Large-load robot control method based on artificial intelligence

Abstract

The invention discloses a heavy-load robot control method based on artificial intelligence, and relates to the technical field of load control, and the method comprises the steps: extracting the local time sequence characteristics of a rotation parameter through a CNN (convolutional neural network), and analyzing the dynamic change and load characteristics of coal rock hardness in combination with a Transform model of an attention mechanism; fusion with real-time geological data and load prediction of a multi-layer perceptron (MLP) are carried out; a deep deterministic strategy gradient (DDPG) algorithm in a deep reinforcement learning DRL framework is adopted, a torque set value is adjusted in real time based on a load prediction result, a strategy network generates an optimal adjustment action through an environment state, and a value network evaluates an action value and quickly responds to load changes under complex working conditions; a feedback mechanism is introduced to further improve the adaptive capacity, the executed torque state and performance data serve as feedback input, parameters of a Transform model and a DRL strategy network are adjusted online, and prediction and actual deviation are continuously optimized through a loss function in a mean square error (MSE) form.

Description

Technical Field

[0001] The invention relates to the technical field of load control, and in particular to a large-load robot control method based on artificial intelligence. Background Art

[0002] During the operation of drilling machines underground in coal mines, sudden changes in geological conditions are often encountered, such as changes in coal rock hardness, the appearance of interlayers or fault structures. These sudden changes will cause an instantaneous increase in load, causing instability in the drilling machine system, and even problems such as drill sticking, drill breakage or equipment damage.

[0003] The current large-load control methods mostly rely on PID control with fixed parameters or compensation strategies based on preset models. When the hardness of coal and rock changes suddenly, PID control will have significant response lag, resulting in the inability to adjust the torque to a safe range in time, causing the instantaneous torque to exceed the allowable range of the equipment, increasing the risk of drill sticking or drill break accidents; at the same time, constant torque control is difficult to cope with sudden changes in hardness. In order to reduce the risk of overload, some solutions adopt a conservative strategy to reduce the drilling speed or limit the torque output, which ensures equipment safety to a certain extent, but at the same time reduces operating efficiency.

[0004] It can be seen that in the environment of load mutation and dynamic interference, the traditional robot large-load control scheme has insufficient adaptability. When dealing with geological changes with strong nonlinear characteristics, it is unable to quickly adjust the torque, which will pose a potential threat to operating efficiency and equipment safety. Therefore, there is an urgent need for an artificial intelligence-based large-load robot control solution to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides an artificial intelligence-based large-load robot control method to solve the problems of traditional large-load robot control methods in a load mutation environment, delayed response, insufficient adjustment ability, and difficulty in quickly adapting to complex dynamic loads.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: The embodiment of the present invention provides a large-load robot control method based on artificial intelligence, which includes: Step S1, real-time collection of the rotation parameters of the robot during operation, pre-processing, denoising and eliminating invalid data; Step S2, based on the preprocessing results of step S1, the convolutional neural network (CNN) is used to extract the features of the rotation parameters, and a deep learning Transformer model combined with an attention mechanism is constructed to analyze the dynamic changes of coal and rock hardness and its impact on load behavior; Step S3, combining the output result of the Transformer model in step S2 with the geological data currently collected in real time to generate a comprehensive load forecast; Step S4, based on the load prediction result generated in step S3, the deep reinforcement learning (DRL) framework is used to perform real-time strategy optimization: The load prediction results are used as the input of the reinforcement learning environment. The policy network based on the deep deterministic policy gradient (DDPG) algorithm learns to generate the optimal decision for torque adjustment. According to the results of the network output, the torque setting value of the robot is adjusted in real time. In step S5, the torque state and performance data after the execution of step S4 are fed back to the prediction model of step S2 and the DRL framework of step S4, and the Transformer model and the DRL strategy network are adjusted online.

[0008] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the rotation parameters include drilling pressure, torque, vibration signal and drilling speed.

[0009] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, in which: in step S2, the output results of the Transformer model include hardness change trend, load characteristics and the torque demand prediction value matching therewith.

[0010] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the steps of using a convolutional neural network CNN to extract features of rotation parameters, constructing a deep learning Transformer model combined with an attention mechanism, and analyzing the dynamic changes in coal and rock hardness and its impact on load behavior are as follows: The acquired rotation parameters are defined as : , in, A dataset representing the rotation parameters, Indicates The multi-dimensional rotation parameter vector of each time step includes drilling pressure, torque, vibration signal and drilling speed. Indicates the total length of the time series data, Preprocessing is performed, and the processing formula is: , in, represents the normalized input data, represents the mean of the data, Indicates the standard deviation of the data; Convolutional neural network CNN is used for feature extraction, and one-dimensional convolutional network is used to extract time series features. The extraction formula is: , in, Indicates The feature vector of the time step, represents the weight matrix of the convolution kernel, represents a one-dimensional convolution operation, Indicates The normalized input vector of time steps, represents the bias of the convolutional layer, represents the linear rectification function.

[0011] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, wherein: in step S2, a Transformer model combined with an attention mechanism is constructed, and the feature sequence is processed by a multi-head attention mechanism, and the processing formula is: , in, Represents the output of multi-head attention, represents the query matrix, represents the key matrix, represents the value matrix, The weight matrices representing queries, keys, and values, respectively, represents the dimension of the key, is the normalized activation function; The Transformer decoding layer uses a feedforward neural network to parse features. The parsing formula is: , in, Represents the output of the Transformer decoding layer, including hardness change trend, load characteristics and torque demand prediction value, Represents the decoding layer structure based on feedforward neural network; Final output : , in, Indicates the hardness change trend. Indicates the load characteristics, Indicates the torque demand prediction value.

[0012] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the prediction content includes the future load peak, the time point of sudden change in coal and rock hardness and its specific impact on torque.

[0013] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the current real-time geological data in step S3 includes drilling depth, changes in coal and rock composition, microcrack distribution, inclusion information, and formation temperature and humidity conditions.

[0014] As a preferred solution of the artificial intelligence-based large-load robot control method of the present invention, the step of combining the output result of the Transformer model in step S2 with the current real-time collected geological data to generate a comprehensive load prediction is: The output of Transformer and real-time geological data Fusion, the fusion formula is: , in, represents the integrated feature set after fusion, Represents the output of the Transformer model, Represents geological data, including drilling depth, coal and rock composition changes, microcrack distribution, inclusion information, and formation temperature and humidity conditions. Represents feature concatenation operation; The comprehensive load is predicted by the multi-layer perceptron MLP, and the prediction formula is: , in, Represents the prediction result, Represents a multi-layer perceptron, including an input layer, two hidden layers and an output layer. The activation function is ReLU. Output load prediction results, the prediction results are expressed as : , in, represents the future load peak, Indicates the time point when the hardness of coal rock changes suddenly. Indicates the specific impact of hardness changes on torque.

[0015] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the step of using the deep reinforcement learning DRL framework to perform real-time strategy optimization based on the load prediction result generated in step S3 is: The reinforcement learning environment is defined as the load prediction result, expressed as: , in, represents the state of the reinforcement learning environment, Indicates load prediction results, including load peak, hardness mutation time point and torque impact; The policy network and value network based on the deep deterministic policy gradient DDPG algorithm are expressed as: , , in, Indicates the current action, i.e. the torque regulation strategy, represents the policy network, represents the action-value function of the value network, Represents the reward value, represents the discount factor, Indicates the next state, represents the optimal action in the next state, and Represent the parameters of the policy network and the value network respectively, Indicates the desired action; Update the torque setting value according to the output of the strategy network. The update formula is: , in, represents the updated torque value, Indicates the current torque value. Represents the adjustment amount, generated by the policy network.

[0016] As a preferred solution of the artificial intelligence-based large-load robot control method described in the present invention, the torque state and performance data after the execution of step S4 are fed back to the prediction model of step S2 and the DRL framework of step S4, and the step of online adjustment of the Transformer model and the DRL strategy network is as follows: Collect post-execution status and performance data as feedback, expressed as: , in, Represents feedback data, Indicates the current torque state. Represents the current reward value, The feedback data is used to adjust the model parameters. The adjustment formula is: , in, Represents model parameters, including parameters of Transformer and policy networks, represents the learning rate, Represents the loss function, specifically: , in, represents the mean square error MSE, represents the number of samples, represents the predicted action value, Indicates actual reward.

[0017] The beneficial effects of the present invention are as follows: the present invention extracts the local time series characteristics of the rotation parameters through the convolutional neural network CNN, and analyzes the dynamic changes of the coal rock hardness and the load characteristics in combination with the Transformer model of the attention mechanism, and accurately estimates the future load peak, the hardness mutation time point and its impact on the torque by integrating the real-time geological data and the load prediction of the multi-layer perceptron MLP; in terms of control strategy optimization, the deep deterministic policy gradient DDPG algorithm in the deep reinforcement learning DRL framework is adopted to adjust the torque set value in real time based on the load prediction result, the policy network generates the optimal adjustment action through the environmental state, and the value network evaluates the action value, so as to quickly respond to the load changes under complex working conditions and ensure the accuracy of the torque regulation; the feedback mechanism is introduced to further improve the adaptability, and the torque state and performance data after execution are used as feedback input, and the parameters of the Transformer model and the DRL policy network are adjusted online, and the deviation between the prediction and the actual is continuously optimized through the loss function in the form of mean square error MSE.

[0018] The present invention improves the robustness and flexibility in complex load environments, significantly reduces the hysteresis risk in traditional methods, and comprehensively improves the operating efficiency of large-load robot control solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0020] Figure 1 It is a flow chart of the artificial intelligence-based large-load robot control method of the present invention. DETAILED DESCRIPTION

[0021] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, reference Figure 1 , this embodiment provides a large-load robot control method based on artificial intelligence, including: Step S1, real-time collection of the rotation parameters of the robot during operation, pre-processing, denoising and eliminating invalid data; Rotational parameters include weight on bit, torque, vibration signal and drilling speed; Step S2, based on the preprocessing results of step S1, the convolutional neural network (CNN) is used to extract the features of the rotation parameters, and a deep learning Transformer model combined with an attention mechanism is constructed to analyze the dynamic changes of coal and rock hardness and its impact on load behavior; In step S2, the output results of the Transformer model include hardness change trend, load characteristics and torque demand prediction values ​​matched thereto; The convolutional neural network (CNN) is used to extract the features of the rotation parameters, and a deep learning Transformer model combined with the attention mechanism is constructed to analyze the dynamic changes of coal and rock hardness and its impact on load behavior. The acquired rotation parameters are defined as : , in, A dataset representing the rotation parameters, Indicates The multi-dimensional rotation parameter vector of each time step includes drilling pressure, torque, vibration signal and drilling speed. Indicates the total length of the time series data, Preprocessing is performed, and the processing formula is: , in, represents the normalized input data, represents the mean of the data, Indicates the standard deviation of the data; Convolutional neural network CNN is used for feature extraction, and one-dimensional convolutional network is used to extract time series features. The extraction formula is: , in, Indicates The feature vector of the time step, represents the weight matrix of the convolution kernel, represents a one-dimensional convolution operation, Indicates The normalized input vector of time steps, represents the bias of the convolutional layer, represents the linear rectification function; In step S2, a Transformer model combined with an attention mechanism is constructed to process feature sequences through a multi-head attention mechanism. The processing formula is: , in, Represents the output of multi-head attention, represents the query matrix, represents the key matrix, represents the value matrix, The weight matrices representing queries, keys, and values, respectively, represents the dimension of the key, is the normalized activation function; The Transformer decoding layer uses a feedforward neural network to parse features. The parsing formula is: , in, Represents the output of the Transformer decoding layer, including hardness change trend, load characteristics and torque demand prediction value, Represents the decoding layer structure based on feedforward neural network; Final output : , in, Indicates the hardness change trend. Indicates the load characteristics, Indicates the torque demand prediction value; Specifically, the CNN and Transformer models are combined to extract and analyze the features of the time series rotation parameters. The CNN network captures the local time series features, and the Transformer uses the attention mechanism to improve the global feature expression ability. The decoding layer analyzes the features and outputs the hardness change trend, load characteristics and torque demand prediction value. Step S3, combining the output result of the Transformer model in step S2 with the geological data currently collected in real time to generate a comprehensive load forecast; The prediction content includes the future load peak, the time point of coal rock hardness mutation and its specific impact on torque; The current real-time geological data in step S3 includes drilling depth, coal rock composition changes, microcrack distribution, inclusion information, and formation temperature and humidity conditions; The steps of combining the output of the Transformer model in step S2 with the current real-time collected geological data to generate a comprehensive load forecast are: The output of Transformer and real-time geological data Fusion, the fusion formula is: , in, represents the integrated feature set after fusion, Represents the output of the Transformer model, Represents geological data, including drilling depth, coal and rock composition changes, microcrack distribution, inclusion information, and formation temperature and humidity conditions. Represents feature concatenation operation; The comprehensive load is predicted by the multi-layer perceptron MLP, and the prediction formula is: , in, Represents the prediction result, Represents a multi-layer perceptron, including an input layer, two hidden layers and an output layer. The activation function is ReLU. Output load prediction results, the prediction results are expressed as : , in, represents the future load peak, Indicates the time point when the hardness of coal rock changes suddenly. Indicates the specific effect of hardness change on torque; Specifically, step S3 combines the dynamic output results of Transformer with real-time geological data and uses the MLP model to generate load prediction results, including load peak, hardness mutation time point and torque influence, which improves the comprehensiveness of load prediction. Step S4, based on the load prediction result generated in step S3, the deep reinforcement learning (DRL) framework is used to perform real-time strategy optimization: The load prediction results are used as the input of the reinforcement learning environment. The policy network based on the deep deterministic policy gradient (DDPG) algorithm learns to generate the optimal decision for torque adjustment. According to the results of the network output, the torque setting value of the robot is adjusted in real time. Based on the load prediction results generated in step S3, the steps of using the deep reinforcement learning DRL framework to optimize the real-time strategy are as follows: The reinforcement learning environment is defined as the load prediction result, expressed as: , in, represents the state of the reinforcement learning environment, Indicates load prediction results, including load peak, hardness mutation time point and torque impact; The policy network and value network based on the deep deterministic policy gradient DDPG algorithm are expressed as: , , in, Indicates the current action, i.e. the torque regulation strategy, represents the policy network, represents the action-value function of the value network, Represents the reward value, represents the discount factor, Indicates the next state, represents the optimal action in the next state, and Represent the parameters of the policy network and the value network respectively, Indicates the desired action; Update the torque setting value according to the output of the strategy network. The update formula is: , in, represents the updated torque value, Indicates the current torque value. represents the adjustment amount, generated by the policy network; Specifically, in this step, the robot torque regulation strategy is optimized through the DDPG algorithm. The strategy network generates actions according to the environmental status, the value network evaluates the action value, and adjusts the torque setting value in real time, so that the robot can quickly adapt and optimize performance under complex working conditions.

[0025] Step S5, feeding back the torque state and performance data after the execution of step S4 to the prediction model of step S2 and the DRL framework of step S4, and adjusting the Transformer model and the DRL strategy network online; The torque state and performance data after the execution of step S4 are fed back to the prediction model of step S2 and the DRL framework of step S4. The steps of online adjustment of the Transformer model and the DRL strategy network are as follows: Collect post-execution status and performance data as feedback, expressed as: , in, Represents feedback data, Indicates the current torque state. Represents the current reward value, The feedback data is used to adjust the model parameters. The adjustment formula is: , in, Represents model parameters, including parameters of Transformer and policy networks, represents the learning rate, Represents the loss function, specifically: , in, represents the mean square error MSE, represents the number of samples, represents the predicted action value, Indicates actual reward; Specifically, through the feedback mechanism, the torque state and reward value after execution are used to update the model parameters online, and the Transformer and policy network are continuously optimized. The loss function evaluates the deviation between prediction and reality in the form of mean square error to improve the adaptability of the model.

[0026] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A large-load robot control method based on artificial intelligence, characterized in that: include, Step S1, real-time collection of the rotation parameters of the robot during operation, pre-processing, denoising and eliminating invalid data; Step S2, based on the preprocessing results of step S1, the convolutional neural network (CNN) is used to extract the features of the rotation parameters, and a deep learning Transformer model combined with an attention mechanism is constructed to analyze the dynamic changes of coal and rock hardness and its impact on load behavior; Step S3, combining the output result of the Transformer model in step S2 with the geological data currently collected in real time to generate a comprehensive load forecast; Step S4, based on the load prediction result generated in step S3, the deep reinforcement learning (DRL) framework is used to perform real-time strategy optimization: The load prediction results are used as the input of the reinforcement learning environment. The policy network based on the deep deterministic policy gradient (DDPG) algorithm learns to generate the optimal decision for torque adjustment. According to the results of the network output, the torque setting value of the robot is adjusted in real time. In step S5, the torque state and performance data after the execution of step S4 are fed back to the prediction model of step S2 and the DRL framework of step S4, and the Transformer model and the DRL strategy network are adjusted online.

2. The artificial intelligence-based large-load robot control method according to claim 1, characterized in that: The rotation parameters include drilling pressure, torque, vibration signal and drilling speed.

3. A method for controlling a large-load robot based on artificial intelligence as claimed in claim 2, characterized in that: In step S2, the output results of the Transformer model include hardness change trend, load characteristics and the corresponding torque demand prediction value.

4. The artificial intelligence-based large-load robot control method as claimed in claim 3, characterized in that: The steps of using the convolutional neural network CNN to extract the features of the rotation parameters, constructing a deep learning Transformer model combined with the attention mechanism, and analyzing the dynamic changes of coal rock hardness and its influence on load behavior are as follows: The acquired rotation parameters are defined as : ， in, A dataset representing the rotation parameters, Indicates The multi-dimensional rotation parameter vector of each time step includes drilling pressure, torque, vibration signal and drilling speed. Indicates the total length of the time series data, Preprocessing is performed, and the processing formula is: ， in, represents the normalized input data, represents the mean of the data, Indicates the standard deviation of the data; Convolutional neural network CNN is used for feature extraction, and one-dimensional convolutional network is used to extract time series features. The extraction formula is: ， in, Indicates The feature vector of the time step, represents the weight matrix of the convolution kernel, represents a one-dimensional convolution operation, Indicates The normalized input vector of time steps, represents the bias of the convolutional layer, represents the linear rectification function.

5. The artificial intelligence-based large-load robot control method according to claim 4, characterized in that: In step S2, a Transformer model combined with an attention mechanism is constructed to process feature sequences through a multi-head attention mechanism. The processing formula is: ， in, Represents the output of multi-head attention, represents the query matrix, represents the key matrix, represents the value matrix, The weight matrices representing queries, keys, and values, respectively, represents the dimension of the key, is the normalized activation function; The Transformer decoding layer uses a feedforward neural network to parse features. The parsing formula is: ， in, Represents the output of the Transformer decoding layer, including hardness change trend, load characteristics and torque demand prediction value, Represents the decoding layer structure based on feedforward neural network; Final output : ， in, Indicates the hardness change trend. Indicates the load characteristics, Indicates the torque demand prediction value.

6. The artificial intelligence-based large-load robot control method according to claim 5, characterized in that: The prediction content includes the future load peak, the time point of coal rock hardness mutation and its specific impact on torque.

7. The artificial intelligence-based large-load robot control method according to claim 6, characterized in that: The current real-time geological data in step S3 includes drilling depth, coal rock composition changes, microcrack distribution, inclusion information, and formation temperature and humidity conditions.

8. The artificial intelligence-based large-load robot control method according to claim 7, characterized in that: The step of combining the output result of the Transformer model in step S2 with the geological data currently collected in real time to generate a comprehensive load forecast is: The output of Transformer and real-time geological data Fusion, the fusion formula is: ， in, represents the integrated feature set after fusion, Represents the output of the Transformer model, Represents geological data, including drilling depth, coal and rock composition changes, microcrack distribution, inclusion information, and formation temperature and humidity conditions. Represents feature concatenation operation; The comprehensive load is predicted by the multi-layer perceptron MLP, and the prediction formula is: ， in, Represents the prediction result, which represents a multi-layer perceptron, including an input layer, two hidden layers and an output layer. The activation function is ReLU. Output load prediction results, the prediction results are expressed as : ， in, represents the future load peak, Indicates the time point when the hardness of coal rock changes suddenly. Indicates the specific impact of hardness changes on torque.

9. The artificial intelligence-based large-load robot control method according to claim 8, characterized in that: The step of using the deep reinforcement learning (DRL) framework to perform real-time strategy optimization based on the load prediction result generated in step S3 is: The reinforcement learning environment is defined as the load prediction result, expressed as: ， in, represents the state of the reinforcement learning environment, Indicates load prediction results, including load peak, hardness mutation time point and torque impact; The policy network and value network based on the deep deterministic policy gradient DDPG algorithm are expressed as: ， ， in, Indicates the current action, i.e. the torque regulation strategy, represents the policy network, represents the action-value function of the value network, Represents the reward value, represents the discount factor, Indicates the next state, represents the optimal action in the next state, and Represent the parameters of the policy network and the value network respectively, Indicates the desired action; Update the torque setting value according to the output of the strategy network. The update formula is: ， in, represents the updated torque value, Indicates the current torque value. Represents the adjustment amount, generated by the policy network.

10. The artificial intelligence-based large-load robot control method according to claim 9, characterized in that: The step of feeding back the torque state and performance data after the execution of step S4 to the prediction model of step S2 and the DRL framework of step S4, and adjusting the Transformer model and the DRL strategy network online is as follows: Collect post-execution status and performance data as feedback, expressed as: ， in, Represents feedback data, Indicates the current torque state. Represents the current reward value, The feedback data is used to adjust the model parameters. The adjustment formula is: ， in, Represents model parameters, including parameters of Transformer and policy networks, represents the learning rate, Represents the loss function, specifically: ， in, represents the mean square error MSE, represents the number of samples, represents the predicted action value, Indicates actual reward.

Images