Scheduling system operation method based on AI model

Abstract

The invention provides a scheduling system operation method based on an AI model, and the method comprises the steps: a central management system receives real-time feedback signals of an AGV and line edge docking equipment, determines a task event needing to be processed, and accesses an AI frame module; the AI framework module calls a deep learning library; the deep learning library acquires content data associated with a current task event from real-time feedback signal data, received by the central management system, of the AGV and the line-side docking equipment according to a request of the central management system, and counts a plurality of pieces of real-time quantitative data corresponding to the task event; inputting the real-time quantitative data and the type of the task event into the trained AI model, and outputting a reasoning result to the central management system by the deep learning library through an AI framework module; and the central management system controls the operation of the AGV and / or the line edge butt joint equipment according to the reasoning result. According to the operation method, the ai technology is utilized to enable the dispatching system to control the AGV more intelligently and efficiently.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent logistics transportation, and in particular to an operation method of a scheduling system based on an AI model. Background Art

[0002] In the existing intelligent logistics transportation operation scenario, a central scheduling system is used to assign handling tasks, charge control, traffic control, etc. to several AGVs. The output commands in the central scheduling system are based on its preset scheduling logic. For example, when assigning handling tasks, when the production line issues a demand command, the central scheduling system assigns an idle vehicle that is relatively close in location to perform the handling work according to the priority level set by the task. Another example is when performing charge control. When an AGV issues a low battery alarm, the central scheduling system arranges it to go for charging. Another example is when performing traffic control. Due to congestion at the intersection in the traffic control area, the central scheduling system will issue instructions to several AGVs according to the set scheduling logic and equipment priority level to clear the intersection. The above control and scheduling of the central scheduling system are all based on the preset scheduling logic, and for abnormal situations and more complex production scenarios, it cannot output special intelligent scheduling methods. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art, provide an operation method of a scheduling system based on an AI model, design a scheduling system based on an AI model, and use AI technology to enable the scheduling system to control AGVs more intelligently and efficiently, making the entire logistics scheduling system more stable, reliable and efficient.

[0004] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0005] An operation method of a scheduling system based on an AI model, comprising:

[0006] Several AGVs;

[0007] Several side docking devices;

[0008] A central management system;

[0009] An AI framework module, which is communicatively connected to the central management system, and a deep learning library is deployed thereon. The deep learning library is used to create an AI model and perform deep learning training for the AI model;

[0010] The AI framework module monitors whether there is data input into the central management system. If so, the data is screened to obtain learning data, and the learning data is input into the deep learning library.

[0011] The operation method includes:

[0012] The central management system receives real-time feedback signals from AGVs and in-line docking devices, determines task events to be processed, and accesses the AI framework module;

[0013] The AI framework module calls the deep learning library;

[0014] The AI framework module selects idle computing hardware resources of the central management system;

[0015] Based on the request of the central management system, the deep learning library obtains content data associated with the current task event from the real-time feedback signal data of AGVs and in-line docking devices received by the central management system, and statistically calculates a number of real-time quantization data corresponding to the task event;

[0016] Input the real-time quantization data and the type of the task event into the trained AI model, and the deep learning library outputs the inference result to the central management system through the AI framework module;

[0017] The central management system sends information to the AGV and / or in-line docking device according to the inference result to control the operation of the AGV and / or in-line docking device, and feedback relevant guidance information.

[0018] Compared with the prior art, a scheduling system operation method based on an AI model of the present invention enables the AI model to access the central management system, uses AI technology to analyze abnormal situations or more complex production scenarios to obtain inference results, enables the central management system to control AGVs more intelligently and flexibly to cope with abnormal situations and more complex production scenarios, improves the operation stability of the entire scheduling system, and improves the processing efficiency of the entire scheduling system after encountering abnormal situations.

[0019] Further, the task event includes equipment life prediction;

[0020] Equipment life prediction means that the AI model predicts the life of AGVs and in-line docking devices based on the operating conditions of the equipment, the no-load and full-load frequencies, the usage frequency, motor performance parameters, and the number of emergency stops under rated conditions;

[0021] After the AI framework module outputs the inference result, the central management system obtains the equipment life prediction result according to the inference result, uses a classification model to classify the health status of the equipment, such as "healthy", "sub-healthy", "fault warning" and other states, and makes a predictive equipment maintenance request to external devices.

[0022] The function of equipment life prediction is to feedback to the superior equipment in advance that the critical life of the components of AGV and the line-side docking equipment is approaching, so as to facilitate subsequent requests for manual processing, replace the key components of AGV and the line-side docking equipment, reduce the impact of the maintenance of AGV and the line-side docking equipment on the operation of the scheduling system, and thus improve the overall reliability of the scheduling system.

[0023] Furthermore, the content data collected for AGV includes the theoretical charge and discharge times of the battery and the actual charge and discharge times, the theoretical mileage, theoretical usage time, the mileage already used and the time already used of the motor, reducer, and wheels, and the number of emergency stops of the equipment.

[0024] The content data collected for the line-side docking equipment includes the actual load and the frequency of the forward and reverse operation of the motor.

[0025] The equipment life prediction result is obtained through calculation.

[0026] Furthermore, after the central management system sends a predictive equipment maintenance request to external equipment, it monitors whether the request is executed.

[0027] If the request is not executed, the central management system adjusts the usage conditions of the corresponding AGV and line-side docking equipment.

[0028] For AGV, the central management system adjusts the moving speed, load capacity, running path, and working frequency of AGV.

[0029] For the line-side docking equipment, the central management system adjusts the load capacity and working frequency of the line-side docking equipment.

[0030] Since the predictive equipment maintenance request is not executed, if the AGV and the line-side docking equipment continue to run in the current state, accidents may occur, affecting the normal operation of the scheduling system and incurring greater problem-solving costs. Therefore, by adjusting (lowering) the working efficiency and workload of the AGV and the line-side docking equipment (such as setting that AGVs with a high battery attenuation degree and a battery approaching the end of its life will undertake some easy and short-distance tasks, while AGVs with insufficient motor life will be arranged for some occasions where they can maintain a uniform speed and travel a long distance), the usage time of the AGV and the line-side docking equipment can be extended until the predictive equipment maintenance request is executed, enabling the service life of this system to continue.

[0031] Furthermore, the task event includes external environment analysis.

[0032] External environment analysis means that the AI model predicts the external environment situation based on the obtained environmental condition parameters and the working conditions of the AGV.

[0033] After the AI framework module outputs the inference results, the central management system obtains the external environment analysis results based on the inference results and makes corresponding adjustments to the use of AGVs;

[0034] If the external environment analysis results show that the air humidity exceeds the preset value, the central management system will reduce the rated handling load of the AGV and increase the number of handling times;

[0035] If the external environment analysis results show that there is water accumulation in a specific area of the work surface, the central management system will reduce the rated handling load of the AGV, increase the number of handling times and reduce the speed, and at the same time make a workplace maintenance request to the external equipment.

[0036] If the external environment analysis results show that a specific area of the work surface is damaged, the central management system controls the AGV to slow down when it drives to that area, and at the same time makes a request for workplace maintenance to the external equipment.

[0037] The beneficial effects of the above external environment analysis results are:

[0038] When the air humidity exceeds the preset value, it means that the AGV is in a humid environment. When working in a humid environment (wet ground), the AGV will slip and stop inaccurately, which enables the AI model to make accurate predictions; by reducing the rated handling load of the AGV, increasing the number of handling times, and reducing the speed, it can avoid turning or emergency stop slipping, and can continue to perform transportation tasks without affecting the production line production rhythm.

[0039] When there is water accumulation in a specific area of the work surface, the AGV will slip and stop inaccurately, and the walking speed (moving position relative to the map) will fluctuate significantly, which enables the AI model to make accurate predictions; by reducing the rated handling load of the AGV, increasing the number of handling times, and performing workplace maintenance, this problem can be solved while ensuring the normal production rhythm of the production line.

[0040] When a specific area of the working surface is damaged, the AGV's walking motor current frequently fluctuates greatly at a certain location on the same section of the road, which enables the AI model to predict that the ground in that area is damaged. By controlling the AGV to slow down when it drives to that area, the AGV can avoid passing through the damaged bottom surface at a high speed, causing materials to fall and the vehicle body to be damaged.

[0041] Further, the task event includes supply and demand analysis;

[0042] Supply and demand analysis refers to the AI model predicting material handling demand based on production schedule, material availability, product shipment priority, and warehouse capacity status;

[0043] After the AI framework module outputs the inference result, the central management system obtains the supply and demand analysis result based on the inference result and sets the temporary production scheduling plan.

[0044] If the material handling demand is speculated based on the production plan schedule, the deep learning library analyzes the time periods of busy production and idle production according to the production plan schedule; when the central management system sets the temporary production scheduling plan, during the busy production period, idle AGVs are arranged in advance around a specific workstation to wait before the operation of the workstation. During the idle production period, the capacity of idle or under-loaded AGVs is utilized to pick up more handling tasks during the movement to perform material distribution.

[0045] If the material handling demand is speculated based on the preparation status of production materials, the deep learning library analyzes the short-board materials according to the preparation status of production materials; when the central management system sets the temporary production scheduling plan, it sends a message to external devices to request the priority production or replenishment of short-board materials.

[0046] If the material handling demand is speculated based on the product shipment priority level, the deep learning library analyzes the level of the goods that need to be shipped out first and the path of the transportation route of the goods according to the product shipment priority level; when the central management system sets the temporary production scheduling plan, it delimits the important transportation channel area according to the transportation route path of the goods to be shipped out first, classifies the AGVs according to the level of the goods to be shipped out first for the distribution task level, restricts the AGVs with a lower distribution task level from entering the important transportation channel area, and enables the AGVs with a higher distribution task level to enter the important transportation channel area for movement.

[0047] If the material handling demand is speculated based on the capacity status of the warehouse, the deep learning library analyzes the utilization rates of several storage positions according to the capacity status of the warehouse; when the central management system sets the temporary production scheduling plan, it determines whether the archived quantity of material A exceeds the preset capacity value of material A relative to the storage position X0 where material A is currently arranged. If so, the central management system locates the storage position Xn with a relatively low utilization rate and allowed placement around the current storage position X0 of material A, and the central management system controls the AGV to temporarily place the material A exceeding the capacity in the storage position Xn.

[0048] Supply and demand analysis coordinates the AGV transportation capacity, warehouse capacity, production plan, material inventory, and product turnover rate, thereby fully mobilizing the operation of all resources within the dispatching system, greatly improving the efficiency of the dispatching system, and enhancing the transportation capacity and productivity. Additionally, based on the transportation route of the goods to be shipped out first, the important transportation channel area is demarcated. The AGVs are classified according to the distribution task levels of the goods to be shipped out first. AGVs with lower distribution task levels are restricted from entering this important transportation channel area, and AGVs with higher distribution task levels are allowed to enter and move in this important transportation channel area, thereby increasing the moving speed of AGVs with heavy tasks and reducing the occurrence of congestion in the important transportation channel area.

[0049] Furthermore, the task event includes unique identification;

[0050] Unique identification means that the AI model performs unique identification of handling tasks, unique identification of items, and unique identification of personnel based on the obtained computer vision data, the medium feature data recognized by the AGV for the handled medium, and the surrounding environment data recognized by the AGV, and infers the analysis result of the AGV working conditions;

[0051] After the AI framework module outputs the inference result, the central management system judges the working status of the AGV and the rationality of the handling task according to the analysis result of the AGV working conditions. If the working status of the AGV is abnormal or the rationality of the handling task is incorrect, feedback such as blocking, warning, and tracking is given to the specific AGV;

[0052] The process of unique identification of handling tasks is as follows: The AI model identifies the type of equipment docked with a specific AGV based on the obtained computer vision data, the medium feature data recognized by the AGV for the handled medium, and the surrounding environment data recognized by the AGV. If the type of the docked equipment belongs to the equipment allowed to be docked by the specific AGV, then the docked equipment passes the unique identification;

[0053] The process of unique identification of items is as follows: The AI model performs one of the following identifications based on the obtained computer vision data, the medium feature data recognized by the AGV for the handled medium, and the surrounding environment data recognized by the AGV:

[0054] i. Identify whether the motion state of the AGV loading a specific item is abnormal compared to its normal motion state when loading this item. If it is judged that the motion state of the AGV is abnormal, then the specific item loaded by the AGV does not pass the unique identification;

[0055] ii. Identify the integrity of the item to be loaded. If it is judged that the integrity of the item to be loaded is abnormal, then the specific item loaded by the AGV does not pass the unique identification;

[0056] The process of uniquely identifying a person is as follows: the AI model identifies whether human intervention has occurred in a specific AGV based on the acquired computer vision data, the characteristic data of the medium being transported identified by the AGV, and the surrounding environment data identified by the AGV. If human intervention has occurred, the physical characteristics of the person who illegally controls or interferes with the normal operation of the AGV are identified, and the person who meets the physical characteristics is identified to obtain the unique data of the person.

[0057] The unique identification of the handling task can ensure that the AGV only docks with the corresponding docking equipment (for example, the AGV dives into the bottom of the material cart and identifies its uniqueness by identifying the mark on the material cart. The AGV in the clean area only docks with the material cart in the clean area, and will not tow and transport the material cart in the non-clean area), avoiding accidents or material contamination caused by the AGV docking with equipment that is not allowed to be docked by the AGV;

[0058] The unique identification of objects can ensure that AGV only transports intact materials, avoids incomplete materials from flowing into the production line and affecting the normal operation of the production line, and reduces the maintenance cost required to locate incomplete materials flowing into the production line (previously, objects were identified by their RFID tags without considering the integrity of the materials. The method of judging the integrity of materials in the present invention is as follows: a. When an AGV in a conventional production line that is transporting materials back and forth transports one of the boxes of materials, the AGV driving data is analyzed, and it is found that the driving current or torque is different from the conventional ones. The AI model will judge that the box of materials may be in an abnormal state; b. Through visual data comparison, it is found that the material box of the transported material is damaged, the label is missing, the bottle cap is loose, or the material box is deformed during the transportation process, or the material box has been opened. The AI model will judge that the box of materials is abnormal, and the central management system will handle the abnormality based on these abnormal information or feedback the corresponding information to the upper system);

[0059] The unique identification of personnel is used to determine the identity of personnel who affect the normal execution of AGV tasks (by analyzing the physical characteristics of personnel, such as hands, walking posture, face, etc.), so as to facilitate subsequent adjustments to the personnel's work path to avoid accidents.

[0060] Furthermore, the mission events include emergency event handling;

[0061] Emergency event handling means that the AI model infers the type of emergency event based on the obtained tracking records without affecting the normal handling work of the AGV and without allowing the emergency event to affect production;

[0062] After the AI framework module outputs the inference results, the central management system obtains the emergency event processing strategy based on the inference results, and the central management system sends messages to external devices to execute the corresponding emergency event processing plan;

[0063] If the tracking record contains information about damaged materials to be loaded, the deep learning library records the flow direction of the AGV that loads the materials to be loaded, analyzes that the damaged materials to be loaded need to be replaced before final assembly; when the central management system outputs the processing strategy for emergency events, it dispatches another AGV to retrieve the damaged materials to be loaded and transport them, and dispatches another AGV to the corresponding parts area to pick up new materials identical to the damaged materials to be loaded, and tracks the AGV that originally loaded the damaged materials to be loaded, and sends the new materials to it;

[0064] If the tracking record contains information about abnormal channels, the deep learning library records the abnormal channel areas, plans a new route for the channel area that can avoid the abnormality and regenerates the driving map; when the central management system outputs the processing strategy for emergency events, it controls the AGV to execute tasks using the new driving map; if the information about abnormal channels in the tracking record is deleted (the channel is restored for use), the deep learning library regenerates the driving map, and the central management system controls the AGV to execute tasks using the new driving map;

[0065] If the tracking record contains newly inserted handling tasks, the deep learning library analyzes the speed-up amount and the additional handling amount of the pre-planned tasks; when the central management system outputs the processing strategy for emergency events, it controls the AGV that executes the pre-planned tasks to perform handling with increased speed or quantity.

[0066] The emergency event handling of the present invention includes the emergency handling of damaged materials to be loaded, the emergency handling of abnormal channels, and the emergency handling of newly inserted handling tasks. By handling the above emergency events, it is possible to prevent damaged materials from flowing into the production line, abnormal channels from affecting the transportation network, at the same time prevent AGVs from entering abnormal channels and aggravating the abnormality of abnormal channels, and prevent newly inserted handling tasks from disrupting the original production plan.

[0067] Further, it further includes a client module. The client module is communicatively connected to the central management system. Operating the client module can monitor the AGV or call the AGV;

[0068] Alternatively, it further includes a client module. The client module is communicatively connected to the central management system. Operating the client module can monitor the AGV or call the AGV; The staff can interact with the central management system through the client module or the AGV, or the staff can directly interact with the central management system. The AI framework module extracts the new transportation task content in the interaction content, and obtains a number of real-time quantization data according to the new transportation task content data;

[0069] The deep learning library outputs an inference result according to the real-time quantization data;

[0070] The central management system obtains a temporary scheduling plan according to the inference result. The execution process of the temporary scheduling plan includes at least one of the following situations:

[0071] i. The central management system assigns the AGV to perform a specific task temporarily;

[0072] ii. The central management system changes the material distribution path or destination of the AGV;

[0073] iii. The central management system adjusts the material distribution speed of the AGV performing the transportation task to achieve early delivery or delayed delivery of materials.

[0074] The setting of the client module realizes the connection to the central management system, realizes the functions of remotely controlling and supervising the AGV or calling the AGV, and can receive the inference analysis structure output by the AI model for the staff to view; the AI model of the present invention is a generative AI, which can generate relevant inference results according to the temporarily added transportation tasks of the staff, so that the central management system outputs the optimal temporary scheduling plan externally, and can complete the execution of the temporary scheduling plan while ensuring the normal progress of the original production plan.

[0075] Furthermore, the central management system includes an operation module, a control module, a data storage module, a data input / output module, a network service module, a map drawing module, an operation module, an AGV control and communication module, a traffic management module, a data recording module, a task management module, a status monitoring module, and an AGV scheduling interface module, wherein the AGV scheduling interface module is used to enable the central management system to support several database formats. Detailed implementation manners

[0076] The following specifically describes the implementation manners of the present invention:

[0077] Introduction to the prior art:

[0078] The AI framework module is a set of standard interfaces, feature libraries, and toolkits for the design, training, and verification of AI algorithm models. It integrates the encapsulation of algorithms, the invocation of data, and the use of computing resources, and includes functions such as data processing, model training, model evaluation, and model optimization, and is used to build, train, and deploy deep learning models.

[0079] The deep learning library includes the core training and inference modules of deep learning (a set of pre-set functions and modules), a basic model library, an end-to-end development kit, and rich tool components, which can simplify the creation and training process of deep neural networks, and more efficiently build and train deep neural networks. These networks usually contain multiple layers of processing units, and discover complex structures and relationships in the data by learning a large amount of data, thereby improving the inference accuracy and efficiency of various tasks.

[0080] Generative AI: Traditional AI relies on manual programming and preset rules to perform tasks. Traditional AI can only learn and optimize existing algorithms based on predefined rules and datasets, and the content it outputs is also based on the datasets input into it. In contrast, generative AI can continuously update algorithms and models during deep learning, and then generate content such as text, images, sounds, and videos based on the algorithms, models, and rules. According to user needs and combining the probabilities of associated words, it can create new data.

[0081] The operation method of the scheduling system based on the AI model in this embodiment includes:

[0082] A number of AGVs;

[0083] A number of in-line docking devices;

[0084] A central management system;

[0085] A client module, which is communicatively connected to the central management system. Operating the client module can monitor the AGVs or call the AGVs;

[0086] An AI framework module, which is communicatively connected to the central management system. A deep learning library is deployed in it, and the deep learning library is used to create an AI model and perform deep learning training for the AI model;

[0087] The AI framework module monitors whether there is data input into the central management system. If so, it screens the data to obtain learning data and inputs the learning data into the deep learning library;

[0088] The operation method includes:

[0089] The central management system receives real-time feedback signals from the AGVs and in-line docking devices, and at the same time determines the task events that need to be processed and accesses the AI framework module;

[0090] The AI framework module calls the deep learning library;

[0091] The AI framework module selects idle computing hardware resources of the central management system;

[0092] The deep learning library obtains content data associated with the current task event from the real-time feedback signal data of the AGVs and in-line docking devices received by the central management system according to the request of the central management system, and counts a number of real-time quantization data corresponding to the task event based on the content data;

[0093] Input the real-time quantization data and the type of task event into the trained AI model, and the deep learning library outputs the inference result to the central management system through the AI framework module;

[0094] The central management system sends information to the AGV and / or the in-line docking device according to the inference result to control the operation of the AGV and / or the in-line docking device, and feedbacks relevant guiding information.

[0095] In this embodiment, the AI framework module can remotely access other deep learning libraries, so as to further improve and optimize the AI model.

[0096] Compared with the prior art, a scheduling system operation method based on an AI model according to the present invention enables the AI model to access the central management system, analyzes abnormal situations or more complex production scenarios using AI technology to obtain inference results, enables the central management system to more intelligently and flexibly control the AGV to cope with abnormal situations and more complex production scenarios, improves the operation stability of the entire scheduling system, and improves the processing efficiency of the entire scheduling system after encountering abnormal situations.

[0097] Further, the task event includes equipment life prediction;

[0098] Equipment life prediction refers to the AI model predicting the life of the AGV and the in-line docking device according to the operating conditions of the equipment, the no-load and full-load frequencies, the usage frequency, the motor performance parameters, and the number of emergency stops under rated conditions.

[0099] After the AI framework module outputs the inference result, the central management system obtains the equipment life prediction result according to the inference result, uses a classification model to classify the health status of the equipment, such as "healthy", "sub-healthy", "fault warning" and other states, and sends a predictive equipment maintenance request to external equipment.

[0100] The role of equipment life prediction is to feedback to the superior equipment in advance that the critical life of the components of the AGV and the in-line docking device is approaching, so as to facilitate subsequent manual processing requests, replace the key components of the AGV and the in-line docking device, and reduce the impact of the repair of the AGV and the in-line docking device on the operation of the scheduling system, thereby improving the overall reliability of the scheduling system.

[0101] Further, the content data collected for the AGV includes the theoretical and actual charge-discharge times of the battery, the theoretical usage mileage, theoretical usage time, the mileage and time already used of the motor, reducer, and wheels, and the number of equipment emergency stops;

[0102] The content data collected for the in-line docking device includes the actual load and the frequency of the motor running forward and backward;

[0103] The equipment life prediction result is obtained through the following formula:

[0104] Introduction to the theoretical and actual charge-discharge cycles of the battery: For example, if the deep charge-discharge cycle of the battery is 2,000 times, and charging is carried out when 20% remains, the predicted actual effective charge-discharge cycle is 2,000 / 20% = 10,000 times.

[0105] Introduction to the theoretical mileage, theoretical service time, mileage already used, and time already used of the motor, reducer, and wheels: The theoretical mileage and theoretical service time are data obtained by the parts manufacturer through life tests, and the mileage already used and time already used are equipment usage data recorded by this system. The above is the current existing solution. In order to make the life as close to the actual situation as possible, this patent is introduced. By using an AI learning robot to determine the service life under actual different working conditions, a more accurate prediction can be obtained.

[0106] The purpose of data collection of the emergency stop times of the equipment is to facilitate this system to make corresponding adjustments according to the working condition coefficient of the reducer. The corresponding relationship between the emergency stop times of the equipment and the working condition coefficient of the reducer belongs to the prior art and can be obtained through several experiments in the laboratory, which is not the main inventive point of this invention. Example: For an Agv with a large number of emergency stops, the system also makes corresponding adjustments to the working condition coefficient of its reducer, such as changing the safety factor from 1.8 to 1.2; this is different from the laboratory working conditions because the emergency stop conditions are also recorded according to whether it is fully loaded or not and whether it is running at high speed, and the set conditions are not fixed.

[0107] The prediction result of the equipment life is obtained through the following method:

[0108] (1) Model objective

[0109] Establish a life prediction model. By monitoring the operating conditions, conveying frequency, usage frequency of the AGV, and the performance parameters of the motor, the life prediction model predicts the remaining life of the mobile robot and the docking equipment.

[0110] The model, through a data-driven approach, combines machine learning and deep learning methods to provide equipment status assessment and life prediction.

[0111] The data types used for driving are as follows:

[0112] Load status data: Whether it runs empty or fully loaded will affect the stress state of the motor;

[0113] Operating time: The continuous working time will affect the fatigue degree of the motor;

[0114] Usage frequency: The frequency of motor startup, stop, acceleration, and deceleration;

[0115] Number of forward and reverse rotations: The mechanical stress and wear caused by motor commutation;

[0116] Temperature: The increase in temperature will accelerate the aging of the insulation layer and the loss of materials.

[0117] Vibration and noise data: The changes in vibration and noise can be used as an omen of mechanical failures.

[0118] (2) Data description and processing

[0119] (2.1) Data sources

[0120] Operating conditions: including the running speed, acceleration, ambient temperature, working load, etc. of the AGV;

[0121] Empty / full load conveying frequency: Record the number and frequency of times the AGV is in the empty or full load state;

[0122] Usage frequency: The running time (daily / weekly working hours), standby time, etc. of the AGV;

[0123] Motor performance parameters: Key data on the working state of the motor such as current, voltage, power, temperature, speed, etc.;

[0124] Among them, the key input variables involved in the model need to be predefined:

[0125] Operating condition data: including external environmental variables such as ambient temperature, humidity, ground conditions, etc., denoted as vector C(t);

[0126] Empty / full load frequency: The frequencies of the AGV transporting in the empty and full load states, denoted as f empty (t) and f full (t) respectively;

[0127] Usage frequency: including the working duration and number of tasks completed by the AGV, denoted as u(t);.

[0128] Motor performance data: including the temperature, speed, current, voltage, etc. of the motor, denoted as M(t);

[0129] Historical maintenance data: including historical data such as the repair records of the equipment and the replacement of components, denoted as H(t).

[0130] (2.2) Data preprocessing

[0131] Data cleaning: Deal with missing values, noise, and outliers.

[0132] Normalization / standardization: Standardize each parameter for input into the machine learning model.

[0133] Time series processing: Treat the operating condition data and motor performance parameters of the AGV as time series data.

[0134] (3) Mathematical Modeling for Life Prediction (can be carried out by one of the following methods)

[0135] (3.1) Health State Modeling

[0136] Model the health state h(t) of the AGV and its related equipment as a function that changes over time, with its state value ranging from 1 to 0, where 1 indicates that the equipment is in the best state and 0 indicates that the equipment's life has ended.

[0137] Define the health state function:

[0138] h(t) = f(C(t), f empty (t), f full (t), u(t), M(t), H(t));

[0139] Among them, the function f(·) is established through the analysis of historical data and trained and predicted through an AI model.

[0140] (3.2) The First Type of Degradation Model

[0141] The degradation process of the equipment is usually non - linear. A degradation model based on exponential decay can be used to describe the change of the health state h(t). Assume that the equipment is in the best state at time t0, then its degradation model is:

[0142]

[0143] Among them, λ(·) is the degradation rate function, which reflects the influence of various factors on the equipment's health state. The degradation rate may fluctuate over time, so it needs to be dynamically adjusted through machine learning algorithms.

[0144] (3.2) The Second Type of Degradation Model

[0145] The physical degradation model describes the degradation behavior of the equipment based on physical processes. For example, assume that the equipment degradation process can be represented by an accumulated damage model:

[0146]

[0147] Among them,

[0148] D(t) is the degree of degradation of the equipment at time t;

[0149] D0 is the initial degradation level;

[0150] r(X(τ)) is the degradation rate under the working parameter X(τ);

[0151] According to a large amount of measured degradation data (such as motor temperature, vibration, etc.), a degradation curve model can be established. When the degree of degradation D(t) reaches a certain threshold Dth When this occurs, the device is considered to have failed, and the life T can be obtained by solving the equation D(T) = D th as follows.

[0152] (3.3) Cox proportional hazards model

[0153] The Cox model is a regression model widely used in survival analysis. It can predict the failure time of a device through multiple variables and can handle right-censored data (i.e., some devices may stop working or be replaced before failure). The expression of this model is:

[0154]

[0155] where

[0156] h(t|X) is the hazard function of the device at time t;

[0157] h0(t) is the baseline hazard function;

[0158] X1, X2,..., X n are variables such as the load of the motor, running time, number of forward and reverse rotations, etc. that affect the life;

[0159] β1, β2,..., β n are the regression coefficients corresponding to each variable.

[0160] (3.4) The first type of life prediction based on statistical regression:

[0161] The multiple regression model is a classic statistical method used to predict the relationship between a target variable (such as life) and multiple independent variables (such as working conditions, frequencies, loads, usage frequencies, motor performance, etc.). Using multiple regression, the influence weights of each variable on life can be obtained by fitting a large number of data samples, and this model can be directly used for the preliminary analysis of life prediction.

[0162] The basic form of this model is:

[0163] Using a regression model to predict life, a regression equation is established by fitting the relationship between the input features and the device life.

[0164] Assume that y represents the remaining service life, and the input features are X = [x1, x2,..., x n , including working conditions, load, usage frequency, and motor performance data, etc. A life prediction equation is established using a linear or non-linear regression model:

[0165]

[0166] w0, w1,..., w nis the regression coefficient, and ∈ is the error term.

[0167] (3.5) The second type of life prediction based on statistical regression:

[0168] Using the historical operation data in big data, the life of the device can be predicted through a statistical regression model. For example, a multiple linear regression is used to establish a linear relationship between the input features and the remaining life. Assuming that there is an approximate linear relationship between the remaining useful life (RUL) of the device and the feature vector X, it mainly fits the training data by the least squares method to obtain w and b, so as to predict the life of the device:

[0169] RUL = w T X + b + ∈

[0170] where,

[0171] X = [X1, X2,..., X n is the input feature, including operating conditions, load information, motor performance, etc.;

[0172] w = [w1, w2,..., w n is the weight vector;

[0173] b is the bias term;

[0174] ∈ is the random error, usually assumed to follow a normal distribution.

[0175] (3.5) The third type of life prediction based on statistical regression:

[0176] Regression analysis can be used to analyze the relationship between the motor life and multiple influencing factors. It mainly collects a large amount of operation data and uses the least squares method or the gradient descent method to estimate the regression coefficients a1, a2, a3, a4, so as to predict the motor life. The linear regression model is as follows

[0177] L = α0 + α1·t load + α2·t unload + α3·f usag + α4·N reversals + ∈

[0178] where,

[0179] L is the motor life;

[0180] t load is the full-load operation time;

[0181] t unload is the no-load operation time;

[0182] f usage is the usage frequency (such as the number of starts and stops per hour)

[0183] N reversals is the number of forward and reverse rotations of the motor;

[0184] ∈ is the error term.

[0185] (3.6) Life prediction based on machine learning

[0186] Regression algorithms in machine learning, such as random forest regression, support vector machine regression, XGBoost, etc., can also be used for life prediction. This type of model does not require assuming the distribution form of the data and can automatically learn the impact of different operating conditions and features on the equipment life from complex multi-dimensional data. Example:

[0187] (3.6.1) Random forest

[0188] Random forest is an ensemble learning method and a machine learning algorithm based on the ensemble of decision trees, suitable for dealing with non-linear relationships (suitable for dealing with diverse and complex input features). Through the feature vectors (such as motor temperature, usage frequency, etc.) in historical data and the corresponding life labels, a random forest model is trained to predict the life of the equipment (it can fit the relationship between the input features X (operating conditions, motor performance, etc.) and the remaining useful life RUL).

[0189] The random forest model consists of multiple decision trees, and the final prediction result is obtained by voting or averaging the results of all decision trees.

[0190] Training stage: Learn multiple decision trees from historical data, and each tree learns different patterns in the feature space

[0191] Prediction stage: Input new feature data, and average the prediction results of multiple trees to obtain the final predicted value of the remaining useful life.

[0192] The formula of random forest can be expressed as:

[0193]

[0194] where M is the number of decision trees, and T m (X) is the prediction result of the m-th tree.

[0195] (3.7) Other mathematical models:

[0196] Markov model

[0197] Model the lives of the AGV and the motor as a Markov process, and each state represents a different health state. For example, the AGV may experience state transitions from "normal" to "mild wear", "moderate wear", and then to "severe wear". The transition probability of each state can be obtained by statistically analyzing historical data, and a Markov chain model is constructed for life prediction.

[0198] Assume that the system has N states and gradually degrades from a healthy state to a failure state. The state transition matrix of the Markov model can be expressed as:

[0199]

[0200] Among them, p ij represents the probability of transitioning from state i to state j. By analyzing historical operation data, the probability of state transition can be obtained, and based on this, the future degradation path and remaining life can be estimated.

[0201] (4) Model evaluation and verification

[0202] Evaluation metrics: To evaluate the accuracy of the model, metrics such as mean squared error (MSE) and mean absolute error (MAE) can be used.

[0203]

[0204] Among them, and are the true life and predicted life respectively.

[0205] Cross-validation: The model is verified through cross-validation techniques to ensure its good generalization performance.

[0206] (5) Model optimization

[0207] Feature selection: Use feature selection algorithms (such as recursive feature elimination, RFE) to find the features that have the greatest impact on life, in order to improve the accuracy and interpretability of the model.

[0208] Model integration: By integrating multiple models (such as combining random forest and LSTM), improve the accuracy and stability of prediction.

[0209] In addition, the prediction results of equipment life can also be obtained through the data collected by noise detection and vibration detection.

[0210] (a) Application process of noise detection: Use noise detection sensors to collect data; after the platform issues a spot inspection task, when the equipment is in a specific scenario, the noise detection sensors collect noise data (the detection range of noise is 30 - 140 DB, and the collection frequency is 30 Hz - 20 KHz); by analyzing the noise data, the wear condition of the mechanical transmission of the equipment can be known; when the mechanical transmission is normal, the noise fluctuates within the noise threshold range, but when there is severe wear and looseness abnormalities, abnormal noise will appear.

[0211] Noise detection process:

[0212] (1) Data collection:

[0213] (1.1) Collect time - series audio signals at a specified location or scenario.

[0214] For example:

[0215] The noise signal can be expressed as a time series: x(t), where x(t) is the noise signal that varies with time t.

[0216] Since the sampling frequency is 30Hz - 20kHz, this continuous signal can be discretely sampled. Assuming the sampling frequency is f s (such as 20kHz), the time interval of each sampling point is

[0217] After discretization, the noise signal is expressed as: nx(nΔt), n = 0, 1, 2,... This discrete signal contains the noise data of the robot.

[0218] x[n]=x(nΔt),

[0219] n = 0, 1, 2,... This discrete signal contains the noise data of the robot.

[0220] (1.2) Define the label of the normal working state.

[0221] (1.3) Define the label of the fault state (mark the types of faults, including mechanical transmission faults, bolt loosening, etc.)

[0222] (2) Data pre - processing:

[0223] Denoising: Extract features in the frequency domain through Fourier transform (FFT, wavelet transform, etc.). The features include frequency distribution, energy spectrum, volume, etc.

[0224] (3) Feature selection:

[0225] According to the extracted frequency and time - domain features, select the features that can effectively distinguish between normal and fault states.

[0226] For example:

[0227] The average energy of the noise;

[0228] The peak value in a specific frequency range (if a fault generates a certain specific frequency of noise);

[0229] The amplitude change of the noise signal;

[0230] The periodic change of the noise.

[0231] (4) Model selection:

[0232] For different situations, different mathematical models can be selected for fault detection:

[0233] Threshold detection: Determine whether a failure has occurred by setting the noise amplitude or energy threshold within a specific frequency range. For example, if the noise energy within a specific frequency band exceeds a certain value, it is considered that a failure has occurred.

[0234] Statistical model: Establish a statistical detection model by analyzing the probability distribution of noise. For example, a Gaussian mixture model can be used to describe the noise distribution under normal conditions, and then the degree of deviation from the normal distribution can be calculated based on new data.

[0235] Machine learning model: Through supervised learning, use labeled data in normal and faulty states to train a classifier, such as a support vector machine, KNN, or neural network. For example, for a support vector machine, the following mathematical model can be constructed:

[0236] Input: Feature vector of noise where n is the dimension of the feature;

[0237] Output: Label y ∈ {0, 1}, where 0 represents normal and 1 represents faulty;

[0238] The goal of the support vector machine model is to find a hyperplane w·x + b = 0 such that: y = sign(w·x + b);

[0239] The model obtains the parameters w and b through training to distinguish between normal and faulty states.

[0240] (b) Application process of vibration detection: Use a three-axis vibration detection sensor to collect data; after the platform issues a task, when the device is in a specific scenario, the sensor collects vibration data (the detection range of vibration is 0 - 98 m / s²: the collection frequency is greater than or equal to 2 KHz / s); by analyzing the vibration data, it is possible to know the abnormal mechanical vibration of the device during operation; during normal use of the device, the vibration frequency of its vehicle body or components fluctuates within the vibration threshold range, and a vibration curve trend can be established based on the vibration threshold. When the vibration value falls within the abnormal range of the vibration curve trend, it indicates that there is a problem with the device (it can be used to warn of abnormal mechanical vibration).

[0241] The process of vibration detection is specifically as follows:

[0242] (1) Model goal: Establish a mathematical model based on vibration data, and detect and warn of abnormal mechanical vibration by analyzing the vibration characteristics of the AGV. This model should identify abnormal vibrations and provide warning information based on the amplitude, frequency, and trend of vibration over time.

[0243] (2) Modeling of vibration data:

[0244] The vibration signal can be expressed as an acceleration signal varying with time, with the unit of m / s2. Assuming the vibration signal is a time series a(t), we discretize it through sampling to obtain a discrete time series:

[0245] a[n] = a(nΔt), n = 0, 1, 2,...

[0246] where Δt is the sampling time interval and the sampling frequency is f s ≥ 2kHz, and the vibration range is from 0 to 98 m / s2.

[0247] (3) Feature extraction

[0248] Before performing vibration anomaly detection, we need to extract relevant features from the time series data. Features can be extracted from two dimensions: the time domain and the frequency domain.

[0249] (3.1) Time domain features:

[0250] Mean: Used to evaluate the overall level of vibration.

[0251]

[0252] Variance: Reflects the volatility of vibration data. Abnormal vibration may lead to an increase in variance.

[0253]

[0254] Peak value: The maximum value of the vibration signal, representing the intensity of vibration

[0255] Root mean square value (RMS): Commonly used to represent the energy of a signal. The larger the RMS value, the greater the vibration intensity.

[0256]

[0257] Shock factor: Represents the relationship between the maximum value and the RMS of the signal, used to evaluate the severity of vibration.

[0258]

[0259] (3.2) Frequency domain features:

[0260] By performing Fourier transform on the vibration signal, we can extract frequency domain features. Frequency domain features can help identify specific frequency components that may be related to abnormal mechanical vibration.

[0261] Main frequency: The main frequency component, which may be related to the vibration of specific mechanical components.

[0262] Frequency peak: The frequency corresponding to the maximum amplitude in the frequency domain.

[0263] Spectrum energy distribution: The vibration energy within a specific frequency range, used to evaluate the energy distribution generated by abnormal vibrations.

[0264] (4) Vibration anomaly detection model:

[0265] (4.1) Detection method

[0266] The threshold detection method is the simplest and most effective vibration anomaly detection method. Set the upper and lower threshold values of the vibration signal. Once this threshold is exceeded, it is considered that vibration anomaly has occurred.

[0267] Assume the threshold is T. When the vibration signal a[n] satisfies the following conditions, an anomaly warning is triggered:

[0268] a[n] ≥ T apper or a[n] < T lower

[0269] where, T upper and T lower are the upper threshold value and the lower threshold value respectively, which can be set according to the actual usage scenario of the device and empirical data.

[0270] (4.2) Curve trend analysis

[0271] In addition to the single-point threshold detection, anomaly detection can also be performed by analyzing the change trend of the vibration signal over time. For example, a continuous increase in the vibration amplitude or a gradual increase in certain frequency components may indicate wear or looseness of mechanical components.

[0272] Use the sliding window method to calculate the trend Trend of the vibration signal, such as calculating the change of the mean value or RMS over time:

[0273]

[0274] where, W is the width of the sliding window and k is the current time point. If the trend value of the vibration signal continuously exceeds a certain set threshold, it is considered that an anomaly has occurred.

[0275] (4.3) Machine learning method

[0276] To improve the detection accuracy, a machine learning model can be used to classify vibration anomalies. Common models include random forest, support vector machine, neural network, etc.

[0277] Input features: Feature vectors X in the time domain and frequency domain;

[0278] Output: Classification result y, 0 indicates normal and 1 indicates abnormal.

[0279] For example, use a support vector machine classifier:

[0280] y = sign(w·X + b)

[0281] The model is trained to learn the characteristic patterns of normal and abnormal vibrations and apply them to the classification of new data.

[0282] (4.4) Adaptive method for anomaly detection

[0283] To adapt to vibration changes under different operating conditions, an adaptive threshold detection or an adaptive model can be used. For example, the vibration signal is predicted and updated by a Kalman filter, and the threshold is dynamically adjusted according to historical data.

[0284] Furthermore, after the central management system issues a predictive equipment maintenance request to an external device, it monitors whether the request is executed;

[0285] If the request is not executed, the central management system adjusts the usage of the corresponding AGV and the line-side docking device;

[0286] For the AGV, the central management system adjusts the moving speed, load capacity, running path, and working frequency of the AGV;

[0287] For the line-side docking device, the central management system adjusts the load capacity and working frequency of the line-side docking device.

[0288] In this embodiment, the central management system adjusts the moving speed, load capacity, running path, and working frequency of the AGV, such as reducing the moving speed of the AGV, reducing the load during transportation of the AGV (performing transportation tasks with lighter loads), shortening the running path of the AGV, and reducing the working frequency of the AGV. The central management system adjusts the load capacity and working frequency of the line-side docking device, such as reducing the workload of the line-side docking device (performing docking tasks with lighter loads) and reducing the working frequency of the line-side docking device.

[0289] Since the predictive equipment maintenance request is not executed, if the AGV and the line-side docking device continue to operate in their current states, accidents may occur, affecting the normal operation of the scheduling system and incurring greater problem handling costs. Therefore, by adjusting (reducing) the working efficiency and workload of the AGV and the line-side docking device (such as setting that AGVs with a high degree of battery attenuation and approaching the end of their battery life will undertake some tasks with easy and short distances, while AGVs with insufficient motor life will be arranged for some occasions where they can maintain a constant speed and travel a long distance), the service life of the AGV and the line-side docking device is extended until the predictive equipment maintenance request is executed, enabling the service life of this system to continue.

[0290] Furthermore, the task event includes external environment analysis;

[0291] External environment analysis means that the AI model predicts the external environment according to the obtained environmental condition parameters and the working conditions of the AGV;

[0292] After the AI framework module outputs the inference result, the central management system obtains the external environment analysis result according to the inference result, makes corresponding adjustments to the use of the AGV, and informs the superior equipment for manual processing;

[0293] If the external environment analysis result shows that the air humidity exceeds the preset value, the central management system reduces the rated handling load of the AGV and increases the number of handling times;

[0294] If the external environment analysis result shows that there is water accumulation in a specific area of the working ground, the central management system reduces the rated handling load of the AGV, increases the number of handling times and reduces the speed, and at the same time requests the external equipment for work site maintenance.

[0295] If the external environment analysis result shows that a specific area of the working ground is damaged, the central management system controls the AGV to reduce its speed when driving to this area, and at the same time requests the external equipment for work site maintenance.

[0296] The beneficial effects of the above external environment analysis results are as follows:

[0297] When the air humidity exceeds the preset value, it means that the AGV is in a humid environment. When working in a humid environment (humid ground), the AGV will experience slipping and inaccurate stopping, so that the AI model can accurately predict; by reducing the rated handling load of the AGV, increasing the number of handling times, and reducing the speed, it can avoid slipping during turning or sudden stop, and can continue to perform the transportation task without affecting the production rhythm of the production line.

[0298] When there is water accumulation in a specific area of the working ground, it will cause the AGV to slip and stop inaccurately, and cause obvious fluctuations in the walking speed (relative moving position on the map), so that the AI model can accurately predict; by reducing the rated handling load of the AGV, increasing the number of handling times, and performing work site maintenance, this problem can be solved while ensuring the normal progress of the production rhythm of the production line.

[0299] When a specific area of the working ground is damaged, the AGV frequently shows large fluctuations in the current of the walking motor at a certain position on the same section of the road, so that the AI model can predict the damage of the ground in this area; by controlling the AGV to reduce its speed when driving to this area, it can avoid the situation of material dropping and vehicle body damage caused by the AGV passing through the damaged bottom surface at a high speed.

[0300] Furthermore, the task event includes supply and demand analysis;

[0301] Supply and demand analysis means that the AI model infers the material handling requirements based on the production plan schedule, the availability of production materials, the priority level of product shipments, and the capacity status of the warehouse;

[0302] After the AI framework module outputs the inference result, the central management system obtains the supply and demand analysis result based on the inference result and sets the temporary production scheduling plan;

[0303] If inferring the material handling requirements based on the production plan schedule, the deep learning library analyzes the time periods of peak production and off-peak production according to the production plan schedule; when the central management system sets the temporary production scheduling plan, during peak production, idle AGVs are arranged in advance to wait around the specific workstations before the operation of the specific workstations, and during off-peak production, the idle or under-loaded AGV capacity is utilized to pick up more handling tasks during the movement for material distribution;

[0304] If inferring the material handling requirements based on the availability of production materials, the deep learning library analyzes the shortage materials according to the availability of production materials; when the central management system sets the temporary production scheduling plan, it sends a message to external devices to request the priority production or replenishment of shortage materials;

[0305] If inferring the material handling requirements based on the priority level of product shipments, the deep learning library analyzes the level of the goods that need to be shipped out first and the path of the transportation route of the goods according to the priority level of product shipments; when the central management system sets the temporary production scheduling plan, it delimits the important transportation channel area according to the transportation route path of the goods to be shipped out first, classifies the AGVs according to the level of the goods to be shipped out first for the distribution task level, restricts the AGVs with a lower distribution task level from entering the important transportation channel area, and enables the AGVs with a higher distribution task level to enter the important transportation channel area for movement;

[0306] If inferring the material handling requirements based on the capacity status of the warehouse, the deep learning library analyzes the utilization rate of several storage positions according to the capacity status of the warehouse; when the central management system sets the temporary production scheduling plan, it determines whether the inventory quantity of material A exceeds the preset capacity value of the storage position X0 where material A is currently arranged for placement. If so, the central management system locates the storage position Xn with a relatively low utilization rate and allowed placement around the storage position X0 where material A is currently arranged for placement, and the central management system controls the AGV to temporarily place the material A that exceeds the capacity in the storage position Xn.

[0307] Supply and demand analysis coordinates the AGV transport capacity, warehouse capacity, production plan, material inventory, and product turnover rate, thus fully mobilizing the operation of all resources within the scheduling system, greatly improving the efficiency of the scheduling system, and enhancing transport capacity and productivity. Additionally, based on the transport route of the goods to be shipped out first, important transport channel areas are demarcated. The AGVs are classified according to the distribution task levels of the goods to be shipped out first. AGVs with lower distribution task levels are restricted from entering these important transport channel areas, while AGVs with higher distribution task levels are allowed to enter and move within these important transport channel areas, thereby increasing the moving speed of AGVs with heavy tasks and reducing congestion in important transport channel areas.

[0308] Furthermore, the task event includes unique identification;

[0309] Unique identification means that the AI model performs unique identification of handling tasks, unique identification of items, and unique identification of personnel based on the obtained computer vision data, the characteristic data of the medium being carried identified by the AGV, and the surrounding environment data identified by the AGV, and infers the analysis result of the AGV's working conditions;

[0310] After the AI framework module outputs the inference result, the central management system determines the working status of the AGV and the rationality of the handling task based on the analysis result of the AGV's working conditions. If the working status of the AGV is abnormal or the rationality of the handling task is incorrect, feedback such as blocking, warning, and tracking is given to the specific AGV;

[0311] The process of unique identification of handling tasks is as follows: The AI model identifies the type of equipment docked with a specific AGV based on the obtained computer vision data, the characteristic data of the medium being carried identified by the AGV, and the surrounding environment data identified by the AGV. If the type of the docked equipment belongs to the equipment allowed to be docked by the specific AGV, then the docked equipment passes the unique identification;

[0312] The process of unique identification of items is as follows: The AI model performs one of the following identifications based on the obtained computer vision data, the characteristic data of the medium being carried identified by the AGV, and the surrounding environment data identified by the AGV:

[0313] i. Identify whether the motion state of the AGV loading a specific item is abnormal compared to its normal motion state when loading the item. If it is determined that the motion state of the AGV is abnormal, then the specific item loaded by the AGV does not pass the unique identification;

[0314] ii. Identify the integrity of the item to be loaded. If it is determined that the integrity of the item to be loaded is abnormal, then the specific item loaded by the AGV does not pass the unique identification;

[0315] The process of uniquely identifying a person is as follows: the AI model identifies whether human intervention has occurred in a specific AGV based on the acquired computer vision data, the characteristic data of the medium being transported identified by the AGV, and the surrounding environment data identified by the AGV. If human intervention has occurred, the physical characteristics of the person who illegally controls or interferes with the normal operation of the AGV are identified, and the person who meets the physical characteristics is identified to obtain the unique data of the person.

[0316] The unique identification of the handling task can ensure that the AGV only docks with the corresponding docking equipment (for example, the AGV dives into the bottom of the material cart and identifies its uniqueness by identifying the mark on the material cart. The AGV in the clean area only docks with the material cart in the clean area, and will not tow and transport the material cart in the non-clean area), avoiding accidents or material contamination caused by the AGV docking with equipment that is not allowed to be docked by the AGV;

[0317] The unique identification of items can ensure that AGV only transports intact materials, avoids incomplete materials from flowing into the production line and affecting the normal operation of the production line, and reduces the maintenance cost required to locate incomplete materials flowing into the production line (previously, items were identified by their RFID tags without considering the integrity of the materials. The method of judging the integrity of materials in the present invention is as follows: a. When an AGV in a conventional production line that is transporting materials back and forth transports one of the boxes of materials, the AGV driving data is analyzed, and it is found that the driving current or torque is different from the conventional ones. The AI model will judge that the box of materials may be in an abnormal state; b. Through visual data comparison, it is found that the material box of the transported material is damaged, the label is missing, the bottle cap is loose, or the material box is deformed during the transportation process, or the material box has been opened. The AI model will judge that there is an abnormality in the box of materials, and the central management system will perform abnormal processing based on these abnormal information or feedback corresponding information to the upper system);

[0318] The unique identification of personnel is used to determine the identity of personnel who affect the normal execution of AGV tasks (by analyzing the physical characteristics of personnel, such as hands, walking posture, face, etc.), so as to facilitate subsequent adjustments to the personnel's work path to avoid accidents.

[0319] Furthermore, the mission events include emergency event handling;

[0320] Emergency event handling means that the AI model infers the type of emergency event based on the obtained tracking records without affecting the normal handling work of the AGV and without allowing the emergency event to affect production;

[0321] After the AI framework module outputs the inference results, the central management system obtains the emergency event processing strategy based on the inference results, and the central management system sends messages to external devices to execute the corresponding emergency event processing plan;

[0322] If the tracking record contains information about the damaged to-be-loaded materials, the deep learning library records the flow direction of the AGV that loads the to-be-loaded materials, analyzes that it is necessary to replace the damaged to-be-loaded materials before final assembly; when the central management system outputs the processing strategy for the emergency event, it dispatches another AGV to retrieve the damaged to-be-loaded materials and transport them, and dispatches another AGV to the corresponding parts area to take out new materials identical to the damaged to-be-loaded materials, and tracks the AGV that originally loaded the damaged to-be-loaded materials, and sends the new materials to it;

[0323] If the tracking record contains information about channel anomalies, the deep learning library records the channel areas where anomalies occur, plans a new route for the channel area that can avoid the anomalies and regenerates the driving map; when the central management system outputs the processing strategy for the emergency event, it controls the AGV to execute tasks using the new driving map; if the information about channel anomalies in the tracking record is deleted (the channel resumes use), the deep learning library regenerates the driving map, and the central management system controls the AGV to execute tasks using the new driving map;

[0324] If the tracking record contains newly inserted handling tasks, the deep learning library analyzes the speed-up amount and the additional handling amount of the pre-planned tasks; when the central management system outputs the processing strategy for the emergency event, it controls the AGV that executes the pre-planned tasks to perform handling with increased speed or quantity.

[0325] The emergency event handling of the present invention includes the emergency handling of damaged to-be-loaded materials, the emergency handling of channel anomalies, and the emergency handling of newly inserted handling tasks. By handling the above emergency events, it is possible to prevent damaged materials from flowing into the production line, abnormal channels from affecting the transportation network, at the same time prevent AGVs from entering abnormal channels and aggravating the abnormal situation of abnormal channels, and prevent newly inserted handling tasks from disrupting the original production plan.

[0326] Example of handling damaged to-be-loaded materials: For example, for a car on the assembly line, all to-be-loaded materials are placed on the material cart accompanying the car. When it is found and notified by the outside that one of the to-be-assembled parts is damaged, record the flow direction of the car, dispatch an AGV to the material cart to retrieve the damaged materials, go to the corresponding parts area to pick up materials, and be responsible for tracking and sending back the new materials.

[0327] Furthermore, the staff can interact with the central management system through the client module or the AGV, or the staff can directly interact with the central management system. The AI framework module extracts the content of the newly added transportation tasks in the interaction content, and obtains a number of real-time quantitative data based on the data of the newly added transportation tasks content;

[0328] The deep learning library outputs an inference result based on the real-time quantitative data;

[0329] The central management system obtains a temporary scheduling plan based on the inference result. The execution process of the temporary scheduling plan includes at least one of the following situations:

[0330] i. The central management system assigns an AGV to temporarily execute a specific task;

[0331] ii. The central management system changes the material delivery path or destination of the AGV;

[0332] iii. The central management system adjusts the material delivery speed of the AGV that is executing a transportation task to achieve early or late delivery of the material.

[0333] The setting of the client module enables connection to the central management system, realizes the functions of remotely controlling and supervising the AGV or calling the AGV, and can receive the inference analysis results output by the AI model for the staff to view; the AI model of the present invention is a generative AI, which can generate relevant inference results according to the temporarily added transportation tasks of the staff, so that the central management system outputs the optimal temporary scheduling plan externally. While ensuring the normal progress of the original production plan, it can also complete the execution of the temporary scheduling plan.

[0334] Furthermore, the central management system includes an operation module, a control module, a data storage module, a data input / output module, a network service module, a map drawing module, an operation module, an AGV control and communication module, a traffic control module, a data recording module, a task management module, a status monitoring module, and an AGV scheduling interface module. The AGV scheduling interface module is used to enable the central management system to support several database formats.

[0335] In this embodiment, the network service module is a network server.

[0336] The setting of the AGV scheduling interface module facilitates data interaction with the logistics management machine.

[0337] The training sample expressions, embedding matrices, etc. used by the AI model for inference based on quantitative data belong to the prior art.

[0338] According to the disclosure and teachings of the above specification, those skilled in the art to which the present invention pertains can also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above. Some modifications and changes to the present invention should also fall within the protection scope of the claims of the present invention. In addition, although some specific terms are used in this specification, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A method for operating a scheduling system based on an AI model, characterized in that, Including: Several AGVs; Several in-line docking devices; Central management system; AI framework module, communicatively connected to the central management system, with a deep learning library deployed thereon, the deep learning library being used to create an AI model and for the AI model to perform deep learning training; The AI framework module monitors whether there is data input to the central management system. If so, the data is screened to obtain learning data, and the learning data is input into the deep learning library; The operation method includes: The central management system receives real-time feedback signals from the AGVs and in-line docking devices, and at the same time determines task events to be processed, and accesses the AI framework module; The AI framework module calls the deep learning library; The AI framework module selects idle computing power hardware resources of the central management system; The deep learning library, according to the request of the central management system, obtains content data associated with the current task event from the real-time feedback signal data of the AGVs and in-line docking devices received by the central management system, and statistically obtains several real-time quantization data corresponding to the task event according to the content data; The real-time quantization data and the type of the task event are input into the trained AI model, and the deep learning library outputs an inference result to the central management system through the AI framework module; The central management system sends information to the AGVs and / or in-line docking devices according to the inference result to control the operation of the AGVs and / or in-line docking devices, and feedback relevant guiding information.

2. The operation method of the scheduling system based on the AI model according to claim 1, characterized in that, The task event includes equipment life prediction; Equipment life prediction means that the AI model predicts the life of the AGVs and in-line docking devices according to the operating conditions of the equipment, the no-load and full-load frequencies, the usage frequency, the motor performance parameters, and the number of emergency stops under rated conditions; After the AI framework module outputs the inference result, the central management system obtains the equipment life prediction result according to the inference result, uses a classification model to classify the health status of the equipment, and makes a predictive equipment maintenance request to an external device.

3. The method for operating a scheduling system based on an AI model according to claim 2, wherein Among them, the content data collected for the AGVs includes the theoretical number of charge and discharge cycles and the actual number of charge and discharge cycles of the battery, the theoretical mileage, theoretical usage time, mileage already used, and time already used of the motor, reducer, and wheels, and the number of equipment emergency stops; Among them, the content data collected for the in-line docking devices includes the actual load and the frequency of forward and reverse operation of the motor; Calculate to obtain the equipment life prediction result.

4. The method for operating a scheduling system based on an AI model according to claim 2, wherein, After the central management system makes a predictive equipment maintenance request to an external device, it monitors whether the request is executed; If the request is not executed, the central management system adjusts the usage of the corresponding AGVs and in-line docking devices; For the AGVs, the central management system adjusts the moving speed, load capacity, running path, and working frequency of the AGVs; For the in-line docking devices, the central management system adjusts the load capacity and working frequency of the in-line docking devices.

5. The operating method of the scheduling system based on the AI model according to claim 1, characterized in that, The task event includes external environment analysis; External environment analysis means that the AI model predicts the external environment situation according to the obtained environmental condition parameters and the working conditions of the AGVs; After the AI framework module outputs the inference result, the central management system obtains the external environment analysis result according to the inference result and makes corresponding adjustments to the use of the AGVs. If the external environment analysis results show that the air humidity exceeds the preset value, the central management system will reduce the rated handling load of the AGV and increase the number of handling times; If the external environment analysis results show that there is water accumulation in a specific area of the work surface, the central management system will reduce the rated handling load of the AGV, increase the number of handling times and reduce the speed, and at the same time make a workplace maintenance request to the external equipment. If the external environment analysis results show that a specific area of the work surface is damaged, the central management system controls the AGV to slow down when it drives to that area, and at the same time makes a request for workplace maintenance to the external equipment.

6. The method for operating a scheduling system based on an AI model according to claim 1, wherein The task events include supply and demand analysis; Supply and demand analysis refers to the AI model predicting material handling demand based on production schedule, material availability, product shipment priority, and warehouse capacity status; After the AI framework module outputs the inference results, the central management system obtains the supply and demand analysis results based on the inference results and sets a temporary production scheduling plan; If the material handling demand is estimated based on the production schedule, the deep learning library will analyze the time intervals of busy and idle production according to the production schedule. When the central management system sets up a temporary production scheduling plan, when production is busy, idle AGVs are arranged around the specific workstations in advance to wait before the operation of the workstations. When production is idle, the idle or partially loaded AGV capacity is used to take on more handling tasks during the movement to carry out material distribution. If the material handling demand is inferred based on the stock situation of production materials, the deep learning library will analyze the short-board materials based on the stock situation of production materials; when the central management system sets a temporary production scheduling plan, it will send a message to the external equipment to request priority production or replenishment of short-board materials; If the material handling demand is inferred based on the product shipment priority level, the deep learning library will analyze the level of the goods that need to be shipped out first and the path of the goods' transportation route based on the product shipment priority level; when the central management system sets up a temporary production scheduling plan, it will delineate important transportation channel areas based on the cargo transportation path of the goods that need to be shipped out first, and classify the AGV's distribution task level according to the level of the goods that need to be shipped out first, restricting AGVs with lower distribution task levels from entering the important transportation channel area, and allowing AGVs with higher distribution task levels to enter and move in the important transportation channel area; If the material handling demand is inferred based on the capacity status of the warehouse, the deep learning library will analyze the utilization rates of several storage locations based on the capacity status of the warehouse. When the central management system sets up a temporary production scheduling plan, it will determine whether the archive volume of material A exceeds the preset capacity value of the storage location X0 where material A is currently placed. If so, the central management system will locate the storage location Xn where material A is currently placed and has a relatively low utilization rate around the storage location X0 where material A is currently placed, and the central management system will control the AGV to temporarily place the material A that exceeds the capacity to the storage location Xn.

7. The operating method of the scheduling system based on the AI model according to claim 1, wherein The task event includes a unique identification; Unique identification means that the AI model conducts unique identification of handling tasks, unique identification of items, and unique identification of personnel based on the obtained computer vision data, the characteristic data of the media being carried identified by the AGV, and the surrounding environment data identified by the AGV, and infers the analysis result of the AGV working conditions; After the AI framework module outputs the inference result, the central management system judges the working status of the AGV and the rationality of the handling task according to the analysis result of the AGV working conditions. If the working status of the AGV is abnormal or the rationality of the handling task is incorrect, feedback such as blocking, warning, and tracking is given to the specific AGV; The process of unique identification of the handling task is as follows: The AI model identifies the type of the device docked with the specific AGV based on the obtained computer vision data, the characteristic data of the media being carried identified by the AGV, and the surrounding environment data identified by the AGV. If the type of the docked device belongs to the devices allowed to be docked by the specific AGV, then the docked device passes the unique identification; The process of unique identification of items is as follows: The AI model conducts one of the following identifications based on the obtained computer vision data, the characteristic data of the media being carried identified by the AGV, and the surrounding environment data identified by the AGV: i. Identify whether the motion state of the AGV loading the specific item is abnormal compared to its normal motion state when loading the item. If it is judged that the motion state of the AGV is abnormal, then the specific item loaded by the AGV does not pass the unique identification; ii. Identify the integrity of the item to be loaded. If it is judged that the integrity of the item to be loaded is abnormal, then the specific item loaded by the AGV does not pass the unique identification; The process of unique identification of personnel is as follows: The AI model identifies whether there is human intervention in the specific AGV based on the obtained computer vision data, the characteristic data of the media being carried identified by the AGV, and the surrounding environment data identified by the AGV. If there is human intervention, then identify the physical characteristics of the personnel who illegally control or interfere with the normal operation of the AGV, and identify the personnel who conform to the physical characteristics to obtain the unique data of the personnel; 8. The operation method of the scheduling system based on the AI model according to claim 1, characterized in that The task event includes emergency event handling; Emergency event handling means that the AI model infers the type of the emergency event based on the obtained tracking record without affecting the normal handling work of the AGV and without allowing the emergency event to affect production; After the AI framework module outputs the inference result, the central management system obtains the handling strategy for the emergency event according to the inference result, and the central management system sends a message to the external device to execute the corresponding emergency event handling plan; If the tracking record contains information about the damaged material to be loaded, the deep learning library records the flow direction of the AGV loading the material to be loaded, analyzes that it is necessary to replace the damaged material to be loaded before final assembly; when the central management system outputs the handling strategy for the emergency event, it dispatches another AGV to retrieve the damaged material to be loaded and transport it, and dispatches another AGV to take out a new material identical to the damaged material to be loaded from the corresponding accessory area, and tracks the original AGV loading the damaged material to be loaded and sends the new material to it; If the tracking record contains information about a channel anomaly, the deep learning library records the channel area where the anomaly occurred, plans a new route for the channel area that can avoid the anomaly, and regenerates the driving map; when the central management system outputs a processing strategy for an emergency event, it controls the AGV to perform tasks using the new driving map; if the information about the channel anomaly in the tracking record is deleted, the deep learning library regenerates the driving map, and the central management system controls the AGV to perform tasks using the new driving map; If the tracking record contains a newly inserted handling task, the deep learning library analyzes the speed increase amount and the additional handling amount of the pre-planned task; when the central management system outputs a processing strategy for an emergency event, it controls the AGV executing the pre-planned task to perform handling with increased speed or quantity.

9. The operation method of the scheduling system based on the AI model according to claim 1, characterized in that, It also includes a client module, which is communicatively connected to the central management system. Operating the client module can monitor the AGV or call the AGV; Alternatively, it also includes a client module, which is communicatively connected to the central management system. Operating the client module can monitor the AGV or call the AGV; the staff can interact with the central management system through the client module or the AGV, or the staff can directly interact with the central management system. The AI framework module extracts the new transportation task content in the interaction content and obtains a number of real-time quantitative data based on the new transportation task content data; The deep learning library outputs an inference result based on the real-time quantitative data; The central management system obtains a temporary scheduling plan based on the inference result. The execution process of the temporary scheduling plan includes at least one of the following situations: i. The central management system assigns the AGV to temporarily execute a specific task; ii. The central management system changes the material distribution path or destination of the AGV; iii. The central management system adjusts the material distribution speed of the AGV that is executing the transportation task to achieve early delivery or delayed delivery of the material.

10. The method for operating a scheduling system based on an AI model according to any one of claims 1 to 9, characterized in that, The central management system includes an operation module, a control module, a data storage module, a data input / output module, a network service module, a map drawing module, an operation module, an AGV control and communication module, a traffic management module, a data recording module, a task management module, a status monitoring module, and an AGV scheduling interface module, where the AGV scheduling interface module is used to enable the central management system to support several database formats.

Images