Robot control method and system combining fuzzy control and neural network

Abstract

The embodiment of the invention relates to the technical field of robot control, in particular to a robot control method and system combining fuzzy control and a neural network. The method comprises the steps of obtaining a multi-dimensional state data set of a robot in a target operation scene, performing fuzzy control feature extraction processing on the multi-dimensional state data set, and calling a pre-trained neural network control model to perform hierarchical decision processing on the fuzzy control feature set, and generating a real-time control instruction set of the robot, adjusting the operation state of the robot in the target operation scene based on the real-time control instruction set, and feeding back the adjusted operation state to the multi-dimensional state data set to execute loop control processing. In this way, the robot can be controlled more accurately, intelligently and dynamically, and the problems that in the prior art, control is not accurate under a complex scene, and adaptability is poor are effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of robot control, and in particular, to a robot control method and system combining fuzzy control and neural network. Background Art

[0002] In the field of robot control, with the increasingly complex and diverse application scenarios of robots, higher requirements are put forward for the intelligent control and adaptability of robots. Traditional robot control technologies often only focus on single-dimensional data information, or have poor adaptability to complex environments and task changes.

[0003] Existing robot control solutions mostly make decisions based on a single type of data when facing complex and changeable operating scenarios, and cannot comprehensively consider various factors in the operation of the robot, resulting in inaccurate decisions and difficulty in coping with the challenges brought by complex environments. In addition, the existing technology lacks an effective processing mechanism for uncertain and fuzzy data, and it is difficult to accurately make control decisions when encountering fuzzy or uncertain data; and traditional control methods are difficult to dynamically adjust control strategies according to real-time changes, making it easy for robots to have problems such as unstable operation and low task execution efficiency during the task execution process. Summary of the Invention

[0004] In order to at least overcome the above deficiencies in the prior art, one of the purposes of the present invention is to provide a robot control method and system combining fuzzy control and neural network.

[0005] An embodiment of the present invention provides a robot control method combining fuzzy control and neural network, including: obtaining a multi-dimensional state data set of the robot in a target operating scenario, where the multi-dimensional state data set includes device attitude data, environmental interference data, and task constraint conditions; performing fuzzy control feature extraction processing on the multi-dimensional state data set to obtain a fuzzy control feature set of the robot in the current operating cycle, where the fuzzy control feature set includes membership distribution features and fuzzy rule matching features; calling a pre-trained neural network control model to perform hierarchical decision-making processing on the fuzzy control feature set to generate a real-time control instruction set of the robot, where the real-time control instruction set includes a motion trajectory adjustment instruction and a device state compensation instruction; adjusting the operating state of the robot in the target operating scenario based on the real-time control instruction set, and feeding back the adjusted operating state to the multi-dimensional state data set to perform loop control processing.

[0006] An embodiment of the present invention also provides a robot control system, including a processor, a memory, and a bus connected to the processor; wherein, the processor and the memory complete communication with each other through the bus; the processor is configured to call program instructions in the memory to execute the above-mentioned robot control method combining fuzzy control and neural network.

[0007] An embodiment of the present invention also provides a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, it implements the above-mentioned robot control method combining fuzzy control and neural network.

[0008] Applying the embodiment of the present invention, first, a multi-dimensional state data set of the robot in the target operation scenario is obtained (covering various aspects of information such as device posture, environmental interference, and task constraint conditions), laying a foundation for comprehensively understanding the operation status of the robot; then, the multi-dimensional state data set is subjected to fuzzy control feature extraction processing, and the generated fuzzy control feature set includes membership degree distribution features and fuzzy rule matching features, which can accurately describe the data characteristics with fuzzy logic and effectively handle the uncertainty and fuzziness in the data; then, a pre-trained neural network control model is called for hierarchical decision-making processing, which can combine the non-linear processing ability of the neural network to generate a real-time control instruction set including motion trajectory adjustment instructions and device state compensation instructions, realizing fine decision-making; finally, based on the real-time control instruction set, the operation state of the robot is adjusted and feedback is formed to form a dynamic optimization mechanism, enabling the robot to continuously adapt to environmental changes and task requirements, significantly improving the operation stability, task execution accuracy, and environmental adaptability of the robot in complex scenarios, and ensuring the efficient and reliable completion of various tasks.

[0009] In summary, the embodiment of the present invention can achieve more accurate, intelligent, and dynamic control of the robot by obtaining a multi-dimensional state data set, performing fuzzy control feature extraction, combining the hierarchical decision-making processing of the neural network control model, and the feedback adjustment mechanism, effectively solving the problems of inaccurate control and poor adaptability in the prior art in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1 It is a flowchart of a robot control method combining fuzzy control and neural network provided by an embodiment of the present invention.

[0012] Figure 2 A block diagram of a robot control system provided by an embodiment of the present invention.

[0013] Icon:

[0014] 100 - Robot control system;

[0015] 101 - Processor; 102 - Memory; 103 - Bus. Detailed implementation manners

[0016] The exemplary embodiments disclosed by the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present invention can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0017] In order to better understand the above technical solutions, the technical solutions of the present invention will be described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solutions of the present invention, rather than limitations on the technical solutions of the present invention. Without conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0018] Figure 1 A flowchart of a robot control method combining fuzzy control and neural network according to an embodiment of the present invention, applied to a robot control system, including steps 110 - 140.

[0019] Step 110: Obtain a multi - dimensional state data set of the robot in the target operation scenario, where the multi - dimensional state data set includes device attitude data, environmental interference data, and task constraint conditions.

[0020] In an embodiment of the present invention, the robot is applied to an industrial warehousing scenario to perform a goods handling task. In this scenario, obtaining the multi - dimensional state data set is crucial for the stable operation and task completion of the robot. The device attitude data can reflect the state of the robot itself, the environmental interference data allows the robot to perceive the surrounding environmental conditions, and the task constraint conditions provide a clear task orientation for the actions of the robot. By comprehensively obtaining these data, the robot can better adapt to environmental changes, make accurate decisions, and ensure the efficient and safe completion of the goods handling task.

[0021] In one embodiment, the obtaining of the multi - dimensional state data set of the robot in the target operation scenario includes:

[0022] Step 111: Collect the device attitude data through the inertial measurement unit of the robot. The device attitude data includes three-dimensional acceleration, angular velocity, and inclination angle.

[0023] In an industrial warehousing scenario, the robot will generate various movements when handling goods, and its attitude will also change continuously. The inertial measurement unit collects these key data in real time. For example, when the robot is driving straight along the warehouse aisle, the three-dimensional acceleration data can reflect its acceleration and deceleration conditions; during the turning process, the angular velocity data can reflect the speed of turning; and the inclination angle data can be used to determine whether the robot is in a balanced state. For instance, when going uphill or downhill, the change in the inclination angle enables the robot to adjust its attitude in a timely manner to prevent the goods from falling. These data are of great significance for accurately evaluating the state of the robot and performing corresponding control subsequently.

[0024] Step 112: Collect the environmental interference data through the environmental sensor of the robot. The environmental interference data includes ground friction coefficient, obstacle distribution density, and air resistance coefficient.

[0025] In an industrial warehousing environment, factors such as the ground condition, obstacle distribution, and air condition will all interfere with the operation of the robot. The environmental sensor can keenly capture this information. For example, the ground materials in different areas are different, so the ground friction coefficient will vary. On a smooth ground and a rough ground, the driving resistance of the robot is different, which will affect its power output and movement trajectory. The obstacle distribution density reflects the distribution of obstacles such as goods and shelves in the warehouse. The robot needs to plan a safe driving path based on this data to avoid collisions. Although the air resistance coefficient is relatively small, it cannot be ignored when the robot is driving at high speed. The air resistance coefficient will affect the energy consumption and speed control of the robot.

[0026] Step 113: Obtain the task constraint conditions through the task scheduling unit of the robot. The task constraint conditions include target path priority, task execution time limit, and safety obstacle avoidance threshold.

[0027] In the task of handling goods in industrial warehousing, the task scheduling unit plays a key commanding role. The target path priority clarifies the execution order of different path segments in a conflict scenario. For example, when multiple robots are performing tasks in the warehouse simultaneously, when there is a path intersection, the robot on the path segment with a higher priority passes first to ensure the efficient execution of the overall task. The task execution time limit restricts the maximum time for the robot to complete the task, thereby enabling reasonable arrangement of warehousing resources and improving work efficiency. The safety obstacle avoidance threshold determines the minimum allowable distance between the robot and the obstacle, ensuring the safety of the robot during operation and preventing damage to equipment and loss of goods caused by collisions.

[0028] In another embodiment, obtaining the task constraint conditions through the task scheduling unit of the robot includes:

[0029] Step 1131: Receive an initial task instruction sent by an external control terminal, and parse the initial task instruction to obtain task path planning data and task execution condition data.

[0030] In an industrial warehousing scenario, the external control terminal is responsible for issuing the initial task instruction. When there is a new goods handling task, the control terminal will send a detailed instruction to the task scheduling unit of the robot. After receiving the instruction, the task scheduling unit parses it. For example, the instruction may include a task of moving goods from area A to area B of the warehouse. The task path planning data will clarify the specific route for the robot to start from area A and reach area B, which may involve information such as which passages and turning points to pass through; the task execution condition data will describe conditions such as the time requirement for completing this task and the weight limit of the goods, providing a basis for determining the task constraint conditions subsequently.

[0031] Step 1132: Determine the target path priority based on the task path planning data, where the target path priority is used to indicate the execution order of different path segments in a conflict scenario.

[0032] In a complex industrial warehousing environment, simultaneous operation of multiple robots may cause path conflicts. According to the task path planning data, the task scheduling unit will comprehensively consider various factors to determine the target path priority. For example, if a path leads to an emergency shipping area, the path segments on this path will have a higher priority. When multiple robots cross paths during operation, the robot on the path segment with a higher priority can pass first, and other robots need to wait or re-plan the path, so as to ensure the efficient progress of the overall task.

[0033] Step 1133: Determine the task execution time limit and the safety obstacle avoidance threshold based on the task execution condition data. The task execution time limit is used to restrict the maximum running duration of the robot, and the safety obstacle avoidance threshold is used to determine the minimum allowable distance between the robot and the obstacle.

[0034] It can be understood that the task execution condition data contains a lot of important information, and the task scheduling unit will extract key content from it. Regarding the task execution time limit, for example, it is stipulated to complete the task of transporting goods from a specific area to a designated location within one hour. This is to ensure the efficient operation of the entire warehousing process and avoid task delays affecting subsequent work. The determination of the safety obstacle avoidance threshold is based on factors such as the type and distribution of obstacles in the warehouse. For example, for large shelves, the safety obstacle avoidance threshold may be set relatively large to prevent the robot from colliding with the shelves during operation, resulting in the falling of goods or damage to equipment; for small obstacles, the safety obstacle avoidance threshold can be relatively small, but sufficient safety distance should also be ensured.

[0035] Step 1134: Associate the target path priority, the task execution time limit, and the safety obstacle avoidance threshold as the task constraint conditions.

[0036] Optionally, the task scheduling unit integrates the target path priority, the task execution time limit, and the safety obstacle avoidance threshold to form complete task constraint conditions. These constraint conditions are interrelated and interact with each other, jointly providing clear guidance for the operation of the robot. In the industrial warehousing scenario, when the robot executes tasks, it needs to meet these constraint conditions simultaneously. It not only needs to complete the tasks within the specified time but also reasonably plan its driving route according to the path priority on the premise of safety.

[0037] Step 120: Perform fuzzy control feature extraction processing on the multi-dimensional state data set to obtain the fuzzy control feature set of the robot in the current operation cycle. The fuzzy control feature set includes membership degree distribution features and fuzzy rule matching features.

[0038] In the industrial warehousing scenario, the multi-dimensional state data set contains a large amount of complex and uncertain information. Performing fuzzy control feature extraction processing on these data can transform them into a more valuable, easier-to-understand, and processable fuzzy control feature set. The membership degree distribution features can help the robot better understand the degree of its own state and environmental factors, and the fuzzy rule matching features can enable the robot to make reasonable decisions based on historical experience and current task requirements.

[0039] Optionally, the performing fuzzy control feature extraction processing on the multi-dimensional state data set to obtain the fuzzy control feature set of the robot in the current operation cycle includes:

[0040] Step 121: Perform membership function mapping processing on the device attitude data to generate the first membership degree distribution feature. The first membership degree distribution feature includes the deviation membership degree of the device attitude data relative to the preset standard attitude interval.

[0041] In the industrial warehousing scenario, the posture of the robot changes continuously during the process of carrying goods. The preset standard posture range is set according to the design of the robot and the requirements of normal operation. For example, when the robot is moving in a straight line, its standard posture is to remain horizontal and the driving direction is stable. The membership function mapping process for the device posture data is to compare the actually collected device posture data, such as three-dimensional acceleration, angular velocity, and inclination angle, with the preset standard posture range. Taking the inclination angle as an example, if the inclination angle of the robot is within the standard range, its deviation membership degree relative to the standard posture range may be low; if the inclination angle exceeds a certain range, the deviation membership degree will become medium or high. The first membership degree distribution feature generated in this way enables the robot to clearly understand the degree of deviation of its own posture from the standard posture.

[0042] Step 122: Perform a fuzzy interval division process on the environmental interference data to generate a second membership degree distribution feature, where the second membership degree distribution feature includes the distribution weights of the environmental interference data in multiple interference intensity intervals.

[0043] In the industrial warehousing environment, the environmental interference factors are complex and diverse. The fuzzy interval division process for the environmental interference data is to divide it into multiple interference intensity intervals according to the characteristics and influence degrees of different interference factors. For example, for the ground friction coefficient, it can be divided into low friction, medium friction, and high friction intervals; for the obstacle distribution density, it can be divided into low density, medium density, and high density intervals. Then, according to the actually collected environmental interference data, determine its distribution weights in each interference intensity interval. For example, when the possibility that the ground friction coefficient in the area where the robot is located is in the medium friction interval is relatively large, the distribution weight of this interval will be relatively high. The second membership degree distribution feature generated in this way enables the robot to have a more intuitive understanding of the environmental interference situation.

[0044] Step 123: Perform a fuzzy rule matching process on the task constraint conditions to generate the fuzzy rule matching feature, where the fuzzy rule matching feature includes the matching degrees between the current task constraint conditions and each rule template in the historical task rule library.

[0045] In the industrial warehousing scenario, the historical task rule library has accumulated a large amount of experience and rules from past tasks. The fuzzy rule matching process for the task constraint conditions is to compare the current task constraint conditions, such as the target path priority, task execution time limit, and safety obstacle avoidance threshold, with each rule template in the historical task rule library. For example, the current task requires the completion of goods handling within a short time and the path passes through an area with a large number of obstacles. Through the fuzzy rule matching process, find a similar task rule template in the historical tasks and calculate the matching degree between the task rule templates. A high matching degree indicates that the current task is similar to the historical task situation, and historical experience can be used for decision-making.

[0046] Step 124: Perform feature fusion on the first membership degree distribution feature, the second membership degree distribution feature, and the fuzzy rule matching feature to obtain the fuzzy control feature set.

[0047] In an industrial warehousing scenario, the first membership degree distribution feature reflects the robot's own posture, the second membership degree distribution feature reflects the environmental interference situation, and the fuzzy rule matching feature provides a decision-making basis based on historical experience. Feature fusion of these three features aims to integrate information from different aspects. For example, a weighted fusion method can be adopted, and different weights are assigned to each feature according to the actual situation. If the current environmental interference has a greater impact on the robot's operation, the weight of the second membership degree distribution feature can be set higher. Through this fusion method, the obtained fuzzy control feature set contains more comprehensive and integrated information, providing a better basis for subsequent decision-making.

[0048] Step 130: Invoke the pre-trained neural network control model to perform hierarchical decision-making processing on the fuzzy control feature set, and generate the real-time control instruction set for the robot. The real-time control instruction set includes a motion trajectory adjustment instruction and a device state compensation instruction.

[0049] In an industrial warehousing scenario, the pre-trained neural network control model can be used for intelligent decision-making. The fuzzy control feature set contains a large amount of complex information. By invoking this model for hierarchical decision-making processing, this information can be transformed into a specific and executable real-time control instruction set. The motion trajectory adjustment instruction allows the robot to adjust its driving route according to the environment and its own state, and the device state compensation instruction can ensure that the robot's device posture and power output are in the best state to better complete the task of handling goods.

[0050] Preferably, the invoking the pre-trained neural network control model to perform hierarchical decision-making processing on the fuzzy control feature set and generate the real-time control instruction set for the robot includes:

[0051] Step 131: Input the fuzzy control feature set into the feature encoding layer of the neural network control model to perform feature space transformation processing, and generate a mapped feature vector.

[0052] In an industrial warehousing scenario, the feature encoding layer of the neural network control model can serve as an information transformer. The fuzzy control feature set contains multi-dimensional complex information, and the feature encoding layer processes this information. For example, the feature encoding layer integrates and transforms different types of features such as the first membership degree distribution feature, the second membership degree distribution feature, and the fuzzy rule matching feature. These features are mapped from the original feature space to another feature space more suitable for neural network processing, generating a mapped feature vector that contains re-encoded and integrated information, providing a better representation form for subsequent decision-making processing.

[0053] Step 132: Invoke the first decision layer of the neural network control model to perform motion trajectory prediction processing on the mapped feature vector, generating an initial trajectory adjustment instruction. The first decision layer predicts the trajectory deviation based on historical trajectory data and current environmental data.

[0054] In an industrial warehousing scenario, the first decision layer performs motion trajectory prediction processing based on the mapped feature vector, historical trajectory data, and current environmental data. For example, historical trajectory data can provide the driving route experience of the robot in the past under similar environments and tasks, and current environmental data includes information such as obstacle distribution and ground conditions. The first decision layer analyzes these data to predict the possible trajectory deviation of the robot at present. If it is found that there are new obstacles ahead, or the change in the ground friction coefficient may cause trajectory deviation, an initial trajectory adjustment instruction will be generated to guide the robot to adjust the driving route in advance.

[0055] Step 133: Invoke the second decision layer of the neural network control model to perform equipment state compensation prediction processing on the mapped feature vector, generating an initial state compensation instruction. The second decision layer predicts the state stability based on equipment historical state data and current interference data.

[0056] In an industrial warehousing scenario, the second decision layer performs equipment state compensation prediction processing based on the mapped feature vector, equipment historical state data, and current interference data. Equipment historical state data records the past equipment postures, power outputs, etc. of the robot, and current interference data such as ground friction and air resistance will affect the equipment state. The second decision layer analyzes these data to predict the stability of the equipment state. If it is found that the current environmental interference may cause the center of gravity of the robot to shift or the power output to be unstable, an initial state compensation instruction will be generated to adjust the equipment posture and power output of the robot.

[0057] Step 134: Normalize the initial trajectory adjustment instruction and the initial state compensation instruction to generate a real-time control instruction set including a motion trajectory adjustment instruction and an equipment state compensation instruction.

[0058] In an industrial warehousing scenario, the initial trajectory adjustment instruction and the initial state compensation instruction may have different dimensions and value ranges. Normalization processing will uniformly process these instructions to make them have the same scale and range. For example, through a normalization algorithm, the initial trajectory adjustment instruction and the initial state compensation instruction are mapped into a standard interval. The resulting real-time control instruction set after such processing contains calibrated and integrated motion trajectory adjustment instructions and equipment state compensation instructions, which can more accurately and effectively guide the operation of the robot.

[0059] In an alternative embodiment, the training process of the pre-trained neural network control model includes:

[0060] Step 210: Obtain a set of historical operation data, where the set of historical operation data includes equipment operation data, environmental data, and corresponding control instruction annotation data for multiple historical operation cycles.

[0061] During the long-term operation of the industrial warehousing scenario, a large amount of historical operation data has been accumulated. These data cover multiple historical operation cycles. The equipment operation data records information such as the posture and speed of the robot at different times. The environmental data includes ground conditions, obstacle distribution, etc. The corresponding control instruction annotation data specifies the correct control instructions that should be taken under these data conditions. For example, in a past operation cycle, it records the equipment posture, speed, and corresponding control instructions of the robot when carrying goods from point P1 to point P2 in an environment with a specific ground friction coefficient and obstacle distribution. These data provide rich materials for the training of the neural network control model.

[0062] Step 220: Perform fuzzy feature extraction processing on the set of historical operation data to obtain a set of historical fuzzy control features.

[0063] Just like processing the current multi-dimensional state data set, the set of historical operation data also needs to be processed by fuzzy feature extraction. For the equipment operation data, the membership degree distribution feature about the equipment posture is generated through membership function mapping processing; for the environmental data, fuzzy interval division processing is performed to obtain the membership degree distribution feature of the environmental interference data; the fuzzy rule matching processing is performed between the control instruction annotation data and the historical task rule base to generate the fuzzy rule matching feature. Then these features are fused to obtain a set of historical fuzzy control features, which contains the fuzzy control information in the historical data and provides appropriate input for model training.

[0064] Step 230: Obtain an initial neural network model, where the initial neural network model includes an input layer, at least two hidden layers, and an output layer.

[0065] It can be understood that the initial neural network model is a basic architecture. The input layer is responsible for receiving external data, which in this scenario is the set of historical fuzzy control features. At least two hidden layers perform complex non-linear transformations and feature extractions on the input data, capable of learning the deep patterns in the data. The output layer outputs the predicted control instructions based on the processing results of the hidden layers. For example, after the input layer receives the set of historical fuzzy control features, it passes them to the hidden layer. The hidden layer analyzes and processes the data through the connections between neurons and weight adjustments, and finally the output layer outputs the set of predicted control instructions.

[0066] Step 240: Input the set of historical fuzzy control features into the initial neural network model for forward propagation processing to generate a set of predicted control instructions.

[0067] In the industrial warehousing scenario, after inputting the set of historical fuzzy control features into the initial neural network model, the data will perform forward propagation according to the structure of the model. The input layer passes the data to the first hidden layer. The neurons in the hidden layer calculate and transform the data according to the preset weights and activation functions, and then pass the results to the next hidden layer. After being processed by multiple hidden layers, finally the output layer generates a set of predicted control instructions.

[0068] Step 250: Generate a loss function based on the difference between the set of predicted control instructions and the control instruction annotation data, and update the weight parameters of the initial neural network model through the backpropagation algorithm until the loss function converges to obtain the pre-trained neural network control model.

[0069] In the industrial warehousing scenario, the difference between the set of predicted control instructions and the control instruction annotation data reflects the accuracy of the model prediction. A loss function is generated by setting an algorithm, and the value of this function represents the gap between the prediction result and the real result. For example, algorithms such as mean squared error can be used to calculate the loss function. Then, through the backpropagation algorithm, starting from the output layer, the loss value is propagated backward along the network structure, and the weight parameters are adjusted according to the loss value. This process is continuously repeated until the loss function converges. At this time, the initial neural network model becomes a pre-trained neural network control model, which can process and make decisions on input data more accurately.

[0070] As an implementation, generating the loss function based on the difference between the set of predicted control instructions and the control instruction annotation data includes:

[0071] Step 251: Calculate the mean squared error between the motion trajectory adjustment instructions in the set of predicted control instructions and the annotated motion trajectory adjustment instructions to obtain the first loss component.

[0072] In the industrial warehousing scenario, mean squared error calculation is a commonly used method to measure the difference between two motion trajectory adjustment instructions. The mean squared error algorithm calculates the sum of the squares of the differences between the predicted motion trajectory adjustment instructions and the labeled motion trajectory adjustment instructions in each dimension, and then takes the average. For example, in a two-dimensional plane, the motion trajectory adjustment instructions may involve adjustments in the x and y directions. Calculate the squares of the differences between the predicted values and the labeled values in these two directions respectively, add them up, and then divide by the number of dimensions. The resulting value is the first loss component, which reflects the prediction error in motion trajectory adjustment.

[0073] Step 252: Calculate the cosine similarity between the device state compensation instructions in the predicted control instruction set and the labeled device state compensation instructions to obtain the second loss component.

[0074] In the industrial warehousing scenario, the device state compensation instructions involve information in multiple dimensions such as the robot device posture compensation amount and the power output compensation amount. Cosine similarity calculation is used to measure the similarity degree between the predicted device state compensation instructions and the labeled device state compensation instructions in the vector space. Consider the device state compensation instructions as vectors and calculate the cosine value of the angle between these two vectors. For example, the device state compensation instruction vector may include the angle vector for device posture adjustment and the power vector for power output adjustment, etc. By calculating the dot product of the two vectors divided by the product of the corresponding vector lengths, the cosine similarity value is obtained. Subtract this cosine similarity value from 1 to get the second loss component. This component can reflect the degree of difference between the prediction result and the true label in device state compensation, providing a basis for model adjustment.

[0075] Step 253: Calculate the sum of absolute values of the overall deviation between the predicted control instruction set and the control instruction labeled data to obtain the third loss component.

[0076] In the industrial warehousing scenario, considering the overall deviation between the predicted control instruction set and the control instruction labeled data, absolute value summation calculation is used to quantify this deviation. The predicted control instruction set includes multiple parts such as motion trajectory adjustment instructions and device state compensation instructions, and the control instruction labeled data also covers the corresponding accurate information. Take the absolute value of the difference between each instruction element in the predicted control instruction set and the corresponding element in the control instruction labeled data, and then add up all these absolute values. The resulting sum is the third loss component. For example, calculate the absolute values of the differences in the adjustment values of the motion trajectory adjustment instructions in the x, y, and z directions and the compensation values of the posture and power output in the device state compensation instructions respectively and accumulate them. This component comprehensively reflects the degree of deviation between the overall prediction result and the labeled data.

[0077] Step 254: Perform weighted summation on the first loss component, the second loss component, and the third loss component to obtain the loss function.

[0078] In an industrial warehousing scenario, different loss components may have different importance for model training. Therefore, weighted summation is required to obtain the loss function. Different weights are assigned to the first loss component, the second loss component, and the third loss component respectively. For example, based on practical experience and the analysis of the importance of the task, if the accuracy of the motion trajectory adjustment is crucial for the robot to complete the task, the weight of the first loss component can be set relatively high; if the accuracy of the equipment state compensation is of secondary importance, the weight of the second loss component is appropriately reduced; the third loss component reflects the overall deviation, and its weight is also set according to the overall situation. Then, each loss component is multiplied by its corresponding weight and added together, that is, the first loss component multiplied by weight 1 plus the second loss component multiplied by weight 2 plus the third loss component multiplied by weight 3, and finally the loss function is obtained. This loss function synthesizes the prediction error information in different aspects and can more comprehensively reflect the difference between the model prediction and the actual situation, providing an accurate basis for updating the model weight parameters through the backpropagation algorithm in the subsequent process.

[0079] Step 140: Adjust the running state of the robot in the target running scenario based on the real-time control instruction set, and feedback the adjusted running state to the multi-dimensional state data set to perform loop control processing.

[0080] In an industrial warehousing scenario, the real-time control instruction set is the key basis for the robot to adjust its running state. Based on these instructions, the robot can accurately adjust its own motion and equipment state to adapt to the changing environment and task requirements. At the same time, the adjusted running state is fed back into the multi-dimensional state data set to form a closed-loop loop control process, enabling the robot to continuously optimize its running and better complete the goods handling task.

[0081] In one embodiment, the adjusting the running state of the robot in the target running scenario based on the real-time control instruction set includes:

[0082] Step 141: Parse the motion trajectory adjustment instruction in the real-time control instruction set, determine the trajectory correction amount of the robot within the target time window, and adjust the driving wheel speed and steering angle of the robot according to the trajectory correction amount.

[0083] In an industrial warehousing scenario, the motion trajectory adjustment instruction is to ensure that the robot can accurately travel to the target location according to the task requirements. After receiving the motion trajectory adjustment instruction in the real-time control instruction set, the robot will parse it. For example, the instruction may indicate that within the next 10 seconds (the target time window), the robot needs to shift a certain distance to the left to avoid a newly emerged obstacle ahead. By parsing the instruction, the trajectory correction amount is determined, which may include information such as lateral displacement and angle change. Then, according to this trajectory correction amount, the rotational speed and steering angle of the drive wheels are adjusted. If a left turn is required, the rotational speed of the left drive wheel will be appropriately reduced, and the rotational speed of the right drive wheel will remain unchanged or be appropriately increased to achieve the turn; at the same time, according to the overall speed requirement and trajectory correction demand, the overall rotational speed of the drive wheels is adjusted to ensure that the robot can accurately reach the target position within the specified time.

[0084] Step 142: Parse the device state compensation instruction in the real-time control instruction set, determine the device attitude compensation amount and power output compensation amount of the robot, adjust the center-of-gravity distribution of the robot according to the device attitude compensation amount, and adjust the motor output power of the robot according to the power output compensation amount.

[0085] In an industrial warehousing scenario, the device state compensation instruction is used to ensure that the robot's device is in the best operating state. After parsing the device state compensation instruction, the device attitude compensation amount and power output compensation amount of the robot are clarified. For example, when the robot is carrying heavy goods, due to the center-of-gravity shift, the attitude may become unstable. The device attitude compensation amount may require adjusting the mechanical structure of the robot to redistribute the center of gravity to a more stable position. The power output compensation amount is determined according to factors such as the weight of the goods and the ground friction. If the ground friction increases, in order to maintain the normal driving speed of the robot, the motor output power needs to be increased; otherwise, it is appropriately reduced. By adjusting the center-of-gravity distribution and the motor output power, it is ensured that the robot can operate stably under different working conditions.

[0086] Step 143: Collect the adjusted rotational speed, steering angle, center-of-gravity distribution, and motor output power of the drive wheels, and generate the adjusted operating state.

[0087] In an industrial warehousing scenario, after the robot completes the adjustment of the rotational speed of the drive wheels, the steering angle, the center of gravity distribution, and the motor output power, it is necessary to collect these parameters. These information are obtained in real-time through sensors installed at various key parts of the robot. For example, a rotational speed sensor is installed on the drive wheel to measure the rotational speed, the steering angle is obtained through an angle sensor, the center of gravity distribution is sensed using a gravity sensor, and the motor output power is collected through a power sensor. The collected data are integrated to form an adjusted operating state, and this operating state data reflects the actual working state of the robot after executing the real-time control instructions, providing accurate data support for subsequent feedback and loop control.

[0088] In the next step, feeding back the adjusted operating state to the multi-dimensional state data set to perform loop control processing includes:

[0089] Step 144: Write the rotational speed of the drive wheels and the steering angle in the adjusted operating state as new device pose data into the multi-dimensional state data set.

[0090] In an industrial warehousing scenario, the rotational speed of the drive wheels and the steering angle are important components of the device pose. Writing these two adjusted parameters as new device pose data into the multi-dimensional state data set can update the robot's information about its own pose in a timely manner. For example, after the robot adjusts its operating state, the rotational speed of the drive wheels is adjusted from the original 100 revolutions per minute to 120 revolutions per minute, and the steering angle changes from the original 0 degrees to 15 degrees to the left. These new data are written into the device pose data part of the multi-dimensional state data set. In this way, when the robot obtains the multi-dimensional state data set next time, it can perform subsequent processing and decision-making based on the latest device pose information.

[0091] Step 145: Write the center of gravity distribution and the motor output power in the adjusted operating state as new device state compensation data into the multi-dimensional state data set.

[0092] In an industrial warehousing scenario, the center of gravity distribution and the motor output power reflect the device state compensation of the robot. Writing the adjusted center of gravity distribution and the motor output power as new device state compensation data into the multi-dimensional state data set can enable the robot to understand the changes in its own device state in a timely manner. For example, through adjustment, the center of gravity position of the robot moves from the original coordinate point to a new position, and the motor output power is adjusted from the original value to a new value. These new data are updated into the multi-dimensional state data set. In this way, during subsequent operation, the robot can better perform state evaluation and control decision-making based on these latest device state compensation data.

[0093] Step 146: Re - execute the fuzzy control feature extraction process and the hierarchical decision - making process based on the updated multi - dimensional state data set to generate a real - time control instruction set for the next operation cycle to achieve cyclic control.

[0094] In an industrial warehousing scenario, the updated multi - dimensional state data set contains the latest operating state information of the robot. Based on this updated data, the fuzzy control feature extraction process is re - executed to process the device attitude data, environmental interference data, and task constraint conditions, generating a new set of fuzzy control features. For example, the new device attitude data may cause a change in the first membership distribution feature, and the update of the environmental interference data will also affect the second membership distribution feature. Then, the new set of fuzzy control features is input into a pre - trained neural network control model for hierarchical decision - making processing to generate a real - time control instruction set for the next operation cycle. This process loops continuously, enabling the robot to continuously adjust its operating state according to changes in the environment and its own state to better complete the goods handling task and achieve efficient and stable cyclic control.

[0095] As a non - limiting embodiment, after the feedback - adjusted operating state is fed back to the multi - dimensional state data set to perform the cyclic control process, it further includes: generating fuzzy rule optimization parameters based on the correspondence between the adjusted operating state and the real - time control instruction set, where the fuzzy rule optimization parameters include a rule weight correction coefficient and a membership function offset; performing real - time update processing on the rule templates in the historical task rule library according to the fuzzy rule optimization parameters to generate an updated set of rule templates, and the real - time update processing includes adjusting the rule trigger threshold and re - distributing the center point coordinates of the membership function; applying the updated set of rule templates to the fuzzy rule matching process for the next operation cycle to optimize the generation accuracy of the fuzzy control feature set.

[0096] In an industrial warehousing scenario, there is an inherent correspondence between the adjusted operating state and the real - time control instruction set. By analyzing this correspondence, fuzzy rule optimization parameters can be generated. For example, if it is observed that in one operating state, the execution effect of an individual fuzzy rule is good, while in other states it is not, then the rule weight correction coefficient can be adjusted according to this situation to adjust the weight of the rule to make it more influential in the appropriate scenario. For the membership function offset, when it is found that the actual situation does not match the preset membership function well, by adjusting the offset, the membership function can better reflect the actual situation.

[0097] Optimize the parameters according to the generated fuzzy rules and perform real-time update processing on the rule templates in the historical task rule library. For example, adjust the rule trigger threshold. When a certain rule is triggered too frequently or too rarely in the current running state, appropriately increase or decrease the trigger threshold to make it more in line with the actual requirements. Reassign the center point coordinates of the membership function so that the membership function can more accurately describe the relationship between data and concepts. After these processes, a set of updated rule templates is generated.

[0098] In the next running cycle, apply the set of updated rule templates to the fuzzy rule matching process. When the robot obtains new task constraint conditions, using the updated rule templates for matching can calculate the matching degree more accurately, thereby optimizing the generation accuracy of the fuzzy control feature set. For example, in a new task, according to the updated rule templates, it is possible to more precisely judge the similarity between the current task constraint conditions and historical experience, providing a more reliable basis for subsequent decisions.

[0099] As another non-limiting embodiment, after feedback-adjusting the running state to the multi-dimensional state data set to perform loop control processing, it further includes: extracting the device pose compensation amount and trajectory correction amount in the adjusted running state to generate a derived training data set; performing spatio-temporal alignment processing on the derived training data set and the corresponding fuzzy control feature set to generate a time-series associated training sample; performing online gradient update on the pre-trained neural network control model based on the time-series associated training sample to adjust the weight distribution of the feature encoding layer and the activation function parameters of the first decision layer.

[0100] In an industrial warehousing scenario, the adjusted running state contains rich information, where the device pose compensation amount and trajectory correction amount reflect the actual adjustment situation of the robot during operation. Extract these data to generate a derived training data set. For example, the device pose compensation amount may include the adjustment values of the robot's center of gravity, angle, etc. at different times, and the trajectory correction amount records the adjustment information of the robot's driving route at each stage.

[0101] Performing spatio-temporal alignment processing on the derived training data set and the corresponding fuzzy control feature set means matching the fuzzy control feature set obtained at the same time or in a related time series with the derived training data. For example, at a specific moment, the robot makes corresponding decisions based on the fuzzy control feature set, resulting in a certain device pose compensation amount and trajectory correction amount. Align these data in chronological order to generate a time-series associated training sample.

[0102] Based on these time-series correlated training samples, online gradient update is performed on the pre-trained neural network control model. Online gradient update is a method of continuously adjusting parameters during the operation of the model. By calculating the gradients of the time-series correlated training samples, the weight distribution of the feature encoding layer is adjusted according to the gradient direction. For example, if it is found that a certain weight parameter causes a large error when processing the current sample, the weight is adjusted according to the gradient information to make it more conducive to accurately processing data. At the same time, the activation function parameters of the first decision layer are adjusted. The activation function determines the output of the neuron. By adjusting the parameters, the neuron can better respond to the input under different circumstances, improving the decision-making accuracy and adaptability of the model.

[0103] As another non-limiting embodiment, after the feedback-adjusted operating state is fed back to the multi-dimensional state data set to perform loop control processing, it further includes: performing sliding window statistical analysis on the fuzzy control feature set of continuous N operating cycles, and extracting the mean vector and variance matrix of each feature; performing dynamic normalization processing on the fuzzy control feature set according to the mean vector and variance matrix to generate a standardized fuzzy feature vector; inputting the standardized fuzzy feature vector into the neural network control model for feature attention allocation between decision layers, and adjusting the priority weights of the motion trajectory adjustment instruction and the device state compensation instruction.

[0104] In an industrial warehousing scenario, in order to better analyze the change trend and stability of the fuzzy control feature set, sliding window statistical analysis is performed on the fuzzy control feature set of continuous N operating cycles. For example, the sliding window size is set to 5 operating cycles, and it moves one cycle each time, and statistics are performed on the fuzzy control feature set within the window. The mean vector of each feature is extracted. The mean vector reflects the average level of each feature within these N cycles. For example, for a certain dimension in the first membership degree distribution feature, its average value within N cycles is calculated to form an element of the mean vector. At the same time, the variance matrix is calculated. The variance matrix can measure the fluctuation of each feature within these N cycles and reflect the degree of dispersion of the data.

[0105] According to the calculated mean vector and variance matrix, dynamic normalization processing is performed on the fuzzy control feature set. Dynamic normalization is to make different features have comparable scales in different operating cycles. For example, for a certain feature, the value of the feature is subtracted by the corresponding value in the mean vector, and then divided by the standard deviation of the corresponding dimension in the variance matrix to obtain the normalized value, thereby generating a standardized fuzzy feature vector.

[0106] Input the standardized fuzzy feature vector into the neural network control model for feature attention allocation among decision layers. In the model, different decision layers may have different attentions to different features. By analyzing the standardized fuzzy feature vector, determine which features are more important for the motion trajectory adjustment instruction and the device state compensation instruction. For example, if it is found that a certain feature has a high correlation with the motion trajectory, appropriately increase the attention weight of this feature in the first decision layer (responsible for motion trajectory prediction), so that the model pays more attention to this feature when generating the motion trajectory adjustment instruction; for the feature related to device state compensation, give corresponding weight adjustment in the second decision layer, so as to adjust the priority weights of the motion trajectory adjustment instruction and the device state compensation instruction, enabling the robot to more reasonably allocate decision-making resources according to the actual situation and improve the control effect.

[0107] In the embodiment of the present invention, the following key processes can also be carried out based on the existing fuzzy control theory, neural network modeling technology, and industrial robot closed-loop control method: For the fuzzy feature extraction link, conventional membership functions (such as triangular and trapezoidal functions) can be used in combination with domain knowledge to preset the standard attitude interval and interference intensity interval, and the deviation membership degree and distribution weight can be fully generated by calculating the membership relationship between the actual data and the preset interval; for fuzzy rule matching, the similarity calculation (such as Euclidean distance, fuzzy closeness) can be used according to the historical task database to realize the quantitative evaluation of the rule template matching degree.

[0108] Optionally, in the neural network decision layer, the feature space conversion can be realized through standard feature encoding techniques (such as fully connected layers, convolutional operations), and the initial adjustment instruction can be generated by referring to the time series prediction model (such as LSTM, GRU) to fuse the historical trajectory / state data; when constructing the loss function, conventional evaluation methods such as mean square error (motion trajectory), cosine similarity (multi-dimensional device state vector), and absolute deviation are comprehensively applied, and their weighting coefficients can be determined by empirical values or grid search. In the closed-loop control, the dynamic normalization process can adopt the sliding window Z-score normalization, and the attention mechanism (such as Additive Attention) can be introduced for feature attention allocation to automatically learn the decision layer weights.

[0109] In addition, for the online update of the rule base, the rule weights and the center points of the membership functions can be dynamically corrected based on the reinforcement learning idea using the state-instruction correlation; for the online fine-tuning of the model, the incremental data after time series alignment is combined with the stochastic gradient descent to optimize the weights of the encoding layer and the parameters of the activation function. In this way, the further optimization processes of feature fuzzification, decision layering, dynamic adaptation, and dimension unification can be effectively realized.

[0110] Applying the embodiments of the present invention, first, a multi-dimensional state data set of the robot in the target operation scenario is obtained (covering various aspects of information such as device attitude, environmental interference, and task constraint conditions), laying a foundation for comprehensively understanding the operation status of the robot; then, the multi-dimensional state data set is processed to extract fuzzy control features, and the generated fuzzy control feature set includes membership degree distribution features and fuzzy rule matching features, which can accurately describe the data characteristics with fuzzy logic and effectively handle the uncertainty and ambiguity in the data; then, a pre-trained neural network control model is called for hierarchical decision-making processing, which can combine the non-linear processing ability of the neural network to generate a real-time control instruction set including motion trajectory adjustment instructions and device state compensation instructions to achieve fine decision-making; finally, based on the real-time control instruction set, the operation state of the robot is adjusted and feedback is formed to establish a dynamic optimization mechanism, enabling the robot to continuously adapt to environmental changes and task requirements, significantly improving the operation stability, task execution accuracy, and environmental adaptability of the robot in complex scenarios, and ensuring the efficient and reliable completion of various tasks.

[0111] In summary, the embodiments of the present invention can achieve more accurate, intelligent, and dynamic control of the robot by obtaining a multi-dimensional state data set, extracting fuzzy control features, combining hierarchical decision-making processing of the neural network control model, and the feedback adjustment mechanism, effectively solving the problems of inaccurate control and poor adaptability in the prior art in complex scenarios.

[0112] The embodiments of the present invention provide a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the robot control method combining fuzzy control and neural network is implemented.

[0113] The embodiments of the present invention provide a processor, and the processor is used to run a program, wherein when the program runs, the robot control method combining fuzzy control and neural network is executed.

[0114] In the embodiments of the present invention, as Figure 2 shown, the robot control system 100 includes at least one processor 101, and at least one memory 102 and a bus 103 connected to the processor 101; wherein, the processor 101 and the memory 102 complete communication with each other through the bus 103; the processor 101 is used to call program instructions in the memory 102 to execute the above-mentioned robot control method combining fuzzy control and neural network.

[0115] The present invention is described with reference to the flowcharts and / or block diagrams of methods, robotic control systems (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0116] In a typical configuration, a robotic control system includes one or more processors (CPUs), a memory, and a bus. The robotic control system may also include an input / output interface, a network interface, and the like.

[0117] The memory may include non-permanent memory in the form of computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory includes at least one storage chip. The memory is an example of computer-readable media.

[0118] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic disk storage or other magnetic storage, computer-readable storage media, or any other non-transmission media that can be used to store information accessible by a robotic control system. As defined herein, computer-readable media does not include transitory computer-readable media, such as modulated data signals and carrier waves.

[0119] It should also be noted that the term "comprise", "include" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, commodity or computer-readable storage medium comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, commodity or computer-readable storage medium. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or computer-readable storage medium comprising the element.

[0120] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, system or computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A robot control method combining fuzzy control and neural network, characterized in that, Including: Obtain a multi-dimensional state data set of the robot in the target operation scenario, where the multi-dimensional state data set includes device attitude data, environmental interference data, and task constraint conditions; Perform fuzzy control feature extraction processing on the multi-dimensional state data set to obtain a fuzzy control feature set of the robot in the current operation cycle, where the fuzzy control feature set includes membership degree distribution features and fuzzy rule matching features; Call a pre-trained neural network control model to perform hierarchical decision-making processing on the fuzzy control feature set, and generate a real-time control instruction set of the robot, where the real-time control instruction set includes motion trajectory adjustment instructions and device state compensation instructions; Based on the real-time control instruction set, adjust the operation state of the robot in the target operation scenario, and feedback the adjusted operation state to the multi-dimensional state data set to perform loop control processing.

2. The robot control method combining fuzzy control and neural network according to claim 1, characterized in that, The performing fuzzy control feature extraction processing on the multi-dimensional state data set to obtain a fuzzy control feature set of the robot in the current operation cycle includes: Perform membership function mapping processing on the device attitude data to generate a first membership degree distribution feature, where the first membership degree distribution feature includes the deviation membership degree of the device attitude data relative to a preset standard attitude interval; Perform fuzzy interval division processing on the environmental interference data to generate a second membership degree distribution feature, where the second membership degree distribution feature includes the distribution weights of the environmental interference data in multiple interference intensity intervals; Perform fuzzy rule matching processing on the task constraint conditions to generate the fuzzy rule matching feature, where the fuzzy rule matching feature includes the matching degrees of the current task constraint conditions with each rule template in the historical task rule base; Fuse the first membership degree distribution feature, the second membership degree distribution feature, and the fuzzy rule matching feature to obtain the fuzzy control feature set.

3. The robot control method combining fuzzy control and neural network according to claim 2, wherein The calling a pre-trained neural network control model to perform hierarchical decision-making processing on the fuzzy control feature set and generate a real-time control instruction set of the robot includes: Input the fuzzy control feature set into the feature encoding layer of the neural network control model to perform feature space transformation processing, and generate a mapped feature vector; Call the first decision-making layer of the neural network control model to perform motion trajectory prediction processing on the mapped feature vector, and generate an initial trajectory adjustment instruction, where the first decision-making layer predicts the trajectory deviation based on historical trajectory data and current environmental data; Call the second decision-making layer of the neural network control model to perform device state compensation prediction processing on the mapped feature vector, and generate an initial state compensation instruction, where the second decision-making layer predicts the state stability based on device historical state data and current interference data; Perform normalization processing on the initial trajectory adjustment instruction and the initial state compensation instruction to generate a real-time control instruction set including motion trajectory adjustment instructions and device state compensation instructions.

4. The robot control method combining fuzzy control and neural network according to claim 3, characterized in that, The training process of the pre-trained neural network control model includes: Obtain a historical operation data set, where the historical operation data set includes device operation data, environmental data, and corresponding control instruction annotation data for multiple historical operation cycles; Perform fuzzy feature extraction processing on the historical operation data set to obtain a historical fuzzy control feature set; Obtain an initial neural network model, where the initial neural network model includes an input layer, at least two hidden layers, and an output layer; Input the historical fuzzy control feature set into the initial neural network model for forward propagation processing to generate a predicted control instruction set; Generate a loss function based on the difference between the predicted control instruction set and the control instruction annotation data, and update the weight parameters of the initial neural network model through the backpropagation algorithm until the loss function converges to obtain the pre-trained neural network control model.

5. The robot control method combining fuzzy control and neural network according to claim 4, wherein, The generating a loss function based on the difference between the predicted control instruction set and the control instruction annotation data includes: Calculate the mean square error between the motion trajectory adjustment instruction in the predicted control instruction set and the annotated motion trajectory adjustment instruction to obtain a first loss component; Calculate the cosine similarity between the device state compensation instruction in the predicted control instruction set and the annotated device state compensation instruction to obtain a second loss component; Calculate the absolute value sum of the overall deviation between the predicted control instruction set and the control instruction annotation data to obtain a third loss component; Perform weighted summation on the first loss component, the second loss component, and the third loss component to obtain the loss function.

6. The robot control method combining fuzzy control and neural network according to claim 1, characterized in that The adjusting the operating state of the robot in the target operating scenario based on the real-time control instruction set includes: Parse the motion trajectory adjustment instruction in the real-time control instruction set, determine the trajectory correction amount of the robot within the target time window, and adjust the driving wheel speed and steering angle of the robot according to the trajectory correction amount; Parse the device state compensation instruction in the real-time control instruction set, determine the device attitude compensation amount and power output compensation amount of the robot, adjust the center of gravity distribution of the robot according to the device attitude compensation amount, and adjust the motor output power of the robot according to the power output compensation amount; Collect the adjusted driving wheel speed, steering angle, center of gravity distribution, and motor output power to generate the adjusted operating state.

7. The robot control method combining fuzzy control and neural network according to claim 6, characterized in that, The feedback the adjusted operating state to the multi-dimensional state data set to perform loop control processing includes: Write the driving wheel speed and steering angle in the adjusted operating state as new device attitude data into the multi-dimensional state data set; Write the center of gravity distribution and motor output power in the adjusted operating state as new device state compensation data into the multi-dimensional state data set; Based on the updated multi-dimensional state data set, re-perform the fuzzy control feature extraction processing and the hierarchical decision-making processing to generate a real-time control instruction set for the next operating cycle to achieve loop control.

8. The robot control method combining fuzzy control and neural network according to claim 1, characterized in that The obtaining the multi-dimensional state data set of the robot in the target operating scenario includes: Collect the device attitude data through the inertial measurement unit of the robot, where the device attitude data includes three-dimensional acceleration, angular velocity, and inclination angle; Collect the environmental interference data through the environmental sensors of the robot, where the environmental interference data includes ground friction coefficient, obstacle distribution density, and air resistance coefficient; Obtain the task constraint conditions through the task scheduling unit of the robot, where the task constraint conditions include target path priority, task execution time limit, and safe obstacle avoidance threshold; The obtaining of the task constraint conditions through the task scheduling unit of the robot includes: Receive the initial task instruction sent by the external control terminal, and parse the initial task instruction to obtain task path planning data and task execution condition data; Determine the target path priority based on the task path planning data, where the target path priority is used to indicate the execution order of different path segments in a conflict scenario; Determine the task execution time limit and the safe obstacle avoidance threshold based on the task execution condition data, where the task execution time limit is used to constrain the maximum running duration of the robot, and the safe obstacle avoidance threshold is used to determine the minimum allowable distance between the robot and the obstacle; Associate the target path priority, the task execution time limit, and the safe obstacle avoidance threshold as the task constraint conditions.

9. A robot control system, characterized in that, It includes a processor, a memory, and a bus connected to the processor; wherein, the processor and the memory complete communication with each other through the bus; the processor is used to call the program instructions in the memory to execute the robot control method combining fuzzy control and neural network according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, A program is stored thereon, and when the program is executed by the processor, it implements the robot control method combining fuzzy control and neural network according to any one of claims 1-8.

Images