After continuous monitoring for 10 complete gait cycles, the current asymmetry of the symmetrical leg pair decreases from 1.7A to 0.22A, remaining below the threshold of 1.0A, and the relative phase deviation decreases from 5.6° to 0.32°, remaining below the threshold of 2.5°. At this point, the compensation parameters are frozen. At the same time, a temperature change rate trigger condition is set. When the temperature change rate exceeds 3℃ / s, the parameter freeze is automatically lifted, and the correction is updated in real time to avoid correction lag under rapid temperature change.

[0055] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for drive correction control of a robot dog under high temperature, characterized in that, Specifically, the following steps are included: S1. Historical Data Acquisition and Preprocessing: Collect multi-source historical data of the robot dog and preprocess it to obtain a standardized dataset; S2. Deep Feature Extraction: Based on the standardized dataset, first-level features, second-level features, and third-level features are extracted sequentially and merged into deep features; S3. Predictive Model Training and Optimization: Train the predictive model with deep features, optimize it, and then deploy it. S4. Pre-operation room temperature calibration: The symmetrical leg walks at a low speed, and records the synthetic current amplitude and rotor angle under the same synchronous phase to establish an ideal symmetrical reference and correct the trigger threshold. S5. Real-time data acquisition: During high-temperature operation, the main motion controller acquires the rotor angle, three-phase current, dq axis current, encoder and winding temperature, and gait phase synchronization stamp of each joint motor driver in real time. S6. Error observation window trigger judgment: Select the symmetrical leg pair and determine whether they are simultaneously in the middle of the support phase; if so, enter the error observation window; otherwise, maintain normal synchronous control. S7. Initial judgment of synchronization error: Calculate the amplitude of the three-phase combined current and the degree of asymmetry of the symmetrical leg pair. If it exceeds the correction trigger threshold, it is determined that there is a synchronization phase error and the dual-path drift estimation is initiated. S8. Dual-path drift estimation: One path inputs real-time operating data into a pre-trained drift prediction model and outputs the encoder zero-point drift prediction value; the other path calculates the relative phase deviation of the symmetrical leg pair and estimates the drift observation value online using the least squares method based on multi-cycle deviation data. S9. Drift fusion and correction compensation: The predicted value and the observed value are adaptively weighted and fused to obtain the final zero-point drift, and the position correction value is generated as the motor position loop feedback input to realize closed-loop compensation. S10, Stable control of correction parameters: Exponential smoothing filter is applied to the position correction amount. When the current asymmetry is lower than the threshold for multiple consecutive cycles, the compensation parameters are frozen to prevent overcorrection.

2. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S1, the historical operating data includes: three-phase current data, rotor electrical angle data, dq axis current data, encoder and winding temperature data, gait phase synchronization stamp data, gait condition data, and encoder zero-point drift annotation data under the corresponding operating conditions obtained through high-precision calibration equipment; preprocessing includes: outlier removal, timing synchronization alignment, dimensional normalization, and hierarchical division of the dataset; Outlier removal: Outliers in current, angle, and temperature data are removed using the 3σ criterion; missing values ​​in the samples are filled using linear interpolation. Timing synchronization alignment: Based on the EtherCAT synchronization clock, all data are aligned to the same sampling interval to ensure that the timestamps of the data of each joint are completely consistent at the same time. Dimensional normalization: z-score normalization is used to normalize all feature dimensions to eliminate the influence of dimensions; Dataset hierarchical partitioning: The dataset is partitioned according to temperature gradient, and then divided into training set, validation set, and test set according to a preset ratio.

3. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S2, the extraction process of the first-level features is as follows: S211. Using a single complete gait cycle as the smallest analysis unit, based on the torque generation mechanism of field-oriented control (FOC), extract time-domain statistical features with physical constraints from the d-axis excitation current, q-axis torque current, composite current amplitude, and rotor angle deviation sequence, including torque domain features, excitation domain features, and angle domain features, and output time-domain intrinsic features. S212. Using the gait period as the fundamental frequency, perform fast Fourier transform and wavelet packet decomposition on the synthetic current and rotor angle deviation sequence within a single period to extract frequency domain features synchronized with the gait, including: fundamental frequency and harmonic features, wavelet packet energy entropy features, and output frequency domain eigenvalues. S213. Based on the thermal expansion and magnet demagnetization mechanism of encoder thermal drift, construct thermal-electric coupling characteristics, including: thermal steady-state characteristics, thermal-electric coupling correlation characteristics, and output thermal-electric coupling intrinsic characteristics; S214. The time-domain intrinsic features, frequency-domain intrinsic features, and thermo-electric coupling intrinsic features are standardized to obtain the first-level features.

4. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S2, the extraction process for the second-level features is as follows: S221. Based on the ideal symmetry relationship of the target gait, perform gait phase alignment feature mapping on the eigenvalue sets of two single joints of the symmetrical leg pair, construct cross-joint coupling features, including: symmetry feature deviation matrix, coupling correlation matrix, cross-cycle temporal consistency features, and output spatiotemporal coupling features. S222. Based on the essential difference between thermal drift error and non-drift error, through orthogonal projection decomposition, the "synchronization error characteristics caused by encoder thermal drift" and "error characteristics caused by non-drift factors" are accurately decoupled, and the drift-specific features and interference features after decoupling are extracted and the decoupled features are output. S2221. Based on the thermal-electric coupling characteristics in the first-level feature set, construct a drift-related projection basis; based on the load fluctuation characteristics in the torque domain features and the transient impact characteristics in the frequency domain features, construct a non-drift projection basis. S2222. Project the spatiotemporal coupling feature set orthogonally onto the drift-related projection base and the non-drift projection base respectively to obtain the drift-specific projection feature set and the non-drift interference projection feature set. S2223. Extract the projection energy ratio, principal component eigenvalues, temporal accumulation of projection features, and correlation with temperature features from the drift-specific projection feature set to form decoupled drift-specific features; extract the projection energy ratio and transient impact amplitude from the non-drift interference projection feature set to form interference degree features. S223. Combine the spatiotemporal coupling features with the decoupling features to form the second-level features.

5. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S2, the extraction process for the third-level features is as follows: S231. To address the thermal creep hysteresis nonlinearity and hysteresis loop characteristics of encoder zero-point drift at high temperatures, a Preisach operator grid is constructed based on the time sequence of the second-level decoupled feature set. Principal component features and hysteresis loop features of the weight matrix are extracted, and hysteresis nonlinearity features are output. S232. Based on the continuous periodic time series features in the second-level feature set, construct drift evolution features with long-term memory, including: long-term and short-term trend features, temperature gradient mapping features, and cross-gait generalization features, and output drift evolution time series memory features. S233. Based on hysteresis and evolutionary features, nonlinear mapping is performed through Gaussian kernel principal component analysis to extract kernel principal component features and form third-level features.

6. The method for driving correction and control of a machine dog under high temperature as described in claim 1, characterized in that: In step S3, the drift prediction model adopts a lightweight temporal convolutional network (TCN) or a lightweight long short-term memory (LSTM) architecture. The model input is deep features, and the output is the zero-point drift amount of the corresponding joint encoder. After the model training is completed, INT8 quantization compression is performed to meet the real-time inference requirements of the embedded controller within the control cycle.

7. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S6, the criteria for determining the mid-support phase are: the reading of the leg Z-axis force sensor is greater than the preset load threshold, and the linear velocity of the leg joint movement is close to zero; symmetrical leg pairs include diagonal leg pairs in trot gait, same-side leg pairs in pace gait, and front and rear leg pairs in bound gait.

8. The method for driving correction and control of a robot dog under high temperature as described in claim 1, characterized in that: In step S7, the formula for calculating the magnitude of the synthesized current is: in, For the first i The combined current amplitude of the joint motor 、 、 These are the three-phase current values ​​of the motor; The formula for calculating current asymmetry is: in, For current asymmetry, , The sum of the current amplitudes of the two joint motors in the symmetrical leg pair.

9. The method for driving correction and control of a machine dog under high temperature as described in claim 1, characterized in that: In step S8, the formula for calculating the relative phase deviation is: in, The relative phase deviation of the symmetrical leg pair. , These are the measured rotor angles of the two motors aligned with the symmetrical legs; The model for online estimation using the least squares method is as follows: in, The zero-point drift of each leg encoder. To observe noise.

10. The method for driving correction and control of a robot dog under high temperature according to claim 1, characterized in that: In step S9, the weights of the adaptive weighted fusion are adaptively adjusted based on the interference characteristics and joint temperature change rate output by the second-level feature extraction: when the interference characteristics exceed a threshold or the temperature change rate is greater than a preset threshold, the weight of the predicted drift value is increased; when the interference characteristics are lower than a threshold and the temperature change rate is less than or equal to a preset threshold, the weight of the observed drift value is increased; the formula for calculating the position correction is: in, Let be the position correction amount for the i-th motor. This represents the measured rotor angle of the motor. This represents the final zero-point drift of the motor.

Images