Neural Network-Based Robot Collision Diagnosis System and Method

Abstract

This invention relates to a robot collision diagnosis system based on a neural network, comprising: a robot arm; an end effector; a control module that generates planned motion data for joints; a detection module that acquires actual motion data and operating status of joints; a calculation module that calculates the motion information deviation of joints; an optimization module that optimizes the motion information deviation of joints; and a prediction module that uses the optimized motion information deviation of joints and operating status labels as a training set to construct a neural network, trains the neural network, and predicts the real-time operating status of the robot's joints based on the real-time motion information deviation of the joints. This invention employs a method of calculating joint motion information deviation combined with a neural network, avoiding the step of performing complex kinematic modeling for a specific robot.

Description

Technical Field

[0001] This invention relates to the fields of robotic arms and robots, and in particular to a robot collision diagnosis system and method based on neural networks. Background Technology

[0002] In complex working conditions, robots may come into contact with or collide with people, environmental objects, or even other robots. These contacts or collisions vary depending on the scenario; some are accidental, while others are anticipated by the operator. For multi-jointed robots, these contacts or collisions invariably cause a motion parameter of at least one joint to deviate from the expected value. In conventional detection methods, any significant deviation from the expected value will cause the robot to stop, thus preventing injury to humans or damage to the robot. In such cases, unnecessary robot downtime often occurs, disrupting smooth production. Therefore, it would be highly beneficial if robots could diagnose the type of contact or collision.

[0003] Traditional diagnostic systems suffer from the following problems: The use of high-precision mechanical detection mechanisms, with tightly meshed mechanical components to achieve higher control accuracy, increases design and manufacturing costs; pressure sensor-based diagnostic systems increase the complexity of the robot system, reduce its flexibility, and raise its operating costs; vision sensor-based diagnostic methods struggle to meet real-time detection needs due to blind spots and low accuracy; sensorless diagnostic methods are dependent on specific robot structures, requiring the design of specific algorithms and lacking universality; complex kinematic models need to be established, and neural networks are introduced to improve accuracy, which also introduces many limitations to neural network methods. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art by providing a robot collision diagnosis system and method based on neural networks. This system can not only effectively solve the common problems of general robot diagnosis systems and reduce the losses caused by unnecessary robot downtime, but also avoid the steps of performing complex kinematic modeling for specific robots by reading joint motion data and combining it with neural networks.

[0005] According to one aspect of the present invention, a robot collision diagnosis system based on a neural network is provided, comprising:

[0006] A robotic arm comprising multiple links and multiple joints, wherein one of the links is connected to another link via a joint;

[0007] End effector;

[0008] The control module generates planned motion data for the joints;

[0009] The detection module acquires the actual motion data and operating status of the joint;

[0010] The calculation module calculates the deviation of joint motion information;

[0011] The optimization module optimizes the deviation of joint motion information.

[0012] The prediction module uses the optimized motion information deviation of the joints and the running state labels as a training set to build a neural network and train the neural network; based on the real-time motion information deviation of the joints, the neural network predicts the real-time running state of the robot's joints.

[0013] The construction of a neural network specifically includes:

[0014] Step 510: Establish a neural network set containing multiple neural networks and randomly initialize the neural network parameters;

[0015] Step 520: Optimize the motion information bias to form a training set and a test set, and then divide the training set and the test set into multiple subsets respectively;

[0016] Step 530: Randomly select two or more subsets of training and test sets, adjust the first or first set of neural networks in the neural network set using the first or first set of subsets, and adjust the second or second set of neural networks in the neural network set using the second or second set of subsets.

[0017] Step 540: Compare the first or first group of neural networks with the second or second group of neural networks, and based on the comparison results, remove the worse one or group of neural networks from the set of neural networks;

[0018] Step 550: Determine whether the current number of comparisons meets the preset number of comparisons, or whether the neural network set contains only one or a group of neural networks. If so, output the best one or a group of neural networks; otherwise, proceed to step 530.

[0019] Preferably, the neural network input is optimized motion information deviation with corresponding labels or classifications, and the output is the running status.

[0020] Preferably, the parameter values ​​of the feature extractor of the neural network and at least one fully connected layer of the neural network that follows the feature extractor are adjusted.

[0021] Preferably, the comparison in step 540 may be based at least in part on the accuracy of the neural network or the model size or computational cost, or a combination of at least two of these.

[0022] Preferably, the constructed neural network includes multiple neuron layers, with 5-10 neuron layers and each neuron layer including multiple neurons, with 15-25 neurons in total.

[0023] Preferably, the joint motion information deviation is the difference between the actual motion data and the planned motion data of the joint.

[0024] Preferably, the motion data of the joint includes, but is not limited to, the actual position, velocity, angular velocity, and torque of the joint.

[0025] Preferably, the operating state of the joint includes the contact status and the location of the contact point. The contact status includes, but is not limited to, "no contact", "human-computer interaction", "minor collision" and "strong collision".

[0026] According to another aspect of the present invention, a robot collision diagnosis method based on a neural network is provided, comprising:

[0027] Generate planned motion data for the robot's joints;

[0028] Acquire motion information and operational status of the robot's joints;

[0029] Obtain motion information deviations and operational status labels for the robot's joints;

[0030] Optimize the deviation of the robot's joint motion information;

[0031] A neural network is constructed, and the optimized motion information deviations of the robot's joints and the operating state labels are used as the training set to train the neural network.

[0032] Based on the deviation of the robot's instantaneous motion information of the joints, the instantaneous operating state of the robot's joints is predicted through a neural network.

[0033] The construction of a neural network specifically includes:

[0034] Step 510: Establish a neural network set containing multiple neural networks and randomly initialize the neural network parameters;

[0035] Step 520: Optimize the motion information bias to form a training set and a test set, and then divide the training set and the test set into multiple subsets respectively;

[0036] Step 530: Randomly select two or more subsets of training and test sets, adjust the first or first set of neural networks in the neural network set using the first or first set of subsets, and adjust the second or second set of neural networks in the neural network set using the second or second set of subsets.

[0037] Step 540: Compare the first or first group of neural networks with the second or second group of neural networks, and based on the comparison results, remove the worse one or group of neural networks from the set of neural networks;

[0038] Step 550: Determine whether the current number of comparisons meets the preset number of comparisons, or whether it is within the neural network set.

[0039] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor of the steps of the aforementioned method. Attached Figure Description

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0041] Figure 1 This is a schematic diagram of the robot collision diagnosis system based on neural networks according to an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of constructing a neural network according to an embodiment of the present invention. Detailed Implementation

[0043] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0044] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0045] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0046] Robot systems are not limited to the structures described below; various modifications may be made, such as omitting some of their components or adding other components. For example, a robot system may also include a base supporting the robot mechanism, with wheels or similar devices mounted at the bottom of the base, enabling the robot to move as a whole.

[0047] This invention proposes a robot collision diagnosis system based on neural networks. Specifically, it utilizes neural networks to make comprehensive judgments based on deviations in the robot's joint motion information, thereby determining whether a collision has occurred and the type of collision.

[0048] Figure 1 A robot collision diagnosis system based on a neural network is shown, comprising:

[0049] A robotic arm comprising multiple links and multiple joints, wherein one of the links is connected to another link via a joint.

[0050] An end effector is a component installed at the end of a robot arm to perform a predetermined task, such as gripping, lifting, suspending, adsorbing a workpiece, or processing the workpiece. An end effector can be a hand (gripping part), a hook, or a suction cup, etc. A robot arm can be equipped with multiple end effectors.

[0051] The control module determines the robot's path and generates planned motion data for the joints as needed. The control module may include an input device that accepts input of target position and posture. The input device may be composed of a switch, button, keyboard, mouse or touch panel, or a communication unit that communicates via a network including at least one wired or wireless connection, thereby accepting input from the operator.

[0052] The detection module is used to acquire the actual motion data and operating status of any of the joints. Specifically, it can detect or record the operating status of the joint by measuring or calculating the joint's motion data. The motion data includes, but is not limited to, the actual position, velocity, angular velocity, and torque of the joint. The operating status includes contact conditions and contact point positions. Contact conditions include, but are not limited to, "no contact," "human-machine interaction," "minor collision," and "strong collision." The contact point position is the location where the robot arm collides or interacts with an external object. Such a contact point can be located, for example, below or above the joint. External objects include, but are not limited to, people, the surrounding environment, other robot arms, and other parts of the robot arm besides the monitored joint. Therefore, the operating status can be further extended to more operating states by combining contact conditions and contact point positions. For example, "no contact," "human-machine interaction," "minor collision," and "strong collision" can be combined with the contact point position to expand to "no contact," "upper arm interaction," "upper arm minor collision," "upper arm strong collision," "lower arm interaction," "lower arm minor collision," and "lower arm strong collision." The operating status is recorded as tag information of the motion data. The tag information can be automatically acquired through image recognition or recorded through teaching or input devices.

[0053] The calculation module calculates the deviation of joint motion information; the deviation of joint motion information is the difference between the actual motion data and the planned motion data of the joint.

[0054] The optimization module optimizes the deviation of joint motion information.

[0055] The prediction module uses the optimized motion information deviation of the joints and the running state labels as a training set to build a neural network and train the neural network; based on the real-time motion information deviation of the joints, the neural network predicts the real-time running state of the robot's joints.

[0056] Although robots come in many varieties, when they are subjected to collisions, these collisions invariably cause at least one joint's motion parameter to deviate from its expected value, resulting in motion information deviation. This motion information deviation is not limited to changes in a single motion parameter, such as changes in joint position, but often involves complex changes in motion parameters, such as changes in joint position accompanied by changes in joint linear velocity and angular velocity. For this motion information deviation, existing measurement devices either cannot meet the accuracy or speed requirements, or require costly and complex sensing systems for monitoring, or require the construction of complex motion models. Furthermore, this motion information deviation is often affected by inherent factors of the robot's mechanism, such as mechanical backlash, non-rigid factors, and mechanical wear, as well as by factors such as temperature and humidity. Therefore, the motion information deviation of robot joints caused by collisions is a highly complex and strongly nonlinear data, making accurate measurement or definition virtually impossible. The neural network of this invention exhibits excellent performance in handling nonlinear problems that are difficult to express through precise mathematical formulas.

[0057] The specific steps to optimize joint motion information bias are as follows:

[0058] Step 100: Initialize the motion information deviation, generate an initial motion information deviation matrix, and normalize the initial motion information deviation matrix to obtain the normalized motion information deviation;

[0059] Step 200: Based on the randomly initialized enhancement and dimensionality reduction parameters, perform preliminary enhancement and dimensionality reduction processing on the normalized motion information deviation to obtain the preliminary optimized motion information deviation;

[0060] Step 300: Determine the enhanced dimensionality reduction parameters through chaotic inversion iteration;

[0061] Step 400: Based on the determined enhancement and dimensionality reduction parameters, perform enhancement and dimensionality reduction processing on the normalized motion information deviation to obtain the optimized motion information deviation.

[0062] In step 200, preliminary enhancement and dimensionality reduction processing is performed on the deviation of the normalized motion information, which specifically includes:

[0063] Step 210: By combining functions:

[0064] The initial data space composed of normalized motion information deviations is mapped to the implicit feature space.

[0065] Among them, F a F b and F c For different mapping functions, f a f bf c These are the corresponding function coefficients; F a Preferably, it is a linear function, F b Preferably, it is a nonlinear local function, F c Preferably, it is a nonlinear global function; in this embodiment, F a F b F c They are respectively:

[0066]

[0067] Where g ≥ 1 and h > 0. By using a combination function, significant distinguishing features in the data can be identified more quickly, making the data linearly separable and capable of mapping to infinite dimensions, resulting in more diverse decision boundaries. Furthermore, the combination function has fewer parameters, which can further reduce the computational cost.

[0068] Step 220: Perform nonlinear feature transformation in the implicit feature space to calculate the class mean kernel matrix D;

[0069] Step 230: Perform partial linear enhancement on the class mean kernel matrix D using the following formula to obtain the enhancement matrix D. s ;

[0070] D s =L×D, where L is

[0071]

[0072] L is a partially linearly enhanced diagonal matrix, where the diagonal elements are randomly generated natural numbers within the range (p, q), where p is preferably 1 and q is preferably 30; among the diagonal elements, the element in the first row and first column L... 1,1 up to the element L in row i and column i i,i The i elements are preferably equal, and the element L in the j-th row and j-th column is... j,j up to the element L in row n column n,n The n-j+1 elements are preferably equal, and the element in the i-th row and i-th column L i,i The element L in row j and column j j,j The j-i+1 elements are mutually independent natural numbers. Through partial linear augmentation, the distinguishing features of the data can be further enhanced, thereby improving data separability and computational efficiency.

[0073] Step 240; using the formula Extract the eigenvalues ​​of the enhancement matrix and sort them in descending order, where the first k eigenvalues ​​are λ1 to λ2. k The corresponding eigenvectors are α1 to α k k is a natural number, preferably k=5;

[0074] Step 250: Based on the extracted feature values ​​and feature vectors, calculate the motion information deviation after enhancement and dimensionality reduction to obtain the preliminary optimized motion information deviation.

[0075] In step 300, the enhanced dimensionality reduction parameters are determined through chaotic inversion iteration, which specifically includes:

[0076] Step 310: Establish the fitness function ;

[0077] Step 320: Establish a multidimensional solution space including multiple enhanced dimensionality reduction parameter solutions, and form a solution set composed of multiple potential problem solutions. In the multidimensional solution space, randomly initialize the solution set; the enhanced dimensionality reduction parameters include combination function parameters and some linear enhanced parameters.

[0078] Step 330: Calculate the fitness value corresponding to each solution in the solution set, and compare them to find the optimal solution;

[0079] Step 340: Update the solution set according to the following formula:

[0080]

[0081] Where X(t+1) is the solution set in the next iteration, and X(t) is the solution set in the current iteration. best X is the best solution in the solution set for the current iteration. r (t) is a randomly selected solution from the current iteration, s is a random number between (0.3, 0.7), μ is a random number between [-2, 2], rand is a randomly generated number within [0, 1], σ is a random number between (0, 1), and a t b t represents the upper and lower limits of the solution set for the current iteration, and m is an adjustment factor, preferably a random natural number between [1, 5].

[0082] Step 350: Recalculate the fitness value for each solution in the solution set and compare them to find the optimal solution;

[0083] Step 360: Calculate the difference between the fitness value of the best solution in this iteration and the fitness value of the best solution in the previous iteration, i.e., in step 330. If the difference is greater than 0, retain this iteration and proceed to the next step; otherwise, discard this iteration and proceed to step 340.

[0084] Step 370: Determine whether the current iteration number meets the preset iteration conditions. If it does, output the best solution, i.e. the best enhanced dimensionality reduction parameters; otherwise, proceed to step 340.

[0085] Chaotic inversion iteration can quickly obtain a position closer to the optimal solution, and avoid getting stuck in the local optimum during the iteration process, thus allowing it to escape the local optimum and approach the global optimum.

[0086] The enhancement and dimensionality reduction processing of the normalized motion information deviation in step 400 is the same as that in step 200. The only difference is that the parameters called in the two processes are different, which will not be described in detail here.

[0087] like Figure 2 As shown, constructing a neural network specifically includes:

[0088] Step 510: Establish a neural network set containing multiple neural networks and randomly initialize the neural network parameters; the input of the neural network is the optimized motion information deviation with corresponding labels or classifications, and the output is the running status;

[0089] Step 520: Optimize the motion information bias to form a training set and a test set, and then divide the training set and the test set into multiple subsets respectively;

[0090] Step 530: Randomly select two or more subsets of training and test sets, adjust the first or first set of neural networks in the neural network set using the first or first set of subsets, and adjust the second or second set of neural networks in the neural network set using the second or second set of subsets.

[0091] The adjustment targets for the neural network include the components, parameters, configuration, or other information of the neural network, preferably adjusting the parameter values ​​of the feature extractor of the neural network and at least one fully connected layer of the neural network that follows the feature extractor.

[0092] Step 540: Compare the first or first group of neural networks with the second or second group of neural networks, and based on the comparison results, remove the worse one or group of neural networks from the set of neural networks;

[0093] The comparison may be based at least in part on the accuracy of the neural network or the model size or the computational cost, or a combination of at least two of these. For example, the first or first group of neural networks may be removed from the set of neural networks if the accuracy of each of the first or first group of neural networks is less than the accuracy of the second or second group of neural networks, or the second or second group of neural networks may be removed from the set of neural networks if the computational cost of each of the first or first group of neural networks is less than the computational cost of the second or second group of neural networks.

[0094] Step 550: Determine whether the current number of comparisons meets the preset number of comparisons, or whether the neural network set contains only one or a group of neural networks. If so, output the best one or a group of neural networks; otherwise, proceed to step 530.

[0095] The constructed neural network comprises multiple neuron layers, with 5-10 neuron layers and each neuron layer comprising 15-25 neurons. In another embodiment, modules and / or layers of the neural network may be optionally added or removed. Optional modules and / or layers include residual layers, attention layers, recurrent layers, squeezing and activation layers, or any other type of neural network layer used to construct the neural network to perform processing, segmentation, or any other neural network function on training data (e.g., motion images or any other type of input data commonly used in neural network operations).

[0096] By constructing a neural network using the aforementioned method, the parameters of the neural network can be determined quickly, and the entire training set does not need to be accessed, which greatly reduces the amount of computation.

[0097] This invention also proposes a robot collision diagnosis method based on neural networks, which includes:

[0098] The planned motion data of the joint is generated. The planned motion data can be obtained through an input device or by calculating the planned path. The input device accepts the input of the target position and posture. It can be composed of a switch, button, keyboard, mouse or touch panel, or a communication unit that communicates through a network including at least one wired or wireless network, thereby accepting input from the operator.

[0099] The system acquires actual joint motion data and operational status. Specifically, it measures or calculates joint motion data to detect or record the joint's operational status. Motion data includes, but is not limited to, the joint's actual position, velocity, angular velocity, and torque. Operational status includes contact conditions and contact point locations. Contact conditions include, but are not limited to, "no contact," "human-machine interaction," "minor collision," and "strong collision." Contact point locations are the points where the robot arm collides or interacts with external objects. Such contact points can be located, for example, below or above the joint. External objects include, but are not limited to, people, the surrounding environment, other robot arms, and other parts of the robot arm besides the monitored joint. Therefore, operational status can be further expanded to more operational statuses through combinations of contact conditions and contact point locations. For example, "no contact," "human-machine interaction," "minor collision," and "strong collision" can be combined with contact point locations to expand to "no contact," "upper arm interaction," "upper arm minor collision," "upper arm strong collision," "lower arm interaction," "lower arm minor collision," and "lower arm strong collision." Operational status is recorded as tag information for motion data. Tag information can be automatically acquired through image recognition or recorded through teaching or input devices.

[0100] Calculate the joint motion information deviation; the joint motion information deviation is the difference between the actual motion data and the planned motion data of the joint.

[0101] Optimize joint motion information deviations;

[0102] The optimized motion information deviation of the joints and the running state labels are used as the training set to construct a neural network, which is then trained. Based on the real-time motion information deviation of the joints, the real-time running state of the robot's joints is predicted through the neural network.

[0103] Although robots come in many varieties, when they are subjected to collisions, these collisions invariably cause at least one joint's motion parameter to deviate from its expected value, resulting in motion information deviation. This motion information deviation is not limited to changes in a single motion parameter, such as changes in joint position, but often involves complex changes in motion parameters, such as changes in joint position accompanied by changes in joint linear velocity and angular velocity. For this motion information deviation, existing measurement devices either cannot meet the accuracy or speed requirements, or require costly and complex sensing systems for monitoring, or require the construction of complex motion models. Furthermore, this motion information deviation is often affected by inherent factors of the robot's mechanism, such as mechanical backlash, non-rigid factors, and mechanical wear, as well as by factors such as temperature and humidity. Therefore, the motion information deviation of robot joints caused by collisions is a highly complex and strongly nonlinear data, making accurate measurement or definition virtually impossible. The neural network of this invention exhibits excellent performance in handling nonlinear problems that are difficult to express through precise mathematical formulas.

[0104] The specific steps to optimize joint motion information bias are as follows:

[0105] Step 100: Initialize the motion information deviation, generate an initial motion information deviation matrix, and normalize the initial motion information deviation matrix to obtain the normalized motion information deviation;

[0106] Step 200: Based on the randomly initialized enhancement and dimensionality reduction parameters, perform preliminary enhancement and dimensionality reduction processing on the normalized motion information deviation to obtain the preliminary optimized motion information deviation;

[0107] Step 300: Determine the enhanced dimensionality reduction parameters through chaotic inversion iteration;

[0108] Step 400: Based on the determined enhancement and dimensionality reduction parameters, perform enhancement and dimensionality reduction processing on the normalized motion information deviation to obtain the optimized motion information deviation.

[0109] In step 200, preliminary enhancement and dimensionality reduction processing is performed on the deviation of the normalized motion information, which specifically includes:

[0110] Step 210: By combining functions:

[0111] The initial data space composed of normalized motion information deviations is mapped to the implicit feature space.

[0112] Among them, F a F b and F c For different mapping functions, f a f b f c These are the corresponding function coefficients; F a Preferably, it is a linear function, F b Preferably, it is a nonlinear local function, F c Preferably, it is a nonlinear global function; in this embodiment, F a F b F c They are respectively:

[0113]

[0114] Where g ≥ 1 and h > 0. By using a combination function, significant distinguishing features in the data can be identified more quickly, making the data linearly separable and capable of mapping to infinite dimensions, resulting in more diverse decision boundaries. Furthermore, the combination function has fewer parameters, which can further reduce the computational cost.

[0115] Step 220: Perform nonlinear feature transformation in the implicit feature space to calculate the class mean kernel matrix D;

[0116] Step 230: Perform partial linear enhancement on the class mean kernel matrix D using the following formula to obtain the enhancement matrix D. s ;

[0117] D s =L×D, where L is

[0118]

[0119] L is a partially linearly enhanced diagonal matrix, where the diagonal elements are randomly generated natural numbers within the range (p, q), where p is preferably 1 and q is preferably 30; among the diagonal elements, the element in the first row and first column L... 1,1 up to the element L in row i and column i i,i The i elements are preferably equal, and the element L in the j-th row and j-th column is... j,j up to the element L in row n column n,n The n-j+1 elements are preferably equal, and the element in the i-th row and i-th column Li,i The element L in row j and column j j,j The j-i+1 elements are mutually independent natural numbers. Through partial linear augmentation, the distinguishing features of the data can be further enhanced, thereby improving data separability and computational efficiency.

[0120] Step 240; using the formula Extract the eigenvalues ​​of the enhancement matrix and sort them in descending order, where the first k eigenvalues ​​are λ1 to λ2. k The corresponding eigenvectors are α1 to α k k is a natural number, preferably k=5;

[0121] Step 250: Based on the extracted feature values ​​and feature vectors, calculate the motion information deviation after enhancement and dimensionality reduction to obtain the preliminary optimized motion information deviation.

[0122] In step 300, the enhanced dimensionality reduction parameters are determined through chaotic inversion iteration, which specifically includes:

[0123] Step 310: Establish the fitness function ;

[0124] Step 320: Establish a multidimensional solution space including multiple enhanced dimensionality reduction parameter solutions, and form a solution set composed of multiple potential problem solutions. In the multidimensional solution space, randomly initialize the solution set; the enhanced dimensionality reduction parameters include combination function parameters and some linear enhanced parameters.

[0125] Step 330: Calculate the fitness value corresponding to each solution in the solution set, and compare them to find the optimal solution;

[0126] Step 340: Update the solution set according to the following formula:

[0127]

[0128] Where X(t+1) is the solution set in the next iteration, and X(t) is the solution set in the current iteration. best X is the best solution in the solution set for the current iteration. r (t) is a randomly selected solution from the current iteration, s is a random number between (0.3, 0.7), μ is a random number between [-2, 2], rand is a randomly generated number within [0, 1], σ is a random number between (0, 1), and a t b t represents the upper and lower limits of the solution set for the current iteration, and m is an adjustment factor, preferably a random natural number between [1, 5].

[0129] Step 350: Recalculate the fitness value for each solution in the solution set and compare them to find the optimal solution;

[0130] Step 360: Calculate the difference between the fitness value of the best solution in this iteration and the fitness value of the best solution in the previous iteration, i.e., in step 330. If the difference is greater than 0, retain this iteration and proceed to the next step; otherwise, discard this iteration and proceed to step 340.

[0131] Step 370: Determine whether the current iteration number meets the preset iteration conditions. If it does, output the best solution, i.e. the best enhanced dimensionality reduction parameters; otherwise, proceed to step 340.

[0132] Chaotic inversion iteration can quickly obtain a position closer to the optimal solution, and avoid getting stuck in the local optimum during the iteration process, thus allowing it to escape the local optimum and approach the global optimum.

[0133] The enhancement and dimensionality reduction processing of the normalized motion information deviation in step 400 is the same as that in step 200. The only difference is that the parameters called in the two processes are different, which will not be described in detail here.

[0134] Constructing a neural network specifically includes:

[0135] Step 510: Establish a neural network set containing multiple neural networks and randomly initialize the neural network parameters; the input of the neural network is the optimized motion information deviation with corresponding labels or classifications, and the output is the running status;

[0136] Step 520: Optimize the motion information bias to form a training set and a test set, and then divide the training set and the test set into multiple subsets respectively;

[0137] Step 530: Randomly select two or more subsets of training and test sets, adjust the first or first set of neural networks in the neural network set using the first or first set of subsets, and adjust the second or second set of neural networks in the neural network set using the second or second set of subsets.

[0138] The adjustment targets for the neural network include the components, parameters, configuration, or other information of the neural network, preferably adjusting the parameter values ​​of the feature extractor of the neural network and at least one fully connected layer of the neural network that follows the feature extractor.

[0139] Step 540: Compare the first or first group of neural networks with the second or second group of neural networks, and based on the comparison results, remove the worse one or group of neural networks from the set of neural networks;

[0140] The comparison may be based at least in part on the accuracy of the neural network or the model size or the computational cost, or a combination of at least two of these. For example, the first or first group of neural networks may be removed from the set of neural networks if the accuracy of each of the first or first group of neural networks is less than the accuracy of the second or second group of neural networks, or the second or second group of neural networks may be removed from the set of neural networks if the computational cost of each of the first or first group of neural networks is less than the computational cost of the second or second group of neural networks.

[0141] Step 550: Determine whether the current number of comparisons meets the preset number of comparisons, or whether the neural network set contains only one or a group of neural networks. If so, output the best one or a group of neural networks; otherwise, proceed to step 530.

[0142] The constructed neural network comprises multiple neuron layers, with 5-10 neuron layers and each neuron layer comprising 15-25 neurons. In another embodiment, modules and / or layers of the neural network may be optionally added or removed. Optional modules and / or layers include residual layers, attention layers, recurrent layers, squeezing and activation layers, or any other type of neural network layer used to construct the neural network to perform processing, segmentation, or any other neural network function on training data (e.g., motion images or any other type of input data commonly used in neural network operations).

[0143] By constructing a neural network using the aforementioned method, the neural network parameters can be determined quickly without needing to access the entire training set, significantly reducing the computational load. Compared to conventional neural network diagnostic systems, the robot collision diagnosis system of this invention has a significantly lower false alarm rate, while simultaneously improving prediction accuracy and speed, and reducing computational load.

[0144] The various methods described above can, in some embodiments, be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on a device via a ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more actions or steps of the methods described above can be performed. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for performing various aspects of this disclosure are loaded. The computer-readable storage medium may be a tangible device capable of holding and storing instructions used by an instruction execution device. The computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof.

Claims

1. A neural network-based robot collision diagnosis system, comprising: a robot arm comprising a plurality of links and a plurality of joints, one of the plurality of links being connected to another link by a joint; an end effector; a control module generating planned motion data of the joints; a detection module acquiring actual motion data and operating state of the joints; a calculation module calculating motion information deviation of the joints; an optimization module optimizing the motion information deviation of the joints; a prediction module constructing a neural network using the optimized motion information deviation and operating state label as a training set, training the neural network; predicting an instant operating state of the joints of the robot by the neural network based on instant motion information deviation of the joints; wherein the neural network is constructed, specifically comprising: step 510: establishing a neural network set comprising a plurality of neural networks and randomly initializing neural network parameters; step 520: grouping the optimized motion information deviation into a training set and a test set, and dividing the training set and the test set into a plurality of subsets respectively; step 530: randomly selecting two groups of training set and test set subsets or multiple groups of training set and test set subsets, adjusting a first or first group of neural networks in the neural network set using a first or first group of subsets in the two groups of subsets or multiple groups of subsets, and adjusting a second or second group of neural networks in the neural network set using a second or second group of subsets in the two groups of subsets or multiple groups of subsets; step 540: comparing the first or first group of neural networks with the second or second group of neural networks, and removing a poorer one or group of neural networks from the neural network set according to a comparison result; step 550: determining whether a preset comparison number is met or only one or a group of neural networks is included in the neural network set, and outputting a best one or group of neural networks if yes, otherwise returning to step 530; wherein the neural network input is the optimized motion information deviation with corresponding labels or classifications, and the output is the operating state; the operating state of the joint comprises a contact condition and a contact point position, and the contact condition comprises no contact, human-robot interaction, slight collision and strong collision.

2. The neural network-based robot collision diagnosis system of claim 1, wherein: The adjusted neural network is the parameter value of the feature extractor of the adjusted neural network and at least one fully connected layer of the neural network following the feature extractor.

3. The neural network-based robotic collision diagnosis system of claim 1, wherein: The comparison in step 540 can be based at least in part on the accuracy or model size or computational amount of the neural network or a combination of at least two thereof.

4. The neural network-based robotic collision diagnosis system of claim 1, wherein: The constructed neural network comprises a plurality of neuron layers, the number of the neuron layers being 5-10, and each of the neuron layers comprises a plurality of neurons, the number of the neurons being 15-25.

5. The neural network-based robotic collision diagnosis system of claim 1, wherein: The motion information deviation of the joint is the difference between the actual motion data and the planned motion data of the joint.

6. The neural network-based robotic collision diagnosis system of claim 1, wherein: The motion data of the joint comprises position, velocity, angular velocity and torque of the joint.

7. A neural network-based robot collision diagnosis method, comprising: generating planned motion data of joints of a robot; acquiring motion information and operating state of the joints of the robot; acquiring motion information deviation and operating state label of the joints of the robot; optimizing the motion information deviation of the joints of the robot; The neural network is constructed, and the optimized motion information deviation of the joints of the robot and the running state label are taken as a training set to train the neural network; Based on the instant motion information deviation of the joints of the robot, the instant running state of the joints of the robot is predicted through the neural network; The neural network is constructed, and the optimized motion information deviation of the joints of the robot and the running state label are taken as a training set to train the neural network; Step 510: a neural network set containing multiple neural networks is established, and neural network parameters are randomly initialized; Step 520: the optimized motion information deviation is divided into a training set and a test set, and the training set and the test set are respectively divided into multiple subsets; Step 530: two sets of training set and test set subsets or multiple sets of training set and test set subsets are randomly selected, a first or first set of neural networks in the neural network set is adjusted using a first or first set of subsets in the two sets of subsets or the multiple sets of subsets, and a second or second set of neural networks in the neural network set is adjusted using a second or second set of subsets in the two sets of subsets or the multiple sets of subsets; Step 540: the first or first set of neural networks and the second or second set of neural networks are compared, and according to the comparison result, a poorer one or set of neural networks is removed from the neural network set; Step 550: whether the current comparison times meet a preset comparison times or whether the neural network set only includes one or one set of neural networks is judged, if yes, the best one or one set of neural networks is output, and if not, the step 530 is entered; The neural network input is the optimized motion information deviation with corresponding labels or classifications, and the output is the running state. The running state of the joint includes a contact condition and a contact point position, and the contact condition includes no contact, human-robot interaction, slight collision and strong collision.

8. A computer-readable storage medium, characterized in that, The computer program is stored on the computer readable medium, and the computer program is executed by the processor to perform the steps of the method according to claim 7.

Images