A gait parameter self-adjusting method based on deep learning

Abstract

The present application relates to the technical field of intelligent robots, and more particularly to a gait parameter self-adjusting method based on deep learning. The method comprises the following steps: obtaining the initial joint angle of a guide robot and performing joint conditioning to determine the joint posture deviation; performing posture calibration according to the joint posture deviation and outputting posture calibration parameters; starting a radar and performing environment sensing to generate an environment map; generating point cloud height information based on the environment map; calculating the local slope using the point cloud height information to determine the terrain slope value; inputting the terrain slope value into a neural network model to output foot trajectory parameters; and driving the guide robot to a preset destination using the foot trajectory parameters and the posture calibration parameters to calculate the joint stability. The present application realizes real-time joint posture calibration and terrain sensing based on intelligent robot technology, achieves adaptive adjustment of the gait parameters of the guide robot, and improves the gait stability and navigation accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot technology, and in particular to a gait parameter self-adjustment method based on deep learning. Background Technology

[0002] Existing guide robots primarily rely on preset gait parameters or rule-based control based on simple sensor feedback for motion adjustment, typically employing fixed stride lengths, fixed stride frequencies, and manually adjusted joint angles for gait planning. In ideal, flat terrain or low-complexity indoor environments (e.g., flat ground, few obstacles, minimal slope variation), basic movement and obstacle avoidance are achievable. However, in complex terrain scenarios (e.g., outdoor slopes, slippery surfaces, uneven gravel roads, stairs, and dynamic obstacles), the environment is diverse and frequently changing, making it difficult for traditional methods to assess joint load, posture deviation, and dynamic stability in real time. As the application scenarios of guide robots expand, they need to traverse various terrain types, adapt to different slopes and obstacle conditions, and maintain stable gait and safe movement. Existing fixed gait parameters and manual adjustment methods cannot effectively address joint vibration, resonance risks, and dynamic terrain changes. They lack real-time gait self-adjustment mechanisms and multi-source sensor data fusion capabilities, leading to gait instability, joint overload, or gait deviations in complex scenarios, severely impacting navigation efficiency. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide a gait parameter self-adjustment method based on deep learning to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a deep learning-based gait parameter self-adjustment method is proposed, comprising the following steps:

[0005] Step S1: Obtain the initial joint angles of the guide robot, perform joint adjustment, and determine the joint posture deviation; perform posture calibration based on the joint posture deviation and output the posture calibration parameters.

[0006] Step S2: Activate the radar and perform environmental perception to generate an environmental map; generate point cloud height information based on the environmental map; calculate the local slope using the point cloud height information to determine the terrain slope value;

[0007] Step S3: Input the terrain slope value into the neural network model and output the foot trajectory parameters; use the foot trajectory parameters and posture calibration parameters to drive the guide robot to the preset destination and calculate the joint stability;

[0008] Step S4: Detect joint angle deviation based on joint stability and generate gait compensation data; correct foot trajectory parameters based on gait compensation data to complete the gait parameter self-adjustment task.

[0009] The beneficial effects of this invention are as follows:

[0010] (1) By collecting and extracting the joint angles, torques and vibration signals of the guide robot in real time and using multi-channel features, the joint posture deviation can be accurately identified. A nonlinear mapping relationship between the joint and the terrain can be constructed using a deep learning model to improve the accuracy and robustness of gait prediction.

[0011] (2) Based on the terrain slope, robot posture characteristics and environmental constraint information, foot trajectory parameters are generated and combined with posture calibration parameters to drive robot movement, so as to realize the precise gait control of guide robot in complex terrain and provide reliable data support for dynamic gait optimization.

[0012] (3) By performing frequency domain analysis and resonance detection on the joint vibration signal, the main frequency and resonance amplitude are extracted, and unstable joints are marked based on dynamic stability index, providing an accurate basis for gait compensation and ensuring the stability of the joint during movement.

[0013] (4) Based on the gait compensation amount, the foot trajectory parameters are corrected in real time to simulate the gait execution effect, form a closed-loop self-adjustment mechanism, realize the gait optimization of the robot in complex environment, and improve gait stability, walking safety and navigation continuity. Attached Figure Description

[0014] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0015] Figure 1 This is a schematic diagram of the steps of a gait parameter self-adjustment method based on deep learning according to the present invention.

[0016] Figure 2 This is a schematic diagram of the steps for determining the joint posture deviation in this invention;

[0017] Figure 3 This is a schematic diagram of the steps for outputting attitude calibration parameters in this invention;

[0018] Figure 4 This is a schematic diagram of the steps for generating an environmental map in this invention;

[0019] Figure 5 This is a schematic diagram of the steps for calculating joint stability in this invention;

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0022] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0023] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0024] To achieve the above objectives, please refer to Figures 1 to 5 This invention provides a gait parameter self-adjustment method based on deep learning, the method comprising the following steps:

[0025] Step S1: Obtain the initial joint angles of the guide robot, perform joint adjustment, and determine the joint posture deviation; perform posture calibration based on the joint posture deviation and output the posture calibration parameters.

[0026] In one embodiment, no-load movements are performed on the 12 joints of the guide robot, including full-range bending, extension, and rotation. The angles and drive currents of each joint are simultaneously recorded at a sampling frequency of 50Hz. The drive currents are converted into output torque values, and a torque-time curve is plotted. Based on curve analysis, it is identified that the torque change rate is greater than... During periods of abnormal resistance, the drive current is reduced to 40% of its original value. The joint resistance value is recorded and compared with a preset joint threshold to calculate the joint posture deviation. For joints with deviations exceeding 5°, a reverse calibration pulse is applied. The pulse amplitude is 0.7 times the deviation, and the pulse duration is 60ms. This is repeated until the residual value of a single calibration is less than 1°. Finally, the number of calibration pulses and residual values ​​for all joints are recorded, and the posture calibration parameter vector is normalized and output. It is used for subsequent gait control and trajectory planning.

[0027] In another embodiment, assuming the robot has 14 main joints, the no-load motion sampling frequency is set to 60Hz. The collected joint torque data shows that 4 joints are... There was abnormal resistance during that period. The drive current of these joints was reduced to its original value. The attitude deviations were calculated as follows:

[0028] joint ,joint ,joint ,joint .

[0029] A reverse calibration pulse is applied to each deviation joint, with the pulse amplitude being the deviation amount. Times, duration After repeating the process 2-3 times, the residuals of each joint were all less than 0.8°. The final output posture calibration parameter vector had a length of 14, providing initial accurate posture information for gait self-adjustment.

[0030] Step S2: Activate the radar and perform environmental perception to generate an environmental map; generate point cloud height information based on the environmental map; calculate the local slope using the point cloud height information to determine the terrain slope value;

[0031] In one embodiment, a 360° full-circle radar scan is initiated, with a scan frequency set to 30Hz. Point clouds with a distance greater than 6m or a reflection intensity lower than the threshold of 22 are filtered out, and the valid point clouds are recorded. The polar coordinate point cloud is converted to Cartesian coordinates and mapped onto a raster map, with the raster size set to [value missing]. For each grid cell, the height difference within a 5×5 neighborhood is calculated to obtain the local slope. Based on the local slope, a terrain gradient value, ranging from 0° to 20°, is calculated. These gradient values ​​are used as input to the neural network to generate foot trajectory parameters, ensuring the robot can dynamically adjust its gait in complex terrain.

[0032] In another embodiment, assuming the radar scanning frequency is set to 40Hz, the distance threshold is 5m, and the reflection intensity threshold is 20, approximately 8000 valid points are retained per frame after filtering. A 50×50 grid map is constructed, and the height difference within a 3×3 neighborhood of each grid point is calculated to obtain the local slope. The slope value range is... On average, approximately 250 slope points are generated per frame. This slope information, combined with posture calibration parameters, provides dynamic input to the neural network model, enabling the robot to adjust its gait in different slope environments.

[0033] Step S3: Input the terrain slope value into the neural network model and output the foot trajectory parameters; use the foot trajectory parameters and posture calibration parameters to drive the guide robot to the preset destination and calculate the joint stability;

[0034] In one embodiment, a neural network model is constructed, including an input layer (slope value, pose features, environmental constraints), three convolutional layers, and a fully connected output layer. The slope sequence is input into the network, and the convolutional layers extract the nonlinear mapping relationship between terrain and pose, outputting the gait cycle, stride length, swing amplitude, and landing point. The output trajectory parameters and pose calibration parameters are used to drive the robot to a preset target position, while simultaneously recording joint vibration signals at a sampling frequency of 200Hz. The dominant joint vibration frequency is extracted using Fourier transform and compared with the natural vibration frequency to calculate the resonance amplitude, thereby evaluating joint stability and providing a basis for subsequent gait compensation.

[0035] In another embodiment, it is assumed that the foot trajectory parameters output by the neural network are a stride length of 0.48m, a swing amplitude of 0.12m, and a swing time of 0.65s. During robot operation, the dominant vibration frequency of the left knee joint is collected at 12.3Hz, and that of the right ankle at 12.1Hz. The preset natural frequency is 12Hz, and the resonance amplitudes are 0.8° and 0.9°, respectively. Joint stability is determined based on a threshold of 1°, confirming that the current gait is stable on the target path, and this can be used to generate gait compensation data.

[0036] Step S4: Detect joint angle deviation based on joint stability and generate gait compensation data; correct foot trajectory parameters based on gait compensation data to complete the gait parameter self-adjustment task.

[0037] In one embodiment, gait compensation data is generated based on joint resonance amplitude, and stride length is adjusted accordingly. Correction: Swing amplitude adjusted ±2cm. The updated trajectory parameters are then used to drive the robot again, repeating stability checks until the resonance amplitude of all joints is reached. The final output is self-adjusting gait parameters, which ensure the robot's gait stability and posture coordination during continuous navigation tasks, and can provide a dynamic correction basis for subsequent navigation path planning.

[0038] In another embodiment, it is assumed that during forward movement, the left knee deviates by 2.4°, the right ankle by 1.9°, and the deviations of other joints are ≤1°. Gait compensation data is generated based on the resonance amplitude, adjusting the stride length from 0.45m to 0.48m and the swing amplitude from 0.11m to 0.12m. After three iterative adjustments, the deviations of the left knee and right ankle joints are both reduced to ≤1°, completing gait self-adjustment and providing stable support for the robot to move in complex environments.

[0039] Of particular importance, step S4, which detects joint angle deviations based on joint stability and generates gait compensation data, includes:

[0040] Instable joints are marked based on joint stability, and the joint angle deviations of the instable joints are statistically analyzed; the gait compensation amount is determined using the joint angle deviations.

[0041] In one embodiment, by installing multi-point angle sensors on the legs of the guide robot, the angle changes of the hip, knee, and ankle joints are collected in real time. Combined with the joint stability results evaluated in the previous step, unstable joints deviating from a preset stability range are identified. Subsequently, the joint angle deviation of each unstable joint is calculated. This refers to the difference between the actual joint angle and the preset ideal angle, expressed in degrees. The deviations of each joint angle are summarized to create an unstable joint angle deviation matrix, and gait compensation is calculated based on the matrix values, for example, for the knee joint. In this step, gait compensation can be defined as adjusting the knee joint swing amplitude by ±5° in the next step to restore balance. This step enables the generation of accurate gait compensation data based on quantitative analysis of joint instability, providing input for subsequent gait simulation and foot vibration detection.

[0042] In another embodiment, assume the collected unstable joints and angular deviations are as follows:

[0043] left knee ,

[0044] right knee ,

[0045] left ankle ,

[0046] Right ankle ,

[0047] Hip joint Based on this data, the calculated gait compensation amounts were for the left knee. , right knee left ankle right ankle Hip joint ± In this way, quantitative analysis of multi-joint cooperative instability can be performed, providing a reliable basis for gait control of robots under complex terrain conditions.

[0048] Gait motion simulation is performed based on gait compensation to determine the landing point; at the landing point, foot vibration is detected; and gait compensation data is generated based on the foot vibration.

[0049] In one embodiment, the compensation values ​​for each joint are input into the gait simulation model. The three-dimensional position coordinates of each step landing point are calculated using inverse kinematics, while simultaneously simulating the dynamic motion trajectory of the legs during the robot's forward movement. Subsequently, at each landing point, accelerometers or pressure sensors installed on the soles of the feet collect foot vibration signals, and the vibration amplitude and frequency characteristics are calculated. Based on the vibration amplitude and a preset safety threshold (e.g., vibration amplitude less than 3mm is considered normal, less than 5mm is considered slightly abnormal, and greater than 5mm is considered severely abnormal), gait compensation data is generated, including adjustment schemes for the next joint angle, stride length, and foot contact angle to optimize walking stability.

[0050] In another embodiment, it is assumed that the landing point coordinates (in cm) during the simulated 10-step process are as follows:

[0051] The foot vibration amplitude (in mm) at each landing point is as follows:

[0052] Based on this vibration data, the generated gait compensation data increased the knee joint height in the third step. Compensation, step 6 ankle joint reduction Compensation, step 7, knee joint increase Compensation, the remaining steps are adjusted within Within this range, continuous dynamic optimization of the gait of guide robots can be achieved, while minimizing foot vibration under complex terrain conditions, thereby improving walking stability and safety.

[0053] Preferably, in step S1, the initial joint angles of the guide robot are obtained, and joint adjustments are performed to determine the joint posture deviation, including:

[0054] Step S11: Obtain the initial joint angles of the guide robot, and use the initial joint angles to perform no-load motion and record the joint torque data; plot the torque change curve based on the joint torque data to identify periods of abnormal resistance.

[0055] In one embodiment, no-load motion is performed on 12 joints of the guide robot. The motion trajectory includes full-stroke bending, extension, and rotation movements. The no-load motion lasts for approximately 15 seconds, and the real-time angle change of each joint is recorded at a sampling frequency of 50Hz. Simultaneously, the joint drive current is collected and converted into an output torque value, and a torque-time curve is plotted. After smoothing the curve and removing instantaneous spikes, the torque change rate for each continuous data segment is calculated, and values ​​with a change rate greater than a certain threshold are marked. The intervals defined are periods of abnormal resistance. This analysis can identify locations of abnormal joint resistance at different stages of motion, providing fundamental data for subsequent posture calibration.

[0056] In another embodiment, assuming the robot has 14 main joints, the no-load motion sampling frequency is set to 60Hz, and the motion trajectory covers the complete flexion and rotation of the shoulder, elbow, knee, and ankle joints. The collected joint torque data shows that 5 joints exhibited abnormal resistance during no-load motion. Specifically, the left knee joint experienced a peak torque of 2.8 Nm for approximately 0.8 seconds, the right shoulder joint experienced a peak torque of 3.1 Nm, and the maximum torque of the remaining joints was within the normal range (≤1.5 Nm). Torque change curves were plotted, and the rate of torque change (in units) was calculated. The study identified abnormal periods in the left knee, right shoulder, left ankle, right elbow, and lumbar joints, which together accounted for approximately 18% of the no-load motion time. This result provides a precise basis for subsequent adjustment of drive current and calculation of posture deviation.

[0057] Step S12: During the period of abnormal resistance, obtain the abnormal drive current value and reduce it to the value of the abnormal drive current. Record the joint resistance value; compare the joint resistance value with the preset joint threshold and calculate the joint posture deviation.

[0058] In one embodiment, during the identified period of abnormal resistance, the joint drive current is reduced to 40% of the original abnormal current value, and the resistance value corresponding to each joint is recorded. Subsequently, the joint resistance value is compared with preset joint thresholds (e.g., 3.0 Nm for the knee joint and 2.5 Nm for the shoulder joint) to calculate the joint posture deviation. The deviation calculation formula is as follows:

[0059] (Actual resistance - threshold) / Torque sensitivity coefficient

[0060] The attitude deviation of each joint is obtained for subsequent attitude calibration. Throughout the process, the acquisition frequency is maintained at 50Hz to ensure high timeliness of attitude deviation calculation. The final output joint attitude deviation vector can be directly used for reverse calibration pulse application and gait self-adjustment.

[0061] In another embodiment, assuming that during the period of abnormal resistance, the abnormal drive current of the left knee is 2.5A and that of the right shoulder is 2.8A, these values ​​are reduced to their original values. The recorded resistance values ​​were 2.1 Nm and 2.3 Nm, respectively. The resistance values ​​for the other three abnormal joints were: left ankle 1.8 Nm, right elbow 2.0 Nm, and lower back 2.2 Nm. The preset thresholds were: left knee 3.0 Nm, right shoulder 2.5 Nm, left ankle 1.7 Nm, right elbow 1.9 Nm, and lower back 2.0 Nm. Based on the formula, the calculated posture deviations were: left knee 4.2°, right shoulder 3.6°, left ankle 2.1°, right elbow 2.8°, and lower back 1.5°. After the entire process was completed, a posture deviation vector was generated. This provides a data foundation for subsequent reverse pulse calibration and gait parameter self-adjustment.

[0062] Preferably, performing no-load motion and recording joint torque data includes:

[0063] The no-load motion trajectory is set according to the initial joint angle, including bending, extension and rotation movements throughout the joint's full range of motion; the joint actuator is activated to perform no-load motion according to the no-load motion trajectory, while simultaneously collecting the changes in joint angle.

[0064] In one embodiment, no-load motion trajectories are set for the 12 main joints of the guide robot. The trajectories cover the bending, extension, and rotation movements of the joints throughout their entire stroke, with a motion cycle of approximately 15 seconds. During the no-load motion, the joint actuators synchronously collect changes in joint angles at a frequency of 50Hz to ensure that even minute displacements and angular velocities are recorded. The collected data includes the real-time angle θ(t) and angular velocity ω(t) of each joint, and noise is removed through low-pass filtering. The entire no-load motion trajectory executes the shoulder, elbow, knee, and ankle joint movements sequentially to ensure that the angles of each joint cover the complete range during the full stroke, providing a sufficient data foundation for subsequent torque calculations. Furthermore, during the motion, the system monitors in real time whether the joint angles reach preset boundaries to prevent mechanical limit impacts or excessive torsion, ensuring the safety and accuracy of the collected data.

[0065] In another embodiment, assuming the robot has 14 joints, with the shoulder, elbow, knee, and ankle joints being the primary sampling joints, the idle motion lasts for 20 seconds, and the sampling frequency is increased to 60Hz. The motion trajectory is set as follows: left shoulder flexion from 0° to 120°, right elbow extension from 0° to 135°, left knee flexion from 0° to 110°, and right ankle rotation from 0° to 45°. Angle data for each joint is recorded during the motion; for example, the angle of the left shoulder at different time points is... The right elbow is By smoothing and interpolating the collected data, a continuous angle change curve is obtained, providing a high-precision input for calculating the drive current and output torque. It can also be used for subsequent attitude deviation analysis and abnormal joint identification.

[0066] Record the driving current during the no-load motion process and convert it into output torque; perform power correction on the output torque based on the change in joint angle, and record the joint torque data.

[0067] In one embodiment, after acquiring the no-load motion trajectory, the current data of each joint actuator is recorded, and the drive current is converted into output torque using a joint dynamics model. The angle change for each joint is then considered. The actual torque is calculated using a dynamic correction method. ,in This is the corrected current value. The torque coefficient of the motor is used. Subsequently, the torque data for each joint is plotted as a time-series curve to observe potential resistance anomalies, friction, or minor mechanical obstruction during no-load movement. By statistically analyzing the maximum, minimum, and average torque values ​​of the curves, the initial mechanical state of the joint is determined, providing a precise basis for posture calibration and gait self-adjustment. Furthermore, abnormal fluctuation regions are marked so that the control signal can be adjusted accordingly in subsequent joint drive calibration.

[0068] In another embodiment, assuming that the maximum drive current of the left knee is 2.5A among 14 joints, the converted output torque is approximately The maximum drive current for the right shoulder joint is 2.8A, and the output torque is approximately 3.1Nm; the maximum torque distribution for other joints is as follows: Between. By calculating the average and standard deviation of the torque curves of each joint throughout its full range of motion, the average torque of the left knee is obtained. Standard deviation Average torque of the right shoulder Standard deviation Further analysis of the torque change rate revealed that the left knee... The segment shows a rapid increase, with a rate of change reaching [missing information]. Right shoulder Segment change rate This information can be used to calculate joint posture deviations and provide detailed reference data for subsequent posture calibration and gait compensation.

[0069] Preferably, based on the joint torque data, a torque variation curve is plotted to identify periods of abnormal resistance, including:

[0070] The joint torque data is smoothed to remove instantaneous spikes, and a torque variation curve is plotted with the sampling time of the joint torque data as the horizontal axis and the joint torque value as the vertical axis.

[0071] In one embodiment, raw torque data collected from each joint of the guide robot is input into a preprocessing module to smooth the data and remove instantaneous spikes and measurement noise. The smoothing process uses a weighted moving average filter or a low-pass filter, with a window length of 5-7 sampling points to ensure that the torque variation trend is preserved while eliminating instantaneous anomalies. The smoothed torque data is plotted as a continuous torque variation curve with sampling time on the x-axis and torque value on the y-axis, with each curve corresponding to one joint. During the plotting process, the torque variation amplitude, maximum value, minimum value, and average value of each joint throughout its entire stroke are recorded, providing a basic reference for subsequent resistance anomaly identification. Observing the curves allows for a visual display of the mechanical state and potential friction or resistance characteristics of the joints at different angular positions, providing accurate data support for further calculation of the torque change rate.

[0072] In another embodiment, assuming the robot has 12 main joints, the no-load motion sampling frequency is 50Hz, and the total sampling time is 20 seconds, 1000 sampling points are obtained for each joint. A 5-point weighted moving average filter is applied to the torque data of each joint to remove errors from the original data. Peak. Taking the left shoulder joint as an example, the torque value sequence before smoothing is: After smoothing, it becomes Then, each joint curve was plotted on the same time axis, with the horizontal axis representing the sampling time. (unit The vertical axis represents torque. (unit This method facilitates comparative analysis of torque variation trends at different joints during motion. It ensures accurate data, free from noise interference, when subsequently calculating torque change rates and identifying abnormal periods.

[0073] During the plotting process, the peak and trough values ​​in the torque change curve are marked, the torque change rate between adjacent points between the peak and trough values ​​is calculated, and the torque change rate that exceeds the preset rate threshold is marked as an abnormal resistance period.

[0074] In one embodiment, on the plotted smooth torque curve, the system automatically marks the peak and valley values ​​of the torque at each joint and calculates the rate of torque change between adjacent sampling points, i.e. A preset rate threshold is used to determine whether torque changes are abnormal. For example, if the rate of torque change on a certain curve exceeds a certain threshold... These periods are marked as abnormal resistance periods. These marked abnormal periods can be used for subsequent joint drive current adjustment and attitude deviation analysis. The system can simultaneously generate a visual representation of the abnormal resistance periods, highlighting peak, trough, and abnormal rate regions, allowing maintenance personnel to intuitively assess the force and mechanical state of each joint during no-load movement, and assisting in determining whether there is increased friction or structural obstruction.

[0075] In another embodiment, assuming that after analyzing the torque curves of 12 joints, the maximum torque change rate of the left knee joint is calculated to be... The maximum rate of change of the right shoulder joint is The rate of change of other joints is distributed in Between. According to the preset rate threshold. Place the left knee joint in The period between t=5.1 and 5.6 s for the right shoulder joint was marked as the abnormal resistance period, while the other joints were considered to be in a normal range. Under this assumption, the abnormal period for the left knee joint lasted approximately 0.6 s, and the abnormal period for the right shoulder joint lasted approximately 0.5 s. These marking results can be used to subsequently reduce the drive current to the 30%-50% range to obtain the actual joint resistance values, further calculate the posture deviation, and develop corrective action strategies, providing a quantitative basis for the safe and reliable gait adjustment of guide robots.

[0076] Preferably, in step S1, attitude calibration is performed based on the joint attitude deviation, and the output attitude calibration parameters include:

[0077] Step S13: If the joint posture deviation is greater than the preset deviation threshold, mark the corresponding joint as the joint to be calibrated; apply a reverse calibration pulse to each joint to be calibrated individually, with the pulse amplitude set to 0.6-0.8 times the joint posture deviation and the pulse duration fixed at 50ms-80ms;

[0078] In one embodiment, the system first acquires the real-time posture data of each joint, compares it with the target standard posture, and calculates the deviation of each joint. If a certain joint Exceeding the preset deviation threshold If the joint is not calibrated, it is marked as a joint to be calibrated. For each joint to be calibrated, a reverse calibration pulse is applied to compensate for the deviation. The pulse amplitude is set to 0.6 to 0.8 times the deviation, and the pulse duration is fixed at [value missing]. to This process ensures that deviations are effectively corrected without causing mechanical shock or excessive oscillation. During pulse application, the system continuously monitors joint angle changes, records the pulse response curve in real time, and outputs angle-time curves for subsequent analysis. This step ensures that the joint to be calibrated can quickly approach the target posture and provides basic data for subsequent calibration residual calculations.

[0079] In another embodiment, assuming the robot has 10 joints, the deviation is calculated in real time. It was later discovered that the left shoulder joint Right knee joint All other joint deviations were less than the threshold. Mark the left shoulder and right knee as joints to be calibrated. Apply a reverse calibration pulse to the left shoulder with an amplitude of [value missing]. The pulse duration is set to A pulse with an amplitude of 4.8 × 0.65 ≈ 3.12° is applied to the right knee, with the same pulse duration. During the application process, the system samples the joint angle every 5ms and records the angle change sequence. This is used to evaluate the pulse effect and for subsequent residual calculations.

[0080] Step S14: After the reverse calibration pulse ends, read the joint feedback angle and calculate the angle difference as the single calibration residual; when the single calibration residual is greater than the residual threshold, repeat the reverse calibration pulse until the single calibration residual is less than or equal to the residual threshold.

[0081] In one embodiment, the joint feedback angle is read immediately after the reverse calibration pulse ends. and the target angle Calculate the angle difference This serves as the residual from a single calibration. If Greater than the preset residual threshold Then, the reverse calibration pulse is applied repeatedly. During each iteration, the pulse amplitude can remain constant or be dynamically adjusted to 0.6-0.8 times the deviation based on the residual, until... This method can quickly converge to the allowable attitude error range through a finite number of iterations, thereby ensuring that the joints are accurately aligned with the target attitude, while avoiding mechanical oscillations or wear caused by excessive pulse application.

[0082] In another embodiment, it is assumed that the feedback angle of the left shoulder joint is after a single calibration. Target angle ,but greater than the residual threshold The system applies the pulse amplitude again. Feedback perspective updated to residual The threshold requirement was met, and the iteration ended. The initial calibration residual for the right knee joint was 1.5°. After applying the second pulse... The calibration is then completed. Under this assumption, the left shoulder requires two pulses to complete the calibration, as does the right knee; the other joints require no iterations. This method effectively quantifies the calibration effect of each joint and provides accurate data for generating the posture calibration vector.

[0083] Step S15: Record the number of calibration pulses and the single calibration residual for all joints to be calibrated, and merge them into an attitude calibration vector; normalize the attitude calibration vector to the [-1,1] interval, and output the attitude calibration parameters.

[0084] In one embodiment, the system combines the number of pulses for all joints to be calibrated and the corresponding single-calibration residuals to form the original attitude calibration vector. ,in For the first Number of pulses per joint This corresponds to the residual. Then, for the vector... Perform normalization to linearly map all elements to The interval is used to obtain the standardized attitude calibration parameter vector. . It can be used as input to the robot posture adjustment module, directly participating in the adaptive correction of motion control strategy to achieve real-time posture calibration.

[0085] In another embodiment, assuming that among the robot's 10 joints, the left shoulder and right knee are the joints to be calibrated, with 2 pulses and 2 pulses respectively, and single-pulse residuals are respectively... and The data is merged to obtain the original vector. Using the linear normalization formula Mapped to After the interval, we get This attitude calibration vector can be directly used in the drive module to achieve rapid attitude adjustment of each joint of the robot and provide quantifiable calibration effect feedback for further optimization of the reverse calibration strategy and gait control algorithm.

[0086] Preferably, step S2, which involves activating the radar and performing environmental perception to generate an environmental map, includes:

[0087] Step S21: Start the radar and operate in 360° full-circle scanning mode, setting the scanning frequency to [value missing]. Real-time reception of raw radar polar coordinates, filtering out ranges greater than [missing information]. Or the reflection intensity is below the threshold For invalid coordinates, record valid polar coordinates;

[0088] In one embodiment, the system activates the radar, operating in a 360° omnidirectional scanning mode with a scanning frequency set between 25Hz and 50Hz. During operation, the radar receives raw polar coordinate point data in real time, including distance and angle information. The system filters the acquired data, removing distant points greater than 5 meters and weakly reflecting points with a reflection intensity below a threshold of 20. The filtered points are recorded as valid polar coordinate points for subsequent projection and map generation. This step ensures the reliability of the radar data and reduces the impact of environmental noise.

[0089] In another embodiment, assuming the radar operates at a frequency of 30Hz, it acquires approximately 2000 polar coordinate points per scan. After filtering out invalid points with a distance greater than 5 meters and a reflection intensity less than 20, approximately 1300 valid points remain. These points mainly originate from surrounding walls, obstacles, and moving targets. The system records these points in chronological order as data input for environmental modeling.

[0090] Step S22: Convert the valid polar coordinates to Cartesian coordinates and project the Cartesian coordinates onto a preset raster map; perform binarization dilation based on the raster map to generate an environment map.

[0091] In one embodiment, valid polar coordinate points are converted into spatial planar coordinates and projected onto a preset grid map. Each valid point corresponds to an occupied cell based on the grid size. The occupied cells are binarized, marked as "obstacles," and empty cells are marked as "passable." Subsequently, the occupied cells are inflated to cover measurement errors and obstacle sizes, generating a final environmental map that can be directly used for navigation and obstacle avoidance.

[0092] In another embodiment, the projected grid map is assumed to be 50×50 grids, with each grid having a side length of 0.1 meters. After projection, effective points occupy approximately 850 grids, leaving approximately 650 empty grids. After an expansion operation, the number of occupied grids increases to 1000, ensuring the complete outline of obstacles while preserving passage space. The generated grid map can be used for path planning and dynamic obstacle avoidance analysis.

[0093] Of particular importance is the conversion of valid polar points to Cartesian coordinates and the projection of those Cartesian coordinates onto a preset raster map, which includes:

[0094] Convert valid polar coordinates to Cartesian coordinates; perform coordinate mapping calculations between the Cartesian coordinates and the preset raster map to generate a raster index;

[0095] In one embodiment, a set of polar coordinate points is acquired from a lidar or depth sensor, wherein each point is represented by an angle. and distance Represented by the formula for converting from standard polar coordinates to Cartesian coordinates. All valid points (distances between 0.2m and 10m, excluding noise points) are converted to two-dimensional Cartesian coordinates. Then, each Cartesian coordinate point is mapped onto a predefined raster map. Assume the raster map resolution is... The raster index corresponding to each coordinate point can be calculated. , This process also allows for boundary constraint checks on coordinate points to ensure that the mapping index falls within the raster map's range, thereby generating a complete raster index sequence and providing a foundation for subsequent occupancy status updates.

[0096] In another embodiment, assuming a total of 1200 polar coordinate points are collected, after conversion, the Cartesian coordinate points are obtained as follows: After mapping these coordinates to a raster map with a resolution of 0.05m, an index sequence is obtained. The repeated index indicates that multiple points fall within the same grid cell. This method enables the discretization of point cloud data from multiple sources, providing accurate grid position references for subsequent grid occupancy status updates.

[0097] Update the occupancy status of raster cells using the raster index to generate raster mapping results; integrate the raster mapping results into the preset raster map.

[0098] In one embodiment, the occupancy probability of the corresponding raster cell is updated according to the raster index. For example, for each observed index cell, its occupancy probability is updated using a Bayesian formula:

[0099]

[0100] Update, among which The probability of sensor measurement (e.g., 0.7). This represents the prior occupancy probability of the grid cell. After the update, the current grid mapping result is generated, which can be represented as a two-dimensional matrix with element values ​​ranging from [0,1], representing the occupancy probability of the corresponding grid cell. Subsequently, the current grid mapping result is integrated with the preset global grid map, merging the new information with the historical map to update the global occupancy status, providing the latest environmental information for path planning and obstacle avoidance.

[0101] In another embodiment, it is assumed that the generated raster mapping matrix is ​​200×200 in size, and the occupancy probability of each raster cell ranges from 0.0 to 1.0. During the update process, after raster index mapping, the distribution of the updated occupancy probability values ​​at the current time is as follows. After being integrated into the global grid map, the probability of some cells in the previous global map was updated from 0.2 to 0.5, and that of others from 0.0 to 0.7. This method enables the dynamic fusion of multiple consecutive frames of grid data, providing a real-time, quantifiable environmental representation for mobile robots or guide robots.

[0102] Preferably, the neural network model in step S3 includes: an input layer, a hidden layer, and an output layer;

[0103] The input layer is used to receive terrain slope values, robot posture features, and environmental constraint information;

[0104] The hidden layer extracts the non-linear mapping relationship between terrain and pose features through multi-layer convolution and fully connected operations;

[0105] The output layer generates the corresponding foot trajectory parameters, including the swing amplitude, stride length, swing time, and landing position in the gait cycle.

[0106] In one embodiment, the collected terrain slope values ​​are input into a pre-trained neural network model. This neural network includes an input layer, hidden layers, and an output layer. The input layer receives the terrain slope, the robot's current posture features (such as tilt angle and torso posture), and environmental constraint information (such as obstacle locations or passable areas). The hidden layer extracts the nonlinear mapping relationship between terrain and posture features through multi-layer convolution and fully connected operations. The output layer generates foot trajectory parameters based on the processing results, including the swing amplitude, stride length, swing duration, and landing point position within the gait cycle. In this embodiment, the neural network model can complete a forward calculation within 50ms, outputting a foot swing amplitude range of approximately 0.15m–0.25m, a stride length of approximately 0.3m–0.5m, a swing duration of approximately 0.4s–0.6s, and a landing point position with a horizontal deviation of no more than 0.05m relative to the current supporting foot position.

[0107] In another embodiment, assuming the input terrain slope range is 0°–15°, the robot's posture characteristics at different slopes include a forward tilt angle of 5°–10° and a side tilt angle of 3°–7°. After neural network processing, the generated foot trajectory parameters are as follows: swing amplitude 0.18m, 0.22m, 0.20m; stride length 0.35m, 0.42m, 0.40m; swing duration 0.45s, 0.50s, 0.48s; and corresponding landing point offsets of 0.03m, 0.04m, and 0.02m, respectively. By processing multiple consecutive terrain slope frames, the system can output a continuous foot trajectory sequence, providing a reference for robot gait control and stable walking.

[0108] Preferably, the hidden layer extracts the nonlinear mapping relationship between terrain and pose features through multi-layer convolution and fully connected operations, including:

[0109] A multi-channel feature map is constructed by combining terrain slope values ​​with robot pose features; weighted operations are performed within the local receptive field using convolutional layers to extract local terrain and pose change features.

[0110] In one embodiment, the collected terrain slope values ​​are integrated with the robot's current posture features (such as forward tilt angle, side tilt angle, and torso rotation angle) into a multi-channel feature map, with each channel representing a feature dimension. This feature map is then input into the convolutional layer of a convolutional neural network, where weighted operations are performed within the local receptive field to extract local features related to terrain slope changes and posture changes. In this embodiment, each convolutional kernel is 3×3 in size, and the number of output channels is 32. After activation, a preliminary local feature representation is obtained, which can capture the impact of local slope changes on the robot's posture. After this step, the local feature map generated by the network can be used for subsequent gait planning or trajectory adjustment.

[0111] In another embodiment, it is assumed that the robot simultaneously collects terrain slope information within a 10m × 10m area in front, totaling 100 × 100 grids. Each grid records the slope value, and the posture features include forward tilt angles of 5°–10°, side tilt angles of 3°–7°, and torso rotation angles of 2°–5°. This information is used to construct a 6-channel feature map (3 channels for terrain slope and 3 channels for posture angles), which is then input into a convolutional network for processing. After the convolutional layer outputs a multi-channel local feature map, the example generates 32 × 100 × 100 local features, reflecting the posture change trend under different local slopes. This output can be used for subsequent multi-step gait generation and stability analysis.

[0112] The multi-channel feature maps output by each convolutional layer are fused at multiple scales to integrate them into terrain-pose information. The terrain-pose information is then nonlinearly transformed using an activation function to output the nonlinear mapping relationship between terrain and pose features.

[0113] In one embodiment, the multi-channel feature maps output by the convolutional layer are fused at multiple scales. For example, pooling layers are used to reduce spatial resolution and integrate information from different receptive fields, thus integrating terrain-pose features into a unified representation. Subsequently, a nonlinear transformation is performed using an activation function (such as ReLU or LeakyReLU) to generate a nonlinear mapping relationship between terrain and pose. This mapping relationship can be directly used to calculate the next foot trajectory or gait adjustment parameters, enabling the robot to walk smoothly in complex terrain.

[0114] In another embodiment, it is assumed that after multi-scale fusion, the feature map is reduced in dimensionality from the original 32×100×100 to 64×25×25, and then a nonlinear mapping feature is generated through an activation function. The output nonlinear feature contains the combined influence of each local terrain unit and the robot's posture. For example, feature values ​​of 0.18, 0.21, and 0.19 for a certain region indicate the changing trend of the robot's swing amplitude or step length on that terrain unit. This output can be used as input for subsequent neural network generation of foot trajectory parameters to achieve adaptation to irregular terrain.

[0115] Preferably, in step S3, the guide robot is driven to a preset destination using foot trajectory parameters and posture calibration parameters, and the calculation of joint stability includes:

[0116] Step S31: Drive the guide robot to the preset destination using foot trajectory parameters and posture calibration parameters, and collect joint vibration signals in real time; perform Fourier transform based on the joint vibration signals to determine the main frequency of joint vibration;

[0117] In one embodiment, the robot is equipped with a high-precision accelerometer and vibration sensor, with a sampling frequency set to 50Hz, enabling it to capture minute vibration changes in the joints. After Fourier transform, the dominant frequency is extracted and the vibration amplitude is calculated for subsequent resonance analysis and stability assessment. This method not only reflects the robot's vibration characteristics under different terrain or path conditions but can also be used to establish joint dynamic response models, providing quantitative basis for gait optimization and fault early warning.

[0118] In another embodiment, assuming the robot walks along a straight path, vibration signals from the left knee, right knee, left ankle, right ankle, and hip joints are collected simultaneously, with 1000 frames of time-series data recorded for each joint. Fourier transform analysis yields the dominant frequencies and corresponding amplitudes (Hz / mm) for each joint: left knee 4.8 / 0.95, right knee 5.1 / 1.02, left ankle 6.0 / 1.15, right ankle 5.8 / 1.10, and hip joint 3.5 / 0.88. These hypothetical data allow for a preliminary analysis of the robot's joint vibration characteristics during movement, providing a reference for stability assessment and control strategy optimization.

[0119] Step S32: Perform resonance analysis based on the joint's dominant vibration frequency and the joint's preset natural vibration frequency, and record the resonance amplitude; assess joint stability based on the resonance amplitude.

[0120] In one embodiment, the dominant frequency of each joint is compared with its corresponding natural vibration frequency. If the dominant frequency is close to the natural frequency and the vibration amplitude is large, resonance is determined to exist, and the resonance amplitude of the joint is recorded. Subsequently, by comparing the resonance amplitudes of different joints, an overall stability index is calculated. The stability of a single joint can be mapped to a rating (e.g., 0–5 points), and the results are summarized to form an overall gait stability evaluation of the robot, which is used to assist in navigation path planning, gait adjustment, and risk warning.

[0121] In another embodiment, assuming the preset natural vibration frequencies of each joint are 5Hz for the left knee, 5Hz for the right knee, 6Hz for the left ankle, 6Hz for the right ankle, and 3.5Hz for the hip joint, the obtained dominant frequency is compared with these natural frequencies to calculate the resonance amplitude (in mm) of each joint: 0.12 for the left knee, 0.07 for the right knee, 0.18 for the left ankle, 0.15 for the right ankle, and 0.05 for the hip joint. Based on a set threshold of 0.1mm, it is determined that the left knee and left ankle have slight resonance, the right knee and right ankle are close to the threshold, and the hip joint is stable. The amplitude is then mapped to a stability level, with an overall stability score of 4.2 / 5, reflecting the robot's dynamic gait performance on the current path, providing a quantitative reference for subsequent control strategy optimization and abnormal gait warning.

[0122] Preferably, the resonance analysis based on the joint's dominant vibration frequency and the joint's preset natural vibration frequency, and the recording of the resonance amplitude, includes the following steps:

[0123] Construct a time series curve of the main frequency based on the main frequency of joint vibration; establish a reference line of natural frequency using the preset natural vibration frequency of the joint; when the frequency difference between the main frequency time series curve and the reference line of natural frequency is less than 2Hz, it is marked as the frequency resonance interval.

[0124] In one embodiment, the collected dominant vibration frequency signals of the knee, ankle, and hip joints are arranged chronologically to form a dominant frequency curve for each joint. The sampling frequency is set to 50Hz, and the time window is set to 10 seconds to cover the dynamic changes within a typical walking cycle. Subsequently, the natural vibration frequency of each joint is input into the calibration data system to plot a time-aligned natural frequency reference line, and the instantaneous difference between the dominant frequency curve and the reference line is calculated. .when Less than When the resonance is moderate, it is identified as a frequency resonance region and thus marked as such. This method allows for the quantitative analysis of potential resonance phenomena at various joints during the movement of a guide robot, providing data for subsequent stability assessments and vibration control.

[0125] In the frequency resonance range, the vibration signal is extracted; the resonance amplitude is calculated based on the vibration signal.

[0126] In one embodiment, the acceleration signal or vibration sensor output signal within the resonance range is truncated at a sampling frequency of 50Hz and subjected to a Fast Fourier Transform (FFT) to obtain the vibration amplitude spectrum. Subsequently, the amplitude corresponding to the dominant frequency is identified as the resonance amplitude, and the amplitude change curve of each joint at different time points is recorded. Based on the amplitude magnitude, joint stability can be quantified into a grade index, such as 0–5 points, where an amplitude less than 1mm is considered stable, 1–3mm is slightly unstable, and greater than 3mm is significantly unstable. This method enables continuous monitoring and anomaly warning of the gait of guide robots.

[0127] In another embodiment, assume the vibration amplitudes (in mm) extracted within the marked resonance range are as follows: left knee [0.8, 1.2, 0.9], right knee [1.0, 0.95, 1.1], left ankle [1.5, 1.8, 1.6], right ankle [1.2, 1.4, 1.3], and hip joint [0.5, 0.6, 0.55]. Based on the above assumed data, it can be determined that the left knee and hip joint are in the stable zone, the right knee is in the slightly unstable zone, the left ankle has obvious resonance, and the right ankle is slightly unstable. By statistically analyzing the resonance amplitudes of each joint, a stability heatmap can be drawn.

[0128] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0129] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A gait parameter self-adjustment method based on deep learning, characterized in that, Includes the following steps: Step S1: Obtain the initial joint angles of the guide robot, perform joint adjustment, and determine the joint posture deviation; perform posture calibration based on the joint posture deviation and output the posture calibration parameters. Step S2: Activate the radar and perform environmental perception to generate an environmental map; generate point cloud height information based on the environmental map; calculate the local slope using the point cloud height information to determine the terrain slope value; Step S3: Input the terrain slope value into the neural network model to output foot trajectory parameters; use the foot trajectory parameters and posture calibration parameters to drive the guide robot to the preset destination, and calculate joint stability, including: Step S31: Drive the guide robot to the preset destination using foot trajectory parameters and posture calibration parameters, and collect joint vibration signals in real time; perform Fourier transform based on the joint vibration signals to determine the main frequency of joint vibration; Step S32: Perform resonance analysis based on the joint's dominant vibration frequency and the joint's preset natural vibration frequency, and record the resonance amplitude; assess joint stability based on the resonance amplitude. Step S4: Detect joint angle deviation based on joint stability and generate gait compensation data; correct foot trajectory parameters based on gait compensation data to complete the gait parameter self-adjustment task.

2. The gait parameter self-adjustment method based on deep learning according to claim 1, characterized in that, In step S1, the initial joint angles of the guide robot are obtained, and joint adjustments are performed to determine the joint posture deviation, including: Step S11: Obtain the initial joint angles of the guide robot, and use the initial joint angles to perform no-load motion and record the joint torque data; plot the torque change curve based on the joint torque data to identify periods of abnormal resistance. Step S12: During the period of abnormal resistance, obtain the abnormal driving current value and reduce it to 30%-50% of the abnormal driving current value, and record the joint resistance value; compare the joint resistance value with the preset joint threshold and calculate the joint posture deviation.

3. The gait parameter self-adjustment method based on deep learning according to claim 2, characterized in that, Perform no-load motion and record joint torque data, including: The no-load motion trajectory is set according to the initial joint angle, including bending, extension and rotation movements throughout the joint's full range of motion; the joint actuator is activated to perform no-load motion according to the no-load motion trajectory, while simultaneously collecting the changes in joint angle. Record the driving current during the no-load motion process and convert it into output torque; perform power correction on the output torque based on the change in joint angle, and record the joint torque data.

4. The gait parameter self-adjustment method based on deep learning according to claim 2, characterized in that, Based on the joint torque data, a torque variation curve is plotted to identify periods of abnormal resistance, including: The joint torque data is smoothed to remove instantaneous spikes, and a torque variation curve is plotted with the sampling time of the joint torque data as the horizontal axis and the joint torque value as the vertical axis. During the plotting process, the peak and trough values ​​in the torque change curve are marked, the torque change rate between adjacent points between the peak and trough values ​​is calculated, and the torque change rate that exceeds the preset rate threshold is marked as an abnormal resistance period.

5. The gait parameter self-adjustment method based on deep learning according to claim 1, characterized in that, In step S1, attitude calibration is performed based on the joint attitude deviation, and the output attitude calibration parameters include: Step S13: If the joint posture deviation is greater than the preset deviation threshold, mark the corresponding joint as the joint to be calibrated; apply a reverse calibration pulse to each joint to be calibrated individually, with the pulse amplitude set to 0.6-0.8 times the joint posture deviation and the pulse duration fixed at 50ms-80ms; Step S14: After the reverse calibration pulse ends, read the joint feedback angle and calculate the angle difference as the single calibration residual; when the single calibration residual is greater than the residual threshold, repeat the reverse calibration pulse until the single calibration residual is less than or equal to the residual threshold. Step S15: Record the number of calibration pulses and the single calibration residual for all joints to be calibrated, and merge them into an attitude calibration vector; normalize the attitude calibration vector to the [-1,1] interval, and output the attitude calibration parameters.

6. The gait parameter self-adjustment method based on deep learning according to claim 1, characterized in that, Step S2 involves activating the radar and performing environmental perception to generate an environmental map, including: Step S21: Start the radar and operate it in 360° full-circle scanning mode, with the scanning frequency set to 25Hz-50Hz; receive the original polar coordinates of the radar in real time, filter out invalid coordinates with a distance greater than 5m or a reflection intensity lower than the threshold of 20, and record the valid polar coordinate points; Step S22: Convert the valid polar coordinates to Cartesian coordinates and project the Cartesian coordinates onto a preset raster map; perform binarization dilation based on the raster map to generate an environment map.

7. The gait parameter self-adjustment method based on deep learning according to claim 1, characterized in that, The neural network model described in step S3 includes: an input layer, a hidden layer, and an output layer; The input layer is used to receive terrain slope values, robot posture features, and environmental constraint information; The hidden layer extracts the non-linear mapping relationship between terrain and pose features through multi-layer convolution and fully connected operations; The output layer generates the corresponding foot trajectory parameters, including the swing amplitude, stride length, swing time, and landing position in the gait cycle.

8. The gait parameter self-adjustment method based on deep learning according to claim 7, characterized in that, The hidden layers extract the non-linear mapping relationship between terrain and pose features through multiple convolutional layers and fully connected operations, including: A multi-channel feature map is constructed by combining terrain slope values ​​with robot pose features; weighted operations are performed within the local receptive field using convolutional layers to extract local terrain and pose change features. The multi-channel feature maps output by each convolutional layer are fused at multiple scales to integrate them into terrain-pose information. The terrain-pose information is then nonlinearly transformed using an activation function to output the nonlinear mapping relationship between terrain and pose features.

9. The gait parameter self-adjustment method based on deep learning according to claim 1, characterized in that, Resonance analysis is performed based on the joint's dominant vibration frequency and the joint's preset natural vibration frequency. The resonance amplitude is recorded through the following steps: Construct a time series curve of the main frequency based on the main frequency of joint vibration; establish a reference line of natural frequency using the preset natural vibration frequency of the joint; when the frequency difference between the main frequency time series curve and the reference line of natural frequency is less than 2Hz, it is marked as the frequency resonance interval. In the frequency resonance range, the vibration signal is extracted; the resonance amplitude is calculated based on the vibration signal.

Images