Torque adaptive control and collision detection method for flexible joint robot and system thereof

Abstract

The application discloses a flexible joint robot torque self-adaptive control and collision detection method and system, and relates to the field of industrial robots.The method comprises the following steps: obtaining joint motor temperature, double-path encoders, actual torque, end acceleration and excitation frequency data; updating self-adaptive stiffness parameters based on temperature data through a thermoelastic degradation model, establishing a benchmark dynamics model; using the model and encoder data to calculate a predicted torque, and calculating a difference between the predicted torque and the actual torque to generate original residual data; performing frequency domain filtering on acceleration and excitation frequency as prior features, eliminating low-frequency shaking and high-frequency noise, and extracting a net residual sequence; projecting the net residual sequence to a three-dimensional assembly task space, calculating spatial mutation energy in a bias direction, and outputting a collision signal when the spatial mutation energy exceeds a threshold value.The application eliminates thermal drift, accurately decouples multiple source disturbances, and avoids false positives caused by normal assembly forces.

Description

Technical Field

[0001] This invention relates to the field of industrial robot technology, and in particular to a method and system for adaptive torque control and collision detection of a flexible joint robot. Background Technology

[0002] With the upgrading of intelligent manufacturing, flexible joint robots are widely used in precision manufacturing scenarios such as aerospace casing assembly and complex component grinding due to their lightweight design and potential for compliant control. However, in actual complex industrial environments (such as those involving high-frequency vibrating tools, handling flexible swaying loads, and requiring long-term continuous operation), traditional sensorless collision detection methods face severe technical bottlenecks.

[0003] First, the dynamic stiffness benchmark of traditional control systems is usually set statically, without considering the decrease in the elastic modulus of the material caused by the heat generated by the motor and transmission components after the robot has been operating continuously for a long time (i.e., thermal drift). This will cause the benchmark dynamic model to gradually become inaccurate as the operating time increases, resulting in a systematic deviation in the calculated torque residual from the bottom layer.

[0004] Secondly, even if the torque residual is obtained, the residual data is often very messy under complex working conditions. It not only includes the actual external impact force, but also heavily contaminated with low-frequency disturbances caused by non-rigid load swaying, as well as high-frequency vibration noise generated by tool processing. Traditional single threshold filtering methods cannot effectively remove these multi-frequency noises, and are very likely to mask the true impact signal.

[0005] Finally, during interference fits or confined space operations, the robot's end effector will inevitably output a large normal contact force in the main assembly direction. Since traditional residual detection only operates at the one-dimensional torque level in the joint space, the system cannot distinguish whether this force is a reasonable assembly force or an unexpected lateral compression. This chain of problems—from model inaccuracies and signal aliasing to spatial force confusion—ultimately leads to frequent false collision alarms in complex, multi-source disturbance scenarios, severely hindering the large-scale application of flexible robots. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for adaptive torque control and collision detection of flexible joint robots, in order to solve the technical problems of frequent false alarms or missed alarms caused by the inaccuracy of the collision detection model reference and the difficulty in extracting real collision signals under the conditions of continuous long-term operation with thermal attenuation, multi-frequency process noise interference and complex assembly stress.

[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for adaptive torque control and collision detection in a flexible joint robot, using a processor as the execution entity, the method comprising:

[0008] S11. The processor acquires the joint motor temperature data, dual encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot.

[0009] S12. The processor inputs the joint motor temperature data into a preset thermoelastic degradation model for mapping calculation, generates a thermal attenuation correction coefficient, and updates the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficient to establish a benchmark dynamic model.

[0010] S13. The processor calculates the predicted torque data based on the reference dynamic model and the dual-channel encoder data, and extracts the difference between the actual torque data and the predicted torque data to generate the original residual data.

[0011] S14. The processor uses the terminal acceleration data and the excitation frequency data as prior feature variables, calls a preset frequency domain filtering algorithm to perform signal decomposition on the original residual data, removes the low-frequency swaying component corresponding to the terminal acceleration data and the high-frequency noise component corresponding to the excitation frequency data, and extracts the net residual sequence.

[0012] S15. The processor projects the net residual sequence into a preset three-dimensional assembly task space, calculates the spatial mutation energy of the net residual sequence in the bias direction, and outputs a collision determination signal when it determines that the spatial mutation energy exceeds a preset safety threshold.

[0013] Further, step S12 includes:

[0014] S21. The processor extracts the absolute temperature parameter at the current moment and the basic environmental reference temperature parameter from the joint motor temperature data.

[0015] S22. The processor obtains the preset material thermal softening constant and basic stiffness parameters, and then uses the equation... Calculate the thermal attenuation correction factor;

[0016] S23. The processor extracts the static stiffness diagonal matrix from the adaptive initial stiffness parameters, replaces the static stiffness diagonal matrix with a dynamic equivalent stiffness matrix containing the thermal attenuation correction coefficient, and completes the establishment of the benchmark dynamic model.

[0017] Further, step S14 includes:

[0018] S31. The processor performs a second time integration operation on the end acceleration data to generate a reference displacement feature vector;

[0019] S32. The processor uses a continuous wavelet transform algorithm to decompose the original residual data into a wavelet coefficient matrix containing multi-scale frequency domain features.

[0020] S33. The processor calculates the Pearson correlation coefficient between each scale row vector in the wavelet coefficient matrix and the reference displacement feature vector, and extracts the scale row vectors whose Pearson correlation coefficient is greater than a preset matching threshold.

[0021] S34. The processor performs wavelet inverse transform reconstruction on the extracted scale row vector to generate the low-frequency sway component.

[0022] Furthermore, step S14 also includes:

[0023] S41. The processor reads the sampling frequency of the control cycle and calculates the normalized digital angular frequency in combination with the excitation frequency data.

[0024] S42. The processor constructs the transfer function model of the infinite impulse response adaptive notch filter based on the normalized digital angular frequency and the preset bandwidth contraction factor.

[0025] S43. The processor inputs the intermediate residual sequence obtained by subtracting the low-frequency jitter component from the original residual data into the transfer function model, filters out the signal frequency band that coincides with the normalized digital angular frequency, and obtains the high-frequency noise component.

[0026] S44. The processor extracts the output sequence of the transfer function model and assigns the output sequence to the net residual sequence.

[0027] Further, step S15 includes:

[0028] S51. The processor reads the current real-time joint angle vector and calculates the pseudo-inverse matrix of the geometric Jacobian matrix under this posture.

[0029] S52. The processor uses the pseudo-inverse matrix to project the net residual sequence existing in the joint space to the Cartesian operation space to generate a three-dimensional external contact force vector.

[0030] S53. The processor obtains the expected assembly principal axis direction vector and normal constraint plane vector in the three-dimensional assembly task space, and constructs mutually orthogonal task projection operators and null space projection operators.

[0031] S54. The processor decomposes the three-dimensional external contact force vector into a first projection force in the expected assembly direction and a second projection force in the offset direction based on the task projection operator and the null space projection operator.

[0032] Furthermore, step S15 also includes:

[0033] S61. The processor extracts the normal velocity component of the end effector in the offset direction based on the positive kinematics model;

[0034] S62. The processor performs real-time time integration on the inner product of the second projected force and the normal velocity component to generate the spatial mutation energy.

[0035] S63. When the processor determines that the spatial mutation energy is greater than the preset safety threshold and the derivative of the second projected force modulus exceeds the preset impact extreme value, it determines that the force is abnormal and outputs the collision determination signal.

[0036] Furthermore, the method also includes:

[0037] S71. After outputting the collision determination signal, the processor obtains the joint drag residual vector within the time window after the collision.

[0038] S72. The processor calculates the time-series decay characteristics of the joint drag residual vector and the eigenvalues ​​of the direction covariance matrix.

[0039] S73. The processor substitutes the time series decay feature and the eigenvalues ​​of the directional covariance matrix into the intention evaluation equation to calculate the intention probability confidence.

[0040] S74. When the confidence level of the intent probability is greater than the preset decision threshold, the processor generates a human-guided decision result and outputs a zero-force drag command.

[0041] Furthermore, the method also includes:

[0042] S81. When the confidence level of the intent probability is less than or equal to the preset decision threshold, the processor generates a mechanical collision determination result.

[0043] S82. The processor retrieves the Cartesian space coordinate vector matrix of the interpolation cycle before the collision occurs as the breakpoint reference parameter.

[0044] S83. The processor uses the current position as the starting point and the breakpoint reference parameters as the ending point to generate a smooth trajectory recovery curve using a polynomial spline interpolation algorithm.

[0045] S84. The processor gradually increases the adaptive initial stiffness parameter according to the slowly varying function equation and outputs a drive control signal that follows the smooth trajectory recovery curve.

[0046] Another object of the present invention is to provide a torque adaptive control and collision detection system for a flexible joint robot, the system comprising:

[0047] The parameter acquisition module is configured to acquire joint motor temperature data, dual encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot.

[0048] The benchmark correction module is configured to input the joint motor temperature data into a preset thermoelastic degradation model for mapping calculation, generate a thermal attenuation correction coefficient, update the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficient, and establish a benchmark dynamic model.

[0049] The residual generation module is configured to calculate the predicted torque data based on the reference dynamic model and the dual encoder data, and extract the difference between the actual torque data and the predicted torque data to generate the original residual data.

[0050] The filtering and decoupling module is configured to use the end acceleration data and the excitation frequency data as prior feature variables, call a preset frequency domain filtering algorithm to decompose the original residual data, remove low-frequency sway components and high-frequency noise components, and extract the net residual sequence.

[0051] The space monitoring module is configured to project the net residual sequence into a preset three-dimensional assembly task space, calculate the spatial mutation energy of the net residual sequence in the bias direction, and output a collision determination signal when it is determined that the spatial mutation energy exceeds a preset safety threshold.

[0052] Furthermore, the filtering and decoupling module includes:

[0053] The low-frequency decoupling unit is configured to perform a second time integral operation on the end acceleration data to generate a reference displacement feature vector, and use a continuous wavelet transform algorithm to extract the components related to the reference displacement feature vector in the original residual data for reconstruction to generate the low-frequency sway component.

[0054] The high-frequency decoupling unit is configured to calculate the normalized digital angular frequency by combining the excitation frequency data, construct an adaptive notch filter, input the sequence after subtracting the low-frequency sway component into the adaptive notch filter, filter out the frequency band that coincides with the normalized digital angular frequency to generate the high-frequency noise component.

[0055] The sequence reconstruction unit is configured to extract the output of the adaptive notch filter, assign values ​​to it, and output it as the net residual sequence.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This application introduces the temperature parameter reflecting the heating effect into the dynamic modeling for the first time. By actively lowering the initial stiffness parameter of the system through thermoelastic mapping, this mechanism accurately compensates for the material thermal decay and softening caused by long-term high-load operation, ensuring that the benchmark dynamic model on which collision detection depends is always absolutely accurate.

[0058] 2. This application uses acceleration and excitation frequency as prior features to guide the directional removal of low-frequency interference caused by non-rigid load swaying and high-frequency noise caused by process tools in the original residual using a frequency domain filtering algorithm, and successfully extracts a pure net residual sequence, which greatly improves the signal-to-noise ratio of the micro-collision signal.

[0059] 3. This application achieves physical separation of force directions in the spatial dimension by projecting the purified one-dimensional net residual sequence onto the three-dimensional assembly task space. By monitoring energy only in the offset direction, the system can accurately accommodate normal insertion and extraction assembly forces and precisely intercept unexpected lateral compression, fundamentally eliminating erroneous shutdowns caused by confusion between process forces and illegal collisions. Attached Figure Description

[0060] Figure 1 This is a flowchart of a method for adaptive torque control and collision detection of a flexible joint robot provided in an embodiment of the present invention;

[0061] Figure 2 This is an architectural block diagram of a flexible joint robot torque adaptive control and collision detection system provided in an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram illustrating the principle of residual frequency domain decoupling and three-dimensional spatial projection under multi-source disturbance environment provided by an embodiment of the present invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Identical components are indicated by the same reference numerals.

[0064] It should be noted that the terms “front,” “back,” “left,” “right,” “up,” and “down” used in the following description refer to the directions shown in the attached diagram, while the terms “inside” and “outside” refer to the directions toward or away from the geometric center of a specific component, respectively.

[0065] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0066] This application provides a method for adaptive torque control and collision detection of a flexible joint robot, the method comprising the following steps:

[0067] Step S11: The processor acquires the joint motor temperature data, dual encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot.

[0068] Explanation: This step provides a comprehensive multi-source sensing foundation for subsequent dynamic physical modeling and signal decoupling. For readers outside this field, flexible joint robots typically contain transmission components prone to elastic deformation, such as harmonic reducers. To accurately sense the robot's internal and external states, the system must simultaneously acquire: joint motor temperature data reflecting joint heating (usually obtained through a thermistor built into the motor stator); dual-encoder data reflecting the degree of deformation within the joint (i.e., the angle difference between the motor input and output linkage); actual torque data reflecting the current motor output (indirectly estimated through the current sensor of the driver or directly measured by the torque sensor); end-effector acceleration data reflecting the macroscopic motion of the load in space (obtained through an IMU inertial measurement unit installed at the end); and excitation frequency data reflecting the inherent vibration characteristics of the external working tool (usually obtained through communication with the tool controller via an industrial bus).

[0069] Example: At the assembly site of an aircraft engine casing, the processor reads in real time that the temperature of a certain core joint is 85℃ (joint motor temperature data). The flexible wire harness that the robot is grasping is slightly swaying in mid-air (characterized by end-effector acceleration data). At the same time, the pneumatic riveting gun attached to the end is generating strong vibrations at a fixed frequency of 100Hz (excitation frequency data).

[0070] Step S12: The processor inputs the joint motor temperature data into the preset thermoelastic degradation model for mapping calculation, generates thermal attenuation correction coefficient, and updates the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficient to establish a benchmark dynamic model.

[0071] Explanation: Following step S11, this step aims to eliminate model inaccuracies caused by heat generation in the physical world. The core principle is that after prolonged high-load operation and subsequent temperature rise, the elastic modulus of the robot's metal and flexible transmission components decreases non-linearly (i.e., the physical properties soften). Traditional control systems typically assume the robot's stiffness is constant, leading to calculation errors. By using a thermoelastic degradation model, the joint motor temperature data is mapped to a thermal attenuation correction coefficient representing the degree of material softening, and the adaptive initial stiffness parameters within the control system are adjusted accordingly. The technical effect of this step is that it proactively adapts to the softening of mechanical stiffness caused by prolonged operation from a physical perspective, ensuring that the established benchmark dynamic model remains highly consistent with the actual physical state.

[0072] Example: After the robot works continuously for 8 hours, the joints reach 85°C. The processor inputs this temperature into the model and calculates that the high temperature causes the actual stiffness of the joint to decrease by 4% (thermal attenuation correction coefficient). Subsequently, the basic stiffness perceived by the system is lowered, and the mechanical prediction model is updated so that the robot's self-perception is completely synchronized with physical reality.

[0073] Step S13: The processor calculates the predicted torque data based on the reference dynamic model and the dual encoder data, and extracts the difference between the actual torque data and the predicted torque data to generate the original residual data.

[0074] Explanation: Building upon the accurate model established in step S12, this step aims to expose unknown external disturbances. The principle is to use the classic Newton-Euler equations of motion to calculate the theoretical torque (i.e., predicted torque data) that the robot should output when it encounters no external obstacles. When the actual torque data measured by sensors is subtracted from the theoretically required predicted torque data for its own motion, the remaining difference is the original residual data. This data represents the sum of all externally applied forces after eliminating the robot's own gravity, inertia, and friction.

[0075] Example: The processor calculates that at the current corrected stiffness of 85°C, lifting the arm should only require a force of 10 Nm (predicted torque data), but the current sensor indicates that 12 Nm was actually used (actual torque data). The difference between the two, resulting in 2 Nm, is the original residual data. This 2 Nm of unknown force may include impact force, harness swaying force, and rivet gun vibration force.

[0076] Step S14: The processor uses the terminal acceleration data and excitation frequency data as prior feature variables, calls a preset frequency domain filtering algorithm to decompose the original residual data, removes the low-frequency swaying component corresponding to the terminal acceleration data and the high-frequency noise component corresponding to the excitation frequency data, and extracts the net residual sequence.

[0077] Explanation: Following step S13, the generated raw residual data is extremely messy, and directly using it for collision detection would lead to a very high false alarm rate. This step is one of the core innovations, utilizing the orthogonality of different physical phenomena in frequency (i.e., frequency features do not overlap). Specifically, the physical swaying of loads (such as wire harnesses) typically exhibits extremely low frequency characteristics, while the vibration of tools (such as rivet guns) exhibits extremely high frequency characteristics. The system uses the acceleration and frequency obtained in step S11 as a priori guide, directing the frequency domain filtering algorithm to precisely remove the low-frequency swaying components and high-frequency noise components from the raw residual data, much like a scalpel. The technical effect of this step is to completely decouple the interference of non-rigid load swaying and high-frequency process vibration on collision detection, extracting only the net residual sequence representing pure external physical contact.

[0078] Example: The processor uses an algorithm to analyze the 2Nm of random forces and finds that 0.5Nm is the force generated by the slow swaying of the flexible wire harness (low-frequency swaying component), and another 0.5Nm is the violent vibration force generated by the 100Hz rivet gun (high-frequency noise component). After accurately removing these two, the remaining 1Nm is the pure force signal (net residual sequence) that we need to focus on monitoring.

[0079] Step S15: The processor projects the net residual sequence into the preset three-dimensional assembly task space, calculates the spatial mutation energy of the net residual sequence in the bias direction, and outputs a collision determination signal when it is determined that the spatial mutation energy exceeds the preset safety threshold.

[0080] Explanation: Following the pure signal extracted in step S14, this step, as another core innovation, solves the technical problem of confusing normal assembly forces with accidental collision forces. The principle is that the torque residual output in step S14 is one-dimensional and cannot distinguish the specific direction of force application. Through mathematical matrix projection technology, the one-dimensional torque of each joint is restored to a linear force in a three-dimensional Cartesian coordinate system. In the three-dimensional assembly task space, the system can distinguish between the normal processing direction where forces should be applied and the offset direction where forces should not be applied. The work quantity integral is performed on the force in the offset direction to obtain the spatial mutation energy. The technical effect of this step is that it perfectly avoids false alarms caused by high contact forces during normal assembly and achieves precise decoupling between process forces and accidental compression in the spatial dimension.

[0081] Example: The processor restores the purified 1Nm net residual to the three-dimensional space inside the casing and finds that the force occurs in the direction of the side wall of the casing (i.e., the offset direction), rather than the perpendicular insertion direction where the force should be applied. Furthermore, the energy generated by this force instantaneously (spatial mutation energy) exceeds the set safety limit. Therefore, it is immediately determined that this is not the insertion force of normal assembly, but an accidental impact on the side wall, and a collision judgment signal is issued.

[0082] In one embodiment of this application, step S12 includes:

[0083] Step S21: The processor extracts the absolute temperature parameter at the current moment and the basic environmental reference temperature parameter from the joint motor temperature data.

[0084] Explanation: This step is a further development of the thermoelastic degradation model in step S12. Since the degree of thermal decay of the robot depends not only on the current absolute high temperature but also on the environmental baseline, it is necessary to extract the current absolute temperature parameter and the baseline environmental reference temperature parameter representing room temperature to provide a reference temperature difference for the calculation.

[0085] Example: The processor confirms from the bus data that the current workshop room temperature is 25°C (basic environmental reference temperature parameter) and the current joint thermistor reading is 85°C (absolute temperature parameter).

[0086] Step S22: The processor obtains the preset material thermal softening constant and basic stiffness parameters, and then uses the equation... Calculate the thermal decay correction factor; Explanation: This step details the implementation of materials physics in the algorithm. In the formula, This represents the corrected current dynamic stiffness; These are the factory-set foundation stiffness parameters; It is the thermal softening constant of the material (which determines how quickly the material softens with temperature); This is an absolute temperature parameter; This serves as the base environmental reference temperature parameter. Using this formula, the system precisely quantifies the attenuation of material properties due to temperature differences, generating a thermal degradation correction coefficient.

[0087] Example: The processor retrieves the thermal softening constant of the special steel for the reducer, and combines it with the temperature difference (85℃-25℃) into the formula to accurately calculate that the current metal stiffness has decreased by 4% compared to the cold state (thermal attenuation correction coefficient).

[0088] Step S23: The processor extracts the static stiffness diagonal matrix from the adaptive initial stiffness parameters, replaces the static stiffness diagonal matrix with the dynamic equivalent stiffness matrix containing the thermal attenuation correction coefficient, and completes the establishment of the benchmark dynamic model.

[0089] Explanation: This step explains the underlying logic for replacing model parameters. The dynamic model in a control system typically represents a series of large matrix equations. The system dynamically overwrites the originally fixed static stiffness diagonal matrix with a dynamic equivalent stiffness matrix that incorporates thermal attenuation characteristics. This ensures that all subsequent mechanical derivations are based on the correct physical foundation.

[0090] Example: The processor enters the underlying computing module of the control system, deletes the original hard-coded factory stiffness matrix, and seamlessly replaces it with a new matrix that has just been calculated and is 4% softer, thus completing the update of the model base.

[0091] In one embodiment of this application, step S14 includes:

[0092] Step S31: The processor performs a second time integration operation on the end acceleration data to generate a reference displacement feature vector.

[0093] Explanation: This step is a detailed expansion of the low-frequency swaying logic in step S14. According to Newton's kinematics, displacement can be obtained by integrating acceleration twice over time. Therefore, integrating the terminal acceleration data is to grasp the true macroscopic physical swaying trajectory and rhythm of the external flexible load, i.e., the reference displacement characteristic vector.

[0094] Example: By performing two calculus calculations on the acceleration data transmitted from the sensor at the top of the wire harness, the processor concludes that the long wire harness is slowly swinging back and forth in space with an amplitude of 2 centimeters (reference displacement eigenvector).

[0095] Step S32: The processor uses the continuous wavelet transform algorithm to decompose the original residual data into a wavelet coefficient matrix containing multi-scale frequency domain features.

[0096] Explanation: This step introduces the core tool of time-frequency domain analysis—the Continuous Wavelet Transform. Because the traditional Fourier Transform lacks time-localization capabilities, the wavelet transform acts like a time-frequency microscope, converting a one-dimensional time-series signal into a two-dimensional matrix (i.e., the wavelet coefficient matrix), thus simultaneously revealing when torque fluctuations occur and their frequencies.

[0097] Example: The processor uses an algorithmic microscope to examine the original force signal containing various noises, and decomposes it into a color matrix spectrum (wavelet coefficient matrix) containing different frequency stripes and time coordinates.

[0098] Step S33: The processor calculates the Pearson correlation coefficient between each scale row vector in the wavelet coefficient matrix and the reference displacement eigenvector, and extracts the scale row vectors whose Pearson correlation coefficient is greater than the preset matching threshold.

[0099] Explanation: The logic of this step lies in signal pattern matching. The Pearson correlation coefficient is used to measure whether the changing trends of two data sequences are synchronized. The system compares the similarity of each frequency band signal in the wavelet matrix with the load swing displacement calculated in step S31. If the change curve of a certain frequency band signal closely matches the load swing curve, then the signal must be the interference force caused by physical shaking.

[0100] Example: The processor takes the rhythm curve of the wire harness swinging back and forth and compares it with each line of the spectrum in the graph. It finds that the fluctuation rhythm of one line has a similarity of up to 95% with the rhythm of the wire harness swinging (that is, the Pearson correlation coefficient is extremely high, exceeding the set matching threshold).

[0101] Step S34: The processor performs inverse wavelet transform reconstruction on the extracted scale row vectors to generate low-frequency sway components.

[0102] Explanation: After locking onto the interference source frequency band, this step uses inverse mathematical operations to restore the selected frequency domain features back to the time domain. The inverse wavelet transform translates the specific spectrum back into the true torque fluctuation waveform. The technical effect is to achieve a perfect mirror extraction of a specific physical disturbance source at the digital signal level.

[0103] Example: The processor takes this highly matched spectrum out of the grid and reverses it to reconstruct a torque curve that fluctuates over time. This is the pure interference force (low-frequency swaying component) caused by the swaying of the wiring harness.

[0104] In one embodiment of this application, step S14 further includes:

[0105] Step S41: The processor reads the sampling frequency of the control cycle and calculates the normalized digital angular frequency by combining it with the excitation frequency data.

[0106] Explanation: This step expands upon the high-frequency noise processing logic in step S14. To design a filter in a digital system, the Hertz (Hz) physical frequencies in the analog physical world must be mapped to a normalized scale that the digital system can recognize. By combining the inherent sampling frequency of the controller, the required normalized digital angular frequency for the filter is calculated.

[0107] Example: The processor knows that the vibration frequency of the pneumatic riveting gun is 100Hz (excitation frequency). The robot's control board calculates 1000 times per second (sampling frequency). Through proportional conversion, this physical frequency is converted into an internal angular frequency parameter that the digital filter can understand.

[0108] Step S42: The processor constructs the transfer function model of the infinite impulse response adaptive notch filter based on the normalized digital angular frequency and the preset bandwidth contraction factor.

[0109] Explanation: This step designs a dedicated digital signal trap. A notch filter is an extremely narrow-band band-stop filter specifically designed to suppress signals of a particular frequency without affecting other useful signals in the surrounding area. The normalized digital angular frequency determines the location of the trap, while the bandwidth contraction factor determines the width of the trap opening.

[0110] Example: The processor constructs an extremely narrow pit (notch filter) centered on the calculated angular frequency, and sets the aperture size specifically to accurately capture strong vibration energy around 100Hz.

[0111] Step S43: The processor inputs the intermediate residual sequence obtained by subtracting the low-frequency jitter component from the original residual data into the transfer function model, and filters out the frequency bands that coincide with the normalized digital angular frequency to generate high-frequency noise components.

[0112] Explanation: This step is a streamlined secondary purification process. Following step S34, the intermediate residual sequence, which has already eliminated low-frequency interference, is fed into the previously constructed notch filter model to completely eliminate the strong vibration interference caused by tool processing.

[0113] Example: The processor passes the remaining data stream after subtracting the wire harness swaying force through this trap, where the intense high-frequency torque caused by the rivet gun falls into the trap and is filtered out (generating high-frequency noise components).

[0114] Step S44: The processor extracts the output sequence of the transfer function model and assigns the output sequence as the net residual sequence.

[0115] Explanation: This step completes the closed-loop decoupling across the entire frequency band. After undergoing the dual process of low-frequency stripping and high-frequency absorption, the data flowing out of the filter is the cleanest, reflecting only the unexpected rigid contact between the robot structure and the external environment.

[0116] Example: After double cleaning, the originally messy forces containing shaking and tremors are now transformed into a set of very smooth, clean force data, which the processor formally names as the net residual sequence and saves.

[0117] In one embodiment of this application, step S15 includes:

[0118] Step S51: The processor reads the current real-time joint angle vector and calculates the pseudo-inverse matrix of the geometric Jacobian matrix under this posture.

[0119] Explanation: This step expands upon the basic preparation for spatial mapping in step S15. In robot kinematics, the transpose of the Jacobian matrix is ​​the only mathematical bridge connecting the forces acting on the end effector in three-dimensional space with the rotational forces acting on the motors of each joint. Finding its pseudo-inverse is to solve the nonlinear equations when the rotational forces of each joint are inversely derived as linear forces in space.

[0120] Example: The robot's arm is currently in a specific bent state (real-time joint angle). Based on this posture, the processor constructs a complex geometric codebook (the pseudo-inverse of the Jacobian matrix) to subsequently translate the rotational torque of the motor into an intuitive three-dimensional spatial thrust.

[0121] Step S52: The processor uses a pseudo-inverse matrix to project the net residual sequence existing in the joint space to the Cartesian operation space to generate a three-dimensional external contact force vector.

[0122] Explanation: Following step S51, since the previously calculated net residuals only represent the rotational torque on each motor bearing and cannot reflect the push-pull direction of the force in real space, the abstract torque is completely restored to a three-dimensional force that can be intuitively understood by humans by multiplying the net residual sequence with the pseudo-inverse matrix and projecting it.

[0123] Example: The processor uses a codebook to calculate the torque at each joint, which in the three-dimensional physical space inside the casing is equivalent to a total force pushing upwards and to the left (a three-dimensional external contact force vector).

[0124] Step S53: The processor obtains the expected assembly principal axis direction vector and normal constraint plane vector in the three-dimensional assembly task space, and constructs mutually orthogonal task projection operators and null space projection operators.

[0125] Explanation: To determine the legality of extracted contact forces, the system must first define spatial rules. This step introduces the concept of a task manifold, artificially dividing the system into legal main paths where forces must be applied (expected assembly main axis direction vector) and illegal lateral no-go zones where forces should absolutely not be applied (normal constraint plane vector). The projection operator is the mathematical tool used to perform this classification.

[0126] Example: The process requires the wire harness to be inserted vertically downwards into the housing. The processor defines the downward Z-axis as the legal direction in which force must be applied, while the horizontal X-axis and Y-axis planes perpendicular to it are defined as lateral no-go zones where force should not be applied. Based on this, the processor generates a classification tool (projection operator).

[0127] Step S54: The processor decomposes the three-dimensional external contact force vector into a first projection force in the expected assembly direction and a second projection force in the offset direction based on the task projection operator and the null space projection operator.

[0128] Explanation: By using the classification operator generated in step S53 to orthogonally decompose the total force, the system breaks down the composite force into legitimate work components and illegitimate disturbance components. The technical effect is that even when the spindle is under extremely high force during normal operation, it will not obscure or interfere with the system's independent monitoring of minor abnormal forces in the lateral offset direction.

[0129] Example: The processor decomposes the calculated composite force and finds that the reasonable insertion and extraction force (first projected force) along the Z-axis downward is large, but at the same time, the unreasonable force (second projected force) that squeezes the side wall of the casing laterally also exists. The system successfully separates the two completely.

[0130] In one embodiment of this application, step S15 further includes:

[0131] Step S61: The processor extracts the normal velocity component of the end effector in the bias direction based on the positive kinematics model.

[0132] Explanation: This step is a refinement of the mutation energy calculation logic in step S15. In physics, the prerequisite for work to generate energy is the presence of a force, and that this force generates velocity (displacement) in its direction. Therefore, in addition to obtaining the abnormal force in the bias direction, it is also necessary to extract the robot's actual movement velocity in that abnormal direction through kinematic equations.

[0133] Example: The processor not only calculates the illegal force that the robot has on the sidewall of the casing, but also calculates the robot end effector that is currently sliding toward the sidewall at a speed of 10 millimeters per second (normal velocity component).

[0134] Step S62: The processor performs real-time time integration on the inner product of the second projected force and the normal velocity component to generate spatial mutation energy.

[0135] Explanation: Following step S61, this step provides the core physical basis for determining whether a collision reaches a dangerous level. The inner product of the force vector and velocity vector in the offset direction (i.e., calculating the power) is performed, and the result is continuously integrated over time to obtain the accumulated work output. This effectively filters out transient torque spikes caused by accidental electromagnetic interference, avoiding false alarms.

[0136] Example: The processor multiplies the lateral squeezing force by the speed of movement toward the sidewall and accumulates it over time to accurately calculate the destructive energy (spatial mutation energy) accumulated by this squeezing action.

[0137] Step S63: When the processor determines that the spatial mutation energy is greater than the preset safety threshold and the derivative of the calculated second projected force modulus exceeds the preset impact extreme value, it determines that the force is abnormal and outputs a collision determination signal.

[0138] Explanation: This step establishes a strict double-insurance judgment standard. The first is the excess of energy due to the inequality, indicating that the destructive force is large enough; the second is the differentiation calculation of the force (called Jerk jerk in physics), the derivative representing the steepness of the force increase. Only when the energy accumulation is large enough and the force burst is sudden and sudden is it ultimately confirmed as a dangerous collision.

[0139] Example: The processor detected that the energy accumulated laterally had exceeded the upper limit that the casing sidewall could withstand, and this force increased by an extremely high value in a fraction of a second (the derivative was extremely steep, exceeding the impact limit). Therefore, it was determined that a violent impact had occurred, and an alarm (collision determination signal) was immediately sent to the system bus.

[0140] In one embodiment of this application, the torque adaptive control and collision detection method for a flexible joint robot further includes:

[0141] Step S71: After outputting the collision determination signal, the processor obtains the joint drag residual vector within the time window after the collision.

[0142] Explanation: This step focuses on the system's intelligent processing extension after a collision. After detecting a collision and automatically entering impedance softening (i.e., the robot's muscles relax and soften) to protect against external forces, the system does not stop sensing. Instead, it continues to collect force data in the softened state for a set observation period (post-collision time window) to assess the subsequent trajectory of the external force.

[0143] Example: After a collision alarm is issued, the robot's arm instantly deflates and softens. Within 2 seconds of softening (the post-collision time window), the processor continues to silently record the passive force data (joint drag residual vector) transmitted from each joint.

[0144] Step S72: The processor calculates the time-series decay characteristics of the joint drag residual vector and the eigenvalues ​​of the direction covariance matrix;

[0145] Explanation: This step quantifies the physical characteristics of residual external forces. The time-series decay characteristics reflect how quickly the force disappears; the eigenvalues ​​of the directional covariance matrix statistically reflect the dispersion of the force direction (human continuous pushing and pulling usually results in stable direction with minimal variance, while the rebound from an impact of an inanimate object shows chaotic direction).

[0146] Example: The processor analyzed the data within these 2 seconds and found that the external force did not instantly return to zero, but rather its magnitude remained very stable (decaying very slowly), and this force continued to push the robot's arm backward steadily (with extremely small directional variance).

[0147] Step S73: The processor substitutes the time series decay features and the eigenvalues ​​of the direction covariance matrix into the intent evaluation equation to calculate the intent probability confidence.

[0148] Explanation: Following step S72, this step uses a specific mathematical evaluation function to integrate the extracted attenuation characteristics and directional stability into a score out of 100. A higher score indicates a greater probability that the external disturbance was intentionally caused by a human operator guiding or pushing / pulling the robot, rather than a falling object hitting the robot.

[0149] Example: The processor substitutes the two features of slow decay and extremely stable orientation into the evaluation formula and calculates a probability score (intention probability confidence) as high as 98%.

[0150] Step S74: When the confidence level of the intent probability is greater than the preset decision threshold, the processor generates a human-guided decision result and outputs a zero-force drag command.

[0151] Explanation: The technical effect of this step is to achieve a seamless intelligent connection between collision safety defense and human-machine collaboration. Once the system confirms that the operator intends to push the robot away, it no longer simply stops passively, but actively cooperates with the operator's pushing force to enter a weightless following state, greatly improving the smoothness of human-machine collaboration.

[0152] Example: Because the score of 98% far exceeds the set passing score, the processor understands that although a side wall collision alarm occurred, the current force is from the assembly worker next to it trying to push the robot's arm away from the danger zone. Therefore, the processor immediately guides the robot to smoothly enter a fully following state (zero-force drag command) following the worker's force.

[0153] In one embodiment of this application, the torque adaptive control and collision detection method for a flexible joint robot further includes:

[0154] Step S81: When the confidence level of the intent probability is less than or equal to the preset decision threshold, the processor generates a mechanical collision determination result.

[0155] Explanation: Corresponding to claim 7, this step deals with another possible branch after the collision: if the post-collision evaluation score is extremely low, it means that it was not actively guided by humans, but merely a brief, unattended physical collision.

[0156] Example: Suppose that within 2 seconds of softening, the processor detects that the external force instantly becomes zero and the direction is chaotic, calculates the intention score as extremely low at 5%, and then determines that it was just that the wiring harness connector accidentally rubbed against the edge of the casing (mechanical impact judgment result), and there were no workers around to intervene.

[0157] Step S82: The processor retrieves the Cartesian space coordinate vector matrix of the interpolation cycle before the collision occurs as the breakpoint reference parameter.

[0158] Explanation: After confirming that it was only a minor collision and there was no human interaction, the system needs to autonomously recover to ensure continuous automated production. The purpose of this step is to use the historical cache of the control system to find the absolutely precise spatial position in the extremely short instant before the accidental collision, as the respawn point for subsequent recovery movements.

[0159] Example: After confirming that the site is safe, the processor immediately retrieves the exact coordinates of the terminal above the chassis 1 millisecond before the alarm occurred from memory, and locks this coordinate as the target resurrection point (breakpoint reference parameter).

[0160] Step S83: The processor uses the current position as the starting point and the breakpoint reference parameters as the ending point to generate a smooth trajectory recovery curve using a polynomial spline interpolation algorithm.

[0161] Explanation: The principle behind this step is that after a collision and softening, the robotic arm typically experiences a slight displacement of a few centimeters due to inertia or its own gravity. To ensure a safe and smooth return to the resurrection point, a high-order mathematical curve (such as a quintic spline algorithm that guarantees absolute continuity of position, velocity, and acceleration) must be used for transitional guidance to avoid secondary mechanical oscillations.

[0162] Example: Because the arm went limp and deviated from its original position, the processor used a smooth, high-level mathematical curve, similar to an S-shape, to plan a return route, ensuring that the robot would not experience any jitter when returning from the drooping point to the resurrection point (smooth trajectory recovery curve).

[0163] Step S84: The processor gradually increases the adaptive initial stiffness parameter according to the slowly varying function equation and outputs a drive control signal that follows the recovery curve of the smooth trajectory.

[0164] Explanation: As the final step in the fully automated processing, this step completes the system's transition from a defensive state to a normal operating state. Utilizing a gradually varying equation similar to a sigmoid curve, it ensures that joint stiffness recovers silently and gradually during the reset process, rather than instantaneously tightening and causing a destructive impact. The technical effect is fully automated, second-level self-healing of faults without human intervention.

[0165] Example: While the processor directs the robot to return to its original position along a smooth curve, the servo motors in the joints, following the rhythm of the gradually changing equation, return to their initial working state from a limp state in just a few seconds, and then automatically continue the assembly action that was not completed before, all in a smooth and seamless manner.

[0166] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it.

[0167] Example 1, refer to Figure 1 This invention provides a method for adaptive torque control and collision detection in a flexible joint robot, with a processor as the main execution unit. This embodiment focuses on the collision detection process under long-term operation and complex vibration conditions. The method specifically includes the following steps:

[0168] Step S11: Obtain the joint motor temperature data, dual-channel encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot. These multi-source physical parameters provide the underlying foundation for subsequently constructing an anti-disturbance system.

[0169] Step S12: Input the joint motor temperature data into the preset thermoelastic degradation model for mapping calculation, generate thermal attenuation correction coefficients, and update the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficients to establish a benchmark dynamic model. This step corrects the softening of mechanical stiffness caused by long-term operation from a physical level.

[0170] Step S13: Calculate the predicted torque data based on the baseline dynamic model and the dual-channel encoder data, and extract the difference between the actual torque data and the predicted torque data to generate the original residual data. The extracted original residual data represents all external unknown forces after removing the power required for the robot's own thermodynamic motion.

[0171] Step S14: Using the end-acceleration data and excitation frequency data as prior feature variables, a preset frequency domain filtering algorithm is called to decompose the original residual data, removing the low-frequency swaying component corresponding to the end-acceleration data and the high-frequency noise component corresponding to the excitation frequency data, and extracting the net residual sequence. This step achieves precise mathematical removal of non-rigid load swaying and process vibration.

[0172] Step S15: Project the net residual sequence into the preset three-dimensional assembly task space, calculate the spatial mutation energy of the net residual sequence in the bias direction, and output a collision determination signal when it is determined that the spatial mutation energy exceeds the preset safety threshold.

[0173] To further explain the core decoupling principle of steps S14 and S15 in this invention, refer to... Figure 3 , Figure 3 A schematic diagram illustrating the principle of residual frequency domain decoupling and three-dimensional spatial projection under multi-source disturbance environment is shown. When the system processes the aliased raw residual data, it first combines the low-frequency characteristics reflected by the end acceleration data to remove the low-frequency sway component caused by load swing in the frequency domain; then, it combines the excitation frequency data to filter out the high-frequency noise component caused by the operating tool.

[0174] After frequency domain purification, the one-dimensional cluttered signal is refined into a net residual sequence. Then, it enters the spatial projection stage: the net residual sequence is mapped to the three-dimensional assembly task space using a projection algorithm. Within this three-dimensional coordinate system, the purified torque is orthogonally decomposed into legitimate process forces (no alarm required) in the expected assembly direction and illegal contact forces in the offset direction. The system calculates the spatial abrupt change energy only for the offset direction, thus accurately outputting a collision determination signal while avoiding normal assembly forces.

[0175] Example 2, refer to Figure 2Based on the same inventive concept as the aforementioned method embodiments, Embodiment 2 of the present invention provides a torque adaptive control and collision detection system for a flexible joint robot. This system is implemented using code logic or integrated circuitry within the processor 100, and specifically includes: a parameter acquisition module 110, executed by the processor 100, configured to acquire joint motor temperature data, dual-channel encoder data, actual torque data, end-effector acceleration data, and excitation frequency data of the flexible joint robot.

[0176] The reference correction module 120, executed by the processor 100, is configured to input the joint motor temperature data into a preset thermoelastic degradation model for mapping calculation, generate thermal attenuation correction coefficients, update the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficients, and establish a reference dynamic model.

[0177] The residual generation module 130, executed by the processor 100, is configured to calculate the predicted torque data based on the reference dynamic model and the dual encoder data, and extract the difference between the actual torque data and the predicted torque data to generate the original residual data.

[0178] The filtering and decoupling module 140, executed by the processor 100, is configured to use the terminal acceleration data and excitation frequency data as prior feature variables, call a preset frequency domain filtering algorithm to decompose the original residual data, remove low-frequency sway components and high-frequency noise components, and extract the net residual sequence.

[0179] The space monitoring module 150, executed by the processor 100, is configured to project the net residual sequence into a preset three-dimensional assembly task space, calculate the spatial mutation energy of the net residual sequence in the bias direction, and output a collision determination signal when it is determined that the spatial mutation energy exceeds a preset safety threshold.

[0180] It should be understood that the above reference numerals are only for illustrating the logical structure of each functional module of the system. In actual applications, the parameter acquisition module 110, the benchmark correction module 120, the residual generation module 130, the filtering and decoupling module 140, and the space monitoring module 150 can all be integrated into the processor chip 100 of the main controller to run together, so as to ensure the millisecond-level real-time response rate under multi-scale complex calculations.

[0181] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for adaptive torque control and collision detection of a flexible joint robot, characterized in that, With the processor as the execution entity, the method includes: S11. The processor acquires the joint motor temperature data, dual encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot. S12. The processor inputs the joint motor temperature data into a preset thermoelastic degradation model for mapping calculation, generates a thermal attenuation correction coefficient, and updates the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficient to establish a benchmark dynamic model. S13. The processor calculates the predicted torque data based on the reference dynamic model and the dual-channel encoder data, and extracts the difference between the actual torque data and the predicted torque data to generate the original residual data. S14. The processor uses the terminal acceleration data and the excitation frequency data as prior feature variables, calls a preset frequency domain filtering algorithm to perform signal decomposition on the original residual data, removes the low-frequency swaying component corresponding to the terminal acceleration data and the high-frequency noise component corresponding to the excitation frequency data, and extracts the net residual sequence. S15. The processor projects the net residual sequence into a preset three-dimensional assembly task space, calculates the spatial mutation energy of the net residual sequence in the bias direction, and outputs a collision determination signal when it determines that the spatial mutation energy exceeds a preset safety threshold.

2. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 1, characterized in that, Step S12 includes: S21. The processor extracts the absolute temperature parameter at the current moment and the basic environmental reference temperature parameter from the joint motor temperature data. S22. The processor obtains the preset material thermal softening constant and basic stiffness parameters, and then uses the equation... Calculate the thermal attenuation correction factor; S23. The processor extracts the static stiffness diagonal matrix from the adaptive initial stiffness parameters, replaces the static stiffness diagonal matrix with a dynamic equivalent stiffness matrix containing the thermal attenuation correction coefficient, and completes the establishment of the benchmark dynamic model.

3. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 1, characterized in that, Step S14 includes: S31. The processor performs a second time integration operation on the end acceleration data to generate a reference displacement feature vector; S32. The processor uses a continuous wavelet transform algorithm to decompose the original residual data into a wavelet coefficient matrix containing multi-scale frequency domain features. S33. The processor calculates the Pearson correlation coefficient between each scale row vector in the wavelet coefficient matrix and the reference displacement feature vector, and extracts the scale row vectors whose Pearson correlation coefficient is greater than a preset matching threshold. S34. The processor performs wavelet inverse transform reconstruction on the extracted scale row vector to generate the low-frequency sway component.

4. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 3, characterized in that, Step S14 further includes: S41. The processor reads the sampling frequency of the control cycle and calculates the normalized digital angular frequency in combination with the excitation frequency data. S42. The processor constructs the transfer function model of the infinite impulse response adaptive notch filter based on the normalized digital angular frequency and the preset bandwidth contraction factor. S43. The processor inputs the intermediate residual sequence obtained by subtracting the low-frequency jitter component from the original residual data into the transfer function model, filters out the signal frequency band that coincides with the normalized digital angular frequency, and obtains the high-frequency noise component. S44. The processor extracts the output sequence of the transfer function model and assigns the output sequence to the net residual sequence.

5. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 1, characterized in that, Step S15 includes: S51. The processor reads the current real-time joint angle vector and calculates the pseudo-inverse matrix of the geometric Jacobian matrix under this posture. S52. The processor uses the pseudo-inverse matrix to project the net residual sequence existing in the joint space to the Cartesian operation space to generate a three-dimensional external contact force vector. S53. The processor obtains the expected assembly principal axis direction vector and normal constraint plane vector in the three-dimensional assembly task space, and constructs mutually orthogonal task projection operators and null space projection operators. S54. The processor decomposes the three-dimensional external contact force vector into a first projection force in the expected assembly direction and a second projection force in the offset direction based on the task projection operator and the null space projection operator.

6. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 5, characterized in that, Step S15 further includes: S61. The processor extracts the normal velocity component of the end effector in the offset direction based on the positive kinematics model; S62. The processor performs real-time time integration on the inner product of the second projected force and the normal velocity component to generate the spatial mutation energy. S63. When the processor determines that the spatial mutation energy is greater than the preset safety threshold and the derivative of the second projected force modulus exceeds the preset impact extreme value, it determines that the force is abnormal and outputs the collision determination signal.

7. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 1, characterized in that, The method also includes: S71. After outputting the collision determination signal, the processor obtains the joint drag residual vector within the time window after the collision. S72. The processor calculates the time-series decay characteristics of the joint drag residual vector and the eigenvalues ​​of the direction covariance matrix. S73. The processor substitutes the time series decay feature and the eigenvalues ​​of the directional covariance matrix into the intention evaluation equation to calculate the intention probability confidence. S74. When the confidence level of the intent probability is greater than the preset decision threshold, the processor generates a human-guided decision result and outputs a zero-force drag command.

8. The method for adaptive torque control and collision detection of a flexible joint robot according to claim 7, characterized in that, The method also includes: S81. When the confidence level of the intent probability is less than or equal to the preset decision threshold, the processor generates a mechanical collision determination result. S82. The processor retrieves the Cartesian space coordinate vector matrix of the interpolation cycle before the collision occurs as the breakpoint reference parameter. S83. The processor uses the current position as the starting point and the breakpoint reference parameters as the ending point to generate a smooth trajectory recovery curve using a polynomial spline interpolation algorithm. S84. The processor gradually increases the adaptive initial stiffness parameter according to the slowly varying function equation and outputs a drive control signal that follows the smooth trajectory recovery curve.

9. A torque adaptive control and collision detection system for a flexible joint robot, applied to the torque adaptive control and collision detection method for a flexible joint robot as described in any one of claims 1-8, characterized in that, include: The parameter acquisition module is configured to acquire joint motor temperature data, dual encoder data, actual torque data, end effector acceleration data, and excitation frequency data of the flexible joint robot. The benchmark correction module is configured to input the joint motor temperature data into a preset thermoelastic degradation model for mapping calculation, generate a thermal attenuation correction coefficient, update the adaptive initial stiffness parameters of the system based on the thermal attenuation correction coefficient, and establish a benchmark dynamic model. The residual generation module is configured to calculate the predicted torque data based on the reference dynamic model and the dual encoder data, and extract the difference between the actual torque data and the predicted torque data to generate the original residual data. The filtering and decoupling module is configured to use the end acceleration data and the excitation frequency data as prior feature variables, call a preset frequency domain filtering algorithm to decompose the original residual data, remove low-frequency sway components and high-frequency noise components, and extract the net residual sequence. The space monitoring module is configured to project the net residual sequence into a preset three-dimensional assembly task space, calculate the spatial mutation energy of the net residual sequence in the bias direction, and output a collision determination signal when it is determined that the spatial mutation energy exceeds a preset safety threshold.

10. The flexible joint robot torque adaptive control and collision detection system according to claim 9, characterized in that, The filtering and decoupling module includes: The low-frequency decoupling unit is configured to perform a second time integral operation on the end acceleration data to generate a reference displacement feature vector, and use a continuous wavelet transform algorithm to extract the components related to the reference displacement feature vector in the original residual data for reconstruction to generate the low-frequency sway component. The high-frequency decoupling unit is configured to calculate the normalized digital angular frequency by combining the excitation frequency data, construct an adaptive notch filter, input the sequence after subtracting the low-frequency sway component into the adaptive notch filter, filter out the frequency band that coincides with the normalized digital angular frequency to generate the high-frequency noise component. The sequence reconstruction unit is configured to extract the output of the adaptive notch filter, assign values ​​to it, and output it as the net residual sequence.

Images