Mechanical arm inverse dynamics fusion control method based on DFNN and EFS

Abstract

The embodiment of the invention discloses a mechanical arm inverse dynamics fusion control method based on DFNN and EFS. The method comprises the steps that a real-time joint state data set of a mechanical arm in the multi-joint movement process is collected; then, constructing a DFNN feedforward control channel, performing nonlinear mapping processing on the mechanical arm inverse dynamic model through a dynamic fuzzy neural network to generate a DFNN feedforward control signal, constructing an EFS error compensation channel, and performing real-time capture and compensation processing on a dynamic error in a mechanical arm movement process through an error feedback compensation algorithm to generate an EFS error compensation signal; a PD feedback control channel is constructed, and closed-loop adjustment processing is conducted on the deviation of the current joint state and the target joint state of the mechanical arm through a proportional differential control algorithm to generate a PD feedback control signal; and the three-channel signals are input into a three-channel parallel fusion network, dynamic fusion processing is conducted through an adaptive weight distribution algorithm, and a mechanical arm joint driving control instruction set is generated and sent to an actuator to drive multiple joints to move cooperatively.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, specifically to a robotic arm inverse dynamics fusion control method based on DFNN and EFS. Background Technology

[0002] In the field of robotic arm control, there are various types of traditional control methods. One type is the analytical control method based on precise modeling. This method constructs analytical dynamic expressions through system identification or dynamic modeling, and then designs a feedforward controller. For example, it uses the Euler-Lagrange equation or the Newton-Euler method to obtain the explicit relationship between joint torque and state variables. This method has strong interpretability and good control performance, but it is highly dependent on system parameters in practical engineering and is difficult to adapt to structural changes, frictional disturbances, and nonlinear dynamics. Once modeling errors exist, high-gain feedback controllers are required for compensation, which can easily cause system oscillations. Furthermore, model updates are difficult, and online adaptability is lacking.

[0003] Another approach is the offline inverse modeling method based on deep neural networks. This method collects system trajectory sample data and uses neural networks to learn the mapping function between joint states and desired control quantities. Representative methods include using fully connected networks (FCN), convolutional neural networks (CNN), or long short-term memory networks (LSTM) for predictive control. These methods alleviate the model dependency problem to some extent, but the model structure is fixed after training, lacks the ability to adapt to new tasks or disturbances, and has limited generalization ability. The performance of networks trained offline degrades under different tasks or conditions, cannot update weights or structures in real time, and is difficult to cope with non-stationary inputs or dynamic environmental changes.

[0004] Another approach is online control and learning based on fuzzy logic, which combines structurally adaptive evolutionary fuzzy systems (EFS) and fuzzy neural networks (FNN) to achieve online modeling and correction of inverse dynamic errors. However, in the initial stage, there is a lack of effective prior models, and the controller structure expands due to feedback errors. The online learning efficiency is low, the rule growth rate is fast but the fusion mechanism is weak, and rule redundancy is easily caused. It usually only achieves error correction rather than complete inverse modeling, making it difficult to use as a standalone feedforward control method.

[0005] In summary, existing robotic arm control technologies struggle to simultaneously achieve modeling accuracy, real-time response, structural flexibility, and generalization ability, failing to meet the dual requirements of trajectory tracking and control performance in complex dynamic environments. Summary of the Invention

[0006] This invention provides a fusion control method for inverse dynamics of a robotic arm based on DFNN and EFS.

[0007] In a first aspect, embodiments of the present invention provide a fusion control method for inverse dynamics of a robotic arm based on DFNN and EFS, applied to a fusion control system for inverse dynamics of a robotic arm based on DFNN and EFS, the method comprising:

[0008] The system collects a set of real-time joint state data during the multi-joint motion of the robotic arm. The set of real-time joint state data includes a sequence of joint angles, a sequence of joint angular velocities, and a sequence of joint angular accelerations within a continuous motion cycle.

[0009] Based on the real-time joint state data set, a DFNN feedforward control channel is constructed. The inverse dynamics model of the robotic arm is nonlinearly mapped through a dynamic fuzzy neural network to generate a DFNN feedforward control signal.

[0010] An EFS error compensation channel is constructed based on the real-time joint state data set. The dynamic error during the movement of the robotic arm is captured and compensated in real time through an error feedback compensation algorithm to generate an EFS error compensation signal.

[0011] Based on the real-time joint state data set, a PD feedback control channel is constructed. The deviation between the current joint state and the target joint state of the robotic arm is adjusted in a closed loop using a proportional-derivative control algorithm to generate a PD feedback control signal.

[0012] The DFNN feedforward control signal, the EFS error compensation signal, and the PD feedback control signal are input into a three-channel parallel fusion network. The network is dynamically fused using an adaptive weight allocation algorithm to generate a set of robotic arm joint drive control commands. The set of robotic arm joint drive control commands is then sent to the robotic arm actuator to drive multi-joint coordinated motion.

[0013] Secondly, embodiments of the present invention provide a robotic arm inverse dynamics fusion control system based on DFNN and EFS, comprising:

[0014] processor;

[0015] Storage device, on which computer programs are stored,

[0016] When the computer program is executed by the processor, the processor implements any of the described DFNN and EFS-based inverse dynamics fusion control methods for robotic arms.

[0017] This invention provides a readable storage medium storing a program or instructions, which, when executed by a processor, implement the steps of the DFNN and EFS-based inverse dynamics fusion control method for a robotic arm.

[0018] This invention, through the acquisition of real-time joint state data sets of multi-joint motion of a robotic arm, constructs a DFNN feedforward control channel, an EFS error compensation channel, and a PD feedback control channel. The three-channel signals are then input into a three-channel parallel fusion network for dynamic fusion processing, generating a set of robotic arm joint drive control commands to drive the coordinated motion of multiple joints. This invention improves the accuracy of robotic arm inverse dynamics modeling, reduces the system's dependence on feedback control, and utilizes a dynamic fuzzy neural network to perform nonlinear mapping on the inverse dynamics model, constructing a feedforward model with higher accuracy and stronger generalization ability. The control method provided by this invention possesses structural adaptive capability, enabling continuous online learning. The error feedback compensation algorithm can capture and compensate for dynamic errors in real time, allowing the controller structure to adaptively evolve according to dynamic error changes. This invention also achieves deep fusion of feedforward modeling and feedback correction, dynamically fusing the three-channel signals through an adaptive weight allocation algorithm to form a closed-loop information flow, enhancing system stability and response efficiency. Furthermore, this invention improves initial stage control performance, reduces system oscillation and learning lag, and uses a proportional-derivative control algorithm to perform closed-loop adjustment of deviations, combined with error feedback compensation, improving the accuracy of initial error estimation and reducing learning time and structural expansion. Attached Figure Description

[0019] Figure 1 The flowchart illustrates a robotic arm inverse dynamics fusion control method based on DFNN and EFS, provided in an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the basic structure of a robotic arm inverse dynamics fusion control system based on DFNN and EFS provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] See Figure 1 As shown, this figure is a flowchart of a robotic arm inverse dynamics fusion control method based on DFNN and EFS provided by an embodiment of the present invention. This method can be applied to a robotic arm inverse dynamics fusion control system based on DFNN and EFS. Figure 1 As shown, the method may include steps 110-150.

[0023] Step 110: Collect a set of real-time joint state data of the robotic arm during multi-joint motion. The set of real-time joint state data includes the joint angle sequence, joint angular velocity sequence and joint angular acceleration sequence within a continuous motion cycle.

[0024] In this embodiment of the invention, high-precision sensors are used to monitor the joint states in real time when the robotic arm performs multi-joint movements. For example, in an industrial production scenario, the robotic arm needs to complete a series of complex grasping and placing actions. During its movement, the sensors continuously record the joint angles, joint angular velocities, and joint angular accelerations within consecutive movement cycles. This data is organized into a sequence, forming a real-time joint state data set. The sensors collect data at a fixed sampling frequency to ensure that the collected data accurately reflects the state changes of the robotic arm throughout the entire movement process. The collected joint angle sequence clearly shows the rotation angle of each joint at different times, the joint angular velocity sequence reflects the changes in joint rotation speed, and the joint angular acceleration sequence reflects the rate of change of joint speed.

[0025] Step 120: Construct a DFNN feedforward control channel based on the real-time joint state data set, and perform nonlinear mapping processing on the inverse dynamics model of the robotic arm through a dynamic fuzzy neural network to generate a DFNN feedforward control signal.

[0026] In this embodiment of the invention, after acquiring the real-time joint state data set, a DFNN feedforward control channel is constructed. The dynamic fuzzy neural network possesses nonlinear mapping capabilities, enabling accurate description of the robotic arm's inverse dynamics model. Taking a scenario where the robotic arm is handling an object as an example, this network can correlate the real-time joint state data with the robotic arm's inverse dynamics model, finding the nonlinear relationship between the two. During this process, the network processes the input real-time joint state data, simulating the robotic arm's dynamic behavior under different states, thereby generating a DFNN feedforward control signal. This signal contains the expected driving torque information required by each joint of the robotic arm in the current state.

[0027] As one embodiment, step 120 includes:

[0028] Step 121: Perform kinematic feature extraction processing on the real-time joint state data set to obtain the kinematic feature set of each joint of the robotic arm. The kinematic feature set includes the joint angle change rate, angular velocity direction consistency parameter, and angular acceleration fluctuation amplitude.

[0029] In this embodiment of the invention, the collected real-time joint state data set needs to be further extracted for kinematic features. Taking the robotic arm performing an assembly task as an example, its joint angle sequence is analyzed to calculate the change in joint angles per unit time, obtaining the joint angle change rate. This rate of change reflects the speed and trend of joint rotation. Simultaneously, by processing the joint angular velocity sequence, the consistency parameter of the angular velocity direction is calculated, which measures the stability of the angular velocity direction during joint movement. Furthermore, statistical analysis of the angular acceleration sequence yields the angular acceleration fluctuation amplitude, which reflects the degree of change in joint angular acceleration. These kinematic features can more deeply describe the motion state of each joint of the robotic arm.

[0030] Step 122: Construct the input layer nodes of the dynamic fuzzy neural network based on the kinematic feature set. The number of input layer nodes is consistent with the number of joint degrees of freedom, and each input layer node corresponds to a combination of kinematic features of a joint.

[0031] In this embodiment of the invention, after obtaining the kinematic feature set of each joint of the robotic arm, the input layer of the dynamic fuzzy neural network is constructed. Taking a robotic arm with multiple joint degrees of freedom as an example, the number of joint degrees of freedom determines the number of input layer nodes. Each input layer node corresponds to a combination of kinematic features of a joint, using features such as the joint angle change rate, angular velocity direction consistency parameter, and angular acceleration fluctuation amplitude of that joint as input. In this way, the input layer can accurately receive the kinematic information of each joint. The setting of input layer nodes ensures that the network can comprehensively acquire the motion state of each joint of the robotic arm, thereby better performing nonlinear mapping of the robotic arm's inverse dynamics model.

[0032] Step 123: The kinematic feature set is processed by the fuzzification layer of the dynamic fuzzy neural network to generate fuzzy feature vectors. The membership function mapping process uses a Gaussian membership function to calculate the fuzzy membership value of each kinematic feature.

[0033] In this embodiment of the invention, after the kinematic feature set enters the fuzzification layer of the dynamic fuzzy neural network, it undergoes membership function mapping processing. Taking the kinematic features of the robotic arm performing a set operation as an example, a Gaussian membership function is used to process each kinematic feature. For the joint angle change rate, the parameters of the Gaussian membership function are determined according to its value range and distribution, and the membership values ​​of the feature belonging to different fuzzy sets are calculated. Similarly, the angular velocity direction consistency parameter and the angular acceleration fluctuation amplitude are processed in a similar way. In this way, each kinematic feature is converted into a fuzzy membership value, and these values ​​are then combined into a fuzzy feature vector. This vector can more accurately describe the fuzziness and uncertainty of the motion state of each joint of the robotic arm.

[0034] In a preferred embodiment, step 123 includes:

[0035] Step 1231: Standardize the joint angle change rate in the kinematic feature set to obtain the normalized angle change rate. The standardization process uses the max-min normalization method to map the angle change rate to a preset numerical range.

[0036] In this embodiment of the invention, the joint angle change rates in the kinematic feature set need to be standardized to make them more suitable for subsequent processing. Taking the joint angle change rates of a robotic arm under different motion modes as an example, a maximum-minimum normalization method is used. First, the maximum and minimum values ​​in the joint angle change rate sequence are found. Then, according to a preset numerical range, each angle change rate value is linearly transformed to map it into that range. The purpose of this is to eliminate the dimensional differences between different joint angle change rates and bring them to the same scale. The normalized angle change rate obtained after standardization can more accurately reflect the relative magnitude of joint angle changes.

[0037] Step 1232: Determine the center parameter and width parameter of the Gaussian membership function based on the normalized angle change rate. The center parameter is calculated based on the mean of the joint angle change rate, and the width parameter is calculated based on the standard deviation of the joint angle change rate.

[0038] In this embodiment of the invention, after obtaining the normalized angle change rate, the parameters of the Gaussian membership function are determined. Taking the normalized angle change rate of a robotic arm during a continuous motion process as an example, the mean of the sequence is calculated and used as the central parameter of the Gaussian membership function. This central parameter represents the average level of the joint angle change rate. Simultaneously, the standard deviation of the normalized angle change rate is calculated and used as the width parameter. The standard deviation reflects the dispersion of the data, and the width parameter determines the shape and width of the Gaussian membership function. Appropriate central and width parameters enable the Gaussian membership function to more accurately describe the distribution of the joint angle change rate, providing a more reasonable basis for calculating fuzzy membership values.

[0039] Step 1233: Calculate the membership value of the normalized angle change rate using the Gaussian membership function to generate the angle change rate membership vector.

[0040] In this embodiment of the invention, after determining the center parameter and width parameter of the Gaussian membership function, the membership value of the normalized angle change rate is calculated. Taking the normalized angle change rate of the robotic arm at different times as an example, each angle change rate value is substituted into the Gaussian membership function to calculate its membership value belonging to different fuzzy sets. These membership values ​​reflect the degree of matching between the angle change rate and different fuzzy concepts. Combining the membership values ​​at all times forms an angle change rate membership vector, which can more clearly show the distribution of the joint angle change rate in different fuzzy sets.

[0041] Step 1234: Perform direction encoding processing on the angular velocity direction consistency parameters in the kinematic feature set to obtain a direction encoding vector. The direction encoding processing uses a direction cosine matrix to convert the three-dimensional angular velocity direction into two-dimensional planar coordinates.

[0042] In this embodiment of the invention, the angular velocity direction consistency parameter in the kinematic feature set needs to be oriented using direction encoding. Taking the angular velocity direction of a robotic arm performing complex motion in space as an example, a direction cosine matrix is ​​used to convert the three-dimensional angular velocity direction into two-dimensional planar coordinates. In three-dimensional space, the angular velocity direction of the robotic arm joint has three dimensions of information. Through the transformation of the direction cosine matrix, it can be projected onto a two-dimensional plane, reducing the dimensionality of the data. This not only simplifies the data representation but also preserves the main features of the angular velocity direction. The converted two-dimensional coordinates are combined into a direction encoding vector, which can be more conveniently used in subsequent processing, providing suitable input for calculating the membership value of the angular velocity direction consistency parameter.

[0043] Step 1235: Determine the boundary parameters of the triangular membership function based on the direction encoding vector, calculate the membership value of the angular velocity direction consistency parameter through the triangular membership function, and generate the direction consistency membership vector.

[0044] In this embodiment of the invention, after obtaining the direction encoding vector, the boundary parameters of the triangular membership function are determined. Taking the direction encoding vectors of the robotic arm at different motion stages as an example, the left and right boundaries and vertex positions of the triangular membership function are determined based on the value range and distribution of the vectors. These boundary parameters determine the shape and range of the triangular membership function. Then, each value in the direction encoding vector is substituted into the triangular membership function to calculate its membership value belonging to different fuzzy sets. These membership values ​​reflect the degree of matching between the angular velocity direction consistency parameter and different fuzzy concepts. All membership values ​​are combined to form a direction consistency membership vector, which can more intuitively show the distribution of the angular velocity direction in different fuzzy sets.

[0045] Step 1236: Perform frequency domain analysis on the angular acceleration fluctuation amplitude in the kinematic feature set to obtain the fluctuation frequency feature vector. The frequency domain analysis uses Fast Fourier Transform to extract the main frequency component of the angular acceleration signal.

[0046] In this embodiment of the invention, the amplitude of angular acceleration fluctuations in the kinematic feature set is analyzed in the frequency domain. Taking the angular acceleration signal of a robotic arm performing a large vibration motion as an example, a Fast Fourier Transform (FFT) is used to convert the time-domain angular acceleration signal to the frequency domain. In the frequency domain, different frequency components of the signal are separated. By analyzing the frequency domain signal, the dominant frequency components are extracted, which represent the main frequency characteristics of the angular acceleration fluctuations. These dominant frequency components are combined into a fluctuation frequency feature vector, which can more clearly show the distribution of angular acceleration fluctuations at different frequencies.

[0047] Step 1237: Determine the slope parameter of the S-shaped membership function based on the fluctuation frequency feature vector, calculate the membership value of the angular acceleration fluctuation amplitude through the S-shaped membership function, and generate the fluctuation amplitude membership vector.

[0048] In this embodiment of the invention, after obtaining the fluctuation frequency feature vector, the slope parameter of the S-shaped membership function is determined. Taking the fluctuation frequency feature vector of a robotic arm under different workloads as an example, the slope of the S-shaped membership function is determined based on the distribution and variation of the dominant frequency component in the vector. The slope parameter determines the rate of change of the S-shaped membership function. Then, the angular acceleration fluctuation amplitude is substituted into the S-shaped membership function to calculate its membership values ​​belonging to different fuzzy sets. These membership values ​​reflect the degree of matching between the angular acceleration fluctuation amplitude and different fuzzy concepts. All membership values ​​are combined to form a fluctuation amplitude membership vector, which can more accurately show the distribution of the angular acceleration fluctuation amplitude in different fuzzy sets.

[0049] Step 1238: The membership vector of the angle change rate, the membership vector of the direction consistency, and the membership vector of the fluctuation amplitude are concatenated along the feature dimension to generate the fuzzy feature vector. The dimension of the fuzzy feature vector is equal to the sum of the dimensions of each membership vector.

[0050] In this embodiment of the invention, after obtaining the membership vectors for the rate of change of angle, the consistency of direction, and the amplitude of fluctuation, they are concatenated along the feature dimensions. Taking these three membership vectors of the robotic arm in one complete motion cycle as an example, they are connected sequentially according to the order of the feature dimensions. This integrates the fuzzy information of different kinematic features into a more comprehensive fuzzy feature vector. The dimension of this vector is equal to the sum of the dimensions of the three membership vectors, and it contains comprehensive fuzzy information about the joint angle changes, angular velocity directions, and angular acceleration fluctuations of each joint of the robotic arm.

[0051] Step 124: Input the fuzzy feature vector into the rule inference layer of the dynamic fuzzy neural network, perform fuzzy inference processing based on the preset fuzzy control rule library, and generate a fuzzy rule activation intensity matrix. The fuzzy rule library contains a set of fuzzy rules that map the joint motion state to the control quantity.

[0052] In this embodiment of the invention, the concatenated fuzzy feature vector is input into the rule inference layer of a dynamic fuzzy neural network. Taking the fuzzy feature vector of a robotic arm performing a set task as an example, the rule inference layer performs fuzzy inference based on a preset fuzzy control rule library. This fuzzy control rule library contains a large number of fuzzy rules, which describe the mapping relationship between joint motion states and control quantities. For example, a rule might stipulate that when the rate of change of joint angle is large and the consistency of angular velocity direction is low, the control quantity of a certain joint needs to be adjusted accordingly. During the inference process, the rule inference layer matches the membership values ​​in the fuzzy feature vector with the antecedent conditions in the rule library to determine which rules are activated. Then, inference calculations are performed based on the activated rules, ultimately generating a fuzzy rule activation intensity matrix. The row dimension of this matrix is ​​equal to the number of fuzzy rules, and the column dimension is equal to the number of joint degrees of freedom, reflecting the activation degree of each rule on each joint.

[0053] In one implementation, step 124 includes:

[0054] Step 1241: Parse the fuzzy control rule base and extract a set of fuzzy rules containing antecedent conditions and consequent conclusions. The antecedent conditions are combinations of fuzzy linguistic variables of kinematic features, and the consequent conclusions are fuzzy linguistic variables of joint control quantities.

[0055] In this embodiment of the invention, a parsing operation is required for the preset fuzzy control rule base. Taking a fuzzy control rule base containing multiple robotic arm motion scenarios as an example, a set of fuzzy rules containing antecedent conditions and consequent conclusions is extracted from it. The antecedent conditions are composed of fuzzy linguistic variables of kinematic features, such as "large joint angle change rate" and "low angular velocity direction consistency," which are fuzzy descriptions of kinematic features. The consequent conclusions are fuzzy linguistic variables of joint control quantities, such as "increase joint driving torque" and "decrease joint rotation speed." The purpose of parsing the rule base is to clearly extract these rules, providing a foundation for subsequent rule matching and inference calculations. By parsing the rule base, the specific content of each rule can be clarified, and the correspondence between antecedent conditions and consequent conclusions can be understood, so as to accurately perform matching and inference when inputting fuzzy feature vectors.

[0056] Step 1242: Calculate the rule matching degree for each membership vector in the fuzzy feature vector, determine the degree of matching between the antecedent conditions of each fuzzy rule and the current kinematic feature, and generate the rule matching degree vector.

[0057] In this embodiment of the invention, after extracting the fuzzy rule set, the rule matching degree is calculated for each membership vector in the fuzzy feature vector. Taking the fuzzy feature vector of a robotic arm at a certain moment as an example, the membership vectors of angle change rate, direction consistency, and fluctuation amplitude are matched with the antecedent conditions of each fuzzy rule. For each rule, the degree of matching between the membership value of the current kinematic feature and the rule requirement is calculated based on the requirements of each fuzzy linguistic variable in its antecedent conditions. For example, if the antecedent condition of the rule is "large joint angle change rate and low angular velocity direction consistency", the degree of matching between the rule and the current kinematic feature is calculated based on the values ​​in the angle change rate membership vector and the direction consistency membership vector. The matching degrees of all rules are combined to form a rule matching degree vector, which can intuitively show the degree of fit between each rule and the current kinematic feature.

[0058] Step 1243: Construct a rule trigger threshold based on the rule matching degree vector. When the rule matching degree is greater than the preset threshold, determine that the fuzzy rule is activated and generate a rule activation flag matrix.

[0059] In this embodiment of the invention, after obtaining the rule matching degree vector, it is necessary to construct a rule triggering threshold. Taking the rule matching degree vector of a robotic arm in different working modes as an example, a preset threshold is determined based on experience and actual needs. This threshold is the standard for determining whether a rule is activated. Each matching degree value in the rule matching degree vector is compared with the preset threshold. When the matching degree of a rule is greater than the threshold, the fuzzy rule is determined to be activated. The activation status of all rules is represented in matrix form to generate a rule activation flag matrix. The elements in the matrix are Boolean values, where 1 indicates that the rule is activated and 0 indicates that the rule is not activated. This matrix can clearly show which rules are triggered under the current kinematic features.

[0060] Step 1244: Perform fuzzy implication processing on the consequents of the activated fuzzy rules, use a product inference engine to calculate the fuzzy set of each rule consequent, and generate the fuzzy set matrix of rule consequents.

[0061] In this embodiment of the invention, after determining the rule activation flag matrix, fuzzy implication processing is performed on the consequent conclusions of the activated fuzzy rules. Taking the case where a robotic arm has activated some rules as an example, a product inference engine is used for calculation. For each activated rule, the fuzzy set of the consequent is calculated based on the matching degree of its antecedent and the fuzzy linguistic variables of its consequent. The product inference engine obtains the fuzzy set of the consequent by multiplying the matching degree of the antecedent with the membership function of the consequent. The fuzzy sets of the consequents of all activated rules are combined to form a rule consequent fuzzy set matrix, which reflects the fuzzy influence of each activated rule on the joint control variables.

[0062] Step 1245: Aggregate the fuzzy set matrix of the rule consequents using a fuzzy synthesis algorithm to generate a comprehensive fuzzy set. The fuzzy synthesis algorithm uses the maximum-minimum synthesis method to calculate the fuzzy set of each control quantity dimension.

[0063] In this embodiment of the invention, after obtaining the fuzzy set matrix of rule consequents, a fuzzy synthesis algorithm is used for aggregation processing. Taking a scenario where the robotic arm is activated by multiple rules as an example, the maximum-minimum synthesis method is adopted. For each control quantity dimension, the fuzzy set corresponding to each rule is found in the fuzzy set matrix of rule consequents, and then the membership values ​​of these fuzzy sets are compared. For each element, the maximum and minimum membership values ​​among all rules are combined to obtain the fuzzy set for that control quantity dimension. The fuzzy sets of all control quantity dimensions are combined to form a comprehensive fuzzy set, which integrates the influence of all activated rules on the joint control quantity and is a fuzzy description of the control quantity in the current motion state of the robotic arm.

[0064] Step 1246: Calculate the activation intensity of each fuzzy rule based on the comprehensive fuzzy set, and generate the fuzzy rule activation intensity matrix. The row dimension of the activation intensity matrix is ​​equal to the number of fuzzy rules, and the column dimension is equal to the number of joint degrees of freedom.

[0065] In this embodiment of the invention, after obtaining the comprehensive fuzzy set, the activation intensity of each fuzzy rule is calculated. Taking the comprehensive fuzzy set of a robotic arm within one motion cycle as an example, based on the fuzzy information of each control dimension in the comprehensive fuzzy set, combined with the consequent information of each rule, the activation intensity of that rule at each joint is calculated. The activation intensity reflects the degree of influence of the rule on the joint control quantity. The activation intensities of all rules at each joint are combined to form a fuzzy rule activation intensity matrix. The row dimension of this matrix is ​​equal to the number of fuzzy rules, and the column dimension is equal to the number of joint degrees of freedom. This matrix can clearly show the effect intensity of each rule at different joints.

[0066] Step 125: The activation intensity matrix of the fuzzy rule is weighted and averaged through the declarative layer of the dynamic fuzzy neural network to generate the DFNN feedforward control signal, which contains the expected driving torque sequence of each joint.

[0067] In this embodiment of the invention, after obtaining the fuzzy rule activation intensity matrix, it is input into the declarative layer of the dynamic fuzzy neural network. Taking the process of a robotic arm completing a set action as an example, the declarative layer performs a weighted average processing on the fuzzy rule activation intensity matrix. For each element in the matrix, a corresponding weight is assigned according to its corresponding fuzzy rule and joint information. Then, the activation intensities of all rules corresponding to each joint are weighted and summed, and divided by the total weight to obtain a declarative value for that joint. The declarative values ​​of all joints are combined to form a DFNN feedforward control signal. This signal contains the expected driving torque sequence of each joint of the robotic arm in the current state. These expected driving torques are calculated based on fuzzy rule inference and weighted average, which can provide more accurate control guidance for the movement of the robotic arm, enabling the robotic arm to perform tasks more precisely.

[0068] Step 130: Construct an EFS error compensation channel based on the real-time joint state data set, and use an error feedback compensation algorithm to capture and compensate for dynamic errors in the movement of the robotic arm in real time, thereby generating an EFS error compensation signal.

[0069] In this embodiment of the invention, after acquiring the real-time joint state data set, the EFS error compensation channel is constructed. Taking a robotic arm performing complex operations as an example, dynamic errors inevitably occur during movement due to the robotic arm's own structural characteristics and external environmental interference. The error feedback compensation algorithm can monitor these errors in real time and perform corresponding compensation processing. The algorithm analyzes the real-time joint state data, compares it with the ideal motion state, and identifies existing errors. Then, based on the magnitude and direction of the error, it calculates the amount that needs to be compensated. During this process, the algorithm continuously updates the compensation amount to ensure timely response to changes in error. Finally, through a series of calculations and processing, an EFS error compensation signal is generated. This signal contains the real-time compensation torque information required by each joint during movement, which can effectively reduce errors during the robotic arm's movement and improve the accuracy and stability of the movement.

[0070] As one embodiment, step 130 includes:

[0071] Step 131: Perform error feature extraction processing on the real-time joint state data set to obtain the dynamic error feature set of each joint of the robotic arm. The dynamic error feature set includes joint angle tracking error, angular velocity fluctuation error and angular acceleration lag error.

[0072] In this embodiment of the invention, error features are extracted from the collected real-time joint state data set. Taking the process of a robotic arm handling an object as an example, the actual collected joint angles are compared with the pre-set target joint angles to calculate the joint angle tracking error. This error reflects the accuracy of the robotic arm in angle control. Simultaneously, the fluctuation of joint angular velocity is analyzed to calculate the angular velocity fluctuation error, which reflects the stability of the joint rotation speed. Furthermore, by comparing the actual joint angular acceleration with the ideal angular acceleration, the angular acceleration lag error is calculated, reflecting the timeliness of the joint acceleration response. These error information are organized into a set to obtain the dynamic error feature set of each joint of the robotic arm. These error features are the key basis for subsequently constructing an error prediction model and performing error compensation.

[0073] Step 132: Construct an error prediction model based on the dynamic error feature set, and use a time series prediction algorithm to predict the error change trend within a preset time window to generate an error prediction sequence.

[0074] In this embodiment of the invention, after obtaining the dynamic error feature set, an error prediction model is constructed. Taking the dynamic error feature set of a robotic arm within a continuous working cycle as an example, a time series prediction algorithm is employed. This algorithm analyzes the changing patterns of error features over time and predicts the error change trend within a preset time window based on historical data. For example, by analyzing the joint angle tracking error over a past period, the possible value of this error in the next few control cycles is predicted. The same method is used to predict different types of errors, such as angular velocity fluctuation error and angular acceleration lag error. The predicted error values ​​of all joints within the preset time window are combined into a sequence to generate an error prediction sequence. This sequence can provide information for error compensation in advance, making error compensation more timely and effective.

[0075] In one implementation approach, step 132 includes:

[0076] Step 1321: Perform stationarity testing on the joint angle tracking errors in the dynamic error feature set. Use the unit root test method to determine whether the error sequence meets the stationarity condition. If it does not meet the condition, perform difference processing until the sequence is stationary. Construct an autoregressive moving average model based on the stationary joint angle tracking error sequence, and determine the autoregressive order and moving average order of the model. The order is calculated and determined by the Akaike Information Criterion.

[0077] In this embodiment of the invention, for the joint angle tracking error in the dynamic error feature set, a stationarity test is first performed. Taking the joint angle tracking error sequence of the robotic arm at different motion stages as an example, the unit root test method is used. This method determines whether the error sequence is stationary, that is, whether the statistical characteristics of the sequence change over time. If the error sequence does not meet the stationarity condition, it is differentially processed. Differential processing eliminates the trend and seasonal components in the sequence by calculating the difference between adjacent data points, making the sequence gradually tend to be stationary. After multiple differential processing, the sequence is kept at stationarity. After obtaining a stationary joint angle tracking error sequence, an autoregressive moving average model is constructed. According to the Akaike information criterion, the information criterion value is calculated under different combinations of autoregressive order and moving average order. The order that minimizes the information criterion value is selected as the autoregressive order and moving average order of the model, thereby enabling the model to better fit the changing pattern of the joint angle tracking error.

[0078] Step 1322: Estimate the parameters of the stationary joint angle tracking error sequence using the autoregressive moving average model to generate a model coefficient vector. The parameter estimation uses the least squares method to calculate the autoregressive coefficients and moving average coefficients.

[0079] In this embodiment of the invention, after determining the order of the autoregressive moving average model, parameter estimation is performed on the stationary joint angle tracking error sequence. Taking a stationary joint angle tracking error sequence of a robotic arm over a relatively long time period as an example, the least squares method is used. This method finds a set of autoregressive coefficients and moving average coefficients that minimizes the sum of squared errors between the model's predicted values ​​and the actual error values. Through analysis and calculation of the error sequence, the autoregressive coefficients and moving average coefficients of the model are determined. These coefficients are combined into a model coefficient vector, which contains the key parameters of the autoregressive moving average model and can accurately describe the variation law of the joint angle tracking error.

[0080] Step 1323: Input the model coefficient vector into the error predictor to perform rolling prediction processing on the joint angle tracking error within the preset time window, and generate an angle error prediction subsequence.

[0081] In this embodiment of the invention, the obtained model coefficient vector is input into the error predictor. Taking the prediction of joint angle tracking error of a robotic arm over several control cycles as an example, the error predictor performs rolling prediction of the joint angle tracking error within a preset time window based on the autoregressive moving average model and the model coefficient vector. Rolling prediction means that at each time point, prediction is made using the latest historical data and model coefficients, and the prediction results are continuously updated over time. The predicted joint angle tracking error values ​​at each time point are combined to form an angle error prediction subsequence. This subsequence can predict the changing trend of the joint angle tracking error in advance, providing predictive information about the angle error for error compensation.

[0082] Step 1324: Using the same processing procedure as for joint angle tracking error, construct autoregressive moving average models for angular velocity fluctuation error and angular acceleration lag error in the dynamic error feature set, and generate angular velocity error prediction subsequence and angular acceleration error prediction subsequence.

[0083] In this embodiment of the invention, the angular velocity fluctuation error and angular acceleration lag error in the dynamic error feature set are processed using the same procedure as for joint angle tracking error. Taking these two errors of the robotic arm under different motion modes as examples, firstly, a stationarity test is performed on them. If the stationarity condition is not met, differential processing is performed to make them stationary. Then, the order of the autoregressive moving average model is determined according to the Akaike information criterion, and the parameters of the model are estimated using the least squares method to obtain the model coefficient vector. Finally, the model coefficient vector is input into the error predictor to perform rolling prediction of the angular velocity fluctuation error and angular acceleration lag error within a preset time window, generating angular velocity error prediction subsequences and angular acceleration error prediction subsequences respectively. These subsequences can predict the changing trends of angular velocity and angular acceleration errors respectively, providing more comprehensive error prediction information for error compensation.

[0084] Step 1325: The angle error prediction subsequence, the angular velocity error prediction subsequence, and the angular acceleration error prediction subsequence are spliced ​​along the time dimension to generate the error prediction sequence. The time length of the error prediction sequence is equal to the number of sampling periods of the preset time window.

[0085] In this embodiment of the invention, after obtaining the angle error prediction subsequence, angular velocity error prediction subsequence, and angular acceleration error prediction subsequence, they are concatenated along the time dimension. Taking these three subsequences of the robotic arm within a complete prediction time window as an example, they are connected sequentially according to time order. This integrates the prediction information of different types of errors, forming a more comprehensive error prediction sequence. The time length of this sequence is equal to the number of sampling periods of the preset time window, and it contains the prediction information of the angle, angular velocity, and angular acceleration errors of each joint of the robotic arm within the preset time.

[0086] Step 133: Input the dynamic error feature set and error prediction sequence into the error compensator, and calculate the error compensation amount in real time through the proportional-integral control algorithm to generate a preliminary error compensation signal.

[0087] In this embodiment of the invention, a dynamic error feature set and an error prediction sequence are input into the error compensator. Taking these two sets of data during the execution of a complex task by a robotic arm as an example, the error compensator uses a proportional-integral control algorithm to calculate the error compensation amount. For joint angle tracking errors, the error compensator calculates the proportional control component and the integral control component based on the current error value and the predicted error trend. The proportional control component is proportional to the current error value and can quickly respond to changes in error; the integral control component processes the accumulation of error over time and can eliminate the steady-state error of the system. Similarly, similar calculations are performed for angular velocity fluctuation errors and angular acceleration lag errors. The error compensation amounts of each joint are combined to generate a preliminary error compensation signal, which contains the preliminary compensation torque information required by each joint in the current state.

[0088] In one embodiment, step 133 includes:

[0089] Step 1331: Calculate the difference between the joint angle tracking error in the dynamic error feature set and the angle error prediction subsequence in the error prediction sequence to obtain the angle error deviation.

[0090] In this embodiment of the invention, the error compensator calculates the difference between the joint angle tracking error in the dynamic error feature set and the angle error prediction subsequence in the error prediction sequence. Taking these two sets of data within one control cycle of the robotic arm as an example, the actual joint angle tracking error is compared with the predicted angle error, and the difference between them is calculated. This difference is the angle error deviation, which reflects the difference between the actual error and the predicted error. The angle error deviation can provide more accurate error information for the subsequent calculation of proportional control components and integral control components, making error compensation more precise.

[0091] Step 1332: Calculate the proportional control component based on the angle error deviation. The proportional control component is equal to the product of the angle error deviation and the proportional coefficient. The proportional coefficient is dynamically adjusted according to the joint stiffness characteristics.

[0092] In this embodiment of the invention, after obtaining the angular error deviation, a proportional control component is calculated. Taking the angular error deviation of a robotic arm under different workloads as an example, the angular error deviation is multiplied by a proportional coefficient, which is dynamically adjusted according to the stiffness characteristics of the joint. Different joint stiffnesses result in different responsiveness to errors; joints with higher stiffness may require larger control inputs to correct errors. Therefore, the proportional coefficient is adjusted in real time according to the actual stiffness of the joint, enabling the proportional control component to respond more effectively to changes in error. The proportional control component can quickly react to the current angular error deviation, providing timely adjustments for error compensation.

[0093] Step 1333: Perform integral calculation on the angle error deviation to obtain the integral control component. The integral calculation uses the trapezoidal integral method to calculate the cumulative sum of the error deviation in the time dimension.

[0094] In this embodiment of the invention, the angle error deviation is processed by integration. Taking the angle error deviation of a robotic arm during a continuous motion process as an example, the trapezoidal integral method is used. This method divides the time dimension into several small intervals. Within each small interval, the error deviation is approximated as a linear change, and the integral value is approximated by calculating the area of ​​the trapezoid. The integral values ​​between each small interval are summed to obtain the cumulative sum of the error deviation in the time dimension, i.e., the integral control component. The integral control component can eliminate the steady-state error of the system, enabling the robotic arm to track the target angle more accurately during long-term operation.

[0095] Step 1334: Perform a weighted summation of the proportional control component and the integral control component to generate an angle error compensation sub-signal.

[0096] In this embodiment of the invention, after obtaining the proportional control component and the integral control component, they are weighted and summed. Taking these two components of the robotic arm within one control cycle as an example, appropriate weights are assigned to the proportional control component and the integral control component according to actual needs and system characteristics. Then, the two components are multiplied by their respective weights and added together to obtain the angle error compensation sub-signal. This signal combines the advantages of proportional control and integral control, enabling both rapid response to error changes and elimination of steady-state errors, thus providing a more effective control signal for joint angle error compensation.

[0097] Step 1335: Using the same processing procedure as for joint angle tracking error, calculate the proportional control component and integral control component for the angular velocity fluctuation error and angular acceleration lag error in the dynamic error feature set, respectively, and generate angular velocity error compensation sub-signal and angular acceleration error compensation sub-signal.

[0098] In this embodiment of the invention, the angular velocity fluctuation error and angular acceleration lag error in the dynamic error feature set are processed using the same procedure as for joint angle tracking error. Taking these two errors of the robotic arm under different motion states as examples, the error deviation is first calculated, and then the proportional control component and integral control component are calculated based on the error deviation. For the angular velocity fluctuation error, the angular velocity error compensation sub-signal is obtained based on its error deviation and the corresponding proportional coefficient and integral operation. Similarly, a similar calculation is performed for the angular acceleration lag error to generate the angular acceleration error compensation sub-signal. These sub-signals can compensate for the angular velocity and angular acceleration errors of the joints respectively, improving the stability and accuracy of the robotic arm's motion.

[0099] Step 1336: The angle error compensation sub-signal, the angular velocity error compensation sub-signal, and the angular acceleration error compensation sub-signal are fused along the joint degree of freedom dimension to generate the preliminary error compensation signal. The dimension of the preliminary error compensation signal is equal to the product of the number of robotic arm joints and the dimension of the control quantity.

[0100] In this embodiment of the invention, after obtaining the angle error compensation sub-signal, angular velocity error compensation sub-signal, and angular acceleration error compensation sub-signal, they are fused along the joint degree of freedom dimension. Taking these three sub-signals of a robotic arm in a complete motion process as an example, the angle, angular velocity, and angular acceleration error compensation sub-signals of each joint are combined according to the order of the joint degrees of freedom. This integrates different types of error compensation information onto each joint, forming a more comprehensive preliminary error compensation signal. The dimension of this signal is equal to the product of the number of robotic arm joints and the dimension of the control quantity, and it contains comprehensive error compensation information for each joint of the robotic arm in terms of angle, angular velocity, and angular acceleration.

[0101] Step 134: Perform dynamic amplitude limiting processing on the preliminary error compensation signal, adjust the amplitude range of the compensation signal based on the maximum driving torque constraint of the robotic arm joint, and generate the amplitude-limited error compensation signal.

[0102] In this embodiment of the invention, the generated preliminary error compensation signal needs to undergo dynamic amplitude limiting processing. Taking the preliminary error compensation signals of a robotic arm under different working scenarios as an example, an amplitude range is determined based on the maximum driving torque constraint of the robotic arm joints. If a value in the preliminary error compensation signal exceeds this range, it is adjusted to the boundary value of the range. This is to ensure that the compensation signal does not cause the driving torque of the joints to exceed their bearing capacity, thus avoiding damage to the robotic arm. After amplitude limiting processing, a limited error compensation signal is generated. The amplitude of this signal is within the safe range of the robotic arm joints, enabling safer and more effective compensation for the motion error of the robotic arm.

[0103] Step 135: The amplitude-limited error compensation signal is subjected to noise suppression processing by an adaptive filtering algorithm to generate the EFS error compensation signal, which contains the real-time compensation torque sequence of each joint.

[0104] In this embodiment of the invention, noise suppression processing is performed on the amplitude-limited error compensation signal. Taking the amplitude-limited error compensation signal of a robotic arm in a complex environment as an example, an adaptive filtering algorithm is adopted. This algorithm automatically adjusts the filtering parameters according to the characteristics of the signal and the noise situation to filter out noise in the signal. During the processing, the algorithm continuously analyzes the changes in the signal, identifies noise components, and removes them. After filtering, an EFS error compensation signal is generated. This signal contains the compensation torque sequence required by each joint in real time. With the noise interference removed, it can more accurately compensate for the motion error of the robotic arm, improving the motion accuracy and stability of the robotic arm.

[0105] Step 140: Construct a PD feedback control channel based on the real-time joint state data set, and perform closed-loop adjustment processing on the deviation between the current joint state and the target joint state of the robotic arm through the proportional-derivative control algorithm to generate a PD feedback control signal.

[0106] In this embodiment of the invention, after acquiring the real-time joint state data set, a PD feedback control channel is constructed. Taking a robotic arm performing a precision assembly task as an example, the proportional-derivative control algorithm performs closed-loop adjustment on the deviation between the current joint state and the target joint state. First, the actual collected joint state data is compared with the pre-set target joint state data to calculate the deviation. Then, based on this deviation, the proportional control component and the derivative control component are calculated. The proportional control component is proportional to the deviation and can quickly respond to changes in the deviation; the derivative control component is proportional to the rate of change of the deviation and can predict the development trend of the deviation, making adjustments in advance. By continuously adjusting the control quantity, the joint state of the robotic arm gradually approaches the target state, forming a closed-loop adjustment system. Finally, a PD feedback control signal is generated, which contains the adjustment torque information required by each joint of the robotic arm in the current state, providing important support for the precise control of the robotic arm.

[0107] As a non-limiting embodiment, step 140 includes:

[0108] Step 141: Obtain the target joint state data set of the robotic arm, which includes the target angle sequence, target angular velocity sequence and target angular acceleration sequence of each joint of the robotic arm.

[0109] In this embodiment of the invention, it is necessary to acquire a target joint state data set for the robotic arm. Taking the robotic arm performing a set production task as an example, the target angle, target angular velocity, and target angular acceleration of each joint of the robotic arm are predetermined according to the task requirements and design scheme. This data is organized into a sequence according to time order to form a target joint state data set. This set clarifies the state that the robotic arm should reach at each moment and serves as a reference standard for subsequent calculation of joint state deviations and generation of PD feedback control signals. For example, in a part assembly task, each joint of the robotic arm has a set target angle and angular velocity. This target data is recorded and used for comparison with the actual joint state.

[0110] Step 142: Perform difference calculation processing on the real-time joint state data set and the target joint state data set to obtain the joint state deviation set of each joint of the robotic arm. The joint state deviation set includes the angle deviation sequence, the angular velocity deviation sequence and the angular acceleration deviation sequence.

[0111] In this embodiment of the invention, the difference between the collected real-time joint state data set and the target joint state data set is calculated. Taking these two sets of data within one motion cycle of the robotic arm as an example, for each joint, the difference between the actual joint angle and the target joint angle is calculated to obtain an angle deviation sequence. Similarly, the difference between the actual joint angular velocity and the target angular velocity is calculated to obtain an angular velocity deviation sequence; the difference between the actual joint angular acceleration and the target angular acceleration is calculated to obtain an angular acceleration deviation sequence. These deviation sequences are organized into a set to form a joint state deviation set for each joint of the robotic arm. This deviation information reflects the gap between the current state and the target state of the robotic arm and is the basis for subsequent calculation of proportional control components and derivative control components.

[0112] Step 143: Calculate the proportional control component based on the joint state deviation set. The proportional control component is equal to the product of the joint state deviation set and the proportional coefficient. The proportional coefficient is dynamically adjusted according to the stiffness characteristics and load conditions of each joint of the robotic arm.

[0113] In this embodiment of the invention, after obtaining the set of joint state deviations, a proportional control component is calculated. Taking the set of joint state deviations of a robotic arm under different workloads and motion modes as an example, the deviation value of each joint is multiplied by a corresponding proportional coefficient. This proportional coefficient is dynamically adjusted according to the stiffness characteristics and load conditions of each joint of the robotic arm. Different joint stiffnesses result in different responses to deviations, and changes in load also affect the motion characteristics of the joints. Therefore, the proportional coefficient is adjusted in real time according to the actual stiffness and load conditions, enabling the proportional control component to respond more effectively to changes in joint state deviations. The proportional control component can quickly react to the current joint state deviations, providing timely adjustments for the control of the robotic arm.

[0114] Step 144: Perform differential operation on the set of joint state deviations to obtain a set of joint state deviation change rates. The differential operation uses the central difference method to calculate the change rate of the deviation sequence in the time dimension.

[0115] In this embodiment of the invention, the set of joint state deviations is processed by differential operations. Taking the set of joint state deviations of a robotic arm in a continuous motion process as an example, the central difference method is used. This method calculates the difference between the joint state deviations at two adjacent time points, and then divides it by the time interval to obtain the rate of change of the deviation sequence in the time dimension. The same processing is performed for angle deviations, angular velocity deviations, and angular acceleration deviations to obtain a set of joint state deviation change rates. This set reflects the development trend of the joint state deviations and can predict the changes in deviations in advance.

[0116] Step 145: Calculate the differential control component based on the set of joint state deviation change rates. The differential control component is equal to the product of the set of joint state deviation change rates and the differential coefficient. The differential coefficient is dynamically adjusted according to the dynamic response characteristics and movement speed of the robotic arm.

[0117] In this embodiment of the invention, after obtaining the set of joint state deviation change rates, differential control components are calculated. Taking the set of joint state deviation change rates of a robotic arm under different movement speeds and dynamic response requirements as an example, the deviation change rate value of each joint is multiplied by a differential coefficient. This differential coefficient is dynamically adjusted according to the dynamic response characteristics and movement speed of the robotic arm. The robotic arm's response capability to deviation changes varies under different movement states; at higher movement speeds, a larger differential control input may be needed for timely adjustment. Therefore, based on the actual dynamic response characteristics and movement speed, the differential coefficient is adjusted in real time, enabling the differential control components to more effectively predict and adjust changes in joint state deviations. The differential control components can react in advance to the development trend of joint state deviations, improving the stability and accuracy of robotic arm control.

[0118] Step 146: Perform a weighted summation on the proportional control component and the derivative control component to generate a preliminary PD feedback control signal. The weighted summation process adopts a dynamic weight allocation strategy based on the movement stage of the robotic arm.

[0119] In this embodiment of the invention, after obtaining the proportional control component and the derivative control component, they are weighted and summed. Taking a complete motion process of the robotic arm as an example, a dynamic weight allocation strategy based on the movement stage of the robotic arm is adopted. The importance of proportional control and derivative control varies at different stages of the robotic arm's motion. For example, at the beginning of the motion, a larger proportional control component may be needed to quickly approach the target state; when approaching the target state, the derivative control component may be more important to avoid overshoot. Therefore, according to the actual movement stage of the robotic arm, weights are dynamically assigned to the proportional control component and the derivative control component, and then they are weighted and summed to generate a preliminary PD feedback control signal. This signal combines the advantages of proportional control and derivative control, enabling more flexible adjustment of the joint state deviation of the robotic arm.

[0120] Step 147: Perform dynamic amplitude limiting processing on the preliminary PD feedback control signal, and limit the amplitude range of the control signal based on the maximum output torque constraint of the robotic arm joint to generate a limited PD feedback control signal.

[0121] In this embodiment of the invention, the generated preliminary PD feedback control signal needs to undergo dynamic amplitude limiting processing. Taking the preliminary PD feedback control signal of the robotic arm under different working scenarios as an example, an amplitude range is determined based on the maximum output torque constraint of the robotic arm joints. If a value in the preliminary PD feedback control signal exceeds this range, it is adjusted to the boundary value of the range. This is to ensure that the control signal does not cause the joint's output torque to exceed its bearing capacity, thus avoiding damage to the robotic arm. After amplitude limiting processing, a limited PD feedback control signal is generated. The amplitude of this signal is within the safe range of the robotic arm joints, enabling safer and more effective adjustment of the robotic arm's joint state.

[0122] Step 148: Perform noise suppression processing on the limited PD feedback control signal using a low-pass filtering algorithm to generate a filtered PD feedback control signal.

[0123] In this embodiment of the invention, noise suppression processing is performed on the PD feedback control signal after amplitude limiting. Taking the PD feedback control signal of a robotic arm in a complex environment as an example, a low-pass filtering algorithm is used. This algorithm allows low-frequency signals to pass through while filtering out high-frequency noise. During processing, the algorithm automatically adjusts the filtering parameters according to the frequency characteristics of the signal to filter out noise in the signal. After filtering, a filtered PD feedback control signal is generated. This signal removes noise interference and can more accurately adjust the joint state of the robotic arm, improving the stability and accuracy of robotic arm control.

[0124] Step 149: Based on the signal format compatible with the control interface of the robotic arm actuator, perform signal format conversion processing on the filtered PD feedback control signal to generate the PD feedback control signal.

[0125] In this embodiment of the invention, after obtaining the filtered PD feedback control signal, signal format conversion processing is required. Taking the filtered PD feedback control signals of a robotic arm on different industrial production lines as an example, the signal is converted into a compatible format according to the requirements of the robotic arm actuator control interface. Different actuators may have different requirements for signal format, such as signal voltage range, encoding method, etc. Therefore, the filtered PD feedback control signal is converted accordingly so that it can be correctly recognized and executed by the robotic arm actuator. After format conversion processing, a PD feedback control signal is generated, which conforms to the control interface requirements of the robotic arm actuator and can effectively drive the joint movement of the robotic arm.

[0126] Step 150: Input the DFNN feedforward control signal, the EFS error compensation signal and the PD feedback control signal into the three-channel parallel fusion network, perform dynamic fusion processing through the adaptive weight allocation algorithm in the three-channel parallel fusion network, generate a set of robotic arm joint drive control commands, and send the set of robotic arm joint drive control commands to the robotic arm actuator to drive multi-joint coordinated motion.

[0127] In this embodiment of the invention, the generated DFNN feedforward control signal, EFS error compensation signal, and PD feedback control signal are input into a three-channel parallel fusion network. Taking these three control signals of a robotic arm performing a complex task as an example, the adaptive weight allocation algorithm dynamically assigns weights to each control signal based on its real-time status. This algorithm considers factors such as signal reliability and accuracy to ensure that each control signal performs optimally under different motion states. Then, the three control signals are fused according to their assigned weights to generate a set of robotic arm joint drive control commands. This set contains the target driving torque values ​​for each joint of the robotic arm in each control cycle. Finally, these control commands are sent to the robotic arm actuator to drive the multi-joint coordinated motion of the robotic arm. This dynamic fusion method fully leverages the advantages of each control channel, improving the accuracy and stability of the robotic arm's motion.

[0128] In an alternative embodiment, the DFNN feedforward control signal, the EFS error compensation signal, and the PD feedback control signal are input into a three-channel parallel fusion network, and dynamically fused using an adaptive weight allocation algorithm in the three-channel parallel fusion network to generate a set of robotic arm joint drive control commands, including:

[0129] Step 151: Perform feature alignment processing on the DFNN feedforward control signal, the EFS error compensation signal, and the PD feedback control signal to ensure that the timestamps of each control signal are consistent with the joint degree of freedom dimension, and generate an aligned control signal set.

[0130] In this embodiment of the invention, feature alignment processing is required before inputting the three control signals into the three-channel parallel fusion network. Taking the three control signals of a robotic arm in a continuous motion process as an example, since they may be collected and generated at different time points, the timestamps and joint degree-of-freedom dimensions may be inconsistent. Therefore, they need to be adjusted to ensure that the timestamps and joint degree-of-freedom dimensions of each control signal are consistent. By calibrating the timestamps and matching the joint degree-of-freedom dimensions, the three control signals are organized to generate an aligned control signal set, in which the control signals are consistent in time and dimension.

[0131] Step 152: Construct a channel reliability assessment model based on the aligned control signal set, and calculate the real-time reliability weight of each control channel through signal quality indicators, including signal-to-noise ratio, signal fluctuation coefficient and signal tracking accuracy.

[0132] In this embodiment of the invention, after obtaining the aligned control signal set, a channel reliability evaluation model is constructed. Taking the aligned control signal set of the robotic arm under different working scenarios as an example, the real-time reliability weight of each control channel is calculated using signal quality indicators. For the DFNN feedforward control signal, its signal-to-noise ratio (SNR), i.e., the proportion of noise energy in the signal power spectral density, reflects the magnitude of noise in the signal; the signal fluctuation coefficient, i.e., the ratio of the standard deviation to the mean of the signal amplitude, reflects the stability of the signal; and the signal tracking accuracy, i.e., the deviation between the signal and the actual motion trajectory of the robotic arm, measures the accuracy of the signal. Similarly, the same calculations are performed for the EFS error compensation signal and the PD feedback control signal. Based on these signal quality indicators, corresponding evaluation functions are constructed to calculate the real-time reliability weight of each control channel. These weights reflect the reliability level of each control channel in the current state.

[0133] In an alternative embodiment, step 152 includes:

[0134] Step 1521: Perform noise analysis on the DFNN feedforward control signal in the aligned control signal set, calculate the noise energy ratio in the signal power spectral density, and generate the feedforward channel noise ratio.

[0135] In this embodiment of the invention, noise analysis is performed on the DFNN feedforward control signal in the aligned control signal set. Taking the DFNN feedforward control signal of a robotic arm in an industrial production scenario as an example, firstly, power spectral density analysis is performed on the signal to convert it to the frequency domain, obtaining the power spectral density distribution. Then, the proportion of noise energy in the power spectral density to the total energy is calculated; this proportion is the feedforward channel noise ratio. The feedforward channel noise ratio reflects the noise content in the DFNN feedforward control signal. A higher noise ratio indicates greater noise interference in the signal and lower signal quality. This indicator provides an important basis for evaluating the reliability of the DFNN feedforward control channel.

[0136] Step 1522: Perform fluctuation feature extraction processing on the DFNN feedforward control signal, calculate the ratio of the standard deviation of the signal amplitude to the mean, and generate the feedforward channel fluctuation coefficient.

[0137] In this embodiment of the invention, fluctuation feature extraction processing is performed on the DFNN feedforward control signal. Taking the DFNN feedforward control signal of a robotic arm under different motion modes as an example, the standard deviation and mean of the signal amplitude are calculated. The standard deviation reflects the dispersion of the signal amplitude, and the mean represents the average level of the signal amplitude. Dividing the standard deviation by the mean yields the feedforward channel fluctuation coefficient, which reflects the fluctuation of the DFNN feedforward control signal. The larger the fluctuation coefficient, the more drastic the amplitude change of the signal, and the worse the signal stability. The feedforward channel fluctuation coefficient provides information about signal stability for evaluating the reliability of the DFNN feedforward control channel.

[0138] Step 1523: Calculate the tracking accuracy of the deviation between the DFNN feedforward control signal and the actual motion trajectory of the robotic arm, and generate the feedforward channel tracking accuracy.

[0139] In this embodiment of the invention, the tracking accuracy is calculated by analyzing the deviation between the DFNN feedforward control signal and the actual motion trajectory of the robotic arm. Taking the DFNN feedforward control signal and the actual motion trajectory of the robotic arm performing a specific task as an example, the joint state indicated by the control signal is compared with the joint state actually reached by the robotic arm, and the deviation between them is calculated. By statistically analyzing these deviations, the tracking accuracy of the feedforward channel is obtained. This accuracy index reflects the tracking ability of the DFNN feedforward control signal to the actual motion trajectory of the robotic arm. The higher the tracking accuracy, the more accurately the control signal can guide the movement of the robotic arm, and the higher the reliability of the channel. The tracking accuracy of the feedforward channel provides information on the accuracy of the signal for evaluating the reliability of the DFNN feedforward control channel.

[0140] Step 1524: Construct a feedforward channel reliability evaluation function based on the feedforward channel noise ratio, feedforward channel fluctuation coefficient and feedforward channel tracking accuracy, and calculate the feedforward channel reliability weight.

[0141] In this embodiment of the invention, after obtaining the feedforward channel noise ratio, feedforward channel fluctuation coefficient, and feedforward channel tracking accuracy, a feedforward channel reliability evaluation function is constructed. Taking these three indicators of the robotic arm under different working conditions as examples, each indicator is assigned a corresponding weight based on its impact on the reliability of the DFNN feedforward control channel. Then, these three indicators are combined with their respective weights to construct an evaluation function. This evaluation function comprehensively considers factors such as signal noise, fluctuation stability, and tracking accuracy to calculate the reliability weight of the feedforward channel. For example, the feedforward channel noise ratio, feedforward channel fluctuation coefficient, and feedforward channel tracking accuracy can be multiplied by their respective weights and then summed to obtain a comprehensive score. This score is then normalized to obtain the feedforward channel reliability weight, which reflects the reliability of the DFNN feedforward control channel under the current state.

[0142] Step 1525: Using the same processing flow as the DFNN feedforward control signal, calculate the noise ratio, fluctuation coefficient and tracking accuracy of the EFS error compensation signal and PD feedback control signal in the aligned control signal set, and generate the reliability weight of the error compensation channel and the reliability weight of the feedback channel.

[0143] In this embodiment of the invention, the EFS error compensation signal and PD feedback control signal in the aligned control signal set are processed using the same procedure as the DFNN feedforward control signal. Taking these two signals within a complete motion cycle of the robotic arm as an example, noise analysis is performed on the EFS error compensation signal to calculate the proportion of noise energy in its signal power spectral density, thus obtaining the noise ratio of the error compensation channel; fluctuation feature extraction is performed on it to calculate the ratio of the standard deviation to the mean of the signal amplitude, thus obtaining the fluctuation coefficient of the error compensation channel; the tracking accuracy is calculated based on the deviation between the error compensation signal and the actual motion trajectory of the robotic arm, thus obtaining the tracking accuracy of the error compensation channel. Then, based on these three indicators, a reliability evaluation function for the error compensation channel is constructed, and the reliability weight of the error compensation channel is calculated. Similarly, the PD feedback control signal is processed in the same way to calculate the feedback channel noise ratio, feedback channel fluctuation coefficient, and feedback channel tracking accuracy, and then a reliability evaluation function for the feedback channel is constructed, and the reliability weight of the feedback channel is calculated. These weights reflect the reliability of the EFS error compensation channel and the PD feedback control channel in the current state, respectively.

[0144] Step 1526: Normalize the reliability weights of the feedforward channel, the error compensation channel, and the feedback channel to generate the real-time reliability weights.

[0145] In this embodiment of the invention, after obtaining the reliability weights of the feedforward channel, the error compensation channel, and the feedback channel, they are normalized. Taking these three weights of the robotic arm at different motion stages as an example, the three weights are added together to obtain a sum. Then, each weight is divided by this sum to obtain the normalized weight value. The purpose of this is to ensure that the sum of the reliability weights of the three channels is 1, ensuring that the weight allocation of each channel is reasonable in the subsequent dynamic weighted fusion process, and that no weight is too large or too small. After normalization, real-time reliability weights are generated. These weights accurately reflect the relative reliability of the DFNN feedforward control channel, the EFS error compensation channel, and the PD feedback control channel at the current moment.

[0146] Step 153: Perform dynamic weighted fusion processing on the aligned control signal set according to the real-time reliability weight to generate a weighted fused control signal. The weighted fusion processing uses a linear weighted summation algorithm to calculate the weighted sum of the control signals of each channel.

[0147] In this embodiment of the invention, after obtaining the real-time reliability weights, the aligned control signal set is dynamically weighted and fused. Taking the aligned control signal set and real-time reliability weights of the robotic arm within one control cycle as an example, a linear weighted summation algorithm is used. The DFNN feedforward control signal is multiplied by its corresponding real-time reliability weight, the EFS error compensation signal is multiplied by its corresponding real-time reliability weight, and the PD feedback control signal is multiplied by its corresponding real-time reliability weight. Then, these three weighted signals are added together to obtain a weighted fused control signal. This signal integrates the information from the three control channels and is reasonably weighted according to the real-time reliability of each channel, enabling more effective use of the advantages of each channel and providing more accurate guidance for the control of the robotic arm. The dynamic weighted fusion processing allows each control channel to play its corresponding role according to its reliability under different motion states, improving the accuracy and stability of the robotic arm control.

[0148] Step 154: Perform dynamic range adjustment processing on the weighted fusion control signal, limit the amplitude range of the control signal based on the torque output constraint of the robotic arm joint, and generate a range-adjusted control signal.

[0149] In this embodiment of the invention, the weighted fusion control signal undergoes dynamic range adjustment processing. Taking the weighted fusion control signal of a robotic arm under different working scenarios as an example, an amplitude range is determined based on the torque output constraints of the robotic arm joints. If a value in the weighted fusion control signal exceeds this range, it is adjusted to the boundary value of the range. This is to ensure that the control signal does not cause the joint's output torque to exceed its bearing capacity, thus avoiding damage to the robotic arm. After dynamic range adjustment processing, a range-adjusted control signal is generated. The amplitude of this signal is within the safe range of the robotic arm joints, enabling safer and more effective control of the robotic arm's joint movements.

[0150] Step 155: Perform time-domain smoothing on the range-adjusted control signal using a control signal smoothing algorithm to generate the set of robotic arm joint drive control commands, which includes the target drive torque value of each joint in each control cycle.

[0151] In this embodiment of the invention, the control signal after range adjustment is smoothed in the time domain. Taking the control signal after range adjustment of a robotic arm in a continuous motion process as an example, a control signal smoothing algorithm is adopted. This algorithm smooths the signal in the time dimension, removing abrupt changes and noise, making the signal more stable. During the processing, the algorithm adjusts the signal values ​​at adjacent time points according to the signal's changing trend, making the signal changes more continuous. After smoothing, a set of robotic arm joint drive control commands is generated. This set contains the target drive torque value of each joint in each control cycle. After smoothing, these values ​​enable the robotic arm's movement to be smoother and more fluid, reducing vibration and errors caused by signal abrupt changes.

[0152] In an alternative embodiment, step 155 includes:

[0153] Step 1551: Perform time-domain segmentation processing on the range-adjusted control signal to divide the continuous control signal into signal segments of multiple control cycles, with each signal segment corresponding to the control signal value of one control cycle.

[0154] In this embodiment of the invention, the control signal after range adjustment is processed by time-domain segmentation. Taking the control signal after range adjustment of a robotic arm during a relatively long movement process as an example, the continuous control signal is divided into multiple signal segments according to the control cycle of the robotic arm. Each signal segment corresponds to the control signal value of one control cycle. This is done to facilitate subsequent processing and analysis of the signal. Through time-domain segmentation, the continuous signal can be discretized, making the signal value of each control cycle more explicit, which facilitates abrupt change detection and signal smoothing.

[0155] Step 1552: Perform abrupt change detection processing on each signal segment. Use the first-order difference method to calculate the rate of change of the signal value between adjacent control cycles. When the rate of change is greater than the preset abrupt change threshold, it is determined that there is abrupt change interference in the signal segment.

[0156] In this embodiment of the invention, abrupt change detection processing is performed on each signal segment. Taking signal segments of a robotic arm at different motion stages as an example, a first-order difference method is used. This method calculates the difference between the signal values ​​of two adjacent control cycles, then divides it by the time interval to obtain the rate of change of the signal value. This rate of change is compared with a preset abrupt change threshold. When the rate of change is greater than the threshold, the signal segment is determined to have abrupt change interference. Abrupt change interference may be caused by external interference, sensor errors, etc., and can affect the motion stability of the robotic arm. Through abrupt change detection, abrupt changes in the signal can be detected in a timely manner.

[0157] Step 1553: Perform sliding window smoothing on signal segments with abrupt interference. Calculate the weighted average of the signal values ​​within the window using a weighted moving average algorithm to generate a smoothed signal segment. The weight coefficients of the weighted moving average algorithm exhibit a Gaussian distribution with respect to the window position.

[0158] In this embodiment of the invention, a sliding window smoothing process is performed on signal segments with abrupt interference. Taking a robotic arm's signal segment with abrupt interference as an example, a weighted moving average algorithm is used. This algorithm sets up a sliding window containing multiple adjacent control cycle signal values. For each signal value within the window, a corresponding weight is assigned based on its position within the window. These weight coefficients follow a Gaussian distribution with respect to the window position, meaning signal values ​​at the center of the window have a larger weight, while those at the edges have a smaller weight. Then, the signal values ​​within the window are multiplied by their respective weights, summed, and divided by the total weight to obtain a weighted average of the signal values ​​within the window. This weighted average is used as a value in the smoothed signal segment. As the window slides, the weighted average is calculated sequentially for each position to generate the smoothed signal segment. The sliding window smoothing process effectively removes abrupt interference from the signal, making the signal more stable.

[0159] Step 1554: For signal segments without sudden interference, perform signal preservation processing directly to maintain the original signal value.

[0160] In this embodiment of the invention, signal segments without abrupt interference are directly processed to preserve the signal. Taking signal segments during the normal movement phase of a robotic arm as an example, since these signal segments are free from abrupt interference, their signal values ​​are inherently stable. Therefore, no additional processing is required; the original signal values ​​are directly maintained. This reduces unnecessary calculations and processing, improves processing efficiency, and ensures signal accuracy.

[0161] Step 1555: The smoothed signal segment and the signal segment after signal preservation processing are spliced ​​together along the time dimension to generate the set of mechanical arm joint drive control instructions. The time resolution of the set of mechanical arm joint drive control instructions is consistent with the control cycle of the mechanical arm.

[0162] In this embodiment of the invention, after obtaining the smoothed signal segment and the signal segment after signal preservation processing, they are spliced ​​together along the time dimension. Taking these two signal segments within a complete motion cycle of the robotic arm as an example, they are connected sequentially according to time order. This integrates the processed signal segments into a complete set of robotic arm joint drive control commands. The time resolution of this set is consistent with the control cycle of the robotic arm, accurately providing the target driving torque value for each control cycle of the robotic arm, enabling the robotic arm to perform precise motion control according to the commands. Finally, this set of robotic arm joint drive control commands is sent to the robotic arm actuator to drive the multi-joint coordinated motion of the robotic arm, achieving efficient and accurate operation of the robotic arm in various tasks.

[0163] Therefore, this invention, through the acquisition of real-time joint state data sets of multi-joint motion of a robotic arm, constructs a DFNN feedforward control channel, an EFS error compensation channel, and a PD feedback control channel. The three-channel signals are then input into a three-channel parallel fusion network for dynamic fusion processing, generating a set of robotic arm joint drive control commands to drive the coordinated motion of multiple joints. This invention improves the accuracy of robotic arm inverse dynamics modeling, reduces the system's dependence on feedback control, and utilizes a dynamic fuzzy neural network to perform nonlinear mapping on the inverse dynamics model, constructing a feedforward model with higher accuracy and stronger generalization ability. The control method provided by this invention possesses structural adaptive capability, enabling continuous online learning. The error feedback compensation algorithm can capture and compensate for dynamic errors in real time, allowing the controller structure to adaptively evolve according to dynamic error changes. This invention also achieves deep fusion of feedforward modeling and feedback correction, dynamically fusing the three-channel signals through an adaptive weight allocation algorithm to form a closed-loop information flow, enhancing system stability and response efficiency. Furthermore, this invention improves initial stage control performance, reduces system oscillation and learning lag, and uses a proportional-derivative control algorithm to perform closed-loop adjustment of deviations, combined with error feedback compensation, improving the accuracy of initial error estimation and reducing learning time and structural expansion.

[0164] See Figure 2 As shown in the figure, this is a schematic diagram of the basic structure of a robotic arm inverse dynamics fusion control system 200 based on DFNN and EFS provided in an embodiment of the present invention. The robotic arm inverse dynamics fusion control system 200 based on DFNN and EFS includes:

[0165] Processor 201;

[0166] Storage device 202, on which computer program 2020 is stored;

[0167] When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the described DFNN and EFS-based inverse dynamics fusion control methods for robotic arms.

[0168] Based on the above, a readable storage medium is provided, on which a program or instructions are stored, and when the program or instructions are executed by a processor, the steps of the above method are implemented.

[0169] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

Claims

1. A fusion control method for inverse dynamics of a robotic arm based on DFNN and EFS, characterized in that, The method is achieved through inverse dynamics fusion control of the robotic arm based on DFNN and EFS, and the method includes: The system collects a set of real-time joint state data during the multi-joint motion of the robotic arm. The set of real-time joint state data includes a sequence of joint angles, a sequence of joint angular velocities, and a sequence of joint angular accelerations within a continuous motion cycle. Based on the real-time joint state data set, a DFNN feedforward control channel is constructed. The inverse dynamics model of the robotic arm is nonlinearly mapped through a dynamic fuzzy neural network to generate a DFNN feedforward control signal. An EFS error compensation channel is constructed based on the real-time joint state data set. The dynamic error during the movement of the robotic arm is captured and compensated in real time through an error feedback compensation algorithm to generate an EFS error compensation signal. Based on the real-time joint state data set, a PD feedback control channel is constructed. The deviation between the current joint state and the target joint state of the robotic arm is adjusted in a closed loop using a proportional-derivative control algorithm to generate a PD feedback control signal. The DFNN feedforward control signal, the EFS error compensation signal, and the PD feedback control signal are input into a three-channel parallel fusion network. The network is dynamically fused using an adaptive weight allocation algorithm to generate a set of robotic arm joint drive control commands. The set of robotic arm joint drive control commands is then sent to the robotic arm actuator to drive multi-joint coordinated motion.

2. The inverse dynamics fusion control method for robotic arms based on DFNN and EFS as described in claim 1, characterized in that, The DFNN feedforward control channel is constructed based on the real-time joint state data set. A dynamic fuzzy neural network is used to perform nonlinear mapping processing on the robotic arm's inverse dynamics model to generate DFNN feedforward control signals, including: The real-time joint state data set is processed by kinematic feature extraction to obtain the kinematic feature set of each joint of the robotic arm. The kinematic feature set includes the joint angle change rate, angular velocity direction consistency parameter and angular acceleration fluctuation amplitude. The input layer nodes of the dynamic fuzzy neural network are constructed based on the kinematic feature set. The number of input layer nodes is consistent with the number of joint degrees of freedom. Each input layer node corresponds to a combination of kinematic features of a joint. The kinematic feature set is processed by a membership function mapping through a fuzzification layer of a dynamic fuzzy neural network to generate a fuzzy feature vector. The membership function mapping process uses a Gaussian membership function to calculate the fuzzy membership value of each kinematic feature. The fuzzy feature vector is input into the rule inference layer of the dynamic fuzzy neural network, and fuzzy inference processing is performed based on the preset fuzzy control rule library to generate a fuzzy rule activation intensity matrix. The fuzzy rule library contains a set of fuzzy rules that map the relationship between joint motion states and control quantities. The activation intensity matrix of the fuzzy rules is weighted and averaged through the declarative layer of the dynamic fuzzy neural network to generate the DFNN feedforward control signal, which contains the expected driving torque sequence of each joint.

3. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 2, characterized in that, The step of performing membership function mapping on the kinematic feature set through the fuzzification layer of a dynamic fuzzy neural network to generate fuzzy feature vectors includes: The joint angle change rate in the kinematic feature set is standardized to obtain the normalized angle change rate. The standardization process uses the max-min normalization method to map the angle change rate to a preset numerical range. The center parameter and width parameter of the Gaussian membership function are determined based on the normalized angle change rate. The center parameter is calculated based on the mean of the joint angle change rate, and the width parameter is calculated based on the standard deviation of the joint angle change rate. The normalized angle change rate is calculated by using the Gaussian membership function to generate an angle change rate membership vector. The angular velocity direction consistency parameters in the kinematic feature set are subjected to direction encoding processing to obtain a direction encoding vector. The direction encoding processing uses a direction cosine matrix to convert the three-dimensional angular velocity direction into two-dimensional planar coordinates. Based on the direction encoding vector, the boundary parameters of the triangular membership function are determined, and the membership values ​​of the angular velocity direction consistency parameters are calculated through the triangular membership function to generate the direction consistency membership vector. The angular acceleration fluctuation amplitude in the kinematic feature set is subjected to frequency domain analysis to obtain the fluctuation frequency feature vector. The frequency domain analysis uses Fast Fourier Transform to extract the main frequency component of the angular acceleration signal. Based on the fluctuation frequency feature vector, the slope parameter of the S-shaped membership function is determined, and the membership value of the angular acceleration fluctuation amplitude is calculated through the S-shaped membership function to generate the fluctuation amplitude membership vector. The angular change rate membership vector, the directional consistency membership vector, and the fluctuation amplitude membership vector are concatenated along the feature dimension to generate the fuzzy feature vector, the dimension of which is equal to the sum of the dimensions of each membership vector.

4. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 2, characterized in that, The step of inputting the fuzzy feature vector into the rule inference layer of the dynamic fuzzy neural network, performing fuzzy inference processing based on a preset fuzzy control rule base, and generating a fuzzy rule activation intensity matrix includes: The fuzzy control rule base is analyzed, and a set of fuzzy rules containing antecedent conditions and consequent conclusions is extracted. The antecedent conditions are combinations of fuzzy linguistic variables of kinematic features, and the consequent conclusions are fuzzy linguistic variables of joint control quantities. The rule matching degree is calculated for each membership vector in the fuzzy feature vector to determine the degree of matching between the antecedent conditions of each fuzzy rule and the current kinematic feature, and a rule matching degree vector is generated. A rule triggering threshold is constructed based on the rule matching degree vector. When the rule matching degree is greater than the preset threshold, it is determined that the fuzzy rule is activated and a rule activation flag matrix is ​​generated. The consequents of the activated fuzzy rules are subjected to fuzzy implication processing. A product inference engine is used to calculate the fuzzy set of each rule consequent, generating a fuzzy set matrix of rule consequents. The fuzzy set matrix of the rule consequent is aggregated by a fuzzy synthesis algorithm to generate a comprehensive fuzzy set. The fuzzy synthesis algorithm uses the maximum-minimum synthesis method to calculate the fuzzy set of each control variable dimension. The activation intensity of each fuzzy rule is calculated based on the comprehensive fuzzy set, and the fuzzy rule activation intensity matrix is ​​generated. The row dimension of the activation intensity matrix is ​​equal to the number of fuzzy rules, and the column dimension is equal to the number of joint degrees of freedom.

5. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 1, characterized in that, The method involves constructing an EFS error compensation channel based on the real-time joint state data set, and using an error feedback compensation algorithm to capture and compensate for dynamic errors during the robotic arm's movement in real time, generating an EFS error compensation signal, including: Error feature extraction processing is performed on the real-time joint state data set to obtain the dynamic error feature set of each joint of the robotic arm. The dynamic error feature set includes joint angle tracking error, angular velocity fluctuation error and angular acceleration hysteresis error. An error prediction model is constructed based on the dynamic error feature set. An error prediction sequence is generated by predicting the error change trend within a preset time window using a time series prediction algorithm. The dynamic error feature set and the error prediction sequence are input into the error compensator, and the error is calculated in real time by the proportional-integral control algorithm to generate a preliminary error compensation signal. The initial error compensation signal is dynamically limited, and the amplitude range of the compensation signal is adjusted based on the maximum driving torque constraint of the robotic arm joint to generate a limited error compensation signal. The amplitude-limited error compensation signal is subjected to noise suppression processing by an adaptive filtering algorithm to generate the EFS error compensation signal, which contains the real-time compensation torque sequence of each joint.

6. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 5, characterized in that, The step of constructing an error prediction model based on the dynamic error feature set, and using a time series prediction algorithm to predict the error change trend within a preset time window to generate an error prediction sequence includes: The joint angle tracking error in the dynamic error feature set is subjected to stationarity test processing. The unit root test method is used to determine whether the error sequence meets the stationarity condition. If it does not meet the condition, the difference processing is performed until the sequence is stationary. An autoregressive moving average model is constructed based on the stationary joint angle tracking error sequence. The autoregressive order and the moving average order of the model are determined. The order is calculated and determined by the Akaike Information Criterion. The parameter estimation of the stationary joint angle tracking error sequence is performed by the autoregressive moving average model to generate a model coefficient vector. The parameter estimation uses the least squares method to calculate the autoregressive coefficients and moving average coefficients. The model coefficient vector is input into the error predictor to perform rolling prediction processing on the joint angle tracking error within a preset time window, generating an angle error prediction subsequence. Using the same processing procedure as for joint angle tracking error, autoregressive moving average models are constructed for angular velocity fluctuation error and angular acceleration hysteresis error in the dynamic error feature set, respectively, to generate angular velocity error prediction subsequence and angular acceleration error prediction subsequence. The angle error prediction subsequence, the angular velocity error prediction subsequence, and the angular acceleration error prediction subsequence are spliced ​​together along the time dimension to generate the error prediction sequence. The time length of the error prediction sequence is equal to the number of sampling periods of a preset time window.

7. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 5, characterized in that, The step of inputting the dynamic error feature set and the error prediction sequence into the error compensator, and calculating the real-time compensation amount of the error through a proportional-integral control algorithm to generate a preliminary error compensation signal includes: The difference between the joint angle tracking error in the dynamic error feature set and the angle error prediction subsequence in the error prediction sequence is calculated to obtain the angle error deviation. The proportional control component is calculated based on the angle error deviation. The proportional control component is equal to the product of the angle error deviation and the proportional coefficient. The proportional coefficient is dynamically adjusted according to the joint stiffness characteristics. The angle error deviation is integrated to obtain the integral control component. The integral operation is performed using the trapezoidal integral method to calculate the cumulative sum of the error deviation in the time dimension. The proportional control component and the integral control component are weighted and summed to generate an angle error compensation sub-signal. Using the same processing procedure as for joint angle tracking error, the proportional control component and integral control component are calculated for the angular velocity fluctuation error and angular acceleration lag error in the dynamic error feature set, respectively, to generate angular velocity error compensation sub-signal and angular acceleration error compensation sub-signal; The angle error compensation sub-signal, the angular velocity error compensation sub-signal, and the angular acceleration error compensation sub-signal are fused along the joint degree of freedom dimension to generate the preliminary error compensation signal. The dimension of the preliminary error compensation signal is equal to the product of the number of robot arm joints and the dimension of the control quantity.

8. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 1, characterized in that, The DFNN feedforward control signal, the EFS error compensation signal, and the PD feedback control signal are input into a three-channel parallel fusion network. Dynamic fusion processing is performed using an adaptive weight allocation algorithm within the three-channel parallel fusion network to generate a set of robotic arm joint drive control commands, including: The DFNN feedforward control signal, the EFS error compensation signal and the PD feedback control signal are subjected to feature alignment processing to ensure that the timestamps of each control signal are consistent with the joint degree of freedom dimension, thereby generating an aligned control signal set. A channel reliability assessment model is constructed based on the aligned control signal set. The real-time reliability weight of each control channel is calculated through signal quality indicators, including signal-to-noise ratio, signal fluctuation coefficient, and signal tracking accuracy. The aligned control signal set is dynamically weighted and fused according to the real-time reliability weight to generate a weighted fused control signal. The weighted fusion process uses a linear weighted summation algorithm to calculate the weighted sum of the control signals of each channel. The weighted fusion control signal is dynamically range adjusted by limiting the amplitude range of the control signal based on the torque output constraint of the robotic arm joint, thereby generating a range-adjusted control signal. The control signal after range adjustment is smoothed in the time domain by a control signal smoothing algorithm to generate the set of mechanical arm joint drive control commands. The set of mechanical arm joint drive control commands includes the target drive torque value of each joint in each control cycle.

9. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 8, characterized in that, The method for constructing a channel reliability assessment model based on the aligned control signal set, and calculating the real-time reliability weight of each control channel using signal quality indicators, includes: Noise analysis is performed on the DFNN feedforward control signals in the aligned control signal set to calculate the noise energy ratio in the signal power spectral density and generate the feedforward channel noise ratio. The DFNN feedforward control signal is subjected to fluctuation feature extraction processing, and the ratio of the standard deviation to the mean of the signal amplitude is calculated to generate the feedforward channel fluctuation coefficient. The tracking accuracy is calculated by analyzing the deviation between the DFNN feedforward control signal and the actual motion trajectory of the robotic arm, and the tracking accuracy of the feedforward channel is generated. Based on the feedforward channel noise ratio, feedforward channel fluctuation coefficient and feedforward channel tracking accuracy, a feedforward channel reliability evaluation function is constructed, and the feedforward channel reliability weight is calculated. Using the same processing flow as the DFNN feedforward control signal, the noise ratio, fluctuation coefficient and tracking accuracy of the EFS error compensation signal and PD feedback control signal in the aligned control signal set are calculated respectively, and the reliability weights of the error compensation channel and the feedback channel are generated. The reliability weights of the feedforward channel, error compensation channel, and feedback channel are normalized to generate the real-time reliability weights.

10. The inverse dynamics fusion control method for a robotic arm based on DFNN and EFS as described in claim 8, characterized in that, The step involves performing time-domain smoothing on the range-adjusted control signal using a control signal smoothing algorithm to generate the set of robotic arm joint drive control commands, including: The control signal after range adjustment is subjected to time-domain segmentation processing to divide the continuous control signal into signal segments of multiple control cycles, with each signal segment corresponding to the control signal value of one control cycle. Each signal segment undergoes abrupt change detection processing. The first-order difference method is used to calculate the rate of change of signal values ​​between adjacent control cycles. When the rate of change is greater than a preset abrupt change threshold, the signal segment is determined to have abrupt change interference. For signal segments with abrupt interference, a sliding window smoothing process is performed. A weighted moving average algorithm is used to calculate the weighted average of the signal values ​​within the window to generate a smoothed signal segment. The weight coefficients of the weighted moving average algorithm are Gaussian distributed with respect to the window position. For signal segments without sudden interference, signal preservation processing is performed directly to maintain the original signal value. The smoothed signal segment and the signal segment after signal preservation processing are spliced ​​together along the time dimension to generate the set of mechanical arm joint drive control instructions. The time resolution of the set of mechanical arm joint drive control instructions is consistent with the control cycle of the mechanical arm.

Images