FPGA-based human-simulating robot joint module full-closed loop control method and system

Abstract

The application discloses a human-imitating robot joint module full-closed loop control method based on FPGA, and relates to the technical field of robot control.The steps of the method comprise the following steps: performing second-order active low-pass filtering and programmable gain amplification conversion on the analog signals of an encoder, a gyroscope and a force sensor to obtain digital signals; constructing a multi-dimensional feature sequence, collecting temperature signals of a joint driver and the encoder through a temperature sensor interface, and constructing a temperature drift model to obtain zero-position drift compensation parameters; then, joint elastic deformation error is applied to the encoder signal to perform online offset correction; based on a lightweight neural network, a reconfiguration region selection signal is output; and according to the region selection signal output by the lightweight neural network, a bit stream local refresh of a corresponding PR Region is triggered.The application solves the problem that a traditional rigid model or a constant parameter model is difficult to reflect deformation characteristics under the combined action of heat-lubrication-load in real time, so that a control loop is difficult to eliminate flexible error.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a closed-loop control method and system for humanoid robot joint modules based on FPGA. Background Technology

[0002] With the rapid development of humanoid robots in service, manufacturing, and medical fields, increasingly higher demands are being placed on the control precision, response speed, and robustness of their joint modules. Traditional servo joint control methods mainly rely on filtering and compensation algorithms with fixed parameters, which struggle to simultaneously meet the following challenges: Existing systems mostly use first-order filtering or fixed-gain amplification, which cannot balance noise suppression and dynamic response between low-speed micro-motion and high-speed large-amplitude motion, resulting in inaccurate quantization in static conditions and easy signal saturation in dynamic conditions.

[0003] Most solutions only perform linear drift correction for single-point temperature, ignoring the temperature coupling effect between the driver and encoder. Under high temperature, alternating hot and cold conditions, or changes in lubrication status, the zero-position offset increases dramatically, directly affecting positioning accuracy. Furthermore, the joint coupling and the machine body connection have non-negligible elastic and damping characteristics, especially under heavy loads, impacts, or rapid turns, which can generate dynamic deformation errors. Traditional rigid models or constant parameter models cannot reflect the deformation characteristics under the combined effects of heat, lubrication, and load in real time, making it difficult for the control loop to eliminate flexibility errors.

[0004] The cutoff frequency, window length, and covariance matrix of IIR / FIR filters and Kalman filters are usually manually adjusted through experience or offline experiments. Once the parameters are fixed, they are difficult to adapt to different operating points, which is not conducive to rapid deployment and large-scale production.

[0005] Therefore, this invention provides a closed-loop control method and system for humanoid robot joint modules based on FPGA. Summary of the Invention

[0006] The purpose of this invention is to provide a closed-loop control method and system for humanoid robot joint modules based on FPGA, so as to solve the existing problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a fully closed-loop control method for humanoid robot joint modules based on FPGA, comprising the following steps: S1. Perform second-order active low-pass filtering and programmable gain amplification on the analog signals from the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through ADC; S2. Construct a multidimensional feature sequence, including calculating the signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting the mechanical resonance frequency; S3. Collect the temperature signals of the joint driver and encoder through the temperature sensor interface, and construct a temperature drift model to obtain the zero-point drift compensation parameters; S4. After obtaining the driving torque, construct an elastic deformation model to obtain the joint elastic deformation error; S5. Apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction; S6. Using the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and part of the reconfiguration region selection signal. S7. Based on the region selection signal output by the lightweight neural network, trigger a local refresh of the bitstream of the corresponding PR Region.

[0008] A further improvement of this invention is that, during the programmable gain amplification process, the gain value G is based on the joint speed of the previous cycle. Dynamic adjustment, the adjustment rules are expressed as follows: ,in This indicates the minimum magnification of the PGA. This indicates the maximum magnification of the PGA, and represents the predefined maximum joint velocity reference value. .

[0009] A further improvement of this invention is that the temperature drift model is obtained by collecting the temperature of the joint actuator. and encoder temperature Simultaneously considering the interactive effect of simultaneous changes in driver and encoder temperatures on drift, the zero-position drift function is obtained. ,in, and This represents the coefficients to be calibrated, where n and m represent the order. This represents the coupling coefficient between the driver temperature and the encoder temperature. During the calibration phase, historical data (N values) are collected through temperature bench experiments. The samples, through regularized fitting, yielded... The real-time temperature vector is then substituted into the zero-point drift function to obtain the real-time zero-point drift compensation parameters. Angles measured using raw joint measurements Joint angles after zero-position drift compensation .

[0010] A further improvement of this invention is that the specific construction process of the elastic deformation model includes: obtaining the driving torque Drt through the motor current of the joint actuator, and calculating the equivalent stiffness online based on the temperature function obtained in advance through the zero-position drift compensation parameters. With viscoelastic damping coefficient To construct a Kelvin-Voigt first-order viscoelastic model reflecting the current thermal lubrication conditions, based on the generated discrete state-space equations, the Kalman filter uses the state estimate from the previous cycle and the torque input Drt from the current cycle to predict and update the current angular displacement error. This error estimate is immediately compared with the joint angle after zero-point drift compensation. Subtraction yields the compensation angle, which has been adjusted to eliminate the effects of static temperature bias and structural flexibility. Then, the commands are sent to the closed-loop controller to generate the next drive instructions.

[0011] A further improvement of this invention is that, in step S6, the multidimensional feature sequence, the original joint angle after zero-position drift compensation, and the flexibility influence compensation angle after removing temperature drift are concatenated into a high-dimensional input vector, which is then input into a lightweight neural network that has been trained offline in advance. The first layer of this network uses one-dimensional convolution to extract temporal features and combines feedback neuron structure to capture dynamic correlations. Subsequently, through a series of quantized fully connected mappings, the hidden features are projected into the FPGA to output control parameters.

[0012] A further improvement of this invention is that the hidden features in step S6 are projected onto four parallel output branches, including: The first branch outputs the optimal cutoff frequency of the IIR filter. , used to adjust the bandwidth of the real-time filter; The second branch outputs the integerized FIR window length L to reconstruct the time-domain response of the FIR filter; The third branch outputs the process noise covariance Q and measurement noise covariance R of the Kalman filter to dynamically update the signal estimation weights of the filter. The fourth branch generates classification probabilities for the reconfigurable logic regions that have undergone IIR, FIR, or EKF filtering through the Softmax layer, and selects the region selection signal Spr by selecting the term with the highest probability.

[0013] A further improvement of the present invention is that step S2 receives the digital signal output in step S1. The variance of the signal residual noise is expressed as: ,in, This represents the filtered estimated value; the mechanical resonant frequency is obtained by... Perform a short-time Fourier transform to identify the frequency corresponding to the amplitude peak. .

[0014] On the other hand, the present invention provides a fully closed-loop control system for a humanoid robot joint module based on FPGA, comprising: The signal acquisition module is used to perform second-order active low-pass filtering and programmable gain amplification on the analog signals from the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through the ADC. The feature extraction and event monitoring module is used to construct multidimensional feature sequences, including calculating signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting mechanical resonance frequency; The temperature drift model construction module is used to acquire temperature signals from the joint actuator and encoder through the temperature sensor interface to obtain zero-point drift compensation parameters. The elastic deformation model construction module is used to construct an elastic deformation model after obtaining the driving torque to obtain the joint elastic deformation error; The compensation application module is used to apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction. The neural network decision module is used to take the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and part of the reconfiguration region selection signal. The PR control module is used to trigger a local refresh of the bitstream of the corresponding PR Region based on the region selection signal output by the lightweight neural network.

[0015] A further improvement of the present invention is that the PR control module further includes a filtering execution unit, which is used to perform IIR, FIR or EKF filtering in the corresponding PRRegion, and transmit the filtered signal to the next-level closed-loop control module through the AXI-Stream bus or DMA channel.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention firstly employs a dual-temperature-coupled polynomial model of driver and encoder to accurately fit the static offset of zero position caused by temperature changes, and calculates the compensation amount online in real time, thereby significantly reducing the impact of thermal expansion and temperature unevenness on measurement accuracy. 2. Based on the temperature function of zero-drift compensation parameter mapping, the equivalent stiffness and damping are dynamically calculated, a Kelvin–Voigt viscoelastic model is constructed, and a Kalman filter is combined to estimate and compensate for elastic deformation error in real time, which significantly suppresses the positioning deviation caused by structural flexibility and load changes. 3. By implementing all key algorithms on the FPGA using fixed-point parallel computing and using partial reconfiguration technology to refresh the corresponding functional areas as needed, the system bandwidth is further improved and the latency is reduced, balancing performance and power consumption, and ensuring rapid response under sudden load and lubrication state changes. Attached Figure Description

[0017] Figure 1 This is a flowchart of the fully closed-loop control method for the joint module of a humanoid robot based on FPGA according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0019] The term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone.

[0020] Example 1 Figure 1 The flowchart of the FPGA-based humanoid robot joint module full closed-loop control method disclosed in this embodiment is shown. The steps are as follows: S1. Perform second-order active low-pass filtering and programmable gain amplification on the analog signals of the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through ADC at an oversampling rate of 8 to 16 times the control bandwidth. During the programmable gain amplification process, the gain value G is based on the joint speed of the previous cycle. Dynamic adjustment, the adjustment rules are expressed as follows: ,in This indicates the minimum amplification factor of the PGA, corresponding to the case of minimum dynamic requirements and maximum signal amplitude in the system. This indicates the maximum amplification factor of the PGA, corresponding to the minimum signal amplitude in the system and the situation requiring maximum gain amplification. A higher angular velocity usually means more intense joint movement, and the corresponding sensor output (such as an encoder or gyroscope) signal amplitude will also be larger; conversely, a lower angular velocity indicates a smaller output amplitude at slow speeds or when stationary, representing a predefined maximum joint velocity reference value. ;when This indicates that the joint is at rest or moving very slowly. , Use minimum gain to prevent over-amplification of background noise; when This indicates that the joint has reached its maximum designed range of motion. , This ensures that the signal amplitude can reach the full input range of the ADC, maximizing the utilization of the quantization stages; through the synergistic effect of active filtering and PGA, high-frequency interference and quantization errors are effectively suppressed; the oversampling ADC improves the signal-to-noise ratio (SNR), providing a high-fidelity signal for subsequent digital algorithms.

[0021] S2. Construct a multidimensional feature sequence, including calculating the signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting the mechanical resonance frequency; Receive the digital signal output in step S1 The variance of the signal residual noise is expressed as: ,in, This represents the filtered estimated value; the mechanical resonant frequency is obtained by... Perform a short-time Fourier transform to identify the frequency corresponding to the amplitude peak. .

[0022] S3. Collect the temperature signals of the joint driver and encoder through the temperature sensor interface, and construct a temperature drift model to obtain the zero-point drift compensation parameters; The temperature drift model is obtained by collecting the temperature of the joint actuator. and encoder temperature Simultaneously considering the interactive effect of simultaneous changes in driver and encoder temperatures on drift, the zero-position drift function is obtained. ,in, and This represents the coefficients to be calibrated, and n and m represent the order, which are generally taken as 2 or 3 to balance fitting accuracy and computational complexity. This represents the coupling coefficient between the driver temperature and the encoder temperature. During the calibration phase, historical data (N values) are collected through temperature bench experiments. The samples, through regularized fitting, yielded... The real-time temperature vector is then substituted into the zero-point drift function to obtain the real-time zero-point drift compensation parameters. Angles measured using raw joint measurements Joint angles after zero-position drift compensation .

[0023] S4. After obtaining the driving torque, construct an elastic deformation model to obtain the joint elastic deformation error; The specific construction process of the elastic deformation model includes: obtaining the driving torque Drt through the motor current of the joint actuator, and calculating the equivalent stiffness online based on the temperature function obtained in advance through the zero-position drift compensation parameters. With viscoelastic damping coefficient To construct a Kelvin-Voigt first-order viscoelastic model reflecting the current thermal lubrication conditions, based on the generated discrete state-space equations, the Kalman filter uses the state estimate from the previous cycle and the torque input Drt from the current cycle to predict and update the current angular displacement error. This error estimate is immediately compared with the joint angle after zero-point drift compensation. Subtraction yields the compensation angle, which has been adjusted to eliminate the effects of static temperature bias and structural flexibility. The data is then sent to the closed-loop controller to generate the next drive command, thereby achieving dynamic real-time compensation for elastic deformation caused by thermal expansion, changes in lubrication conditions, and load.

[0024] By using a single-term and coupled-term polynomial model that simultaneously considers the temperatures of the driver and encoder, the static shift of the zero position due to temperature changes can be accurately fitted, achieving zero-position drift compensation at the microradian level and significantly reducing measurement errors caused by thermal expansion or temperature inhomogeneity. By using the zero-position drift compensation parameter mapping to calculate the equivalent stiffness and viscoelastic damping coefficient online, the Kelvin–Voigt model can reflect the joint dynamic characteristics under the current thermal-lubrication conditions in real time, avoiding the distortion problem of a single constant model when the temperature is high, low, or the lubrication state changes.

[0025] S5. Apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction; S6. Using the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and part of the reconfiguration region selection signal. The multidimensional feature sequence, the original joint angle after zero-position drift compensation, and the flexibility influence compensation angle after removing temperature drift are concatenated into a high-dimensional input vector, which is then input into a lightweight neural network that has been trained offline. The first layer of the network uses one-dimensional convolution to extract temporal features and combines feedback neuron structure to capture dynamic correlations. Then, through a series of quantized fully connected mappings, the hidden features are projected into the FPGA to output control parameters.

[0026] The hidden features in step S6 are projected onto four parallel output branches, including: The first branch outputs the optimal cutoff frequency of the IIR filter. , used to adjust the bandwidth of the real-time filter; The second branch outputs the integerized FIR window length L to reconstruct the time-domain response of the FIR filter; The third branch outputs the process noise covariance Q and measurement noise covariance R of the Kalman filter. It usually only outputs the diagonal elements and ensures positive definiteness through Softplus to dynamically update the signal estimation weights of the filter. The fourth branch generates classification probabilities for the reconfigurable logic regions that perform IIR, FIR, or EKF filtering through the Softmax layer, and selects the region selection signal Spr by selecting the term with the highest probability; this enables the local reconfiguration of the corresponding functional units at the hardware level. The entire inference process is completed on the FPGA using fixed-point arithmetic, significantly reducing each control cycle. This ensures rapid response to sudden load changes or temperature variations, while achieving fully automatic, real-time adaptive filtering and state estimation module parameters with extremely low storage and computational overhead, thereby continuously optimizing joint positioning accuracy and system stability.

[0027] S7. Based on the region selection signal output by the lightweight neural network, trigger a local refresh of the bitstream of the corresponding PR Region.

[0028] The threshold and weight settings can be set by default according to the present invention, or they can be set by those skilled in the art.

[0029] Example 2 Based on the same inventive concept as Embodiment 1, this invention provides a fully closed-loop control system for humanoid robot joint modules based on FPGA, comprising: The signal acquisition module is used to perform second-order active low-pass filtering and programmable gain amplification on the analog signals from the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through the ADC. The feature extraction and event monitoring module is used to construct multidimensional feature sequences, including calculating signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting mechanical resonance frequency; The temperature drift model construction module is used to acquire temperature signals from the joint actuator and encoder through the temperature sensor interface to obtain zero-point drift compensation parameters. The elastic deformation model construction module is used to construct an elastic deformation model after obtaining the driving torque to obtain the joint elastic deformation error; The compensation application module is used to apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction. The neural network decision module is used to take the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and part of the reconfiguration region selection signal. The PR control module is used to trigger a local refresh of the bitstream of the corresponding PR Region based on the region selection signal output by the lightweight neural network.

[0030] The PR control module also includes a filtering execution unit, which performs IIR, FIR or EKF filtering in the corresponding PR Region and transmits the filtered signal to the next-level closed-loop control module via the AXI-Stream bus or DMA channel.

[0031] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0032] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0033] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0034] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0035] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A fully closed-loop control method for humanoid robot joint modules based on FPGA, characterized in that: Includes the following steps: S1. Perform second-order active low-pass filtering and programmable gain amplification on the analog signals from the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through ADC; S2. Construct a multidimensional feature sequence, including calculating the signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting the mechanical resonance frequency; S3. Collect the temperature signals of the joint driver and encoder through the temperature sensor interface, and construct a temperature drift model to obtain the zero-point drift compensation parameters; S4. After obtaining the driving torque, construct an elastic deformation model to obtain the joint elastic deformation error; S5. Apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction; S6. Using the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and reconfiguration region selection signal. S7. Based on the region selection signal output by the lightweight neural network, trigger a local refresh of the bitstream of the corresponding PR Region; Step S6 concatenates the multidimensional feature sequence, the original joint angle after zero-position drift compensation, and the flexibility influence compensation angle after removing temperature drift into a high-dimensional input vector, which is then input into a lightweight neural network that has been trained offline in advance. The first layer of the network uses one-dimensional convolution to extract temporal features and combines a feedback neuron structure to capture dynamic correlations. Then, through a series of quantized fully connected mappings, the hidden features are projected onto the FPGA to output control parameters. The hidden features in step S6 are projected onto four parallel output branches, including: The first branch outputs the optimal cutoff frequency of the IIR filter. , used to adjust the bandwidth of the real-time filter; The second branch outputs the integerized FIR window length L to reconstruct the time-domain response of the FIR filter; The third branch outputs the process noise covariance Q and measurement noise covariance R of the Kalman filter to dynamically update the signal estimation weights of the filter. The fourth branch generates classification probabilities for reconfigurable logic regions that perform IIR, FIR, or EKF filtering through the Softmax layer, and selects the region selection signal Spr by selecting the term with the highest probability. Step S7 also includes a filtering execution unit, which performs IIR, FIR or EKF filtering in the corresponding PR Region and transmits the filtered signal to the next-level closed-loop control module via the AXI-Stream bus or DMA channel.

2. The FPGA-based closed-loop control method for humanoid robot joint modules according to claim 1, characterized in that: During the programmable gain amplification process, the gain value G is based on the joint speed of the previous cycle. Dynamic adjustment, the adjustment rules are expressed as follows: ,in This indicates the minimum magnification of the PGA. This indicates the maximum magnification of the PGA, and represents the predefined maximum joint velocity reference value. .

3. The FPGA-based closed-loop control method for humanoid robot joint modules according to claim 1, characterized in that: The temperature drift model is obtained by collecting the temperature of the joint actuator. and encoder temperature Simultaneously considering the interactive effect of simultaneous changes in driver and encoder temperatures on drift, the zero-position drift function is obtained. ,in, and This represents the coefficients to be calibrated, where n and m represent the order. This represents the coupling coefficient between the driver temperature and the encoder temperature. During the calibration phase, historical data (N values) are collected through temperature bench experiments. The samples, after being fitted using regularization, yielded... The real-time temperature vector is then substituted into the zero-point drift function to obtain the real-time zero-point drift compensation parameters. Angles measured using raw joint measurements Joint angles after zero-position drift compensation .

4. The FPGA-based closed-loop control method for humanoid robot joint modules according to claim 3, characterized in that: The specific construction process of the elastic deformation model includes: obtaining the driving torque Drt through the motor current of the joint actuator, and calculating the equivalent stiffness online based on the temperature function obtained in advance through the zero-position drift compensation parameters. With viscoelastic damping coefficient To construct a Kelvin-Voigt first-order viscoelastic model reflecting the current thermal-lubrication conditions, based on the generated discrete state-space equations, the Kalman filter uses the state estimate from the previous cycle and the torque input Drt from the current cycle to predict and update the current angular displacement error. This error estimate is immediately compared with the joint angle after zero-point drift compensation. Subtraction yields the compensation angle, which has been adjusted to eliminate the effects of static temperature bias and structural flexibility. Then, the commands are sent to the closed-loop controller to generate the next drive instructions.

5. The FPGA-based closed-loop control method for humanoid robot joint modules according to claim 1, characterized in that: Step S2 receives the digital signal output in step S1. The variance of the signal residual noise is expressed as ,in, This represents the filtered estimated value; the mechanical resonant frequency is obtained by... Perform a short-time Fourier transform to identify the frequency corresponding to the amplitude peak. .

6. A fully closed-loop control system for a humanoid robot joint module based on FPGA, used to execute the fully closed-loop control method for a humanoid robot joint module based on FPGA as described in any one of claims 1-5, characterized in that: include: The signal acquisition module is used to perform second-order active low-pass filtering and programmable gain amplification on the analog signals from the encoder, gyroscope and force sensor, and convert the processed signals into digital signals through the ADC. The feature extraction and event monitoring module is used to construct multidimensional feature sequences, including calculating signal residual noise variance, signal power, signal-to-noise ratio, joint velocity, acceleration, error and error change rate, and detecting mechanical resonance frequency; The temperature drift model construction module is used to acquire temperature signals from the joint actuator and encoder through the temperature sensor interface to obtain zero-point drift compensation parameters. The elastic deformation model construction module is used to construct an elastic deformation model after obtaining the driving torque to obtain the joint elastic deformation error; The compensation application module is used to apply the zero-position drift compensation parameters and elastic deformation error to the encoder signal for online offset correction. The neural network decision module is used to take the multidimensional feature sequence and the joint angle and the compensation angle for the influence of flexibility after zero-position drift compensation as input, and based on the lightweight neural network, the output includes the filter cutoff frequency, FIR window length, Kalman filter covariance matrix and reconfiguration region selection signal. The PR control module is used to trigger a local refresh of the bitstream of the corresponding PR Region based on the region selection signal output by the lightweight neural network.

Images