Man-machine interaction method and system for mechanical arm

Abstract

The invention provides a mechanical arm man-machine interaction method and system. The method comprises the steps that force data, collected by a force sensor, of external force applied to the tail end of a mechanical arm are obtained; pre-processing the force data, wherein the pre-processing comprises the steps of extracting time domain features and carrying out linear fitting on the time domain features; inputting the preprocessed force data into a pre-trained random forest model, and outputting the type of human motion as a motion intention recognition result; the random forest model is obtained through training by taking force data corresponding to multiple types of human motions as input and taking the types of the human motions as output. According to the mechanical arm man-machine interaction method provided by the embodiment of the invention, the random forest method is adopted to construct the human body upper limb subjective motion intention recognition algorithm based on the force sensor, multiple human body upper limb motion behaviors are recognized, certain predictability is achieved, the human body upper limb motion intention can be recognized in advance, and the real-time performance of a man-machine interaction system is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of man-machine interaction of a robot arm, and in particular to a man-machine interaction method and system of a robot arm. Background Art

[0002] With the transformation and upgrading of the manufacturing industry and the innovative development of robotics technology, the working environment and work tasks of robotic arms are becoming more and more complex and diversified. Many complex tasks cannot be completed effectively by robotic arms alone and require human participation to complete the tasks. Human-Robot Collaboration (HRC) is a working mode in which humans and robotic arms complete tasks together, in which the robotic arms and human collaborative partners share the workspace and work closely together. Robots can undertake repetitive, high-load or high-precision work, while humans are responsible for tasks that require creativity, judgment or decision-making. The two complement each other's strengths, improve overall productivity, and achieve common goals through coordination.

[0003] Human-machine collaboration combines the advantages of high precision of the robot arm with human creativity to efficiently complete complex tasks. In the process of human-machine collaboration, people and robot arms need physical contact to complete interactive tasks. In order to ensure that the robot arm can complete tasks with high precision, the real-time, stability and safety of the system are particularly important. A high real-time system can improve the accuracy of the robot arm's operation, and high stability can increase the operating time of the robot arm system. At the same time, the robot arm does not require complex programming, and can be taught through manual guidance. It cooperates with humans in real time in a highly dynamic working environment and adapts to changing task requirements.

[0004] The human-computer interaction system of the robotic arm has put forward increasingly higher requirements on the accuracy, stability and safety of the robotic arm's task execution. The existing human-computer interaction solutions cannot meet the needs of the robotic arm to perform complex and diverse tasks. Summary of the invention

[0005] To solve the above problems, an embodiment of the present invention provides a robotic arm human-computer interaction method, including: obtaining force data of external force applied to the end of the robotic arm collected by a force sensor; preprocessing the force data, and the preprocessing includes: extracting time domain features and performing straight line fitting on the time domain features; inputting the preprocessed force data into a pre-trained random forest model, and outputting the type of human movement as a motion intention recognition result; the random forest model is trained with force data corresponding to multiple types of human movements as input and the type of human movement as output, and the types of human movement include at least one of the following: extension, flexion, and rotation.

[0006] The robotic arm human-computer interaction method provided in an embodiment of the present invention adopts the random forest method to construct a human upper limb subjective movement intention recognition algorithm based on a force sensor, recognizes a variety of human upper limb movement behaviors, has a certain predictive ability, can recognize human upper limb movement intention in advance, and improve the real-time performance of the human-computer interaction system.

[0007] Optionally, the time domain feature extraction includes: extracting time domain features from the force data of a single axis, and detecting a rising edge and a falling edge in the time domain; if it is detected that the curve corresponding to the time domain feature maintains a rising edge or a falling edge state within a preset time length, performing a straight line fitting on the rising edge or the falling edge.

[0008] In the embodiment of the present invention, time domain feature extraction and rising edge or falling edge straight line fitting are performed on the force data, thereby avoiding vibration errors caused by noise.

[0009] Optionally, the preprocessing further includes: filtering the force data using a Kalman filter; or removing the end execution gravity contained in the force data through gravity calibration.

[0010] In the embodiment of the present invention, the force data is filtered and gravity calibrated to improve the accuracy of the force data.

[0011] Optionally, the decision tree of the random forest model adopts a classification and regression tree algorithm; the classification and regression tree algorithm adopts minimum mean square error as a partitioning principle; the classification and regression tree algorithm adopts Pearson correlation coefficient to measure vector similarity.

[0012] In the embodiment of the present invention, a random forest model is used to identify action intentions, which specifically uses a classification regression tree algorithm, which can improve the real-time performance of the human-computer interaction system.

[0013] Optionally, the force data includes multi-axis force data, and the force data of each axis is used to train the random forest model and identify motion intention.

[0014] In the embodiment of the present invention, the data of each axis of the multi-dimensional force data are used for model training and recognition, thereby improving the training and recognition efficiency.

[0015] Optionally, the method further comprises: compensating for the zero drift error of the force sensor by using a force sensor zero drift compensation algorithm constructed by a Pi-Sigma fuzzy neural network.

[0016] In the embodiment of the present invention, a compensation algorithm is provided for the zero drift of the force sensor using a Pi-Sigma fuzzy neural network, thereby reducing the impact of the zero drift on the force data, extending the single startup time of the six-dimensional force sensor, and improving the stability of the human-computer interaction system.

[0017] Optionally, the Pi-Sigma fuzzy neural network uses the zero drift error of the force sensor that changes with usage time, force size and temperature data as a training set, uses the usage time, force size and temperature data of the force sensor as input, and uses the zero drift compensation of the force sensor as output.

[0018] In the embodiment of the present invention, Pi-Sigma fuzzy neural network is used for training. When the training error reaches the error performance index, the Pi-Sigma fuzzy neural network is used as the complete model.

[0019] Optionally, the Pi-Sigma fuzzy neural network includes an input layer, a fuzzification layer, a fuzzy reasoning layer and an output layer;

[0020] The output of the fuzzification layer is:

[0021]

[0022] In the formula is the membership, i is the membership function set of the i-th input, j is the j-th membership function of the i-th input, is the center of the jth membership function of the ith input, b j is the membership function width;

[0023] Early reasoning for building rules:

[0024]

[0025] Where ω k is the kth weight ω;

[0026] The output of each inference rule is The inference conclusion of each rule is l is the number of previous reasoning layers;

[0027] The output layer result is:

[0028]

[0029] After the forward channel of the Pi-Sigma fuzzy neural network is constructed, the reverse learning algorithm is constructed, and the output layer weight adjustment method is as follows:

[0030]

[0031]

[0032] The learning algorithm for weight p is:

[0033]

[0034] Where η is the learning rate and α is the momentum factor.

[0035] The embodiment of the present invention provides a specific structure of a Pi-Sigma fuzzy neural network and expressions defined by each structure, which can realize the zero drift error compensation of a force sensor.

[0036] An embodiment of the present invention provides a robot arm human-computer interaction system, which is used to execute any of the methods described above.

[0037] Optionally, a random forest model is included, which is trained using force data corresponding to multiple types of human motion as input and types of human motion as output, and the types of human motion include at least one of the following: extension, flexion, and rotation.

[0038] The robotic arm human-computer interaction system provided by the embodiment of the present invention can achieve the same technical effect as the above-mentioned robotic arm human-computer interaction method. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A structural block diagram of a human-computer interaction system provided by an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a flow chart of a man-machine interaction method for a robot arm provided in an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the Pi-Sigma fuzzy neural network structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] The existing human-computer interaction information acquisition mostly adopts force sensors, visual sensors, voice sensors, etc. Force sensors are widely used in the field of human-computer interaction due to their fast response speed, good stability, and data accuracy. Therefore, the embodiment of the present invention selects force sensors as the device for acquiring human behavior operation data during the human-computer interaction process.

[0045] The embodiment of the present invention is applied to the end-mounted force sensor and the end effector of a multi-DOF manipulator. The end effector serves as the contact medium between the operator and the manipulator. The operator applies force to the force sensor by holding the end effector, and the force sensor obtains force data. For example, the embodiment of the present invention adopts a six-dimensional force sensor.

[0046] Different from the existing research technology, the embodiment of the present invention identifies the subjective movement intention of the human upper limbs based on the force data of the six-dimensional force sensor, and predicts the zero-point drift mechanism of the six-dimensional force sensor. It designs and implements a human-computer interaction method for a robotic arm based on the six-dimensional force sensor, realizing a high-reliability and high-real-time human-computer interaction method applied to the robotic arm.

[0047] Figure 1 The structure diagram of the human-computer interaction system provided by the embodiment of the present invention is shown, which mainly includes a mechanical arm, a lower computer, a host computer, a force sensor and an end effector in terms of hardware. The mechanical arm and the lower computer, the lower computer and the host computer, and the host computer and the force sensor all use Ethernet communication.

[0048] Among them, the upper computer includes a human-computer interaction main program, which can obtain force data, robot arm status and Pi-Sigma fuzzy neural network zero drift compensation, as well as motion intention recognition, trajectory planning and compliance control; the lower computer can perform motion command reception, robot arm status acquisition and robot arm motion control; the multi-degree-of-freedom robot arm is connected to the lower computer, controlled by the lower computer, and performs position response and tracking. For example, the six-dimensional force sensor is mounted on the robot end effector, the human upper limb force acts on the end effector, and the six-dimensional force sensor can collect the corresponding force data.

[0049] In the embodiment of the present invention, the random forest method is used to construct a human upper limb subjective movement intention recognition algorithm based on a force sensor, which can recognize a variety of human upper limb movement behaviors, has a certain predictive ability, can recognize the human upper limb movement intention in advance, and improve the real-time performance of the human-computer interaction system.

[0050] Furthermore, in the embodiment of the present invention, a Pi-Sigma fuzzy neural network is used to construct a force sensor zero drift compensation algorithm, thereby reducing the impact of zero drift on force data, extending the single startup time of the force sensor, and improving the stability of the human-computer interaction system.

[0051] The present invention provides a method for man-machine interaction of a robot arm. Figure 2 The schematic diagram of the process flow of the manipulator human-computer interaction method shown in FIG. 1 mainly includes the following steps:

[0052] S202, acquiring force data of an external force applied to the end of the robotic arm collected by a force sensor.

[0053] S204, preprocessing the force data, including: extracting time domain features and performing linear fitting on the time domain features.

[0054] The time domain feature extraction in this embodiment includes: extracting the time domain features of the single-axis force data, and detecting the time domain rising edge and falling edge; if it is detected that the corresponding curve of the time domain feature maintains the rising edge or falling edge state within a preset time length, a straight line fitting is performed on the rising edge or falling edge.

[0055] The force data of each axis of the multi-dimensional force sensor needs to be extracted in the time domain, and the rising edge and falling edge need to be fitted from them.

[0056] Furthermore, the above preprocessing may also include: using Kalman filtering to filter the force data; or removing the end execution gravity contained in the force data through gravity calibration.

[0057] In this embodiment, Kalman filtering can be used to process the initial data, and the end execution gravity is removed through gravity calibration. Then, the time domain features of the filtered force data are extracted, and the rising and falling edges in the time domain are detected. The rising and falling edges of the force data are fitted by the least squares method combined with the residual square and partial derivative function, and a variable length moving time window is set to avoid fitting distortion of the rising and falling edges.

[0058] The rising and falling edges of the single-axis data are fitted for the filtered force sensor data to avoid vibration errors caused by noise. According to the force sensor data, if it is detected that the force data curve remains in a rising or falling edge state for a period of time, fitting is performed.

[0059] Exemplarily, this embodiment adopts the least squares fitting method, taking the minimum sum of squares of the residuals between the dependent variable value on the fitted straight line and the actual value as the optimization target, and determining the correlation coefficient of the fitted straight line.

[0060] S206, input the preprocessed force data into a pre-trained random forest model, and output the type of human body movement as a movement intention recognition result.

[0061] The random forest model is trained using force data corresponding to various types of human motion as input and types of human motion as output, and the types of human motion include at least multiple of the following: extension, flexion, and rotation.

[0062] This embodiment uses the random forest (RF) machine learning method to perform pattern recognition on the six-axis force data of the sensor (Fx, Fy, Fz, Mx, My, Mz), takes a variety of upper limb movements (extension, flexion, rotation) as training sets, identifies the human upper limb movement intention, and finally uses the RF model to decode the movement intention, infer the sensor data in real time, and output the predicted upper limb movement intention.

[0063] The RF model uses multiple decision trees for integration. Multiple regression trees together form a random forest regression model, and the decision trees in the random forest model are independent of each other. Each decision tree in the model jointly determines the final output of the regression model.

[0064] Optionally, the decision tree of the random forest model adopts the Classification And Regression Tree (CART) algorithm. CART is a binary tree that can only choose yes or no when making a decision at each node. For each node to be split, the optimal value of the optimal feature in the data set is selected as the condition for splitting.

[0065] Optionally, the classification and regression tree algorithm uses minimum mean square error (MSE) as a division principle. The classification and regression tree algorithm uses Pearson correlation coefficient (Pearson correlation) to measure vector similarity.

[0066] In this embodiment, multiple upper limb movement (extension, flexion, rotation) force data can be used as training sets to input into the RF model to build a complete random forest model. The model takes six-dimensional force sensor data as input and human upper limb subjective movement intention as output.

[0067] It should be noted that the above-mentioned force data may include multi-axis force data, and the force data of each axis may be used to train the above-mentioned random forest model and identify motion intention.

[0068] The robotic arm human-computer interaction method provided in an embodiment of the present invention adopts the random forest method to construct a human upper limb subjective movement intention recognition algorithm based on a force sensor, recognizes a variety of human upper limb movement behaviors, has a certain predictive ability, can recognize human upper limb movement intention in advance, and improve the real-time performance of the human-computer interaction system.

[0069] The rising and falling edges of the single-axis data are fitted for the filtered force sensor data. To avoid vibration errors caused by noise, if the force data curve is detected to maintain a rising or falling edge state for a period of time, fitting is performed. Exemplarily, this embodiment adopts the least squares fitting method. The minimum sum of squares of the residuals between the dependent variable value and the actual value on the fitted straight line is taken as the optimization goal, and the correlation coefficient of the fitted straight line is determined.

[0070] The objective function is defined as follows:

[0071]

[0072] In the formula, x i is the value of the random variable x at position i; y i is the value of the random variable y at point i; α and β are the fitting straight line coefficients.

[0073] The objective function is partial derivatived with respect to α and β, and the partial derivative is:

[0074]

[0075]

[0076] Through the zero point of partial derivative, the fitting straight line coefficient is obtained as:

[0077]

[0078]

[0079] A random forest machine learning model is constructed by performing multi-axis fitting (Fx, Fy, Fz, Mx, My, Mz) on the force sensor data.

[0080] This embodiment can use the minimum mean square error (MSE) as the CART decision tree partitioning principle. The partition point S corresponding to the feature C in the data sets D1 and D2 is found by the minimum mean square error:

[0081]

[0082] Where c1 is the sample output mean of data set D1; c2 is the sample output mean of data set D2.

[0083] The Pearson correlation coefficient is used to measure vector similarity:

[0084]

[0085] In the formula, x i is the value of the random variable x at position i; y i is the value of the random variable y at position i; is the mean of the random variable y; The mean of the random variable x.

[0086] The Pearson coefficient ranges from (-1 to 1). When the correlation coefficient is 0, it means there is no correlation between the two. A negative value means a negative correlation between the two. A positive value means a positive correlation between the two. The correlation degree |ρ| is: 0.0-0.2 means very weak correlation or no correlation; 0.2-0.4 means weak correlation; 0.4-0.6 means moderate correlation; 0.6-0.8 means strong correlation; 0.8-1.0 means extreme correlation.

[0087] The force data of various upper limb movements (extension, flexion, rotation) are used as training sets to input into the RF model, and a complete random forest model is constructed. The model takes the six-dimensional force sensor data as input and the subjective movement intention of the human upper limb as output. Prediction through the six-dimensional force sensor data can identify the movement intention of the human upper limb in advance and improve the real-time performance of the human-computer interaction system.

[0088] Since the force sensor will have zero drift during actual use, in order to ensure the stability of the human-computer interaction system, the embodiment of the present invention also constructs a Pi-Sigma fuzzy neural network zero drift compensation algorithm based on the principle of zero drift of the force sensor. The algorithm can perform feedback compensation on the force sensor, reduce the impact of zero drift on force data, extend the single startup time of the force sensor, and improve the stability of the human-computer interaction system.

[0089] Based on this, the above method also includes the following steps: using a force sensor zero drift compensation algorithm constructed by a Pi-Sigma fuzzy neural network to compensate for the force sensor zero drift error.

[0090] Among them, the Pi-Sigma fuzzy neural network uses the zero drift error of the force sensor that changes with the use time, force size and temperature data as the training set, the use time, force size and temperature data of the force sensor as the input, and the zero drift compensation of the force sensor as the output.

[0091] For example, the six-dimensional force sensor may have zero drift due to the use time, force magnitude and temperature. In order to improve the stability of the human-computer interaction system and extend the single start-up time of the six-dimensional force sensor, a six-dimensional force sensor zero drift compensation algorithm based on Pi-Sigma fuzzy neural network is constructed. The Pi-Sigma fuzzy neural network includes an input layer, a fuzzification layer, a fuzzy reasoning layer, and an output layer. The structure of the Pi-Sigma fuzzy neural network is as follows: Figure 3 shown.

[0092] The force sensor data of usage time, force magnitude and temperature are used as the input layer of the fuzzy neural network, and the zero drift compensation of the six-dimensional force sensor is used as the output layer. The fuzzification layer uses the Gaussian function as the membership function. The output of the fuzzification layer is:

[0093]

[0094] In the formula is the degree of membership, x i is the membership function set of the ith input, j is the jth membership function of the ith input, is the center of the jth membership function of the ith input, b j is the membership function width.

[0095] Early reasoning for building rules:

[0096]

[0097] Where ω k is the kth weight ω.

[0098] The output of each inference rule is The inference conclusion of each rule is l is the number of previous reasoning layers, p is the weight, and h is the neuron output, which is also the reasoning output.

[0099] The output layer result is:

[0100]

[0101] After the forward channel of the Pi-Sigma fuzzy neural network is constructed, the reverse learning algorithm is constructed, and the output layer weight adjustment method is as follows:

[0102]

[0103]

[0104] In the formula, E p is the error, which is the difference between the actual output and the expected output of the training process, and η is the learning rate;

[0105] The learning algorithm for weight p is:

[0106]

[0107] In the formula, α is the momentum factor.

[0108] The actual six-dimensional force sensor zero drift error according to the change value of usage time, force size and temperature data is used as the training set of the Pi-Sigma fuzzy neural network. When the training error reaches the error performance index E, the Pi-Sigma fuzzy neural network is used as a complete model.

[0109] The usage time, force magnitude and temperature data of the force sensor are used as input of the fuzzy neural network, and the output of the fuzzy neural network is used to compensate for the zero drift error of the force sensor, thereby reducing the impact of zero drift on force data, extending the single startup time of the force sensor, and improving the stability of the human-computer interaction system.

[0110] In view of the fact that a manipulator human-computer interaction system based on a six-dimensional force sensor is prone to real-time and stability problems, the embodiment of the present invention constructs a random forest motion intention recognition algorithm and a Pi-Sigma fuzzy neural network zero drift compensation algorithm based on the subjective motion intention recognition of the human upper limbs and the zero drift mechanism of the force sensor, thereby realizing real-time intention recognition of the human-computer interaction system and improving the stability of the human-computer interaction system. The real-time and stability indicators reach a leading level.

[0111] An embodiment of the present invention provides a robot arm human-computer interaction system, which is used to execute the steps of the method provided in the above embodiment.

[0112] Optionally, the system includes a random forest model, which is trained using force data corresponding to multiple types of human motion as input and types of human motion as output, and the types of human motion include at least one of the following: extension, flexion, and rotation.

[0113] The above-mentioned robotic arm human-computer interaction system provided by the embodiment of the present invention has the same implementation principle and technical effects as those of the aforementioned embodiment. For the sake of brief description, for matters not mentioned in the system embodiment, reference may be made to the corresponding contents in the aforementioned method embodiment.

[0114] An embodiment of the present invention provides an electronic device, which includes a processor and a storage device. The storage device stores a computer program that can be run on the processor. When the processor executes the computer program, the steps of the method provided in the above embodiment are implemented.

[0115] An embodiment of the present invention provides a computer-readable medium, wherein the computer-readable medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method described in the above embodiment.

[0116] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the control device through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it may include the processes of the above-mentioned method embodiments, wherein the storage medium may be a memory, a disk, an optical disk, etc.

[0117] In this article, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A robot arm human-computer interaction method, characterized in that: include: Obtain force data of external force applied to the end of the robotic arm collected by the force sensor; Preprocessing the force data, the preprocessing comprising: extracting time domain features and performing linear fitting on the time domain features; The preprocessed force data is input into a pre-trained random forest model, and the type of human movement is output as a result of movement intention recognition; the random forest model is trained using force data corresponding to multiple types of human movements as input and the types of human movement as output, and the types of human movement include at least one of the following: extension, flexion, and rotation.

2. The method according to claim 1, characterized in that The time domain feature extraction comprises: Extracting time-domain features of the force data of a single axis, and detecting rising edges and falling edges in the time domain; If it is detected that the time domain characteristic corresponding curve maintains a rising edge or a falling edge state within a preset time length, a straight line fitting is performed on the rising edge or the falling edge.

3. The method according to claim 1, characterized in that The pre-processing further comprises: Using Kalman filtering to filter the force data; or, The force data is then subjected to gravity calibration to remove the end-effect gravity.

4. The method according to claim 1, characterized in that: The decision tree of the random forest model adopts a classification regression tree algorithm; The classification and regression tree algorithm adopts the minimum mean square error as the division principle; the classification and regression tree algorithm adopts the Pearson correlation coefficient to measure the vector similarity.

5. The method according to claim 1, characterized in that The force data includes multi-axis force data, and the force data of each axis is used to train the random forest model and identify movement intention.

6. The method according to claim 1, characterized in that The method further comprises: A force sensor zero drift compensation algorithm constructed by a Pi-Sigma fuzzy neural network is used to compensate for the zero drift error of the force sensor.

7. The method according to claim 6, characterized in that The Pi-Sigma fuzzy neural network uses the zero drift error of the force sensor that changes with the use time, force size and temperature data as a training set, uses the use time, force size and temperature data of the force sensor as input, and uses the zero drift compensation of the force sensor as output.

8. The method according to claim 6, characterized in that The Pi-Sigma fuzzy neural network includes an input layer, a fuzzification layer, a fuzzy reasoning layer and an output layer; The output of the fuzzification layer is: In the formula is the membership, i is the membership function set of the i-th input, j is the j-th membership function of the i-th input, is the center of the jth membership function of the ith input, b j is the membership function width; Early reasoning for building rules: Where ω k is the kth weight ω; The output of each inference rule is The inference conclusion of each rule is l is the number of previous reasoning layers, p is the weight, and h is the neuron output; The output layer result is: After the forward channel of the Pi-Sigma fuzzy neural network is constructed, the reverse learning algorithm is constructed, and the output layer weight adjustment method is as follows: In the formula, E p is the error, which is the difference between the actual output and the expected output of the training process, and η is the learning rate; The learning algorithm for weight p is: In the formula, α is the momentum factor.

9. A robot arm human-computer interaction system, characterized in that: Used to perform the method according to any one of claims 1 to 8.

10. The system according to claim 9, characterized in that It includes a random forest model, which is trained by using force data corresponding to multiple types of human movements as input and types of human movements as output, and the types of human movements include at least multiple of the following: extension, flexion, and rotation.

Images