A multi-machine cooperative work industrial robot dynamic task scheduling method and system

Abstract

The application discloses a kind of multi-machine cooperative work's industrial robot dynamic task scheduling method and system, the method is according to preset sampling frequency to the state parameter of multiple industrial robots is real-time sampling, original data is divided into sampling sequence according to time window;The sampling sequence of current time window is normalized, and the distance of each robot to feature sample mean is calculated to determine the leading scheduling robot;Extract the historical sampling data of leading scheduling robot, construct the deep learning prediction model based on GRU neural network, predict the task completion state of next time window;According to the prediction result, dynamically adjust the task allocation and execution parameters of the remaining robots, realize the consistent work progress of cooperation.The application can effectively improve the cooperative work efficiency and environmental adaptability of multi-robot system, and has strong practical value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial automation control, and relates to a dynamic task scheduling method and system for industrial robots in multi-machine cooperative work. BACKGROUND

[0002] A multi-robot cooperative work system has been widely applied in the fields of automobile manufacturing, electronic assembly, logistics and warehousing, and multiple industrial robots need to cooperatively complete tasks such as assembly, carrying and processing. The overall efficiency of the system is largely dependent on the advantages and disadvantages of the task scheduling strategy. Traditional multi-robot task scheduling methods are mainly based on static rules or heuristic algorithms, such as first-come-first-served (FCFS) and shortest processing time first (SPT). These methods often perform poorly when facing dynamic production environments and are difficult to deal with complex situations such as sudden failures, changes in task priorities and resource conflicts.

[0003] Although existing dynamic task scheduling methods can adapt to environmental changes to some extent, they still have problems such as scheduling decision lag and insufficient global optimization capability. On the one hand, these methods usually make scheduling decisions based on the current system state, lacking the ability to predict future states, resulting in a lack of foresight in scheduling strategies. On the other hand, traditional methods often treat all robots as equal scheduling objects, ignoring the importance of different robots in cooperative work, and cannot achieve true collaborative optimization.

[0004] Therefore, how to design an efficient collaborative scheduling dynamic task scheduling method has become a technical problem to be solved. SUMMARY

[0005] The present application aims to solve the technical problems of insufficient prediction ability, poor collaboration and low real-time performance of existing multi-robot cooperative work system task scheduling methods. A multi-machine cooperative work industrial robot dynamic task scheduling method and system based on deep learning are provided. By introducing the concept of leading scheduling robots and GRU neural network prediction models, the future work state of the robots is accurately predicted, and dynamic task scheduling is performed based on the prediction results, thereby improving the collaborative work efficiency and adaptability of the multi-robot system. To achieve the above purpose, the specific technical solutions of the present application are as follows:

[0006] In a first aspect, the present application provides a dynamic task scheduling method for industrial robots in multi-machine cooperative work, which is used for dynamic task scheduling of a multi-robot cooperative work system. The method comprises the following steps:

[0007] Step 1: Sample the state parameters of n industrial robots in real time according to the preset sampling frequency to obtain raw data; divide the collected raw data into preset time windows to form a sampling sequence based on the time window.

[0008] Step 2: Normalize the sampling sequence of the current time window to obtain the normalized feature dataset. Find the element in the feature dataset whose distance from the mean of the feature samples is the smallest difference from the baseline value of the feature samples, and mark the robot corresponding to it as the dominant scheduling robot.

[0009] Step 3: Extract historical sampling data of the dominant scheduling robot from the time window-based sampling sequence as a sample set, and construct a deep learning-based prediction model to predict the task completion status of the dominant scheduling robot in the next time window.

[0010] Step 4: Based on the prediction result of the dominant scheduling robot at the next moment, dynamically adjust the task allocation and execution parameters of the remaining n-1 robots so that the operation progress of the remaining n-1 robots is coordinated with the operation progress of the dominant scheduling robot in the next time window.

[0011] Step 5: Repeat step 4, and then perform real-time adjustments to the tasks of the remaining n-1 robots in the subsequent time window; after the preset time, repeat steps 1-4, re-mark the dominant scheduling robot and perform real-time adjustments to the tasks of the remaining robots.

[0012] Furthermore, the robot's state parameters include: the real-time completion efficiency V of the current process, the number of workpieces buffered in the current process L, and the expected waiting time S for the downstream collaborative robot; the robot's state parameters are sampled based on a preset sampling frequency of 10-20 times per minute, with a sampling interval T of 3-6 seconds, and the collected raw data is divided according to a preset time window M. ,in Indicates the number of samples. .

[0013] The sampling sequence based on the time window is represented as follows: ,in, Indicates the first The robot in the first Sampled sequence samples within a time window, This indicates the number of robots. Each time window's sampling sequence consists of k sample data points. , Indicates the first time within a time window The sampling data from the second sampling, , .

[0014] Further, the method for normalizing the sampling sequence of the current time window to obtain the normalized feature dataset is as follows:

[0015] Let the tthtime window be the current time window, and the normalized feature dataset is represented as:

[0016] ; wherein is an element in the feature dataset, and represents the sampling sequence sample of the ithrobot in the current time window The normalized data, , respectively represent the real-time completion efficiency, the workpiece buffer quantity, and the expected waiting time of the downstream collaborative robot of the sampling sequence sample of the ithrobot in the current time window after normalization.

[0017] Further, the calculation formula of the normalization is as follows: ; ; ; wherein are the average values of the real-time completion efficiency, the workpiece buffer quantity, and the expected waiting time of the downstream collaborative robot in the current time window, respectively.

[0018] Further, the specific method for finding the element in the feature dataset whose distance to the feature sample mean is the smallest difference from the feature sample reference value, and marking the robot corresponding to the element as the leading scheduling robot is as follows:

[0019] The feature sample mean of the feature dataset in the current time window is: ; ; ;

[0020] ; the distance of the element to the feature sample mean is an independent and identically distributed random variable, and the variance is ; the average distance of the element to the feature sample mean is ; and is taken as the feature sample reference value ;

[0021] The minimum difference between the distance of the element to the feature sample mean and the distance of the element to the feature sample reference value is , and the robot corresponding to the minimum difference is the leading scheduling robot.

[0022] ​​​Further, the prediction model based on deep learning is a GRU neural network model, and the prediction model is composed of an input layer, a convolution layer, a pooling layer, a bidirectional gated recurrent layer, a Dropout layer, a full connection layer and an output layer connected in sequence.

[0023] The position coordinates, moving speed, current load and task execution progress at time x are taken as inputs, and the task completion state at time x+1 is taken as output.

[0024] The convolution layer performs one-dimensional convolution operation on the input sample data, the pooling layer simplifies the data dimension after one-dimensional convolution, the bidirectional gated recurrent layer extracts sequence features of state parameters by using bidirectional GRU, and the data after secondary feature extraction by the GRU layer is randomly discarded in the Dropout layer to prevent overfitting and transmit the processed data to the full connection layer. The full connection layer maps the output features to the label space of the sample by nonlinear combination, and outputs the prediction result of the task completion state at time x+1 by linear regression operation of the output layer.

[0025] Further, in the bidirectional gated recurrent layer, represents the state of the input layer, represents the state of the output layer, represents the hidden layer state in the transmission process of the bidirectional GRU layer, and:

[0026] ; ; ;

[0027] In the formula, represents the feature state learned by the forward layer in the one-way transmission process, is the feature state learned by the backward layer in the one-way transmission process; and are the weights of the forward layer and the backward layer to the output layer; represents the bias vector added in the output layer.

[0028] Further, after a preset time, steps 1-4 are repeated, and the preset time is 1-7 days.

[0029] Further, 80% of the sample set is randomly selected as a training set for training the neural network, and the remaining 20% of the sample set is taken as a validation set to verify the trained neural network.

[0030] The prediction result output by the neural network model is compared with the measured data in the validation set, and the fitting index R2 is used to describe the prediction performance of the model. ​​,when If the training stops when the condition is met, the backpropagation algorithm is used to adjust the model's bias and the loop continues.

[0031] goodness-of-fit index The formula for calculation is: In the formula, The first in the representative verification set The actual value of the task completion status at the next time step for each sample. Let be the predicted value of the task completion status at the next time step for the i-th sample. To verify the average completion status of the tasks in the set, m is the total number of samples.

[0032] In a second aspect, the present invention provides a dynamic task scheduling system for industrial robots that perform multi-robot collaborative operations, for executing the method of the first aspect. The system includes: a data sampling module, a dominant scheduling robot marking module, a neural network prediction module, and a dynamic task scheduling module.

[0033] Furthermore, the data sampling module is connected to the dominant scheduling robot marking module and the neural network prediction module respectively, and is used to sample the state parameters of n industrial robots in real time according to a preset sampling frequency to obtain raw data; the collected raw data is divided according to a preset time window to form a sampling sequence based on the time window.

[0034] Furthermore, the dominant scheduling robot labeling module is connected to the neural network prediction module and the dynamic task scheduling module respectively. It is used to normalize the sampling sequence of the current time window to obtain the normalized feature dataset, find the element in the feature dataset whose distance from the mean of the feature samples is the smallest difference from the baseline value of the feature samples, and label the robot corresponding to it as the dominant scheduling robot.

[0035] Furthermore, the neural network prediction module is connected to the dynamic task scheduling module, and extracts historical sampling data of the dominant scheduling robot as a sample set from the sampling sequence based on time windows, and constructs a deep learning-based prediction model to predict the task completion status of the dominant scheduling robot in the next time window.

[0036] Furthermore, the dynamic task scheduling module dynamically adjusts the task allocation and execution parameters of the remaining n-1 robots based on the prediction result of the dominant scheduling robot at the next time window, so that the operation progress of the remaining n-1 robots is coordinated with the operation progress of the dominant scheduling robot in the next time window.

[0037] Compared with the prior art, the beneficial effects of this invention are:

[0038] The application introduces the concept of a dominant scheduling robot, establishes a robot importance evaluation mechanism based on statistical principles, converts the complex multi-robot cooperative scheduling problem into a hierarchical scheduling mode with the dominant robot as the core, realizes accurate prediction of the future operation state of the robot based on a deep learning method, significantly reduces the computational complexity, and improves the scheduling efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A flow chart of a multi-robot cooperative operation industrial robot dynamic task scheduling method of the application.

[0040] Figure 2 A composition schematic diagram of a multi-robot cooperative operation industrial robot dynamic task scheduling system of the application. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions and advantages of the application clearer, the technical solutions in the application are described below in a clear and complete manner. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0042] Embodiment 1

[0043] As shown in the figure, a multi-robot cooperative operation industrial robot dynamic task scheduling method of the application is used for dynamic task scheduling of a multi-robot cooperative operation system, and the method comprises the following steps: Figure 1 Step 1: Real-time sampling of state parameters of n industrial robots according to a preset sampling frequency to obtain original data; and dividing the collected original data according to a preset time window to form a sampling sequence based on the time window.

[0044] The state parameters of the robot include real-time completion efficiency V of the current process, workpiece buffer quantity L of the current process and expected waiting time S of the downstream cooperative robot; the state parameters of the robot are sampled based on a preset sampling frequency, the sampling frequency is 10-20 times per minute, the sampling interval T is 3-6 s, and the collected original data is divided according to a preset time window M,

[0045] wherein represents the sampling number, .

[0046] The sampling sequence based on the time window is represented as wherein, represents the i th robot in the j th time window, . ​Sampled sequence samples within a time window, This indicates the number of robots. Each time window's sampling sequence consists of k sample data points. , Indicates the first time within a time window The sampling data from the second sampling, , .

[0047] In an automobile manufacturing production line, suppose six industrial robots work together to complete the assembly of car doors, including welding robots, handling robots, and adhesive application robots. What is the work efficiency of each robot? This reflects the speed at which it completes the current process, such as the welding point completion rate of a welding robot being 85 points / minute; the number of workpieces in buffer. This indicates the number of workpieces waiting to be processed, such as 12 parts piled up in front of the handling robot; downstream waiting time. This reflects the closeness of collaboration; for example, the glue-applying robot needs to wait for the upstream welding to complete, with an expected waiting time of 45 seconds. Through high-frequency sampling of 15 times per minute, the system can promptly capture minute changes in production status, ensuring the real-time nature and accuracy of scheduling decisions.

[0048] Step 2: Normalize the sampling sequence of the current time window to obtain the normalized feature dataset. Find the element in the feature dataset whose distance from the mean of the feature samples is the smallest difference from the baseline value of the feature samples, and mark the robot corresponding to it as the dominant scheduling robot.

[0049] The method for normalizing the sampling sequence of the current time window to obtain the normalized feature dataset is as follows:

[0050] Let the t-th time window be the current time window, then the normalized feature dataset is represented as:

[0051] ;in Let be an element in the feature dataset, representing the sampling sequence sample of the i-th robot in the current time window. Normalized data, , They represent the first The real-time completion efficiency of the robot in the current time window after sample sequence normalization, the number of workpieces in the buffer, and the expected waiting time of the downstream collaborative robot.

[0052] The normalization formula is: ; ; In the formula, The average value of the real-time completion efficiency of the current time window, the number of workpiece caches and the expected waiting time of the downstream collaborative robot, respectively.

[0053] Further, the specific method for marking the robot corresponding to the minimum difference between the distance of the feature data set to the feature sample mean and the feature sample reference value as the leading scheduling robot is as follows:

[0054] The feature sample mean of the feature data set of the current time window is: ; ; ;

[0055] ; the distance to the feature sample mean is an independent and identically distributed random variable, and the variance is ; the average distance of the element to the feature sample mean is ; and the feature sample reference value is ; ; ;

[0056] The minimum difference between the distance of the element to the feature sample mean and the distance of the feature sample reference value is , The robot corresponding to the minimum difference is the leading scheduling robot. Taking three robots as an example, the normalized feature vectors are , , , the feature mean is . It is found through distance calculation that R1 is closest to the mean reference value, so the robot corresponding to R1 is marked as the leading scheduling robot.

[0057] Step 3: Extract the historical sampling data of the leading scheduling robot from the sampling sequence based on the time window as a sample set, and construct a prediction model based on deep learning to predict the task completion state of the leading scheduling robot in the next time window.

[0058] The prediction model based on deep learning is a GRU neural network model, and the prediction model is composed of an input layer, a convolution layer, a pooling layer, a bidirectional gated recurrent layer, a Dropout layer, a full connection layer and an output layer connected in turn.

[0059] The position coordinates, moving speed, current load and task execution progress at time x are taken as inputs, and the task completion state at time x+1 is taken as output.

[0060] The convolutional layer performs one-dimensional convolution operation on the input sample data, the pooling layer simplifies the dimension of the one-dimensional convolution data, the bidirectional gated recurrent layer extracts the sequence features of the state parameters using bidirectional GRU, the data after the secondary feature extraction of the GRU layer is randomly discarded in the Dropout layer to prevent overfitting, and the processed data is transmitted to the full connection layer, the full connection layer maps the output features to the label space of the sample through nonlinear combination, and the prediction result of the task completion state at x+1 moment is output through the output layer linear regression operation.

[0061] In the bidirectional gated recurrent layer, at the training time of x moment, the input of the layer is represented by , the state of the input layer is represented by , the state of the output layer is represented by , and the hidden layer state in the bidirectional GRU layer transmission process is represented by

[0062] ; ; ;

[0063] In the formula, represents the feature state learned by the forward layer in the one-way transmission process, is the feature state learned by the backward layer in the one-way transmission process; and are the weights of the forward layer and the backward layer to the output layer in turn; , and represents the bias vector added in the output layer.

[0064] When the position trajectory, speed change, load fluctuation and progress data of the leading robot in the past 30 minutes are input, the model extracts local features through the convolutional layer, reduces the calculation complexity through the pooling layer, learns the timing pattern from the front and back directions through the bidirectional GRU layer, and finally predicts the task completion probability in the next 5-minute time window. During the training process, when R2 reaches 0.987, the training is stopped, at this time the prediction accuracy of the model has met the requirements of industrial application, and can accurately predict the task completion state of more than 95%.

[0065] Randomly select 80% of the sample set as the training set for training the neural network, and use the remaining 20% of the sample set as the verification set to verify the trained neural network;

[0066] Compare the prediction results output by the neural network model with the measured data in the verification set, and use the fitting index to describe the prediction performance of the model, when , the model training stop condition is met, if not, adjust the bias of the model using the back propagation algorithm and continue the cycle.​

[0067] Fitting index The calculation formula is: ; In the formula, represents the actual value of the task completion state of the next time corresponding to the i-th sample in the verification set, is the predicted value of the task completion state of the next time corresponding to the i-th sample, is the average value of the task completion state in the verification set, and m is the total number of samples.

[0068] Step 4, according to the prediction result of the leading scheduling robot in the next time, dynamically adjust the task allocation and execution parameters of the remaining n-1 robots, so that the work progress of the remaining n-1 robots is consistent with the work progress of the leading scheduling robot in the next time window.

[0069] Based on the prediction result of the leading robot, the system calculates the adjustment strategy of other robots. For example, when it is predicted that the completion progress of the leading robot in the next period will be improved by 15%, the system automatically increases the moving speed of the transfer robot by 10%, starts the downstream glue coating robot in advance, and adjusts the working tempo of the assembly robot, so that the work progress of all robots is kept synchronized. This predictive scheduling avoids production bottlenecks caused by passive response, and improves the overall efficiency of the system by 12-18%.

[0070] Step 5, repeat step 4, and adjust the task execution of the remaining n-1 robots in the subsequent time window; after a preset time, repeat steps 1-4, re-mark the leading scheduling robot and adjust the task execution of the remaining robots in real time.

[0071] The preset time is 1-7 days, and the dynamic re-marking mechanism should respond to long-term changes in the production environment. In a continuous 7-day production cycle, due to equipment wear, process adjustment or task type change, the representativeness of the original leading robot may decrease. Through regular re-evaluation and marking, it is ensured that the leading robot can always accurately reflect the overall state of the system. For example, on the 3rd day, due to the increase of welding tasks, the welding robot is promoted from the 2nd leading position to the new leading scheduling robot, and the system re-optimizes the scheduling strategy accordingly.

[0072] When the position trajectory, speed change, load fluctuation and progress data of the leading robot in the past 30 minutes are input, the model extracts local features through convolutional layers, reduces computational complexity through pooling layers, and learns time series patterns from both directions through bidirectional GRU layers. Finally, the task completion probability in the next 5-minute time window is predicted. During the training process, when R² reaches 0.987, the training is stopped, at which time the prediction accuracy of the model has met the requirements of industrial applications and can accurately predict the task completion state of more than 95%. ​

[0073] Embodiment 2

[0074] As shown in Figure 2 A composition diagram of a multi-machine cooperative industrial robot dynamic task scheduling system for executing the method of the first aspect, the system comprising: a data sampling module, a leading scheduling robot marking module, a neural network prediction module, and a dynamic task scheduling module.

[0075] The data sampling module is connected with the leading scheduling robot marking module and the neural network prediction module respectively, and is used for sampling the state parameters of n industrial robots in real time according to a preset sampling frequency to obtain original data; and the original data collected is segmented according to a preset time window to form a sampling sequence based on the time window. The module is connected with each robot controller through an industrial Ethernet, and realizes standardized data communication by using an OPC UA protocol. In each sampling period, the module simultaneously monitors the three-dimensional state parameters of all robots to ensure data synchronization and integrity. For example, in an electronic product assembly line, the module can synchronously collect 30 state parameters of 10 robots, and the data transmission delay is controlled within 50 ms.

[0076] The leading scheduling robot marking module is connected with the neural network prediction module and the dynamic task scheduling module respectively, and is used for normalizing the sampling sequence of the current time window to obtain a normalized feature data set, finding an element in the feature data set whose distance to the mean of the feature sample and the difference value of the feature sample reference value are the smallest, and marking the robot corresponding to the element as the leading scheduling robot.

[0077] The neural network prediction module is connected with the dynamic task scheduling module, extracts historical sampling data of the leading scheduling robot from the sampling sequence based on the time window as a sample set, constructs a prediction model based on deep learning for predicting the task completion state of the leading scheduling robot in the next time window. The module adopts a distributed computing architecture, deploys a lightweight GRU model on an edge computing device, and the inference time is controlled within 100 ms. In actual operation, the module can accurately predict the task completion within 15 minutes in the future based on the historical 5-hour work data of the leading robot

[0078] The dynamic task scheduling module dynamically adjusts the task allocation and execution parameters of the remaining n-1 robots according to the prediction result of the leading scheduling robot at the next time, so that the work progress of the remaining n-1 robots is consistent with the work progress of the leading scheduling robot in the next time window.

[0079] In the application of the logistics sorting center, when the prediction module prompts that the next cycle processing capacity of the leading sorting robot will increase by 20%, the scheduling module immediately adjusts the sorting strategy of the other 6 robots: increases the feeding frequency of the feeding robot, optimizes the path planning of the carrying robot, and activates the standby packaging robot in advance to ensure that the processing capacity of the entire sorting system is synchronized to increase, avoiding the impact of local bottlenecks on overall efficiency.

[0080] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A dynamic task scheduling method for multi-robot collaborative industrial robots, used for dynamic task scheduling of multiple industrial robot collaborative operation systems, characterized in that, The method includes the following steps: Step 1: Sample the state parameters of n industrial robots in real time according to the preset sampling frequency to obtain raw data; divide the collected raw data according to the preset time window to form a sampling sequence based on the time window; Step 2: Normalize the sampling sequence of the current time window to obtain the normalized feature dataset. Find the element in the feature dataset whose distance from the mean of the feature samples is the smallest difference from the baseline value of the feature samples, and mark the robot corresponding to it as the dominant scheduling robot. Let the t-th time window be the current time window, then the normalized feature dataset is represented as: ;in For each element in the feature dataset, the element represents the first element. The sampling sequence samples of the robot in the current time window Normalized data, , They represent the first The real-time completion efficiency of the robot in the current time window after normalization of the sampled sequence samples, the number of workpieces in the buffer, and the expected waiting time of the downstream collaborative robot; The specific method for finding the element in the feature dataset whose distance from the mean of the feature samples is the smallest difference from the baseline value of the feature samples, and marking the robot corresponding to that element as the dominant scheduling robot, is as follows: The mean of the feature samples in the feature dataset for the current time window is: ;element to the mean of the feature samples distance Let be independent and identically distributed random variables, and its variance be... ;element The average distance to the mean of the feature samples is Then As a baseline value for feature samples ; The minimum difference between the distance from the element to the mean of the feature samples and the distance to the baseline of the feature samples. , The corresponding robot is the dominant scheduling robot; Step 3: Extract historical sampling data of the dominant scheduling robot from the time window-based sampling sequence as a sample set, and construct a deep learning-based prediction model to predict the task completion status of the dominant scheduling robot in the next time window. Step 4: Based on the prediction result of the dominant scheduling robot at the next moment, dynamically adjust the task allocation and execution parameters of the remaining n-1 robots so that the operation progress of the remaining n-1 robots is coordinated with the operation progress of the dominant scheduling robot in the next time window. Step 5: Repeat step 4, and then perform real-time adjustments to the tasks of the remaining n-1 robots in the subsequent time window; after the preset time, repeat steps 1-4, re-mark the dominant scheduling robot and perform real-time adjustments to the tasks of the remaining robots.

2. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 1, characterized in that, The robot's status parameters include: the real-time completion efficiency of the current process. Number of workpieces buffered in the current process Expected waiting time for downstream collaborative robots The robot's state parameters are sampled based on a preset sampling frequency of 10-20 times per minute, with a sampling interval T of 3-6 seconds. The collected raw data is then processed according to a preset time window. Segmentation, ,in Indicates the number of samples. ; The sampling sequence based on the time window is represented as follows: ,in, Indicates the first The robot in the first Sampled sequence samples within a time window, This indicates the number of robots; the sampling sequence samples for each time window are all... It consists of individual sample data. , Indicates the first time within a time window The sampling data from the second sampling, , .

3. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 2, characterized in that, The prediction model based on deep learning is a GRU neural network model, which consists of an input layer, a convolutional layer, a pooling layer, a bidirectional gated recurrent layer, a Dropout layer, a fully connected layer, and an output layer connected in sequence. The inputs are the position coordinates, movement speed, current load, and task execution progress at time x, and the output is the task completion status at time x+1. The convolutional layer performs one-dimensional convolution operations on the input sample data. The pooling layer simplifies the dimension of the data after one-dimensional convolution. The bidirectional gated recurrent layer uses a bidirectional GRU to extract the sequence features of the state parameters. After the data undergoes secondary feature extraction by the GRU layer, some nodes of the data are randomly discarded in the Dropout layer to prevent overfitting. The processed data is then passed to the fully connected layer. The fully connected layer maps the output features to the label space of the samples through nonlinear combination. The output layer performs linear regression to output the prediction result of the task completion state at time x+1.

4. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 3, characterized in that, In the bidirectional gated recurrent layer, during training... At any time, with Represents the state of the input layer. Represents the output layer state. If we represent the hidden layer state during the bidirectional GRU layer propagation process, then: ； ； ； In the formula, This represents the feature states learned by the feedforward layer during the unidirectional propagation process. These are the feature states learned by the backward layer during the unidirectional transmission process; and These are the weights from the forward layer, backward layer, and so on up to the output layer; This represents the bias vector added to the output layer.

5. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 4, characterized in that, After a preset time, repeat steps 1 to 4, where the preset time is 1 to 7 days.

6. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 5, characterized in that, 80% of the sample set is randomly selected as the training set to train the neural network, and the remaining 20% ​​of the sample set is used as the validation set to validate the trained neural network. The prediction results output by the neural network model are compared with the measured data in the validation set, using a goodness-of-fit index. To describe the predictive performance of the model, ,when If the training stops when the condition is met, the backpropagation algorithm is used to adjust the model's bias and the loop continues.

7. The dynamic task scheduling method for multi-machine collaborative industrial robots according to claim 6, characterized in that, The process of adjusting the task allocation and execution parameters of the remaining n-1 robots includes: calculating the difference between the target execution progress and the actual execution progress based on the predicted task completion status of the dominant scheduling robot in the next time window; calculating the corresponding task allocation weight and execution path adjustment amount based on the magnitude and direction of the difference using a dynamic scheduling algorithm; and sending the adjustment amount to the remaining n-1 robots for real-time task scheduling.

8. A dynamic task scheduling system for industrial robots operating in a multi-robot collaborative manner, used to execute the method described in any one of claims 1-7, characterized in that, The system includes: a data sampling module, a dominant scheduling robot labeling module, a neural network prediction module, and a dynamic task scheduling module; The data sampling module is connected to the dominant scheduling robot marking module and the neural network prediction module respectively. It is used to sample the state parameters of n industrial robots in real time according to a preset sampling frequency to obtain raw data. The collected raw data is divided according to a preset time window to form a sampling sequence based on the time window. The dominant scheduling robot labeling module is connected to the neural network prediction module and the dynamic task scheduling module respectively. It is used to normalize the sampling sequence of the current time window to obtain the normalized feature dataset, find the element in the feature dataset whose distance from the mean of the feature sample is the smallest difference from the baseline value of the feature sample, and label the robot corresponding to it as the dominant scheduling robot. The neural network prediction module is connected to the dynamic task scheduling module. It extracts historical sampling data of the dominant scheduling robot as a sample set from the sampling sequence based on time windows, and constructs a deep learning-based prediction model to predict the task completion status of the dominant scheduling robot in the next time window. The dynamic task scheduling module dynamically adjusts the task allocation and execution parameters of the remaining n-1 robots based on the prediction results of the dominant scheduling robot in the next time window, so that the operation progress of the remaining n-1 robots is coordinated with the operation progress of the dominant scheduling robot in the next time window.

Images