Task scheduling method for inspection robot of two-for-one twister based on large language model

Abstract

The invention relates to a two-for-one twister inspection robot task scheduling method based on a large language model, which integrates natural language understanding ability of the large language model, a traditional scheduling algorithm and statistical data, and realizes high-speed acquisition, preprocessing and noise reduction of two-for-one twister inspection data. A distributed stream processing platform is utilized, an edge calculation module is introduced at an inspection robot end, and transmission delay is effectively shortened through data compression and network protocol optimization; intelligent analysis, dynamic decomposition and real-time optimization are performed on complex tasks and data through a large language model, and efficient utilization of real-time data is ensured in combination with an adaptive scheduling algorithm, an online learning model and a data time window aggregation technology. Natural language interaction is achieved based on the fact that a robot management end is connected into a large language model, man-machine cooperation is achieved through the system, the complexity of multi-robot task scheduling is remarkably reduced according to dynamic planning of broken yarn statistical data, and the response capacity to the dynamic environment and user requirements is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of textile doubling frame detection, and particularly relates to a task scheduling method for a doubling frame inspection robot based on a large language model. Background Art

[0002] A textile doubling frame is a key device in the textile industry, mainly used for twisting single-strand yarns into double-strand or multi-strand yarns. This process is crucial for improving the strength, elasticity, and uniformity of the yarns, and also affects the quality and performance of the final textiles. In modern textile production, the stable operation of the doubling frame and product quality control are directly related to the efficiency of the entire production line and product quality. The inspection mode based on inspection robots can greatly improve productivity, but there is also a problem of a large number of machines in the actual doubling frame factory, so multiple robots are required to complete the scheduling of the entire factory.

[0003] Common inspection robots in the industry generally adopt the following several scheduling methods:

[0004] 1) Centralized scheduling: Use a central controller, such as a server, to receive task information and assign tasks to robots. The central controller has the status and task information of all robots.

[0005] 2) Distributed scheduling: Each robot has a certain ability of autonomous decision-making, collaborates with each other to complete tasks, and realizes task scheduling through local information exchange.

[0006] 3) Method based on something similar to market auction: Robots regard tasks as commodities and obtain tasks by means of task bidding among robots.

[0007] 4) Behavior-based method: Task allocation and execution are carried out through predefined behaviors or rules, and robots select appropriate behaviors according to perception and environmental information.

[0008] However, the above scheduling methods all have problems such as poor scalability, high communication overhead, and low task allocation efficiency. As the number of robots and the complexity of tasks increase, it is often difficult to maintain system performance; frequent information exchange may cause network congestion; insufficient global optimization leads to low task allocation efficiency, especially in a dynamic environment, where the system has poor adaptability to changes in task priorities. Complex scheduling algorithms also bring problems of high computational overhead and insufficient real-time performance. In terms of resource utilization, uneven task allocation among robots often results in some robots being overloaded while others are idle; the initial configuration of the system is complex and highly dependent on the environment, making the scheduling method unstable in cross-scenario applications. In addition, centralized scheduling also has the following limitations: 1. Single-point failure: If the central system crashes, the entire system may be scrapped; 2. Communication bottleneck: As the number of robots increases, the communication load will increase; 3. Poor scalability: After increasing the number of robots or adding task points, the system performance may decline. Distributed scheduling also has the following disadvantages: 1. Insufficient global optimization: Due to the lack of global information, suboptimal solutions may be obtained; 2. Complexity issues: Frequent communication between robots may cause delays or conflicts; 3. Consistency issues: Task failures are likely to occur due to decision conflicts among different robots. The method based on a market-auction-like approach also has the following deficiencies: 1. Excessive calculation: Complex bidding strategies may increase the computational burden; 2. Fairness issues: It may lead to some robots being overloaded while others are idle; 3. Communication cost: Frequent information exchange is required during the bidding process. The behavior-based method also has the following problems: 1. Low flexibility: Predefined rules limit the adaptability of the system; 2. Poor global performance: Duplicate work or resource waste is likely to occur; 3. Difficult debugging: Interactions between behaviors may lead to unpredictable results. Summary of the Invention

[0009] The technical problem to be solved by the present invention is: In order to overcome the above technical problems, the present invention provides a task scheduling method for a doubling frame inspection robot based on a large language model, aiming to solve the problems of weak autonomy, high communication cost, insufficient global optimization, and poor scalability existing in traditional multi-robot task scheduling. By using a large language model, the robot can autonomously plan the inspection route, independently judge and allocate tasks to multiple machines, and achieve orderly operation.

[0010] The technical solution adopted by the present invention to solve its technical problems is: A task scheduling method for a doubling frame inspection robot based on a large language model, comprising the following steps:

[0011] Step S1: On-site data collection and edge data preprocessing:

[0012] When an abnormality occurs during the operation of the doubling twister, sensors on the doubling twister site and / or on the inspection robot capture relevant abnormal data. After the inspection robot obtains the relevant abnormal data, the edge computing module on the inspection robot performs preliminary preprocessing on the data;

[0013] Step S2: Real-time data transmission and distributed stream processing:

[0014] The preprocessed data is uploaded to the central server through the optimized network protocol; with the help of distributed data stream transmission, it is ensured that each piece of data can reach in time. After receiving the data stream, the central server uses time window aggregation and real-time computing technology to further integrate and analyze the data; the optimized network protocol refers to setting the communication address as a static IP to make the communication between devices stable, and the data transmission uses the HTTP protocol; it also ensures the security in data transmission.

[0015] Step S3: Intelligent data parsing and task dynamic decomposition:

[0016] The data processed in step S2 is sent to the large language model integrated in the central server. The large language model uses natural language understanding ability to deeply analyze the monitoring data, identify abnormal patterns and key indicators, and dynamically decompose specific inspection tasks according to the abnormal pattern and key indicator information, clarifying the priority and processing requirements of each task;

[0017] Step S4: Adaptive scheduling and task optimization:

[0018] Based on the parsing results of the large language model, the scheduling system uses an adaptive scheduling algorithm combined with an online learning model to formulate an optimal task scheduling plan in real time; among them, the adaptive scheduling algorithm is to identify the map of the entire factory through the large language model, assign values to each path segment, and add the specific coordinates of each inspection robot to the map of the factory, and allocate tasks according to natural language;

[0019] Step S5: Task distribution and natural language interaction:

[0020] The task scheduling plan is conveyed to the operator and the inspection robot through the natural language interface. The operator directly interacts with the scheduling system through voice or text to confirm or fine-tune the scheduling plan; after receiving the detailed instructions, the inspection robot immediately enters the task execution state to ensure that on-site abnormal problems are quickly responded to;

[0021] Step S6: Inspection task execution and on-site feedback:

[0022] The inspection robot goes to the fault or monitoring area according to the received task instructions to carry out on-site inspection and maintenance work; during the task execution process, the inspection robot continuously collects new on-site data and feeds it back to the central server in real time;

[0023] Step S7: Real-time Monitoring and Dynamic Task Adjustment:

[0024] The central server monitors in real time all the uploaded data, the status of the inspection robots, and the progress of task execution; when the scheduling system detects a new anomaly or a change in the on-site environment, it immediately triggers an alarm and, with the help of a large language model and a scheduling algorithm, re-plans the tasks and sends the new scheduling plan to the relevant inspection robots.

[0025] Before step S1, it also includes step S0: Scheduling System Initialization and On-site Equipment Startup:

[0026] Start the scheduling system, complete the initialization configuration of the inspection robots and tasks; collect the current status of the inspection robots; create a task pool; the user inputs natural language through the management platform, the large language model integrated in the central server reads the information, extracts keywords from the natural language input by the user for analysis and processing, and the task pool receives the task requests.

[0027] In step S0, during the calibration process of the sensor, the calibration error Ecal is calculated:

[0028] Ecal = ∣Smeasured - Sreference∣

[0029] where Smeasured represents the measured value, Sreference represents the reference value, and when Ecal is less than the preset allowable error, the sensor is considered calibrated successfully.

[0030] In step S1, the edge computing module performs preliminary preprocessing on the data, including the following steps: First, perform noise filtering; then compress the data; at the same time, perform preliminary anomaly detection to ensure that only key data is transmitted to the central server. In the operating environment of multiple robots, it is easy for multiple robots to inspect the same location simultaneously. If the broken yarn in this area is not processed in time, it will lead to the transmission of duplicate data. The preliminary anomaly detection here is equivalent to removing duplicate data from a large amount of repeated data.

[0031] For noise filtering, an exponential smoothing filter is used to reduce the noise of the original data, and the smoothing filter formula can be written as:

[0032] y(t) = αx(t) + (1 - α)y(t - 1)

[0033] where x(t) is the current sampled data, y(t) is the filtered data, and α is the smoothing factor.

[0034] In step S2, Apache Kafka is used to achieve distributed data stream transmission. The Apache Flink platform of the central server receives the data stream and aggregates, groups, and calculates time windows for the data in real time. The average calculation of the time window data is as follows:

[0035]

[0036] where N is the time window size, and x(t - i) is the data value at each moment.

[0037] In step S3, the following steps are included:

[0038] Step S31: Sub - task decomposition is carried out according to four decomposition rules: region, device, special requirements, and dynamic tasks. The user's natural language tasks are decomposed from four aspects: region, device, special requirements, and dynamic tasks, and the priorities are determined according to four aspects: urgency, importance, timeliness, and path optimization, and the final priority results are sorted; if there are specific tasks that need to be processed first, they are manually designated as the highest priority;

[0039] The priority calculation formula: P = w1·U + w2·I + w3·T - w4·D

[0040] U: Urgency weight; I: Importance weight; T: Timeliness weight; D: Path distance, a negative value indicates a lower priority for a longer distance; w1, w2, w3, and w4 are all weight values;

[0041] Step S32: Allocate task resources according to performance: According to the performance indicators of the inspection robot itself, the consumption of task requirements needs to be evaluated for performance, and the workload of the task is matched with the performance consumption of the inspection robot, and the allocation is carried out in combination with the user's natural language requirements;

[0042] The task priority sorting algorithm in step S31 includes greedy algorithm, dynamic programming, comprehensive sorting algorithm, and priority queue.

[0043] In step S7, the adaptive scheduling algorithm dynamically adjusts the task allocation based on the performance evaluation of the inspection robot according to the changes in the task pool or the changes in the state of the inspection robot.

[0044] After step S7, there is also step S8: Closed - loop feedback and continuous optimization:

[0045] After each task is completed, the collected data, task records, and operation feedback are archived by the scheduling system. Through continuous analysis of historical data and online learning, the scheduling system continuously optimizes the parsing ability of the large - language model and the response strategy of the scheduling algorithm, forming a continuously self - improving closed - loop management process.

[0046] In step S8, the scheduling system continuously adjusts the parameters of the large language model and the scheduling algorithm using an online learning method. The parameter update is optimized using the gradient descent method, and its update formula is: where θ new represents the updated model parameters, θ old represents the model parameters before update, η is the learning rate, L(θ) is the loss function, and θ is the model parameter.

[0047] The beneficial effects of the present invention are as follows. A task scheduling method for a doubling frame inspection robot based on a large language model of the present invention constructs an efficient data stream processing and intelligent monitoring system on the basis of the original central server management end, integrating the natural language understanding ability of the large language model, traditional scheduling algorithms, and statistical data, realizing high-speed acquisition, preprocessing, and noise reduction of doubling frame inspection data, especially broken yarn data; using a distributed stream processing platform and introducing an edge computing module at the inspection robot end to effectively shorten the transmission delay through data compression and network protocol optimization; at the same time, the system intelligently analyzes, dynamically decomposes, and real-time optimizes complex tasks and data through the large language model, and combines an adaptive scheduling algorithm, an online learning model, and a data time window aggregation technology to ensure the efficient utilization of real-time data, thereby endowing the scheduling system with stronger flexibility, adaptability, and intelligence; based on the access of the large model to the robot management end to achieve natural language interaction, the system not only realizes human-machine collaboration, but also significantly reduces the complexity of multi-robot task scheduling and enhances the response ability to dynamic environments and user requirements according to the dynamic planning of broken yarn statistical data; specifically including the following advantages:

[0048] 1. Efficient task parsing

[0049] Using the large language model to perform semantic parsing on natural language, extracting key task information (such as goals, priorities, deadlines, etc.), and converting the parsed tasks into structured data (such as JSON) for subsequent scheduling processing, enabling users to directly describe tasks in natural language without complex scheduling rules or technical backgrounds; the system quickly parses task goals and constraints, avoiding the cumbersome steps of manual configuration.

[0050] 2. Dynamicity and flexibility

[0051] By collecting robot status (battery level, location, workload) and environmental change information, combining task requirements and real-time data, dynamically updating task allocation and priorities; the real-time feedback of task status and user instructions act together to trigger task adjustment, enabling task scheduling to adapt to environmental changes and reducing system stagnation caused by new tasks or robot failures; multiple robots can flexibly adjust task allocation to achieve efficient cooperation.

[0052] 3. Intelligent optimization ability

[0053] Calculate the optimal solution for task allocation based on task priorities and robot resources; analyze historical task patterns through the context memory ability of the large language model to optimize the current scheduling and improve the overall task completion efficiency; reduce task conflicts and resource waste, and maximize robot utilization.

[0054] 4. Simplified user interaction

[0055] The large language model converts task execution data into easily understandable language descriptions; user queries are quickly parsed through the language model, and the feedback is presented in natural language. Users can query the task status and robot execution situation in real time. Task operations are intuitive and simple; friendly feedback information allows users to easily understand the progress without frequent intervention.

[0056] 5. Heterogeneous system support

[0057] Model the performance of robots, establish performance indicators for each robot (such as load capacity, endurance, task capabilities); based on the matching degree between task requirements and robot performance, intelligently allocate tasks; through real-time monitoring of robot status, adopt a load balancing algorithm to ensure uniform task distribution. In this way, the system can simultaneously schedule robots with different functions and performances, efficiently complete complex tasks, accurately allocate tasks according to robot characteristics, and reduce resource waste.

[0058] 6. Improve system robustness

[0059] The large language model has strong context reasoning ability, which can complete fuzzy information or handle ambiguity. The system can process incomplete or fuzzy task descriptions to ensure that task parsing is error-free; through real-time monitoring, identify robot failures and trigger reallocation, quickly resume scheduling when a failure occurs, and the distributed scheduling architecture ensures that the failure of a single task will not affect the overall system, reducing downtime and task delays.

[0060] 7. Environmental adaptability:

[0061] Adopt network remote communication. Users do not need to go to the factory site and can achieve remote control only on the management side; users use "0" code and operate in natural language throughout the process; the environment of the doubling frame textile factory is extremely noisy, and the air environment and temperature environment are extremely harsh. Through dynamic deployment, it adapts to the complex factory environment.

[0062] 8. Real-time data

[0063] The inspection data of the doubling machine, especially the yarn breakage data, is characterized by a large amount and real-time nature. During the inspection process, the doubling machine inspection robot continuously sends data to the central server through network communication, and there will be a delay in the data update speed. By collecting and preprocessing the inspection data of the doubling machine at high speed, especially the yarn breakage data, filtering out environmental interference and eliminating invalid data, and at the same time using the Apache Kafka and Flink distributed stream processing platforms, and introducing an edge computing module at the inspection robot end, the transmission delay is effectively shortened through data compression and network protocol optimization; through the large language model, intelligent parsing, dynamic decomposition and real-time optimization of complex tasks and data are carried out, and combined with the adaptive scheduling algorithm, online learning model and data time window aggregation technology, the efficient utilization of real-time data is ensured, and the real-time nature of data utilization is guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The present invention will be further described below in conjunction with the drawings and embodiments.

[0065] Figure 1 It is a detailed flowchart of the task scheduling method of the doubling machine inspection robot based on the large language model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] The present invention will now be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0067] A task scheduling method for a doubling machine inspection robot based on a large language model of the present invention includes the following steps:

[0068] Step S1: On-site data collection and data edge preprocessing:

[0069] When an abnormality occurs during the operation of the doubling machine, the most common abnormality of the doubling machine is yarn breakage. The on-site sensors of the doubling machine and / or the sensors on the inspection robot capture relevant abnormal data. After the inspection robot obtains the relevant abnormal data, the edge computing module on the inspection robot performs preliminary preprocessing on the data.

[0070] Step S2: Real-time data transmission and distributed stream processing:

[0071] The preprocessed data is uploaded to the central server through the optimized network protocol; with the help of distributed data stream transmission, it is ensured that each piece of data can reach in time. After the central server receives the data stream, time window aggregation and real-time computing technologies are used to further integrate and analyze the data; the optimized network protocol refers to setting the communication address as a static IP to make the communication between devices stable, and the data transmission uses the HTTP protocol; it also ensures the security in data transmission.

[0072] Step S3: Intelligent data parsing and task dynamic decomposition:

[0073] The data processed in Step S2 is sent to the large language model integrated in the central server. The large language model uses natural language understanding capabilities to deeply analyze the monitoring data, identify abnormal patterns and key metrics, and automatically generate an analysis report. According to the abnormal patterns and key metric information, specific inspection tasks are dynamically decomposed, and the priority and processing requirements of each task are clarified;

[0074] Step S4: Adaptive scheduling and task optimization:

[0075] Based on the parsing results of the large language model, the scheduling system adopts an adaptive scheduling algorithm combined with an online learning model to formulate an optimal task scheduling plan in real time; among them, the adaptive scheduling algorithm is to identify the map of the entire factory through the large language model, assign values to each section of the path, and add the specific coordinates of each inspection robot to the map of the factory, and perform task allocation according to natural language; for example: the first inspection robot is supposed to inspect the first ten machines, but now the first inspection robot is not in the area of the first ten machines, so it is necessary to calculate the shortest arrival path according to the assigned values of each section of the path. At the same time, the results of other inspection robots during the inspection, such as there are workers working on a certain section of the path or there are obstacles on the path, can also be used as reference factors in path planning.

[0076] Step S5: Task distribution and natural language interaction:

[0077] The task scheduling plan is conveyed to the operator and the inspection robot through the natural language interface. The operator directly interacts with the scheduling system through voice or text to confirm or fine-tune the scheduling plan; after receiving the detailed instructions, the inspection robot immediately enters the task execution state to ensure that on-site abnormal problems are quickly responded to;

[0078] Step S6: Inspection task execution and on-site feedback:

[0079] The inspection robot goes to the fault or monitoring area according to the received task instructions to carry out on-site inspections and maintenance work; during the task execution process, the inspection robot continuously collects new on-site data and real-time feedbacks information such as detection results and processing status to the central server, so that the scheduling system can track and monitor the overall operation situation and make timely adjustments;

[0080] Step S7: Real-time monitoring and dynamic task adjustment:

[0081] The central server monitors all uploaded data, the status of the inspection robots, and the progress of task execution in real time; when the scheduling system detects new anomalies or changes in the on-site environment, it immediately triggers an alarm, and uses a large language model and scheduling algorithms to re-plan tasks, and sends the new scheduling plan to the relevant inspection robots to ensure that production line anomalies can be processed in a timely manner.

[0082] Before step S1, it also includes step S0: Initialization of the scheduling system and startup of on-site equipment:

[0083] Start the scheduling system, complete the initialization configuration of the inspection robots and tasks; collect the current status of the inspection robots, such as location, battery level, load capacity, etc.; create a task pool to record the basic information of all tasks to be executed, including task location, priority, timeliness, etc. Before production, the scheduling system first completes the initialization; all on-site equipment (sensors, inspection robots, edge computing units) perform self-checks, calibrations, and networking in sequence to ensure the accuracy of data collection and clock synchronization; after the sensors installed on-site are started, they begin to monitor the operating status of the twister in real time, paying special attention to anomalies (such as broken yarn, etc.); at the same time, the edge computing module of the inspection robot is activated and enters the standby state, ready to process on-site data at any time; the user inputs natural language through the management platform, the large language model integrated in the central server reads the information, extracts keywords from the natural language input by the user for analysis and processing, and the task pool receives the task request.

[0084] In step S0, during the calibration process of the sensor, calculate the calibration error Ecal:

[0085] Ecal = ∣Smeasured - Sreference∣

[0086] Where Smeasured represents the measured value, Sreference represents the reference value, and when Ecal is less than the preset allowable error, the sensor is considered calibrated. This step ensures that subsequent data processing is based on reliable inputs.

[0087] In step S1, during the operation of the twister, high-precision sensors installed at key positions collect abnormal data such as broken yarn in real time, or the inspection robot's own sensors or cameras collect abnormal data, and the edge computing module performs preliminary preprocessing on the data, including the following steps: First, perform noise filtering to filter out environmental interference and eliminate invalid data; then compress the data to reduce the data transmission volume and ensure that the network load remains within a reasonable range; at the same time, perform preliminary anomaly detection to ensure that only key data is transmitted to the central server. In a multi-robot operating environment, it is easy for multiple robots to inspect the same location at the same time. If the broken yarn in this area is not processed in time, it will cause duplicate data transmission. The preliminary anomaly detection here is equivalent to deduplicating a large amount of duplicate data.

[0088] For noise filtering, an exponential smoothing filter is used to reduce the noise of the original data. The smoothing filter formula can be written as:

[0089] y(t) = αx(t) + (1 - α)y(t - 1)

[0090] where x(t) is the current sampled data, y(t) is the filtered data, and α is the smoothing factor.

[0091] In step S2, Apache Kafka is used to achieve distributed data stream transmission. The Apache Flink platform of the central server receives the data stream, and the scheduling system aggregates, groups, and calculates the time window of the data in real time for subsequent intelligent parsing. The average calculation of the time window data is as follows:

[0092]

[0093] where N is the time window size, and x(t - i) are the data values at each moment.

[0094] Step S3 includes the following steps:

[0095] Step S31: Sub - task decomposition is carried out according to four decomposition rules: region, device, special requirements, and dynamic tasks. The user's natural language tasks are decomposed from four aspects: region, device, special requirements, and dynamic tasks, and the priority is determined according to four aspects: urgency, importance, timeliness, and path optimization, and the final priority results are sorted; if there are specific tasks that need to be processed first, manually specify them as the highest priority;

[0096] The priority calculation formula: P = w1·U + w2·I + w3·T - w4·D

[0097] U: Urgency weight; I: Importance weight; T: Timeliness weight; D: Path distance, a negative value indicates a lower priority for a longer distance; w1, w2, w3, and w4 are all weight values and can be adjusted according to actual needs;

[0098] Step S32: Allocate task resources according to performance: According to the performance indicators of the inspection robot itself, the consumption of task requirements needs to be evaluated for performance, and the workload of the task is matched with the performance consumption of the inspection robot, and the allocation is carried out in combination with the user's natural language requirements;

[0099] Performance evaluation: For example, when there are sixty machines that need to be inspected and five inspection robots are idle, but the remaining battery levels of the five inspection robots and their stopping positions are different, we need to analyze the performance of each inspection robot and analyze how much power is consumed by relevant environmental factors such as the distance traveled to inspect the sixty machines. Here, the user's requirements are the top priority. However, if there is a performance deficiency, such as when the user asks the No. 1 inspection robot to inspect the first ten machines, but at this time the No. 1 inspection robot has insufficient battery power and cannot complete the task, then resource allocation becomes the top priority, and the inspection robot that is closest to the first ten machines and has sufficient battery power will be given priority to work.

[0100] The task priority sorting algorithms in step S31 include the greedy algorithm, dynamic programming, comprehensive sorting algorithm, and priority queue. Greedy algorithm: When the large model initially recognizes natural language, in order to meet the requirements of efficient and real-time requirement changes, the greedy algorithm will be used to achieve the initial planning of tasks. Dynamic programming: In the task sorting of robots, there will be path planning implementations. Through dynamic programming, the robot can reach the target location or return to the charging pile. Comprehensive sorting algorithm: Combine the robot's own performance, obstacles on the path, and on-site working conditions to sort tasks. Priority queue: Mainly consider that the administrator temporarily changes tasks and adjusts priorities during the robot's work.

[0101] The triggering conditions for the task priority sorting algorithm include when the scheduling system receives a new task request; when the resource status of the inspection robot itself changes; when the scheduling system has an emergency task; and triggering task re-sorting at regular intervals.

[0102] The generation speed of the task priority sorting algorithm is related to the complexity of the algorithm itself. The complexity of the greedy algorithm is O(nlogn), and it can still quickly generate sorting results when the task scale is large, which is suitable for scenarios with high real-time requirements; the complexity of the dynamic programming algorithm is usually O(n 2 ), which is applicable to situations with complex task dependencies, but the speed is relatively slow; the comprehensive sorting algorithm can comprehensively consider various factors under the complexity of O(nlogn) and generate sorting results with excellent trade-off performance; the priority queue method can insert or remove tasks within O(logn) time, which is very suitable for scenarios where the task pool changes dynamically.

[0103] The update frequency of task sorting is usually determined according to specific application scenarios and requirements.

[0104] The greedy algorithm is suitable for scenarios where there are no dependencies between tasks; the dynamic programming algorithm is applicable to scenarios where there are complex dependencies between tasks; the comprehensive sorting algorithm is applicable to scenarios where it is necessary to balance task priorities and execution costs; the priority queue algorithm is suitable for scenarios where the task pool changes dynamically.

[0105] Calculation basis for task sorting: First is the task priority, which is determined by the importance or urgency of the task; second is the timeliness of the task, that is, the degree of time urgency for the task to be completed; in addition, the distance between the task location and the current location of the inspection robot, as well as the current resource status of the inspection robot (such as battery power, load capacity, etc.) also need to be considered.

[0106] Timeliness of the algorithm: One is the response speed; the other is the timeliness of scheduling. The priority queue algorithm can quickly adjust the task order to cope with the dynamic changes of the task pool; at the same time, the ability to reorder in combination with priorities can quickly handle sudden tasks and ensure the timely completion of high-priority tasks.

[0107] Detailed calculation process of the basic timeliness of task sorting:

[0108] Task completion timeliness calculation formula: Ttotal = Ttravel + Texecution

[0109] Ttravel: The travel time from the current location to the task location,

[0110] Ttravel = D / V, where D is the distance and V is the average speed of the inspection robot;

[0111] Texecution: The time required for task execution, which is determined by the task complexity.

[0112] In step S7, the adaptive scheduling algorithm dynamically adjusts the task allocation based on the performance evaluation of the inspection robot according to the changes in the task pool or the changes in the state of the inspection robot, specifically according to the current broken yarn statistics data, the location of the inspection robot, the workload, and the changes in the on-site environment, (ensuring efficient collaboration among inspection robots and minimizing the fault response time. According to the changes in the task pool (new tasks, cancelled tasks) or the changes in the state of the inspection robot (battery exhaustion, failure, etc.), the task sorting and resource allocation are adjusted in real time, and the current unfinished tasks are re-evaluated and optimized.

[0113] After step S7, it also includes step S8: Closed-loop feedback and continuous optimization:

[0114] After each task is completed, the collected data, task records, and operation feedback are archived by the scheduling system for subsequent analysis and model optimization. Through continuous analysis of historical data and online learning, the scheduling system continuously optimizes the parsing ability of the large language model and the response strategy of the scheduling algorithm, thereby further improving the inspection efficiency and the system intelligence level in subsequent operations, forming a continuously self-improving closed-loop management process.

[0115] In step S8, the scheduling system continuously adjusts the parameters of the large language model and the scheduling algorithm by using an online learning method to improve the subsequent task scheduling efficiency. The parameter update is optimized by using the gradient descent method, and its update formula is as follows: where θ new represents the updated model parameters, θ old represents the model parameters before update, η is the learning rate, L(θ) is the loss function, and θ is the model parameter.

[0116] Taking the ideal embodiments of the present invention as the above inspiration, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A task scheduling method for a two-for-one twister inspection robot based on a large language model, characterized in that: The following steps are involved: Step S1: Field data collection and data edge preprocessing: When an abnormality occurs during the operation of the twister, the on-site sensors of the twister and / or the sensors on the inspection robot capture the relevant abnormal data. After the inspection robot obtains the relevant abnormal data, the edge computing module on the inspection robot performs preliminary preprocessing on the data. Step S2: Real-time data transmission and distributed stream processing: The pre-processed data is uploaded to the central server through the optimized network protocol; with the help of distributed data stream transmission, after the central server receives the data stream, it further integrates and analyzes the data using time window aggregation and real-time computing technology; the optimized network protocol refers to setting the communication address to a static IP, and data transmission using the HTTP protocol; Step S3: Intelligent data analysis and dynamic task decomposition: The data processed in step S2 is sent to the large language model integrated in the central server. The large language model uses natural language understanding capabilities to deeply analyze the data, identify abnormal patterns and key indicators, and dynamically decompose specific inspection tasks based on abnormal patterns and key indicator information, clarifying the priority and processing requirements of each task; Step S4: Adaptive scheduling and task optimization: Based on the analysis results of the large language model, an adaptive scheduling algorithm is used in combination with an online learning model to formulate the optimal task scheduling plan in real time. The adaptive scheduling algorithm uses the large language model to identify the map of the entire factory, assign values ​​to each path, and add the specific coordinates of each inspection robot to the factory map, and assign tasks based on natural language. Step S5: Task distribution and natural language interaction: Communicate the task scheduling plan to operators and inspection robots through a natural language interface; Step S6: Inspection task execution and on-site feedback: The inspection robot goes to the fault or monitoring area according to the received task instructions to carry out on-site inspection and maintenance work. During the task execution, the inspection robot collects new on-site data and feeds it back to the central server in real time. Step S7: Real-time monitoring and dynamic task adjustment: The central server monitors the uploaded data, inspection robot status and task execution progress in real time; when new anomalies are detected or changes in the on-site environment occur, the task is re-planned with the help of a large language model and scheduling algorithm, and the new scheduling plan is sent to the relevant inspection robots.

2. The task scheduling method for the inspection robot of the twister based on the large language model as claimed in claim 1 is characterized in that: Before step S1, step S0 is also included: initializing the dispatching system and starting the on-site equipment: Start the scheduling system, complete the initial configuration of the inspection robot and tasks; collect the current status of the inspection robot; create a task pool; Users input natural language through the management platform, the large language model integrated in the central server reads the information, extracts keywords from the natural language input by the user, analyzes and processes them, and the task pool receives the task request.

3. The task scheduling method for the inspection robot of the twister based on the large language model as claimed in claim 2 is characterized in that: In step S0, the sensor calculates the calibration error Ecal during the calibration process: Ecal=∣Smeasured-Sreference| Wherein, Smeasured represents the measured value, Sreference represents the reference value, and when Ecal is less than the preset allowable error, the sensor is considered to be calibrated properly.

4. The task scheduling method for a two-for-one twister inspection robot based on a large language model as claimed in claim 1, characterized in that: In step S1, the edge computing module performs preliminary preprocessing on the data, including the following steps: first, noise filtering is performed; then the data is compressed; at the same time, preliminary anomaly detection is performed to ensure that only critical data is transmitted to the central server.

5. The task scheduling method for the inspection robot of the twister based on the large language model as claimed in claim 4 is characterized in that: For noise filtering, an exponential smoothing filter is used to reduce the noise of the original data. The smoothing filter formula is: y(t)=αx(t)+(1-α)y(t-1) Where x(t) is the current sampling data, y(t) is the filtered data, and α is the smoothing factor.

6. The task scheduling method for a two-for-one twister inspection robot based on a large language model as claimed in claim 1, characterized in that: In step S2, distributed data stream transmission is implemented with the help of Apache Kafka. The Apache Flink platform of the central server receives the data stream, and the scheduling system aggregates, groups and calculates the time window of the data in real time. The average calculation of the time window data is as follows: Among them, N is the time window size, and x(ti) is the data value at each moment.

7. The task scheduling method for a two-for-one twister inspection robot based on a large language model as claimed in claim 1, characterized in that: Step S3 includes the following steps: Step S31: Decompose the subtasks according to the four decomposition rules of region, equipment, special needs and dynamic tasks. Decompose the user's natural language tasks from the four aspects of region, equipment, special needs and dynamic tasks, and determine the priority according to the four aspects of urgency, importance, timeliness and path optimization, and sort the final priority results; if there is a specific task that needs to be processed first, manually specify it as the highest priority; Priority calculation formula: P = w1·U+w2·I+w3·T-w4·D U: urgency weight; I: importance weight; T: timeliness weight; D: path distance, negative value means the longer the distance, the lower the priority; w1, w2, w3 and w4 are all weight values; Step S32: Allocate task resources based on performance: Based on the performance indicators of the inspection robot itself, the consumption of the task requirements needs to be evaluated, the workload of the task is matched with the performance consumption of the inspection robot, and the allocation is carried out in combination with the natural language requirements of the user; The task priority sorting algorithm in step S31 includes a greedy algorithm, a dynamic programming, a comprehensive sorting algorithm and a priority queue.

8. The task scheduling method for a two-for-one twister inspection robot based on a large language model as claimed in claim 1, characterized in that: In step S7, the adaptive scheduling algorithm dynamically adjusts the task allocation based on the performance evaluation of the inspection robot according to the changes in the task pool or the state of the inspection robot.

9. The task scheduling method for a two-for-one twister inspection robot based on a large language model as claimed in claim 1, characterized in that: After step S7, step S8 is also included: closed-loop feedback and continuous optimization: After each task is completed, the collected data, task records and operation feedback are archived. Through continuous analysis of historical data and online learning, the parsing capabilities of the large language model and the response strategy of the scheduling algorithm are continuously optimized.

10. The task scheduling method for the inspection robot of the twister based on the large language model as claimed in claim 9, characterized in that: In step S8, the parameters of the large language model and the scheduling algorithm are continuously adjusted using an online learning method, and the parameter update is optimized using a gradient descent method, and the update formula is: θ new =θ old -η where θ new represents the updated model parameters, θ old represents the model parameters before updating, η is the learning rate, L(θ) is the loss function, and θ is the model parameter.

Images