A bionic hand self-adaptive anti-skid control method based on multi-source sensor fusion

Abstract

A bionic hand self-adaptive anti-skid control method based on multi-source sensing fusion, comprising: establishing a kinematics forward and inverse solution model based on the geometric parameters of the metal link mechanism in the bionic hand, calculating a real-time Jacobian matrix, and constructing a system state space equation; a finger dynamics model is established, a loss current is calculated through a reference compensation function, and an operating space impedance control compliance reference based on an expected stiffness matrix and an expected damping matrix is established; when the bionic hand finger approaches the target, start double-link monitoring; after contact is established, the nonlinear feature extraction capability of the deep neural network model is used to output the slip probability reflecting the object micro-slip state; a nonlinear gain function based on the slip probability is constructed to dynamically correct the expected stiffness matrix and the expected damping matrix; after entering the steady-state grabbing stage, switch to the virtual self-locking holding mode. The present application solves the calculation explosion problem of the prior art, reduces the perception lag, and balances the compliance and anti-slip ability of grabbing.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically to a bionic hand adaptive anti-slip control method based on multi-source sensor fusion. Background Technology

[0002] Existing metal bionic hands face two major technical challenges when performing digital twin simulations and physical object grasping: First, the explosion of mechanical numerical calculations caused by multi-point contact (i.e., "pressure explosion") makes the underlying simulation engine extremely prone to crashing; second, existing control algorithms often struggle to balance the compliance and anti-slip capability of grasping when handling multi-dimensional force feedback, leading to the generation of false forces of hundreds of Newtons due to overcompensation when grasping lightweight targets, and even damaging the target object. Furthermore, traditional solutions heavily rely on large six-dimensional force sensors on the wrist, which not only increases system weight and hardware costs but also exhibits severe perception lag in complex grasping topologies. Summary of the Invention

[0003] In view of this, the present invention provides a bionic hand adaptive anti-slip control method based on multi-source sensor fusion to solve the computational explosion problem of existing technologies, reduce perception lag, and balance the compliance of grasping and anti-slip capability.

[0004] A bionic hand adaptive anti-slip control method based on multi-source sensor fusion includes: Step S1: The initial angular displacement of each DC motor of the bionic hand fingers is read by the joint encoder, and the original output signal of the tactile sensor array set on the fingertip surface is collected simultaneously to establish the environmental noise baseline. Then, the sensor zeroing calibration is performed. Then, the kinematic forward and inverse kinematic model is established based on the geometric parameters of the metal linkage mechanism in the bionic hand, and the real-time Jacobian matrix of the associated joint space and operation space is calculated to construct the system state space equation. Step S2: Establish a finger dynamics model including static friction, Coulomb friction and viscous friction terms, monitor the motor armature current in real time, calculate the loss current caused by the mechanism's self-weight and internal friction through the benchmark compensation function, and establish an operating space impedance control compliance benchmark based on the desired stiffness matrix and desired damping matrix. Use the real-time Jacobian matrix to map the desired command to the joint torque space. Step S3: When the bionic hand's fingers approach the target, dual-link monitoring is activated. In the first link, the first derivative of the motor's net current after removing the loss current is extracted for judgment. In the second link, the deviation of the normal stress of the tactile sensor array relative to the environmental noise baseline is extracted for judgment. When contact is determined to occur, the cooperative generalized momentum observer based on the system state space equation is activated. Then, the unified momentum residual vector of the whole hand is calculated, and the grasping matrix is ​​established by combining the contact topology relationship between the fingertip spatial pose and the target object determined by the forward and inverse kinematics model. The force on the fingertip is decoupled into the internal force term for maintaining grasping and the external force term for resisting environmental disturbances, thereby realizing a sensorless pure external force feedback closed loop. Step S4: After establishing contact, the normal stress and tangential stress fed back by the tactile sensing array are collected in real time to form a high-dimensional feature sequence. The temporal feature tensor is constructed through sliding window processing and input into the pre-trained deep neural network model. The nonlinear feature extraction capability of the deep neural network model is used to output the slip probability that reflects the micro-slip state of the object. Step S5: Construct a nonlinear gain function based on slip probability, dynamically correct the desired stiffness matrix and desired damping matrix, and when the slip probability exceeds the set safety threshold, activate the exponential amplification law to adjust the motor drive current, thereby increasing the fingertip centripetal gripping force to achieve instantaneous anti-slip lock-up compensation. Step S6: After entering the steady-state grasping stage, switch to the virtual self-locking holding mode; based on the expected damping matrix and the real-time Jacobian matrix, calculate the spatial consistency damping matrix mapped from the operating space to the joint space, introduce the LM adjustment factor in the damping mapping solution, and then output the high-frequency differential damping torque, combined with the micro-holding current to enter the low-power holding stage; synchronously maintain the monitoring of the tactile sensing array, and when the external stress change or slip characteristics are sensed, wake up the full closed-loop exponential compensation program.

[0005] The bionic hand adaptive anti-slip control method based on multi-source sensor fusion provided by the present invention has the following beneficial effects: 1. This invention establishes a finger dynamics model that includes static friction, Coulomb friction, and viscous friction terms, and establishes an operating space impedance control compliance benchmark based on the desired stiffness matrix and damping matrix by determining the loss current caused by the mechanism's self-weight and internal friction. This achieves a clean impedance control benchmark, allowing the contact force calculation to smoothly return to the physical reality range, effectively solving the calculation explosion problem and improving system stability.

[0006] 2. This invention innovatively integrates dual-link sensing of the first derivative of the motor net current and the deviation of the normal stress relative to the environmental noise baseline, and incorporates a collaborative generalized momentum observer. This enables the realization of a pure external force feedback closed loop with extremely low hysteresis without increasing the hardware cost of the wrist six-dimensional force sensor, thereby reducing the sensing hysteresis present in complex grasping topologies.

[0007] 3. This invention combines a nonlinear gain function based on slip probability with an exponential amplification law and a virtual self-locking mode to achieve a smooth transition from compliant envelope to instantaneous anti-slip lock-up compensation, effectively balancing the compliance of gripping and anti-slip capability. In addition, the low-power sleep mode and the full closed-loop mechanism of instant wake-up upon sensing stress mutation achieve effective energy saving and safety, and can take into account both high response and disturbance resistance and low power consumption operation. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the bionic hand adaptive anti-slip control method based on multi-source sensor fusion provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a bionic hand; Figure 3 The figure shows a comparison of the anti-slip robustness of the present invention and the traditional algorithm under a 1kg step load disturbance. Figure 4 The graph shows the response curve of normal stress versus rate of change of current under external force disturbance. Figure 5 This is the desired control current response curve under external disturbance. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0010] Please see Figure 1 The bionic hand adaptive anti-slip control method based on multi-source sensor fusion provided by the present invention includes steps S1 to S6: Step S1: The initial angular displacement of each DC motor of the bionic hand fingers is read by the joint encoder, and the original output signal of the tactile sensor array set on the fingertip surface is collected simultaneously to establish the environmental noise baseline. Then, the sensor zeroing calibration is performed. Then, the forward and inverse kinematic model is established based on the geometric parameters of the metal linkage mechanism in the bionic hand, and the real-time Jacobian matrix of the associated joint space and operation space is calculated to construct the system state space equation.

[0011] Please see Figure 2 In this embodiment, a metal bionic hand with 16 degrees of freedom is used as the experimental platform. The bionic hand 10 includes a metal palm base 20 and a plurality of bionic hand fingers 30 hinged to the metal palm base 20.

[0012] The bionic hand finger 30 includes a metal linkage skeleton, a DC drive motor located at the joint and its matching first joint encoder 31, second joint encoder 32, third joint encoder 33, and a composite sensing unit integrated on the surface of the fingertip 40.

[0013] Specifically, the composite sensing unit includes a high-resolution tactile sensing array 41 and a miniature six-dimensional torque sensor 42.

[0014] A high-resolution tactile sensor array 41 is used to sense the normal pressure distribution and tangential micro-displacement signal of the contact surface.

[0015] A miniature six-dimensional torque sensor 42 is positioned between the fingertip 40's distal skeleton and the tactile sensing array 41, forming a heterogeneous torque calibration link with the current closed loop of the DC drive motor. During offline calibration or under transient extremely high-frequency disturbances, the measured data from the miniature six-dimensional torque sensor 42 and the momentum residual vector output by the cooperative generalized momentum observer are compared. A heterogeneous verification link is constructed for online correction of the observer gain matrix and friction compensation coefficient.

[0016] In this embodiment, the control system of the bionic hand adopts a heterogeneous processing architecture based on FPGA+ARM, and realizes a 1 kHz control closed loop through the EtherCAT industrial bus. The tactile sensing array 41 adopts a flexible piezoelectric sensing film, and transmits an 8×8 stress distribution matrix through a high-frequency sampling chip; the miniature six-dimensional torque sensor 42 communicates with the main controller through the SPI bus, providing a sub-Newton level zero-drift calibration benchmark for algorithm observation.

[0017] After the control system is powered on, the main controller acquires the real-time angular displacement through the encoders of each joint. This serves as the initial vector for the joint. The raw electrical signal of the tactile sensing array 41 under no-load conditions for 200ms is synchronously acquired at high frequency, and its mean value is calculated as the environmental noise floor benchmark. A forward and inverse kinematics model is established based on the DH parameter table, and the real-time Jacobian matrix is ​​solved. This establishes the mapping relationship between the joint space angular velocity and the end-effector velocity in the operating space, thereby completing the parameterization of the system state space equations.

[0018] Step S2: Establish a finger dynamics model that includes static friction, Coulomb friction and viscous friction terms, monitor the motor armature current in real time, calculate the loss current caused by the mechanism's self-weight and internal friction through the benchmark compensation function, and establish an operating space impedance control compliance benchmark based on the desired stiffness matrix and desired damping matrix. Use the real-time Jacobian matrix to map the desired command to the joint torque space.

[0019] In this embodiment, for the linkage mechanism of the bionic hand finger 30, a recursive Newton-Euler method (RNEA) is used to establish a finger dynamics model that includes terms of static friction, Coulomb friction, and viscous friction. To avoid chattering caused by traditional sign functions at the zero-crossing point of velocity, this embodiment uses a hyperbolic tangent function. A smooth friction model is constructed by replacing sgn, and the baseline compensation function is obtained:

[0020] in, To reduce current loss, The torque constant of the motor. This represents the gravitational torque offset as attitude changes. The coefficient of friction is 0.015 in this embodiment; It is the hyperbolic tangent function. This represents the real-time angular displacement measured by the joint encoder. Represents the real-time joint angular velocity. As a smoothing factor for the basic system, The tactile stress coupling coefficient is... The deviation of the normal stress of the tactile sensing array from the environmental noise floor reference is expressed as an adaptively smoothed denominator. It can dynamically adjust the smooth gradient of friction compensation based on the contact stress intensity. It is the coefficient of viscous friction. , The link gravity offset parameters in the dynamic model were obtained by collecting motion data from the control system under no-load conditions and identifying them offline using the least squares method.

[0021] In step S2, during the process of mapping the desired command to the joint torque space using the real-time Jacobian matrix, the following equation is satisfied:

[0022] in, Indicates the motor output torque. Represents the real-time Jacobian matrix. Indicates transpose. Let be the desired stiffness matrix. This refers to the positional deviation of the fingertip in the operating space. This refers to the speed deviation of the fingertip in the operating space.

[0023] Step S3: When the bionic hand's fingers approach the target, dual-link monitoring is activated. In the first link, the first derivative of the motor's net current after removing the loss current is extracted for determination. In the second link, the deviation of the tactile sensor array's normal stress relative to the environmental noise baseline is extracted for determination. When contact is determined to occur, a cooperative generalized momentum observer based on the system's state-space equation is activated. Then, the unified momentum residual vector of the entire hand is calculated, and a grasping matrix is ​​established by combining the contact topology between the fingertip spatial pose determined by the kinematic forward and inverse kinematics model and the target object. The force on the fingertip is decoupled into an internal force term to maintain grasping and an external force term to resist environmental disturbances, thereby achieving a sensorless pure external force feedback closed loop.

[0024] Among them, when the bionic hand finger 30 approaches the target, high-frequency dual-link monitoring is activated.

[0025] In the first link, the first derivative of the motor net current is monitored in real time. When it exceeds the threshold of 2.0A / s, a preliminary judgment is triggered. The motor net current is obtained by subtracting the loss current from the detected total motor current, which can reduce the error caused by the motor's own weight and friction.

[0026] Specifically, in the first link, the first derivative of the motor net current after removing the loss current is extracted for determination, including: A time-invariant augmented state-space model based on the electromagnetic dynamics of a motor is constructed, and a steady-state Kalman filter is used to estimate the state of the high-frequency current signal. Innovation sequences between the filter-predicted observations and actual sampled values ​​are extracted online in real time. When the Mahalanobis distance of the innovation sequence exceeds a set test threshold... When this occurs, it is determined that an unmodeled transient external force collision or physical slip has occurred; Wherein, the test threshold The statistical anomaly determination boundary is based on the chi-square distribution and a preset confidence level, and its degrees of freedom match the dimensionality reduction observation dimension of the time-invariant augmented state-space model.

[0027] The deviation of the tactile sensing array 41 is monitored in the second link. When it breaks through the physical dead zone of 0.1N, the two-way fusion determination contact occurs.

[0028] Upon contact, a cooperative generalized momentum observer based on the system's state-space equations is activated, using the momentum residual vector. The force on the fingertip 40 is calculated, and heterogeneous verification is performed using measured data from the miniature six-dimensional torque sensor 42 to correct the observer gain matrix online. .

[0029] The momentum residual vector is calculated using the following formula:

[0030] in, Let be the momentum residual vector. Represents the observer gain matrix. This represents the generalized momentum vector. Indicates runtime. Represents the real-time Jacobian matrix. Indicates transpose. Indicates the motor output torque. The Jacobian matrix represents the local mapping from the fingertip coordinate system to the joint coordinate system. This represents the real-time rotational feedback value of the external force at the contact end, acquired by a six-dimensional torque sensor integrated into the fingertip surface, and is calculated by introducing a compensation term. This enables preliminary cancellation of unmodeled transient high-frequency disturbances; The sum of nonlinear terms, including nonlinear terms such as centrifugal force, Coriolis force, and gravity of the system; To represent the differential, To represent a time variable, it should be noted that... , , , All are about The variables can be detected or calculated using sensors.

[0031] Through momentum residual vector It can calculate the equivalent external load force acting on the fingertip 40.

[0032] Step S4: After establishing contact, the normal stress and tangential stress fed back by the tactile sensing array are collected in real time to form a high-dimensional feature sequence. The temporal feature tensor is constructed through sliding window processing and input into a pre-trained deep neural network model. The nonlinear feature extraction capability of the deep neural network model is used to output the slip probability that reflects the micro-slip state of the object.

[0033] Specifically, the mean normal stress of the tactile sensing array is extracted within a set sampling window (in this embodiment, the length is 20 sampling points, i.e., 40ms). tangential stress variance and stress change rate As feature vectors , The feature vector is input into a pre-trained deep neural network model containing two layers of LSTM units (64 hidden nodes per layer). The deep neural network model outputs a real number between 0 and 1 as the glide probability. This is used to reflect the risk of shedding.

[0034] Step S5: Construct a nonlinear gain function based on slip probability, dynamically correct the desired stiffness matrix and desired damping matrix, and when the slip probability exceeds the set safety threshold, activate the exponential amplification law to adjust the motor drive current, thereby increasing the fingertip centripetal gripping force to achieve instantaneous anti-slip lock-up compensation.

[0035] In this embodiment, a safety threshold is set. When the deep neural network model outputs At that time, according to the nonlinear exponential amplification law:

[0036] in, This represents the dynamically corrected desired stiffness matrix. Represents the basic compliance stiffness matrix. This represents the increment of the foundation stiffness (set to 12.0 N / m). This represents the exponential amplification penalty coefficient (set to 5.0).

[0037] This invention can increase the stiffness of the operating space in milliseconds and output extremely high centripetal anti-slip gripping force through Jacobian mapping.

[0038] Step S6: After entering the steady-state grasping stage, switch to the virtual self-locking holding mode; based on the expected damping matrix and the real-time Jacobian matrix, calculate the spatial consistency damping matrix mapped from the operating space to the joint space, introduce the LM adjustment factor in the damping mapping solution, and then output the high-frequency differential damping torque, combined with the micro-holding current to enter the low-power holding stage; synchronously maintain the monitoring of the tactile sensing array, and when the external stress change or slip characteristics are sensed, wake up the full closed-loop exponential compensation program.

[0039] Specifically, the spatial consistency damping matrix is ​​calculated using the following formula:

[0040] in, The spatially consistent damping matrix, Let be the desired damping matrix. As the LM regulating factor, The slip probability sensitivity coefficient, This represents the identity matrix that matches the joint degrees of freedom dimension.

[0041] In step S6, the monitoring of the tactile sensing array is maintained synchronously. When a sudden change in external stress or slippage characteristics is detected, the full closed-loop exponential compensation program is activated, specifically including: During the low-power maintenance phase, the stress standard deviation of the tactile sensing array is continuously monitored. When external disturbances are detected, causing Exceeding the wake-up threshold or slip probability When a transition occurs, the physical self-locking state is forcibly exited through a hardware interrupt, and the system instantly switches to the anti-slip lock-up compensation in step S5.

[0042] In this embodiment, after the steady-state grasping is maintained for more than 2 seconds, the system enters a low-power mode. This is based on the desired damping matrix. With real-time Jacobian matrix And introduce the LM (Levenberg-Marquardt) modulator. To prevent singularity values ​​from diverging, the current is reduced to 15% of its peak value, and the orientation is maintained using the physical self-locking property of the worm gear mechanism. The stress standard deviation of the tactile sensor array 41 is continuously monitored in the background. Once an external disturbance is detected that causes the value to exceed 2.5 times the initial steady-state value, the system immediately wakes up the CPU via a hardware interrupt and returns to the anti-slip lock-up compensation in step S5 to achieve high stiffness control.

[0043] Furthermore, in this embodiment, the method further includes: During the dynamic correction of the desired stiffness matrix, the full-range normal stress of the tactile sensing array and the angular velocity of the joint encoder are monitored simultaneously. When the sum of the normal stress drops to the noise floor threshold within a set time window and is accompanied by a sudden change in the joint angular velocity, it is determined that the grasping physical failure has occurred. At this time, a hardware-level circuit breaker is triggered, the desired stiffness matrix of the operating space is forced to zero, and peak sustaining damping is injected into the joint space.

[0044] To further verify the transient disturbance rejection capability of this invention in complex grasping environments, a 1kg step load extreme test was conducted on a physical prototype platform. To strictly ensure the purity of the excitation signal and eliminate low-frequency timing noise introduced by artificially applied disturbances, this embodiment designed and built a standardized transient excitation platform based on an electromagnetic release mechanism. Specifically, a DC electromagnet is rigidly fixed directly above the experimental platform to attract a 1kg standard weight; the lower end of the weight is connected to the vertical position of the center of mass of the grasped object through an inextensible polyarylamide (Kevlar) wire with a low elastic modulus, and the wire is initially kept slightly slack.

[0045] After the steady-state grasping is established, the main control system sends a hardware interrupt power-off signal to the electromagnet through a high-frequency I / O port, achieving sub-millisecond release of the weight without mechanical recoil. The level transition edge of this power-off signal is synchronously connected to the global data acquisition card as a hardware trigger reference at absolute zero. This physical excitation scheme ensures that the external force applied to the system is an ideal step function, providing a rigorous experimental mechanical verification closed loop for the underlying millisecond-level response evaluation.

[0046] The actual test results are as follows Figures 3 to 5 As shown.

[0047] Please see Figure 3 Under the same 1kg step load, the traditional single impedance control algorithm, due to its slow response and insufficient compensation, causes the object to continuously slip, with displacement exceeding 20mm and accompanied by obvious stick-slip effect, resulting in the failure of the grasping task. In contrast, the method of this invention achieves transient mechanical locking through a huge centripetal force only when the object slips by 1.8mm after the response is initiated, the curve is instantly flattened and accompanied by minimal underdamped oscillation.

[0048] Please see Figure 4 At time t0 (specifically 1.00 seconds), a disturbance is suddenly applied to the target object. Figure 4 When the curve representing the tactile normal stress just begins to rise, thanks to the first link of this invention, a peak value can be instantly triggered by the electromagnetic mutation characteristic, breaking the physical conduction hysteresis of traditional elastomer sensors. Furthermore, please refer to... Figure 5 After the system detects a disturbance, it expects the control current to increase exponentially within a very short response time (Δt=5ms), which proves that the method of this invention achieves extremely low response latency at the underlying hardware level.

[0049] In summary, the bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to the above embodiments has the following beneficial effects: 1. This invention establishes a finger dynamics model that includes static friction, Coulomb friction, and viscous friction terms, and establishes an operating space impedance control compliance benchmark based on the desired stiffness matrix and damping matrix by determining the loss current caused by the mechanism's self-weight and internal friction. This achieves a clean impedance control benchmark, allowing the contact force calculation to smoothly return to the physical reality range, effectively solving the calculation explosion problem and improving system stability.

[0050] 2. This invention innovatively integrates dual-link sensing of the first derivative of the motor net current and the deviation of the normal stress relative to the environmental noise baseline, and incorporates a collaborative generalized momentum observer. This enables the realization of a pure external force feedback closed loop with extremely low hysteresis without increasing the hardware cost of the wrist six-dimensional force sensor, thereby reducing the sensing hysteresis present in complex grasping topologies.

[0051] 3. This invention combines a nonlinear gain function based on slip probability with an exponential amplification law and a virtual self-locking mode to achieve a smooth transition from compliant envelope to instantaneous anti-slip lock-up compensation, effectively balancing the compliance of gripping and anti-slip capability. In addition, the low-power sleep mode and the full closed-loop mechanism of instant wake-up upon sensing stress mutation achieve effective energy saving and safety, and can take into account both high response and disturbance resistance and low power consumption operation.

[0052] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A bionic hand adaptive anti-slip control method based on multi-source sensor fusion, characterized in that, include: Step S1: The initial angular displacement of each DC motor of the bionic hand fingers is read by the joint encoder, and the original output signal of the tactile sensor array set on the fingertip surface is collected simultaneously to establish the environmental noise baseline. Then, the sensor zeroing calibration is performed. Then, the kinematic forward and inverse kinematic model is established based on the geometric parameters of the metal linkage mechanism in the bionic hand, and the real-time Jacobian matrix of the associated joint space and operation space is calculated to construct the system state space equation. Step S2: Establish a finger dynamics model including static friction, Coulomb friction and viscous friction terms, monitor the motor armature current in real time, calculate the loss current caused by the mechanism's self-weight and internal friction through the benchmark compensation function, and establish an operating space impedance control compliance benchmark based on the desired stiffness matrix and desired damping matrix. Use the real-time Jacobian matrix to map the desired command to the joint torque space. Step S3: When the bionic hand fingers approach the target, dual-link monitoring is activated. In the first link, the first derivative of the net motor current after removing the loss current is extracted for determination. In the second link, the deviation of the normal stress of the tactile sensor array relative to the environmental noise baseline is extracted for judgment. When contact is determined to occur, the cooperative generalized momentum observer based on the system state space equation is activated. Then, the momentum residual vector of the whole hand is calculated, and the grasping matrix is ​​established by combining the contact topology relationship between the fingertip spatial pose and the target object determined by the kinematic forward and inverse solution model. The force on the fingertip is decoupled into the internal force term to maintain grasping and the external force term to resist environmental disturbance, thereby realizing a sensorless pure external force feedback closed loop. Step S4: After establishing contact, the normal stress and tangential stress fed back by the tactile sensing array are collected in real time to form a high-dimensional feature sequence. The temporal feature tensor is constructed through sliding window processing and input into the pre-trained deep neural network model. The nonlinear feature extraction capability of the deep neural network model is used to output the slip probability that reflects the micro-slip state of the object. Step S5: Construct a nonlinear gain function based on slip probability, dynamically correct the desired stiffness matrix and desired damping matrix, and when the slip probability exceeds the set safety threshold, activate the exponential amplification law to adjust the motor drive current, thereby increasing the fingertip centripetal gripping force to achieve instantaneous anti-slip lock-up compensation. Step S6: After entering the steady-state grasping stage, switch to the virtual self-locking holding mode; based on the expected damping matrix and the real-time Jacobian matrix, calculate the spatial consistency damping matrix mapped from the operating space to the joint space, introduce the LM adjustment factor in the damping mapping solution, and then output the high-frequency differential damping torque, combined with the micro-holding current to enter the low-power holding stage; synchronously maintain the monitoring of the tactile sensing array, and when the external stress change or slip characteristics are sensed, wake up the full closed-loop exponential compensation program.

2. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 1, characterized in that, In step S2, the expression for the reference compensation function is: in, To reduce current loss, The torque constant of the motor. This represents the gravitational torque offset as attitude changes. The coefficient of friction is Coulomb. It is the hyperbolic tangent function. This represents the real-time angular displacement measured by the joint encoder. Represents the real-time joint angular velocity. As a smoothing factor for the basic system, The tactile stress coupling coefficient is... This represents the deviation of the normal stress of the tactile sensing array relative to the ambient noise floor reference. It is the coefficient of viscous friction.

3. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 2, characterized in that, In step S2, during the process of mapping the desired command to the joint torque space using the real-time Jacobian matrix, the following equation is satisfied: in, Indicates the motor output torque. Represents the real-time Jacobian matrix. Indicates transpose. Let be the desired stiffness matrix. This refers to the positional deviation of the fingertip in the operating space. This refers to the speed deviation of the fingertip in the operating space.

4. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 1, characterized in that, In the first link, the first derivative of the motor net current after removing the loss current is extracted for determination, specifically including: A time-invariant augmented state-space model based on the electromagnetic dynamics of a motor is constructed, and a steady-state Kalman filter is used to estimate the state of the high-frequency current signal. Innovation sequences between the filter-predicted observations and actual sampled values ​​are extracted online in real time. When the Mahalanobis distance of the innovation sequence exceeds a set test threshold... When this occurs, it is determined that an unmodeled transient external force collision or physical slip has occurred; Wherein, the test threshold The statistical anomaly determination boundary is based on the chi-square distribution and a preset confidence level, and its degrees of freedom match the dimensionality reduction observation dimension of the time-invariant augmented state-space model.

5. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 3, characterized in that, In step S3, the momentum residual vector is calculated using the following formula: in, Let be the momentum residual vector. Represents the observer gain matrix. This represents the generalized momentum vector. Indicates runtime. The Jacobian matrix represents the local mapping from the fingertip coordinate system to the joint coordinate system. This indicates the real-time rotational feedback value of the external force at the contact end, acquired by a six-dimensional torque sensor integrated into the fingertip surface. The sum of nonlinear terms, To represent the differential, Represents a time variable.

6. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 5, characterized in that, Step S4 specifically includes: The mean normal stress, variance of tangential stress, and rate of change of stress of the tactile sensing array within a set sampling window are extracted as feature vectors. These feature vectors are then input into a pre-trained deep neural network model containing a long short-term memory (LSM) layer or a fully connected layer. The deep neural network model outputs a real number between 0 and 1 as the slip probability. .

7. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 6, characterized in that, In step S6, the spatially consistent damping matrix is ​​calculated using the following formula: in, The spatially consistent damping matrix, Let be the desired damping matrix. As the LM regulating factor, The slip probability sensitivity coefficient, This represents the identity matrix that matches the joint degrees of freedom dimension.

8. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 7, characterized in that, In step S6, the monitoring of the tactile sensing array is maintained synchronously. When a sudden change in external stress or slippage characteristics is detected, the full closed-loop exponential compensation program is activated, specifically including: During the low-power maintenance phase, the stress standard deviation of the tactile sensing array is continuously monitored. When external disturbances are detected, causing Exceeding the wake-up threshold or slip probability When a transition occurs, the physical self-locking state is forcibly exited through a hardware interrupt, and the system instantly switches to the anti-slip lock-up compensation in step S5.

9. The bionic hand adaptive anti-slip control method based on multi-source sensor fusion according to claim 8, characterized in that, The method further includes: During the dynamic correction of the desired stiffness matrix, the full-range normal stress of the tactile sensing array and the angular velocity of the joint encoder are monitored simultaneously. When the sum of the normal stress drops to the noise floor threshold within a set time window and is accompanied by a sudden change in the joint angular velocity, it is determined that the grasping physical failure has occurred. At this time, a hardware-level circuit breaker is triggered, the desired stiffness matrix of the operating space is forced to zero, and peak sustaining damping is injected into the joint space.

Images