An AI model empowered sorting management method, system, product and medium

Abstract

An AI model enabled sorting management method, system, product and medium. The method comprises: obtaining initial information of a target object and a dynamic running speed of a conveying line at a preset position of the conveying line, calculating a reference prediction trigger time; performing sorting according to the time and obtaining a position deviation value of the execution result, continuously counting the deviation to form a real-time sorting hit rate within a period; when the hit rate is lower than a preset threshold, generating a running calibration parameter in combination with the hit rate, the position deviation and the running speed, and dynamically adjusting the prediction trigger time of the subsequent sorting action by using the parameter to perform closed-loop calibration. The technical solution provided by the application improves the accuracy of sorting positioning under the speed change of the conveying line.

Description

Technical Field

[0001] This application relates to the field of advanced control, and more particularly to an AI model-enabled sorting management method, system, product, and medium. Background Technology

[0002] Currently, with the development of intelligent manufacturing and modern logistics technology, automated sorting systems play a crucial role in production, warehousing and distribution, which can significantly improve material flow efficiency and reduce labor costs.

[0003] In related technologies, predictive control methods based on real-time feedback are employed to continuously improve the accuracy and operational efficiency of sorting systems. This method uses a vision sensor to capture the instantaneous position of the target object at the entrance, and combines this with an encoder mounted on the conveyor drive shaft to continuously acquire high-frequency displacement pulse signals. The motion controller performs real-time integration calculations based on the encoder's displacement signals, thereby tracking the dynamic speed changes of the conveyor line and continuously updating the prediction of the target object's position to trigger the actuator to complete the sorting at a preset calculation point.

[0004] However, the aforementioned predictive control based on real-time feedback is difficult to adapt effectively when the production environment changes or the equipment experiences performance degradation (e.g., the slippage rate between the belt and roller changes slightly due to wear or the actuator cylinder experiences sluggish response due to air pressure fluctuations and oil circuit aging). The control logic can only follow the physical speed of the conveyor line, which leads to a decrease in the accuracy of the sorting system. Summary of the Invention

[0005] This application provides an AI model-enabled sorting management method, system, product, and medium for improving the accuracy of sorting positioning under varying conveyor line speeds.

[0006] The first aspect of this application provides an AI model-enabled sorting management method, the method comprising: At a preset detection position on the conveyor line, the initial position of the moving target object, the corresponding initial timestamp, and the dynamic operating speed of the conveyor line are acquired. Based on the initial position, initial timestamp, dynamic operating speed, and preset execution position, the baseline predicted trigger time for the target object to reach the execution position is calculated. When the baseline predicted trigger time arrives, a sorting action instruction is generated and issued to the execution mechanism at the corresponding execution position, and the real-time execution result position deviation value is acquired. Within a preset statistical period, for multiple sorting actions that have been executed, the corresponding execution result position deviation value is continuously acquired, and the real-time sorting hit rate within the corresponding statistical period is calculated. When the real-time sorting hit rate is lower than the preset hit rate threshold, operation calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic operating speed. Combining the operation calibration parameters, the baseline predicted trigger time is adjusted to the actual predicted trigger time in subsequent sorting actions within the corresponding statistical period.

[0007] In the above embodiments, by continuously acquiring the positional deviation value of the execution result and monitoring the real-time sorting hit rate, a closed-loop feedback directly addresses the final result, overcoming the shortcomings of related technologies that rely solely on input signals and cannot detect systematic errors such as belt slippage and actuator delay. When the hit rate decreases, operational calibration parameters can be dynamically generated based on the statistical relationship between deviation and speed, quantifying the deviation between the ideal model and physical reality. Ultimately, this parameter is used to correct the predicted trigger time, achieving adaptive optimization of the control model and ensuring improved sorting positioning accuracy when the conveyor line speed changes.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, when the real-time sorting hit rate is lower than a preset hit rate threshold, operation calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic running speed, specifically including: When the real-time sorting hit rate is lower than the preset hit rate threshold, the real-time sorting hit rate, the real-time execution result position deviation value and the corresponding dynamic running speed are input into the preset decision model to generate running calibration parameters.

[0009] In the above embodiments, the neural network decision model, pre-trained with historical data, can learn and deeply understand the complex nonlinear coupling relationship between sorting hit rate, positional deviation, and conveyor line speed. Therefore, when the real-time hit rate decreases, the current operating state can be used as an overall feature input, and the decision model can output optimized operating calibration parameters based on its experience learned from massive amounts of historical operating conditions. These parameters dynamically adapt to the complex system errors caused by various factors such as equipment wear and environmental changes, thereby adjusting the predicted trigger time and ultimately improving the sorting positioning accuracy under changes in conveyor line speed.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, when the real-time sorting hit rate is lower than a preset hit rate threshold, operation calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic running speed, specifically including: Acquire the category information of the target object and the workstation information of the actuator corresponding to each sorting action; within the corresponding statistical period, when the real-time sorting hit rate is lower than the preset hit rate threshold, calculate the workstation hit rate of each independent workstation and the category hit rate of each independent category based on the workstation information and category information; among all workstation hit rates and category hit rates, determine the dimension with the lowest value as the core analysis object; extract all execution result position deviation values ​​corresponding to the core analysis object within the corresponding statistical period and calculate the average deviation value; extract all dynamic running speeds corresponding to the core analysis object within the corresponding statistical period and calculate the average running speed; calculate the ratio of the average deviation value to the average running speed as the running calibration parameter.

[0011] In the above embodiment, attribution analysis of sorting errors is performed by introducing workstation and product category dimensions. When the overall hit rate decreases, the root cause of the problem is located by identifying the dimension with the lowest value, whether it is a mechanical performance degradation at a specific workstation or a systematic deviation caused by the physical characteristics of a specific product category. Subsequently, the solution only performs statistical analysis on the data of this core analysis object to calculate highly targeted operational calibration parameters. This targeted calibration mechanism avoids blindly adjusting the entire system and can compensate for the main contradictions. Therefore, even when the conveyor line speed is constantly changing, sorting positioning accuracy can be improved.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, after determining the dimension with the lowest value among all workstation hit rates and category hit rates as the core analysis object, the method further includes: When the lowest value in the workstation hit rate is the same as or the absolute value of the lowest value in the category hit rate is less than the preset coupling judgment threshold, the lowest workstation with the lowest workstation hit rate and the lowest category with the lowest category hit rate will be combined into a coupling analysis object. The position deviation values ​​of all coupling execution results generated by the coupling analysis object within the corresponding statistical period will be extracted and combined with the corresponding coupling dynamic running speed to form deviation speed data pairs. Based on all deviation speed data pairs, the sum of all coupling dynamic running speeds, the sum of all coupling execution result position deviation values, the sum of the squares of all coupling dynamic running speeds, and the sum of the product of speed and deviation within each deviation speed data pair will be calculated. These will be substituted with the total number of deviation speed data pairs into the preset least squares slope calculation formula to obtain the corrected running calibration parameters. The corrected running calibration parameters will then replace the original running calibration parameters.

[0013] In the above embodiments, by identifying and pinpointing the coupled errors caused by the combined effects of specific workstations and product categories, a deep diagnosis of the root cause of the problem is achieved. Instead of simple average calculations, the least squares method is used to perform regression analysis on the deviation-speed data in this coupled scenario, thereby revealing the inherent law of system error variation with speed. The resulting corrected calibration parameters can more accurately and proactively compensate for dynamic errors, ultimately achieving high accuracy in sorting and positioning even with constantly changing conveyor line speeds.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, based on operational calibration parameters, the baseline predicted trigger time is adjusted to the actual predicted trigger time in subsequent sorting actions within the corresponding statistical period, specifically including: The system acquires the operating calibration parameters and the baseline predicted trigger time of the target object; it collects and stores the dynamic operating speed of the conveyor line at a preset acquisition frequency during the current sorting action window, which is the current speed and the speed of the immediately preceding time; it multiplies the instantaneous speed change with the operating calibration parameters to generate a dynamic compensation time value; it subtracts the dynamic compensation time value from the baseline predicted trigger time to obtain the actual predicted trigger time, and generates and issues sorting action instructions based on the actual predicted trigger time.

[0015] In the above embodiments, by acquiring the conveyor line speed at high frequency, subtle instantaneous speed changes can be captured. This change is then combined with operational calibration parameters to generate a dynamic compensation time value that accurately reflects the impact of the current acceleration. By subtracting this dynamic compensation value from the baseline prediction trigger time, real-time fine-tuning of the prediction can be completed before the sorting instruction is issued, offsetting the positioning error caused by speed fluctuations. Ultimately, even under complex operating conditions with frequent changes in conveyor line speed, the accuracy of sorting positioning is improved.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after multiplying the instantaneous velocity change by the operating calibration parameters to generate the dynamic compensation time value, the method further includes: When the current target object belongs to the preset pneumatically sensitive category and the instruction type of the immediately preceding sorting action is pneumatic sorting, the preset pneumatic disturbance factor bound to the pneumatically sensitive category is additionally obtained from the decision model; the pneumatic disturbance factor is summed with the dynamic compensation time value to obtain the final compensation time value, and the dynamic compensation time value is replaced by the final compensation time value.

[0017] In the above embodiments, the error sources that cause air disturbances to subsequent pneumatically sensitive product categories due to previous pneumatic sorting actions are identified. By introducing a product category-specific pneumatic disturbance factor and superimposing this factor on the original dynamic compensation time, the final compensation model can not only cope with conveyor speed fluctuations but also proactively incorporate environmental impacts under specific operating conditions into the calculation. This targeted compensation for both mechanical and environmental variables can offset complex compound errors, ultimately improving the accuracy of sorting and positioning even in dynamic scenarios where conveyor speeds continuously change.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, after generating and issuing a sorting action instruction to the actuator at the corresponding execution position when the baseline prediction trigger time arrives, and obtaining the real-time execution result position deviation value, the method further includes: When the position deviation value of the execution result exceeds the preset instantaneous failure threshold, the target object that failed to be sorted in real time is marked as a dynamic obstacle. Based on the last known obstacle position of the dynamic obstacle and the current real-time dynamic running speed of the conveyor line, the real-time obstacle position trajectory of the dynamic obstacle on the conveyor line is continuously predicted and updated. If the predicted future motion trajectory of the next target object and the real-time obstacle position trajectory overlap within the preset sorting execution time and space area, the sorting action instruction corresponding to the next target object is canceled until the predicted trajectory of the dynamic obstacle leaves the sorting execution time and space area, and then the sorting action is recalculated for the next target object.

[0019] In the above embodiments, by marking failed sorting objects as dynamic obstacles in real time and continuously predicting their movement trajectories based on the real-time speed of the conveyor line, it is possible to determine whether there is a risk of collision or interference with subsequent objects during sorting. Based on this, by intelligently canceling potential conflicting instructions, chain collisions and errors caused by a single failure point are proactively avoided. Even in environments where the conveyor line speed is constantly changing, this fundamentally ensures that subsequent sorting actions have uninterrupted execution space, ultimately improving the overall accuracy of sorting positioning.

[0020] In a second aspect, embodiments of this application provide a sorting management system, which includes: one or more processors and a memory; the memory is coupled to one or more processors and is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the sorting management system to perform the methods described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a sorting management system, cause the sorting management system to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a sorting management system, cause the sorting management system to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the sorting management system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the AI ​​model-enabled sorting management method provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This application constructs a closed-loop feedback system that directly addresses the final result by continuously acquiring the positional deviation value of the execution result and monitoring the real-time sorting hit rate. This overcomes the shortcomings of related technologies that rely solely on input signals and cannot detect systematic errors such as belt slippage and actuator delay. When the hit rate decreases, it can dynamically generate operational calibration parameters based on the statistical relationship between deviation and speed, quantifying the deviation between the ideal model and physical reality. Ultimately, by correcting the predicted trigger time using these parameters, it achieves adaptive optimization of the control model, ensuring improved sorting positioning accuracy when the conveyor line speed changes.

[0025] 2. This application utilizes a neural network decision-making model pre-trained with historical data, enabling it to learn and deeply understand the complex nonlinear coupling relationship between sorting hit rate, positional deviation, and conveyor line speed. Therefore, when the real-time hit rate decreases, the current operating state can be used as an overall feature input, allowing the decision-making model to output optimized operating calibration parameters based on its experience learned from massive amounts of historical operating conditions. These parameters dynamically adapt to the complex system errors caused by various factors such as equipment wear and environmental changes, thereby adjusting the predicted trigger time and ultimately improving sorting positioning accuracy under varying conveyor line speeds.

[0026] 3. This application employs workstation and product category dimensions to perform attribution analysis on sorting errors. When the overall hit rate declines, the root cause is located by identifying the dimension with the lowest value, whether it's a mechanical performance degradation at a specific workstation or a systematic deviation caused by the physical characteristics of a specific product category. Subsequently, the solution only performs statistical analysis on the data of this core analysis object, calculating highly targeted operational calibration parameters. This targeted calibration mechanism avoids blindly adjusting the entire system and compensates for the main problems, thus improving sorting positioning accuracy even when the conveyor line speed changes continuously. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an AI model-enabled sorting management method in an embodiment of this application. Figure 2 This is another flowchart illustrating the sorting management method powered by an AI model in the embodiments of this application; Figure 3 This is a schematic diagram of an exemplary hardware structure of the sorting management system in this application embodiment. Detailed Implementation

[0028] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0029] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0030] In related technologies, the control methods of sorting systems typically rely on predictive mechanisms based on ideal physical models. This method acquires the initial position of objects using sensors at the entrance and combines this with real-time speed data provided by the conveyor encoder to calculate the theoretical trigger time, thus instructing the actuator to perform sorting at a predetermined position. This open-loop or semi-open-loop control method performs reasonably well when the system is in an ideal state. However, it is essentially based on a static, lossless system model, and its core flaw lies in its inability to perceive and quantify changes in the physical state of the system during long-term operation. For example, the increased slippage rate of the drive belt due to wear, or the response delay of pneumatic components due to aging and air supply fluctuations. When these unpredictable physical deviations accumulate, a disconnect arises between the theoretical model and physical reality, causing predictive calculations based solely on input speed to become inaccurate. Ultimately, this leads to a significant decrease in sorting positioning accuracy as equipment ages and operating conditions change.

[0031] In this embodiment, a dynamic closed-loop self-calibration system based on actual sorting results is constructed, thus avoiding the shortcomings of related technologies where the control model is fixed and cannot adapt to changes in the physical state of the system. Instead of relying solely on input operating parameters, the focus shifts to the final execution result's positional deviation value and the real-time sorting hit rate calculated from it. When the hit rate is detected to be lower than a preset threshold, it is confirmed that the current control model is insufficient to match the actual working conditions, and a calibration mechanism is immediately initiated. By deeply analyzing the intrinsic correlation between massive positional deviation values ​​and corresponding dynamic operating speeds, quantitative operating calibration parameters are dynamically generated. These parameters reflect the comprehensive deviation between the theoretical model and physical reality at the current moment, and are used as a basis to adjust the benchmark prediction trigger time, thereby achieving continuous self-optimization and iteration of the control model. This ensures that even under conditions of system performance degradation or operating condition fluctuations, high-precision sorting operations can still be maintained.

[0032] Figure 1 This is a flowchart illustrating a sorting management method powered by an AI model as described in this application, including the following steps: S101. At a preset detection position on the conveyor line, obtain the initial position of the moving target object, the corresponding initial timestamp, and the dynamic running speed of the conveyor line.

[0033] Here, a conveyor line refers to mechanical equipment used to continuously transport goods on a fixed route, such as a belt conveyor; a preset detection position refers to a data acquisition area that is fixed in advance in physical space. The position is set based on a comprehensive engineering design considering factors such as the distance to the downstream actuator, the maximum response time, and the performance of the sensors; a moving target object refers to an independent unit that needs to be sorted while moving on the conveyor line, such as a package, bottle, or component; the initial position refers to the two-dimensional or three-dimensional coordinates of the key physical feature points (such as the geometric center) of the target object in the conveyor line coordinate system when it first fully enters the preset detection position; the initial timestamp is the moment when the initial position is captured, usually provided by a high-precision system clock, serving as the reference zero point for all subsequent time calculations; and the dynamic running speed represents the real-time linear speed of the conveyor belt or rollers at the same instant the initial timestamp is captured.

[0034] Specifically, when each new target object enters the sensing range of the sorting system, as the target object moves along the conveyor line and enters the preset detection position, a trigger signal (e.g., generated by the detection of the object by a photoelectric sensor or vision system) activates the entire data acquisition sequence. At this time, the core sensing device (such as an industrial camera or laser profile scanner) captures complete information about the target object, and calculates its position in the coordinate system (initial position) through image processing or point cloud analysis algorithms; records the system time for executing the first step of the operation as the initial timestamp; and synchronously reads pulse data from the encoder connected to the conveyor line drive motor or driven wheel, converting it into the current actual operating speed (dynamic operating speed).

[0035] S102. Based on the initial position, initial timestamp, dynamic running speed, and preset execution position, calculate the baseline predicted trigger time for the target object to reach the execution position.

[0036] The preset execution position refers to the spatial point or area of ​​action where the sorting execution mechanism (such as a swing wheel, flap, or push rod) can exert a physical effect on the target object. This position is measured and fixed in the system parameters during the system installation and commissioning phase. The baseline predicted trigger time represents the future time point calculated based on the initial state information, at which theoretically the center of the target object will just reach the preset execution position.

[0037] Specifically, the most basic calculation model is based on the assumption of uniform linear motion. First, the straight-line distance from the target object's initial position to its execution position is calculated; this is typically the displacement difference along the conveyor line's direction of travel. Then, this distance is divided by the dynamic running speed obtained in S101 to obtain an estimated travel time. Finally, this travel time is added to the initial timestamp obtained in S101; the sum is the baseline predicted trigger time. Theoretically, this calculation process is instantaneous, providing the first time anchor for subsequent execution actions. However, in more complex systems, this calculation may consider more advanced motion models, such as those capable of predicting short-term trends in conveyor line speed (e.g., simple linear regression based on historical data), which would employ variable acceleration motion formulas to calculate a more accurate travel time.

[0038] In some embodiments, the calculation of the baseline prediction trigger time in this step can be implemented in a variety of ways: Optionally, a simplified linear model is adopted: read the preset execution position coordinates P_exec=(x_exec, y_exec) and the initial position coordinates P_init=(x_init, y_init) obtained in S101; calculate the displacement Δx=x_exec-x_init along the main motion direction of the conveyor line (assuming it is the X-axis); use the initial timestamp t_init and dynamic running speed v_init obtained in S101 to calculate the baseline prediction trigger time through the formula t_trigger_base=t_init+Δx / v_init.

[0039] It is understandable that other methods can be used to generate this predicted time, such as using a nonlinear model that takes into account the elastic deformation of the conveyor belt, or introducing a simple speed prediction model based on historical data to optimize the calculation of travel time. No limitation is made here.

[0040] S103. When the baseline prediction trigger time arrives, generate and issue sorting action instructions to the corresponding execution mechanism, and obtain the real-time execution result position deviation value.

[0041] Among them, sorting action instructions refer to digital signals or data packets sent to hardware that can be parsed by the controller (such as PLC) and drive the actuator to perform actions, such as high-level pulses or CAN bus messages containing action parameters; actuators refer to the equipment that actually performs physical sorting actions, such as pneumatic push rods, steering wheel, cross belt trolleys, etc.; the real-time execution result position deviation value is used to represent the physical distance difference between the actual final position of the target object after the sorting action is completed and the ideal target position that it should theoretically reach, and is an indicator for quantifying sorting accuracy.

[0042] Specifically, the system is triggered when the internal clock reaches the baseline predicted trigger time calculated by S102, generating an instruction associated with the target object and its execution position. This instruction is sent to the lower-level I / O controller or directly to the corresponding actuator driver. Upon receiving the instruction, the actuator performs a preset mechanical action, such as a push rod extending and retracting momentarily, thereby changing the trajectory of the target object and guiding it to the designated sorting slot. After the action is completed, the result detection module begins operation. Typically, another vision sensor or light curtain is installed downstream of the sorting area. When the object being sorted passes through this detection area, its actual position is measured. This actual position is then compared with the preset ideal center position that the sorting action should achieve. The difference between the two (usually a millimeter-level value with a plus or minus sign, indicating advance or lag) is recorded as the real-time execution result position deviation value for this sorting operation.

[0043] S104. Within a preset statistical period, continuously obtain the corresponding execution result position deviation value for multiple sorting actions that have been executed, and calculate the real-time sorting hit rate within the corresponding statistical period.

[0044] The preset statistical period refers to a time window or event counting window set in advance by the system administrator based on production cycle time, equipment stability, and optimization frequency requirements, such as every 15 minutes or every 1000 objects processed; the multiple sorting actions executed represent the set of all sorting events with feedback results that occurred within the statistical period; the real-time sorting hit rate is a dynamically updated performance indicator used to represent the percentage of sorting actions in the current statistical period where the position deviation value of the execution result falls within the acceptable range, out of the total number of sorting actions.

[0045] Specifically, the system runs continuously in the background as a sliding or fixed window throughout the entire system operation. A buffer is maintained to store the positional deviation values ​​of all completed sorting actions within the most recent statistical period. For each newly completed sorting action and after acquiring the deviation value, it is determined whether the absolute value of the deviation is less than a preset hit tolerance threshold. This threshold is preset based on the width of the sorting slot, the object size, and process requirements. If the deviation is within the threshold, it is considered a hit; otherwise, it is considered a miss. Subsequently, the total number of sorting operations and the total number of hits within the period are updated, and the real-time sorting hit rate is recalculated (hit rate = total hits / total sorting operations).

[0046] In some embodiments, the hit rate statistics and calculation in this step can be implemented in a variety of ways: Optionally, statistics based on a sliding event window can be used: A fixed-length queue (FIFO, First-In-First-Out) is maintained, with the queue length being the preset statistical period, such as 1000 events. This queue stores the execution results (hit or miss) of the most recent 1000 sorting operations. Each time a new sorting operation is completed, the system determines whether it was a hit and pushes the result (a boolean value, true for a hit, false for a miss) to the end of the queue. If the queue length exceeds 1000, the oldest result at the head of the queue is popped. After each queue update, the number of hit results in the entire queue is recalculated and divided by the current queue length to obtain a continuously and smoothly updated real-time sorting hit rate.

[0047] It is understandable that other methods can be used to quantify this performance metric, such as using an exponentially weighted moving average (EWMA) to calculate the hit rate, which gives higher weight to more recent events and thus more sensitively reflects recent changes in the system state. No limitation is made here.

[0048] S105. When the real-time sorting hit rate is lower than the preset hit rate threshold, generate operation calibration parameters based on the real-time sorting hit rate, the position deviation value of the real-time execution result and the corresponding dynamic running speed.

[0049] Among them, the preset hit rate threshold is a pre-set performance baseline defined by process engineers according to production quality requirements. When the actual hit rate is lower than this value, it indicates that the system performance has deteriorated unacceptably. The running calibration parameters are used to represent one or a set of values ​​used to correct subsequent prediction calculations, which are dynamically adjusted coefficients or factors.

[0050] Specifically, when the real-time sorting hit rate first falls below the preset hit rate threshold, it is considered that there is a persistent, non-random systematic deviation (such as physical stretching of the conveyor belt, sensor zero-point drift, or actuator wear). This deviation cannot be resolved by a single adjustment and model calibration is necessary. At this time, a calibration algorithm is invoked. The input to this algorithm is a series of bad data accumulated within the current statistical period, that is, detailed records of sorting events that led to the drop in hit rate. This includes at least the real-time execution result position deviation value and the corresponding dynamic running speed for each event, and sometimes also the real-time sorting hit rate itself as a macroscopic state input. The goal of the algorithm is to analyze the potential relationship between these deviations and variables such as speed, and generate operating calibration parameters that can compensate for this relationship. The essence of these parameters is to correct the ideal physical model (distance = speed × time) to adapt to the real physical world full of disturbances.

[0051] In some embodiments, the generation of calibration parameters in this step can be achieved in several ways: Optionally, a statistical analysis method based on linear regression can be used: when the hit rate is below a threshold, all data pairs (Δp, v) of (deviation value, running speed) within the current statistical period are extracted; these data pairs are treated as points on a two-dimensional plane, and a simple least-squares linear regression algorithm is applied to attempt to fit a straight line of the form Δp = k * v + b; the calculated slope k is regarded as the running calibration parameter. The physical meaning of this parameter is how much systematic drift the position deviation will produce for every unit change in speed, directly quantifying the sensitivity of the deviation to speed.

[0052] It is understandable that other methods can be used to implement the decision-making process for this calibration parameter, such as using more traditional control theory methods, like a PID controller, with the hit rate as the controlled variable and the calibration parameter as the output, for iterative adjustment. No limitation is made here.

[0053] S106. Based on the operating calibration parameters, adjust the baseline predicted trigger time to the actual predicted trigger time in subsequent sorting actions within the corresponding statistical period.

[0054] Among them, the newly generated calibration parameters are applied to the original calculation process by combining mathematical or logical operations; the subsequent sorting action refers to all sorting events after the new running calibration parameters are generated until the end of the next statistical cycle or the next calibration occurs; the actual predicted trigger time refers to the moment when the command is finally issued, which is closer to the real physical situation and is obtained after dynamic correction of the running calibration parameters based on the benchmark predicted time.

[0055] Specifically, the operational calibration parameters generated in S105 are applied to every subsequent sorting action. After generating the operational calibration parameters, a fine-tuning step is added after the calculation process in S102 as a dynamic compensation factor. For each new target object, the baseline predicted trigger time is still calculated first according to the method in S102. However, after that, the dynamic compensation time value is calculated using the operational calibration parameters generated in S105 and real-time dynamic information related to the current target object (such as the instantaneous speed change of the conveyor line). This compensation time value may be positive or negative, representing the amount of time that needs to be advanced or delayed for triggering. Finally, the dynamic compensation time value is subtracted (or added) from the baseline predicted trigger time to obtain the actual predicted trigger time. This final dynamically corrected time is used by S103 as the moment for issuing the sorting instruction.

[0056] In some embodiments, the dynamic adjustment of the prediction time in this step can be achieved in a variety of ways: Optionally, velocity-based linear compensation is employed: after calculating the baseline predicted trigger time t_trigger_base for an object, the dynamic running velocity v_init of the object at the detection point is obtained; using the running calibration parameter k generated by S105 (whose physical meaning is time deviation / velocity), the compensation time Δt_comp = k * v_init is calculated; the final actual predicted trigger time is t_trigger_actual = t_trigger_base - Δt_comp. This means that the faster the velocity, the more time compensation (determined by parameter k) is introduced.

[0057] Understandably, other methods can also be used to correct the final trigger time. For example, the calibration parameters can be used as an input, along with various features of the object (such as weight and size), to input a complex function to calculate a more personalized compensation time. This is not limited here.

[0058] In some embodiments, when the sorting hit rate deteriorates, a pre-trained AI decision model can be used to combine high-frequency dynamic compensation for instantaneous speed changes of the conveyor line with correction for specific physical disturbances (such as pneumatics) to achieve dual compensation for systematic deviations and instantaneous physical disturbances, thereby improving the adaptability and robustness of the sorting system under complex dynamic conditions.

[0059] This is triggered when the system detects that the real-time sorting hit rate is lower than a preset hit rate threshold. The preset hit rate threshold is a performance red line pre-set by process engineers based on Quality of Service (QoS) standards. Falling below this value indicates a systemic deviation requiring intervention. At this point, a preset decision model is invoked. This decision model is essentially a pre-trained neural network model (e.g., a multilayer perceptron or recurrent neural network), which packages three types of core data from the current statistical period into a feature vector as input to the model: 1) Macro performance indicators, i.e., the real-time sorting hit rate, reflecting the overall health of the system; 2) Micro error distribution, i.e., the set of all real-time execution result positional deviations within the period, revealing the specific error pattern (e.g., overall early or late, or increased dispersion); 3) Dynamic operating environment background, i.e., the dynamic operating speed corresponding to each deviation, providing the physical environment parameters at the time of the error.

[0060] This decision-making model, through its internally complex, trained nonlinear connection weights, is able to autonomously learn and deeply understand the potential relationships between these three factors. For example, it can identify a slight decrease in hit rate and a deviation primarily manifested as lag during high-speed operation, and generate optimal operational calibration parameters accordingly. These parameters are not fixed time or distance compensation values, but rather sensitivity coefficients used for subsequent dynamic adjustments. They represent the model's diagnostic conclusion regarding the root cause of the current system deviation.

[0061] To ensure the decision-making model possesses accurate decision-making capabilities, in one embodiment of the invention, the dataset used to train the decision-making model is obtained as follows: During system debugging or maintenance, operators can simulate different operating conditions (e.g., adjusting conveyor belt tension to simulate slippage, or adjusting air pressure to simulate execution delay). Under a certain operating condition, when the system experiences sorting deviation, experienced engineers manually adjust calibration parameters (e.g., fine-tuning trigger delay or position offset in the control system) and conduct small-batch tests until the sorting hit rate returns to the ideal level. At this point, the current operating condition data (hit rate, deviation value, speed) and this verified effective calibration parameter are recorded, and they are used as a set of training data. By repeating this process, a large number of problem-solution data pairs are accumulated for training the AI ​​decision-making model.

[0062] After obtaining new operational calibration parameters generated by the AI ​​model, these parameters are applied for real-time compensation in every subsequent sorting operation. The core idea is that systematic deviations are often strongly correlated with the dynamic characteristics of the conveyor line (especially speed variations). Therefore, a sorting operation window (the period from when the object is detected to when the instruction is issued) is established for the upcoming target object. Within this window, the dynamic operating speed of the conveyor line is continuously collected at a preset acquisition frequency (this frequency is determined based on the system's required reaction speed and control cycle, and is usually much higher than the frequency at which the object passes through, such as 1 kHz).

[0063] For each data acquisition point, the current speed and the speed of the immediately preceding moment are recorded, and the difference between the two is calculated to obtain the instantaneous speed change. This value is essentially a direct reflection of the conveyor line's acceleration at that instant. Subsequently, the instantaneous speed change is multiplied by the operational calibration parameters generated by the AI ​​model to obtain the dynamic compensation time value. The advantage of this approach is that it enables feedforward control: when the conveyor line suddenly accelerates (positive change), the trigger time is shortened proactively; when the conveyor line decelerates (negative change), the trigger time is extended. The magnitude of the compensation is proportional to the acceleration and is scaled by the AI ​​parameters, thereby offsetting the positional errors caused by speed fluctuations.

[0064] To address more nuanced physical interference, a scenario-based compensation mechanism is introduced. This involves conditional judgment: First, determining whether the current target object belongs to a pre-defined pneumatically sensitive category. Pneumatically sensitive categories are pre-classified and entered into the system's material list by those skilled in the art based on the item's physical characteristics (e.g., lightweight, large surface area, susceptibility to airflow), such as empty plastic bottles and potato chip bags. Second, querying the instruction type of the immediately preceding sorting action to determine if it is a pneumatic sorting operation (e.g., using a high-pressure air nozzle).

[0065] When both conditions are met, it is determined that the current object is highly likely to be disturbed by residual airflow from previous pneumatic sorting. At this point, a pre-defined aerodynamic disturbance factor, bound to this aerodynamically sensitive product category, is requested from the decision model again. This factor is also a specific compensation time value learned through targeted experimental data during the model training phase. This aerodynamic disturbance factor is summed with the dynamic compensation time value calculated in the previous stage to obtain the final compensation time value.

[0066] Finally, the actual predicted trigger time is obtained by subtracting the final compensation time value, which takes into account three major factors: systematic deviation, instantaneous speed fluctuation, and specific physical disturbance, from the baseline predicted trigger time of the object. Based on this time, sorting action instructions are generated and issued.

[0067] The aforementioned technical solution, by introducing an AI decision-making model, can learn from massive amounts of data and cope with complex, nonlinear system performance degradation patterns. Furthermore, by applying calibration parameters to compensate for conveyor line acceleration, it achieves feedforward-based precise control of dynamic disturbances, far superior to traditional static or velocity-based compensation. Finally, by superimposing disturbance factors for specific scenarios (such as aerodynamic disturbances), the compensation model can delve into the level of specific physical interactions, ultimately enabling the system to maintain a high sorting hit rate under various complex and dynamically changing real-world operating conditions.

[0068] In the above embodiments, by continuously acquiring the positional deviation value of the execution result and monitoring the real-time sorting hit rate, a closed-loop feedback directly addresses the final result, overcoming the shortcomings of related technologies that rely solely on input signals and cannot detect systematic errors such as belt slippage and actuator delay. When the hit rate decreases, operational calibration parameters can be dynamically generated based on the statistical relationship between deviation and speed, quantifying the deviation between the ideal model and physical reality. Ultimately, this parameter is used to correct the predicted trigger time, achieving adaptive optimization of the control model and ensuring improved sorting positioning accuracy when the conveyor line speed changes. In some other embodiments of this application, when the minimum workstation hit rate and the minimum category hit rate are very close, it may be difficult to distinguish a single source of failure, leading to deviations in the calibration model. The sorting management method powered by the AI ​​model provided in this application can combine the two into a coupled analysis object, locating and correcting systematic deviations caused by specific interactions between workstations and categories.

[0069] like Figure 2 The diagram shown illustrates another process flow of the AI-powered sorting management method provided in this application embodiment, including the following steps: S201. At a preset detection position on the conveyor line, obtain the initial position of the moving target object, the corresponding initial timestamp, and the dynamic running speed of the conveyor line.

[0070] S202. Based on the initial position, initial timestamp, dynamic running speed, and preset execution position, calculate the baseline predicted trigger time for the target object to reach the execution position.

[0071] S203. When the baseline prediction trigger time arrives, generate and issue sorting action instructions to the corresponding execution mechanism, and obtain the real-time execution result position deviation value.

[0072] In some embodiments, if the positional deviation value of the result of a single sorting action exceeds a preset instantaneous failure threshold, the failed object can be virtualized into a dynamic obstacle in real time and its trajectory predicted. This allows for proactive avoidance and intelligent adjustment of the sorting timing of other objects immediately following it, thereby preventing chain collisions, blockages, and damage caused by sorting failures and enhancing the fault tolerance and operational stability of the entire sorting system.

[0073] Specifically, after S203 obtains the positional deviation value of each sorting execution result, it first compares the absolute value of the deviation value with a preset instantaneous failure threshold. This preset instantaneous failure threshold is a strict error boundary pre-set by the system engineer based on physical conditions and process requirements. The threshold is typically set by subtracting the maximum width of the target object from the physical width of the sorting slot, dividing by two, and multiplying by a safety factor less than 1. Once the deviation value exceeds this threshold, it means that the target object has missed its opportunity to enter the designated slot and will remain on the main conveyor line continuing to move downstream.

[0074] At this point, the target object, which has become out of control in the physical world, is immediately marked as a dynamic obstacle in the internal digital twin model. This is not merely a change in state; it triggers specialized tracking and prediction logic. The last known position of the obstacle at this moment (i.e., the result detection position in S203) and the current timestamp are recorded.

[0075] A dynamic obstacle is generated, initiating a high-frequency prediction loop. This loop runs at the basic clock cycle of the conveyor control system (e.g., every 10 milliseconds). In each cycle, the prediction algorithm calculates and updates the obstacle's predicted position at the current moment based on the obstacle's last known position, the timestamp of that position, and the real-time dynamic operating speed of the conveyor obtained from the encoder. The core theory is based on the kinematic formula displacement = velocity × time. By continuously extrapolating the obstacle's position using the latest real-time speed, a real-time obstacle position trajectory highly synchronized with an out-of-control object in the physical world is generated in digital space. The advantage of this is that even if the conveyor changes speed or starts / stops, the virtual position of the obstacle remains extremely accurate.

[0076] When preparing to calculate the sorting action for the next target object adjacent to the obstacle (i.e., executing processes S202 and S203), the aforementioned safety subsystem intervenes to conduct a proactive risk assessment. First, it calculates the complete predicted future trajectory of the next target object. Simultaneously, it defines a preset sorting execution spatiotemporal region. This preset region is a four-dimensional virtual danger zone. The three-dimensional spatial boundary is determined by the physical motion envelope of the actuator (such as a balance wheel), while the time dimension is defined by the start and end times of the execution action. This region is pre-calibrated and fixed in the system based on the actuator's design parameters and motion simulation.

[0077] The core task of collision detection algorithms is to determine whether the predicted future trajectory of the next target object and the real-time obstacle position trajectory of the dynamic obstacle will intersect within the sorting execution spatiotemporal region corresponding to the target object. If any form of intersection or overlap exists, a high risk of collision is determined.

[0078] Once a collision risk is detected, avoidance measures will be taken immediately: the sorting action command originally planned for the next target object will be cancelled. This means that even if the baseline or actual predicted trigger time is reached, no action command will be issued to the corresponding actuator. This fundamentally avoids physical collisions between the actuator and dynamic obstacles drifting upstream after it extends, thus protecting the equipment itself. Simultaneously, it avoids pushing the next target object into a space already occupied by obstacles, preventing collisions and damage between goods.

[0079] After the instruction is canceled, it enters an intelligent waiting state. It continuously monitors the predicted trajectory of dynamic obstacles. Only when the trajectory has completely left the sorting execution space-time area required by the next target object, a safety window is confirmed to appear. The sorting action for the next target object is recalculated, a new, delayed trigger time is generated, and sorting is finally completed.

[0080] The core benefit of the aforementioned technical solution lies in transforming isolated sorting failures into manageable and avoidable dynamic events through real-time capture and proactive trajectory prediction. This predictive collision detection and proactive instruction intervention mechanism can break the chain reaction caused by a single point of failure (a package not being successfully sorted) (subsequent packages colliding with it, piling up, or even causing the entire line to stop), minimizing losses. It elevates the system's fault tolerance from passive post-event alarms to proactive pre-event avoidance, enhancing the reliability of continuous operation of automated sorting lines.

[0081] S204. Within a preset statistical period, continuously obtain the corresponding execution result position deviation value for multiple sorting actions that have been executed, and calculate the real-time sorting hit rate within the corresponding statistical period.

[0082] Steps S201-S204 and Figure 1 Steps S101-S104 in the illustrated embodiment are similar and can be found in the descriptions of steps S101-S104, which will not be repeated here.

[0083] S205. Obtain the category information of the target object and the workstation information of the actuator corresponding to each sorting action.

[0084] Category information refers to classification data used to describe the physical or commercial attributes of the target object, such as brand, weight range, size specifications, or material (e.g., "fragile items", "clothing", "bottled beverages"). This information is usually preset in the warehouse management system and associated with specific objects through barcodes or visual identification. The actuator refers to the equipment that actually performs the physical sorting action, such as pneumatic push rods, steering wheels, etc. The workstation information is used to represent the physically and logically unique identifier assigned to a specific actuator, such as "sorting port 3" or "Area A 05 chute". This information is preset according to the physical layout of the sorting line during the system deployment phase.

[0085] Specifically, when a target object is detected in S201, not only are its physical motion parameters acquired, but its carrying identifiers (such as barcodes, QR codes, or RFID tags on packages) are also used to query the backend business database (such as a WMS or ERP system). This query returns detailed category information for the object. Simultaneously, based on the object's destination (e.g., a package destined for Beijing), the specific sorting exit is retrieved from the routing table. This exit corresponds to the workstation information for this sorting operation. Category and workstation information are dynamically appended to the target object's digital profile.

[0086] S206. Within the corresponding statistical period, when the real-time sorting hit rate is lower than the preset hit rate threshold, the workstation hit rate and the category hit rate of each independent workstation and each independent category are calculated based on the workstation information and category information.

[0087] Among them, the workstation hit rate refers to the percentage of sorting times successfully performed at an independent workstation within a specific statistical period, which is an indicator of sorting performance at a specific physical location; the category hit rate is used to represent the percentage of sorting times successfully performed for an independent category of objects within the same period, which is an indicator of the system's ability to process specific types of items.

[0088] Specifically, when the real-time sorting hit rate calculated by S204 falls below the preset performance alarm line, a step-by-step diagnostic process is initiated, from macro-level issues to micro-level root causes. Once triggered, detailed logs of all sorting events recorded within the current statistical period are retrieved. These logs not only include the success or failure of each sorting attempt (i.e., whether the deviation value is within the tolerance), but also the workstation and category information obtained in S205. Two independent data aggregation operations are performed: First, based on workstation information, the total number of attempts and successes for each workstation are calculated, and their respective hit rates are determined; second, based on category information, the same operation is performed to calculate the hit rate for each category. Through this step, the system performance degradation problem is broken down into two more specific and comparable lists of performance indicators.

[0089] S207. Among all workstation hit rates and product category hit rates, identify the dimension with the lowest value as the core analysis object.

[0090] In this context, "dimension" refers to two basic directions or perspectives for data analysis and problem segmentation: the workstation dimension based on physical location and the category dimension based on item type. The core analysis object refers to the specific instance that performs the worst among all sub-items and is most likely to be the root cause of the overall system performance decline, such as "workstation number 5" or "ultra-light small item category".

[0091] Specifically, from the two sets of hit rate data generated in the previous step, the unique, worst-performing performance point is identified. First, the minimum hit rate is found across all workstations, and the workstation that produced this minimum is recorded. Then, another minimum hit rate is found across all product category hit rates, and the corresponding product category is recorded. Finally, these two minimums are compared. If the lowest hit rate in the workstation dimension is lower than the lowest hit rate in the product category dimension, the problem is more likely at the physical equipment level, and the worst-performing workstation is identified as the core analysis object. Conversely, if they are different, the system is considered to have a problem with the processing of a specific product category, and that product category is identified as the core analysis object. The theoretical basis for this step is the bottleneck theory, which states that prioritizing the resolution of the worst-performing link in the system can maximize the improvement of overall performance.

[0092] In some embodiments, when the identified minimum workstation hit rate and minimum category hit rate are highly similar, the two dimensions can be locked together as a coupled analysis object, and linear regression analysis can be performed on all relevant data points using the least squares method. This allows for the identification and quantification of complex interactive failures that only manifest when a particular device processes a particular type of item, thereby generating corrected operation calibration parameters that are more statistically robust and predictively accurate than simple averages.

[0093] Specifically, this step is a supplement and enhancement to the single-dimensional fault location logic in S207, and is specifically designed to handle performance degradation problems caused by multiple factors rather than a single factor.

[0094] After calculating the hit rate for all workstations and product categories, S206 first identifies the lowest hit rate and its corresponding workstation (denoted as W_min) in the workstation dimension, and the lowest hit rate and its corresponding product category (denoted as C_min) in the product category dimension. Next, the absolute value of the difference between these two lowest hit rate values ​​is calculated and compared with a preset coupling judgment threshold. This preset coupling judgment threshold is a small value pre-set by the data analyst based on statistical principles and historical operational experience, defining an interval where there is no statistically significant difference.

[0095] When the absolute value of the hit rate difference is less than the threshold (including cases where the two are exactly equal), it is determined that the root cause of the current system performance degradation is most likely a coupling fault, which is a problem that only erupts when the specific workstation W_min processes the specific product category C_min. Therefore, these two are combined to form a coupling analysis object, such as "workstation 5 processes small and light parts".

[0096] After identifying the objects of the coupling analysis, a targeted data filtering process is initiated. All sorting logs within the current statistical period are iterated again, but the filtering criteria become more stringent: a log record is selected only if its workstation information is simultaneously equal to W_min and its product category information is simultaneously equal to C_min. The advantage of this approach is that it isolates all raw data that reflects the essence of the coupling fault.

[0097] For each selected record, the positional deviation value of the coupled execution result and the corresponding coupled dynamic running speed are extracted and stored as a deviation-speed data pair. After traversing all logs and collecting all relevant data pairs, statistical calculations are performed on this dataset. It calculates the following five core statistics: the sum of coupled dynamic running speeds in all data pairs (Σv); the sum of coupled execution result positional deviation values ​​in all data pairs (ΣΔp); the sum of the squares of coupled dynamic running speeds in all data pairs (Σ(v^2)); the sum of the products of speed and deviation within each data pair (Σ(v*Δp)); and the total number of deviation-speed data pairs (n).

[0098] Finally, the five statistics calculated in the previous step are directly substituted into the preset least squares slope calculation formula. This preset formula is a standard formula derived from linear regression theory in mathematical statistics; it does not require online calculation and is embedded in the algorithm. The formula is typically: Corrected running calibration parameter (k) = [n*Σ(v*Δp)-(Σv)*(ΣΔp)] / [n*Σ(v^2)-(Σv)^2].

[0099] The essence of this formula is to find the best-fitting straight line Δp = k*v + b in a scatter plot of all deviation-velocity data points, minimizing the sum of the squared distances from all data points to this line. The calculated slope k, i.e., the corrected operational calibration parameter, represents the optimal linear estimate of the effect of velocity changes on position deviation. This parameter will then replace the operational calibration parameters based on simple averages that may have been calculated in S208-S210, and will be used for subsequent compensation calculations.

[0100] Compared to analyzing faults from only a single dimension, the above-mentioned technical solution can identify and quantify the complex interactions between equipment and material characteristics, revealing deeper fault mechanisms. More importantly, by employing least squares regression analysis on all relevant data points, rather than simply calculating the ratio of averages, the generated calibration parameters consider the overall distribution trend and dispersion of the data, avoiding the excessive influence of a few outliers on the results. Therefore, the resulting parameter model is more robust, and the predicted compensation effect is more accurate.

[0101] S208. Extract the positional deviation values ​​of all execution results corresponding to the core analysis object within the corresponding statistical period, and calculate the average deviation value.

[0102] Among them, the core analysis object refers to the specific workstation or product category that performs the worst in S207; the position deviation value of the execution result is used to represent the accuracy of a single sorting action, that is, the difference in physical distance between the actual landing point and the theoretical landing point; the average deviation value is the result obtained by arithmetically averaging all independent deviation values ​​related to the core analysis object, reflecting the statistical average error direction and magnitude of this core issue.

[0103] Specifically, the sorting logs for the current statistical period are accessed again, but instead of focusing on all data, only records relevant to the core analysis object are extracted. For example, if the core analysis object is workstation 5, all sorting events with workstation 5 as the target workstation are filtered out. Then, the positional deviation values ​​of the execution results corresponding to these filtered events are extracted to form a numerical set. Finally, the arithmetic mean of all values ​​in this set is calculated. The calculated average deviation reveals a systematic deviation pattern: whether it is consistently too early (negative average), consistently too late (positive average), or random fluctuation (average close to zero but large variance).

[0104] In some embodiments, the calculation of the average deviation value in this step can be achieved in a variety of ways: Optionally, a two-step filtering calculation can be used: After identifying the core analysis object (e.g., category ID "SKU-123"), a second traversal of the logs for the current statistical period is performed. While traversing each entry, its category ID is checked to see if it is "SKU-123"; if so, the positional deviation value of the execution result is accumulated into a sum variable, and a counter is incremented. After the traversal is complete, the average value is obtained by dividing the sum by the count value.

[0105] S209. Extract all dynamic operating speeds of the core analysis object within the corresponding statistical period and calculate the average operating speed.

[0106] The average operating speed refers to the arithmetic mean of the instantaneous dynamic operating speeds of the conveyor line when all sorting events related to the core analysis object (the worst-performing workstation or category) occur. It represents the average operating speed of the sorting events that caused the problem.

[0107] Specifically, while S208 focuses on the error results (deviation values), S209 focuses on the input conditions (running speed) that generated these errors. The same data filtering logic as S208 is executed, focusing only on sorting logs that match the core analysis object. For each filtered log, the dynamic running speed value at the initial stage of that sorting (S201) is recorded. After collecting all these speed values, their arithmetic mean is calculated. This average running speed, together with the average deviation value from S208, constitutes a symptom-condition data pair describing the core problem.

[0108] In some embodiments, the average running speed in this step can be calculated in a variety of ways: Optionally, merge the traversal with S208: During the log traversal to calculate the average deviation value, complete two tasks at once; when a log record is identified as related to the core analysis object, the deviation value is simultaneously added to the total deviation variable and the dynamic running speed is added to the total speed variable; after the traversal is completed, divide the two sums by the same count value to obtain the average deviation value and average running speed at once, improving calculation efficiency.

[0109] S210. The ratio of the average deviation value to the average running speed is the running calibration parameter.

[0110] Specifically, the average deviation value obtained from S208 and S209 is linked to the average operating speed. The calculation formula is: Operating calibration parameter = Average deviation value / Average operating speed. The theoretical basis of this calculation is a first-order linear approximation model, which assumes that the positional deviation (Δp) generated at a specific problem (a specific workstation or product category) is proportional to the current operating speed (v), i.e., Δp ≈ k * v. By estimating this proportionality coefficient k using the average value, a calibration parameter that can specifically compensate for this core problem is obtained. For example, if the calculated parameter is -0.05 (unit: seconds), it means that for this core analysis object, the trigger time can be dynamically adjusted according to the current speed during subsequent sorting (Δt = -0.05 * v), thereby systematically eliminating this deviation.

[0111] In some embodiments, the generation of calibration parameters in this step can be achieved in a variety of ways: Optional regression analysis method: Instead of using the mean, this method uses all data pairs (deviation value, running speed) selected from S208 and S209 as sample points; it applies least squares to perform univariate linear regression analysis to fit a straight line Δp = k*v + b; and uses the fitted slope k as the running calibration parameter. This method considers the distribution of all data points, not just the mean, and is therefore generally more robust and accurate when the data dispersion is large.

[0112] It is understandable that other methods can be used to calculate this correction factor, such as using a more complex nonlinear model to fit the relationship between deviation and speed, which is not limited here.

[0113] S211. Based on the operation calibration parameters, adjust the baseline predicted trigger time to the actual predicted trigger time in subsequent sorting operations within the corresponding statistical period.

[0114] Step S211 and Figure 1 Step S106 in the illustrated embodiment is similar and can be found in the description of step S106, which will not be repeated here.

[0115] In the above embodiment, workstation and product category information are first introduced to provide orthogonal dimensions for fault analysis. When the sorting hit rate decreases, by identifying the dimension with the lowest hit rate as the core analysis object, the general performance problem can be pinpointed to the root cause—whether it is a specific workstation or a specific product category. Subsequently, the algorithm only performs statistical analysis on historical data from this core source, calculating the ratio of average deviation to average speed. The resulting operational calibration parameters are dynamically generated for the most prevalent fault mode, rather than being statically corrected based on global data. This achieves adaptive and accurate calibration of the root cause of the fault, improving the system's self-diagnosis and dynamic optimization capabilities.

[0116] The following describes an exemplary sorting management system 300 provided in an embodiment of this application. Figure 3 This is an exemplary hardware structure diagram of the sorting management system 300 provided in this application embodiment.

[0117] In some embodiments, the sorting management system 300 is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can also be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0118] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements. The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0119] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0120] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0121] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A sorting management method powered by an AI model, characterized in that, include: At a preset detection position on the conveyor line, the initial position of the moving target object, the corresponding initial timestamp, and the dynamic running speed of the conveyor line are obtained; Based on the initial position, the initial timestamp, the dynamic running speed, and the preset execution position, calculate the baseline predicted trigger time for the target object to reach the execution position; When the baseline prediction trigger time arrives, a sorting action instruction is generated and sent to the execution mechanism corresponding to the execution position, and the real-time execution result position deviation value is obtained; When the position deviation value of the execution result exceeds the preset instantaneous failure threshold, the target object that failed to be sorted in real time is marked as a dynamic obstacle. Based on the last known obstacle position of the dynamic obstacle and the current real-time dynamic running speed of the conveyor line, the real-time obstacle position trajectory of the dynamic obstacle on the conveyor line is continuously predicted and updated. If the predicted future trajectory of the next target object and the real-time obstacle location trajectory overlap within the preset sorting execution time and space area, the sorting action instruction corresponding to the next target object is canceled until the predicted trajectory of the dynamic obstacle leaves the sorting execution time and space area, and then the sorting action is recalculated for the next target object. Within a preset statistical period, for multiple sorting actions that have been executed, the corresponding execution result position deviation value is continuously obtained and the real-time sorting hit rate within the corresponding statistical period is calculated. When the real-time sorting hit rate is lower than the preset hit rate threshold, running calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value and the corresponding dynamic running speed. Based on the aforementioned operational calibration parameters, the baseline predicted trigger time is adjusted to the actual predicted trigger time in subsequent sorting operations within the corresponding statistical period.

2. The method according to claim 1, characterized in that, When the real-time sorting hit rate is lower than a preset hit rate threshold, operation calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic running speed. Specifically, these parameters include: When the real-time sorting hit rate is lower than the preset hit rate threshold, the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic running speed are input into the preset decision model to generate running calibration parameters. The decision model is obtained by pre-training a neural network framework with multiple sets of historical sorting hit rates, historical execution result position deviation values, historical running speeds, and corresponding historical running calibration parameters.

3. The method according to claim 1, characterized in that, When the real-time sorting hit rate is lower than a preset hit rate threshold, operation calibration parameters are generated based on the real-time sorting hit rate, the real-time execution result position deviation value, and the corresponding dynamic running speed. Specifically, these parameters include: Obtain the category information of the target object and the workstation information of the actuator corresponding to each sorting action; Within the corresponding statistical period, when the real-time sorting hit rate is lower than the preset hit rate threshold, the workstation hit rate and the category hit rate of each independent workstation and each independent category are calculated based on the workstation information and the category information, respectively. Among all the workstation hit rates and product category hit rates, the dimension with the lowest value is determined as the core analysis object; the core analysis object is one of the workstation information or the product category information. Extract the positional deviation values ​​of all execution results corresponding to the core analysis object within the corresponding statistical period, and calculate the average deviation value. Extract all dynamic operating speeds of the core analysis object within the corresponding statistical period, and calculate the average operating speed. The ratio of the average deviation value to the average operating speed is calculated as the operating calibration parameter.

4. The method according to claim 3, characterized in that, After determining the dimension with the lowest numerical value among all the workstation hit rates and the product category hit rates as the core analysis object, the analysis also includes: When the lowest value in the workstation hit rate is the same as or the absolute value of the difference between the lowest value in the category hit rate is less than the preset coupling judgment threshold, the lowest workstation with the lowest workstation hit rate and the lowest category with the lowest category hit rate will be combined together as a coupling analysis object. Extract all position deviation values ​​of the coupling execution results generated by the coupling analysis object within the corresponding statistical period, and combine them with the corresponding coupling dynamic running speed to form a deviation speed data pair; Based on all the aforementioned deviation speed data pairs, calculate the sum of all coupled dynamic running speeds, the sum of all coupled execution result position deviation values, the sum of the squares of all coupled dynamic running speeds, and the sum of the product of speed and deviation within each of the aforementioned deviation speed data pairs. Substitute these values ​​and the total number of the aforementioned deviation speed data pairs into a preset least squares slope calculation formula to obtain the corrected running calibration parameters. Use the corrected running calibration parameters to replace the original running calibration parameters.

5. The method according to claim 2, characterized in that, The step of adjusting the baseline predicted trigger time to the actual predicted trigger time in subsequent sorting operations within the corresponding statistical period, based on the aforementioned operational calibration parameters, specifically includes: Obtain the running calibration parameters and the baseline predicted trigger time of the target object; The dynamic operating speed of the conveyor line during the current sorting action window is collected and stored at a preset collection frequency, which is the current speed and the speed of the next adjacent previous moment. The instantaneous speed change is multiplied by the running calibration parameter to generate a dynamic compensation time value; the instantaneous speed change is the difference between the current speed and the speed at the previous moment. The actual predicted trigger time is obtained by subtracting the dynamic compensation time value from the baseline predicted trigger time, and sorting action instructions are generated and issued based on the actual predicted trigger time.

6. The method according to claim 5, characterized in that, After multiplying the instantaneous velocity change by the operational calibration parameters to generate the dynamic compensation time value, the method further includes: When the current target object belongs to a preset pneumatically sensitive category and the instruction type of the immediately preceding sorting action is pneumatic sorting, a preset pneumatic disturbance factor bound to the pneumatically sensitive category is additionally obtained from the decision model. The aerodynamic disturbance factor is summed with the dynamic compensation time value to obtain the final compensation time value, and the final compensation time value replaces the dynamic compensation time value.

7. A sorting management system, characterized in that, The sorting management system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the sorting management system to perform the method as described in any one of claims 1-6.

8. A computer program product containing instructions, characterized in that, When the computer program product is run on the sorting management system, the sorting management system performs the method as described in any one of claims 1-6.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the sorting management system, the sorting management system performs the method as described in any one of claims 1-6.

Images