Method, device and system for scheduling and controlling a production line

Abstract

The application provides a production line scheduling and control method, device and system. In the application, load state data of a target station is collected, and the load state data is synchronized to adjacent stations of the target station to establish a shared state between the target station and the adjacent stations; a motion adjustment strategy of the target station is determined based on the shared state; a motion control instruction including a motion speed and a speed curve is generated according to the motion adjustment strategy, and a robot of the target station is controlled to execute the motion control instruction; and a coordination instruction is sent to the adjacent stations of the target station, so that the adjacent stations of the target station adjust motion control signals of robots of the adjacent stations of the target station based on the coordination instruction. The application solves the problem of low scheduling and control efficiency of the existing production line.

Description

Technical Field

[0001] This application relates to the field of industrial automation control technology, and more specifically, to a production line scheduling and control method, a production line scheduling and control device, and a production line scheduling and control system. Background Technology

[0002] Traditional automated production lines often employ fixed cycle times or simple logic control, making it difficult to handle the dynamic fluctuations in multi-product mixed-flow production. When a workstation experiences a brief delay, such as due to workpiece abnormalities, sensor noise, or a sudden increase in load, upstream and downstream equipment continues to operate as planned, easily leading to material accumulation, idle waiting, or even line blockage and shutdown. Existing scheduling methods rely on a central host computer periodically issuing instructions, resulting in poor real-time performance and a lack of dynamic optimization of the collaborative postures between robots, thus limiting overall production efficiency. Summary of the Invention

[0003] The main objective of this application is to provide a production line scheduling and control method, a production line scheduling and control device, and a production line scheduling and control system, so as to at least solve the problem of low scheduling and control efficiency of production lines in the prior art.

[0004] To achieve the above objectives, according to one aspect of this application, a method for scheduling and controlling a production line is provided. The method is applied to an automated production line including multiple workstations and multiple robots. The method includes: collecting load status data of a target workstation and synchronizing the load status data to adjacent workstations to establish a shared state between the target workstation and the adjacent workstations; determining a motion adjustment strategy for the target workstation based on the shared state; generating motion control instructions including motion speed and acceleration / deceleration curves according to the motion adjustment strategy, and controlling the robot at the target workstation to execute the motion control instructions; and sending coordination instructions to adjacent workstations of the target workstation, causing the adjacent workstations of the target workstation to adjust the motion control signals of their robots based on the coordination instructions.

[0005] Optionally, when multiple robots are located within the working area of ​​the same target workstation, the method includes: generating a predicted path for each robot based on the motion adjustment strategy, wherein the predicted path represents the motion trajectory within a preset time period in the future; exchanging the predicted paths of each robot with other robots within the target workstation; generating a collision-free interpolation trajectory for each robot based on the exchanged predicted paths using a path planning algorithm, and controlling each robot to move along the collision-free interpolation trajectory.

[0006] Optionally, before generating motion control commands including motion speed and acceleration / deceleration curves according to the motion adjustment strategy, the method further includes: predicting whether the target workstation will experience a processing timeout based on the load state data; if the processing timeout is predicted to occur at the target workstation, transmitting adjustment data to the adjacent workstations of the target workstation in advance for a preset time period, so that the adjacent workstations of the target workstation adjust the operating state of their respective workstations according to the adjustment data; wherein, the adjustment data includes a gripping offset and / or a conveyor belt speed adjustment coefficient.

[0007] Optionally, the load status data includes visual recognition time. Based on the load status data, predicting whether the target workstation will experience a processing timeout includes: obtaining the visual recognition time of the target workstation, whereby the visual recognition time represents the duration of processing the current frame image; predicting the time required to process the next frame image based on the visual recognition time; and determining that the target workstation will experience a processing timeout when the predicted time required to process the next frame image exceeds a preset time threshold.

[0008] Optionally, determining the motion adjustment strategy for the target workstation based on the shared state includes: calculating the comprehensive load coefficient of the target workstation according to the load state data; determining the target action using a pre-trained reinforcement learning model based on the comprehensive load coefficient and the shared state, and generating an execution instruction for the motion adjustment strategy according to the target action. The pre-trained reinforcement learning model is trained using multiple sets of data through machine learning. Each set of data includes: a state input and a label corresponding to the state input. The state input includes the comprehensive load coefficient of the target workstation and the shared state. The label represents at least one of the following control behaviors: adjusting the motion speed of the robot at the target workstation; issuing a speed curve to control the acceleration and deceleration process of the robot at the target workstation; generating a control signal to be sent to adjacent workstations of the target workstation, the control signal including a deceleration factor or a waiting compensation value.

[0009] Optionally, based on the comprehensive load coefficient and the shared state, a pre-trained reinforcement learning model is used to determine the target action, including: discretizing the comprehensive load coefficient and the shared state to obtain the current state; obtaining an action value mapping table, which is established by the reinforcement learning model through trial and error learning during the training phase and is used to store the mapping relationship between the discretized state and the action value of the corresponding candidate action; querying the action value mapping table, filtering the action values ​​of all the candidate actions corresponding to the current state to obtain multiple candidate action values; and determining the candidate action corresponding to the maximum value among the multiple candidate action values ​​as the target action.

[0010] Optionally, calculating the comprehensive load coefficient of the target workstation based on the load status data includes: acquiring the real-time sampled values ​​of each sensor index and the corresponding nominal reference values ​​in the load status data, wherein the load status data includes the workpiece arrival frequency, joint torque, and visual recognition time collected by each sensor; calculating the ratio between the real-time sampled values ​​of each sensor index and the corresponding nominal reference values ​​to obtain the normalized load ratio of each sensor index; inputting the normalized load ratio of each sensor index into a saturation limiting function to obtain the limiting value of each sensor index; and weighting and summing the limiting values ​​according to the dynamic weights of each sensor index to obtain the comprehensive load coefficient.

[0011] Optionally, synchronizing the load status data to adjacent workstations of the target workstation to establish a shared state between the target workstation and the adjacent workstations includes: sending the load status data to the adjacent workstations of the target workstation; receiving the load status data of the adjacent workstations from the adjacent workstations; constructing the shared state by combining the load status data of the adjacent workstations and the load status data of the target workstation, and storing the shared state in the local storage unit of the target workstation.

[0012] According to another aspect of this application, a production line scheduling and control device is provided, comprising: a data acquisition unit, configured to acquire load status data of a target workstation and synchronize the load status data to adjacent workstations of the target workstation to establish a shared state between the target workstation and the adjacent workstations; a determination unit, configured to determine a motion adjustment strategy for the target workstation based on the shared state; a first generation unit, configured to generate motion control instructions including motion speed and acceleration / deceleration curves according to the motion adjustment strategy, and control the robot of the target workstation to execute the motion control instructions; and a sending unit, configured to send a coordination instruction to the adjacent workstations of the target workstation, so that the adjacent workstations of the target workstation adjust the motion control signals of the robots of the adjacent workstations of the target workstation based on the coordination instruction.

[0013] According to another aspect of this application, a production line scheduling and control system is provided, the system comprising: multiple edge controllers deployed at each workstation, the edge controllers being used to execute any of the production line scheduling and control methods described above; and multiple robots disposed at each of the workstations, the robots being communicatively connected to the edge controllers at their respective workstations, and being used to execute motion control signals issued by the edge controllers.

[0014] By applying the technical solution of this application, load status data of the target workstation is collected and synchronized to adjacent workstations to establish a shared state between the target workstation and its neighbors. Based on this shared state, a motion adjustment strategy for the target workstation is determined. According to the motion adjustment strategy, motion control commands, including motion speed and acceleration / deceleration curves, are generated, and the robot at the target workstation is controlled to execute these commands. Coordination commands are sent to adjacent workstations to enable them to adjust the motion control signals of their respective robots. In this solution, by collecting load status data of the target workstation and synchronizing it to its neighbors to establish a shared state, dynamically determining the motion adjustment strategy and generating corresponding motion control commands based on this shared state, and simultaneously sending coordination commands to adjacent workstations to adjust their operating signals, coordination and adaptive adjustment between workstations are achieved. This effectively eliminates global cycle mismatch caused by local load fluctuations, reduces waiting time and energy consumption during multi-robot collaboration, and thus solves the problem of low scheduling and control efficiency in existing production lines. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0016] Figure 1 A flowchart illustrating a production line scheduling and control method according to an embodiment of this application is shown.

[0017] Figure 2 An architecture diagram of a production line scheduling and control system provided according to an embodiment of this application is shown;

[0018] Figure 3 A flowchart illustrating the dynamic cycle time adjustment of a production line scheduling and control method according to an embodiment of this application is shown.

[0019] Figure 4 A schematic diagram of a multi-robot cooperative trajectory provided according to an embodiment of this application is shown;

[0020] Figure 5 An anomaly pre-compensation timing diagram of a production line scheduling and control method provided according to an embodiment of this application is shown;

[0021] Figure 6 A structural block diagram of a production line scheduling and control device according to an embodiment of this application is shown. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] As described in the background section, the scheduling and control of production lines in the prior art rely on the periodic issuance of instructions by a central host computer, which has poor real-time performance and lacks dynamic optimization of the collaborative postures between robots, resulting in limited overall production efficiency. In order to solve the problem of low scheduling and control efficiency of production lines in the prior art, the embodiments of this application provide a scheduling and control method, a scheduling and control device, and a scheduling and control system for production lines.

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0027] Figure 1 This is a flowchart illustrating a production line scheduling and control method according to an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0028] Step S101: Collect the load status data of the target workstation and synchronize the load status data to the adjacent workstations of the target workstation to establish a shared status between the target workstation and the adjacent workstations.

[0029] Specifically, in this embodiment, each workstation is equipped with an edge computing unit, which performs the execution. The load status data of the target workstation (i.e., the current workstation) is collected in real time. This load status data reflects the current operating load of the workstation, including key indicators such as workpiece arrival frequency, robot joint torque, and visual recognition time. To achieve distributed collaborative control, the edge computing unit of the target workstation synchronously sends the collected load status data to its upstream and downstream neighboring workstations via a low-latency communication link (such as Wi-Fi 6 Mesh or 5G URLLC machine-to-machine communication). After receiving this data, the neighboring workstations combine it with their own load status data, thereby establishing a shared state containing operational status information from both sides. This mechanism allows each workstation to not only perceive its own local load but also receive real-time load fluctuations from neighboring workstations, providing a data foundation for subsequent collaborative decision-making.

[0030] Step S102: Determine the motion adjustment strategy for the target workstation based on the shared state described above.

[0031] Specifically, after establishing a shared state, the current shared state is used as input to determine a motion adjustment strategy through a preset local control strategy or model. This strategy aims to solve the production cycle mismatch problem caused by local load fluctuations (such as processing delays or sudden load increases at a certain workstation). Based on the upstream and downstream pressure situation reflected in the shared state (such as upstream backlog or downstream idleness), it determines how the target workstation should adjust its operating rhythm, such as deciding whether to accelerate through, decelerate and wait, or maintain the current cycle, thereby generating a specific motion adjustment strategy. This process does not require periodic instructions from a central host computer, achieving rapid autonomous decision-making based on local information.

[0032] Step S103: Based on the above motion adjustment strategy, generate motion control instructions including motion speed and acceleration / deceleration curves, and control the robot at the target workstation to execute the above motion control instructions.

[0033] Specifically, after determining the motion adjustment strategy, it is translated into specific executable instructions. These motion control instructions include not only the final speed of the target robot but also detailed acceleration and deceleration curves (such as S-shaped velocity curves). Generating these curves aims to prevent the robot from abruptly stopping and starting due to sudden acceleration or deceleration commands, thereby reducing mechanical shock, protecting the equipment, and extending its lifespan. The target workstation sends these motion control instructions to the robot at the target workstation via a hard real-time motion bus (such as EtherCAT). The robot then smoothly adjusts its trajectory and speed according to these instructions to execute the current production task.

[0034] Step S104: Send a coordination instruction to the adjacent workstations of the target workstation so that the adjacent workstations of the target workstation adjust the motion control signals of the robots of the adjacent workstations of the target workstation based on the coordination instruction.

[0035] Specifically, to maintain the synchronization of the entire production line, the target workstation, while adjusting its own operation, also needs to issue coordination instructions to adjacent workstations. These instructions include the target workstation's adjustment requirements for the adjacent workstations' operating states (e.g., requesting the upstream to slow down the material supply speed, or requesting the downstream to extend the waiting time). Upon receiving the coordination instructions, the adjacent workstations parse the adjustment parameters and adjust the motion control signals of their corresponding robots accordingly (e.g., reducing conveyor belt speed, adjusting gripping timing, or changing movement speed). Based on this coordination mechanism, each workstation can respond to local load fluctuations in real time, preventing disturbances from accumulating along the production line, thereby ensuring the continuity of the production process and overall operating efficiency.

[0036] In this embodiment, by collecting the load status data of the target workstation and synchronizing it with its adjacent workstations to establish a shared state, the motion adjustment strategy is dynamically determined based on the shared state and corresponding motion control commands are generated. At the same time, collaborative commands are sent to adjacent workstations to adjust their running signals in a coordinated manner, thereby realizing the collaboration and adaptive adjustment between workstations. This effectively eliminates the global cycle mismatch caused by local load fluctuations, reduces the waiting time and energy consumption in the multi-robot collaboration process, and solves the problem of low scheduling and control efficiency in existing production lines.

[0037] In some embodiments of this application, when multiple robots are located within the working area of ​​the same target workstation, the method includes: generating a predicted path for each robot based on the motion adjustment strategy, wherein the predicted path represents the motion trajectory within a preset time period in the future; exchanging the predicted paths of each robot with other robots within the target workstation; generating a collision-free interpolation trajectory for each robot using a path planning algorithm based on the exchanged predicted paths, and controlling each robot to move along the collision-free interpolation trajectory.

[0038] Specifically, when multiple robots are located within the same target workstation and need to work collaboratively, to avoid collisions and optimize motion efficiency, the first step is to independently generate the future motion trajectory for each robot within the workstation based on the currently determined motion adjustment strategy. The predicted path characterizes the robot's possible motion trend and spatiotemporal distribution within a preset time period (e.g., 0.3 seconds). This step aims to predict the spatial occupancy of each robot within a short time window, providing input data for subsequent collaborative obstacle avoidance calculations.

[0039] To achieve cooperative obstacle avoidance, each robot in the workstation broadcasts or sends its generated predicted path to other robots in the same workstation via a local communication network. Through this exchange of path information, each robot can receive the expected motion trajectories of other robots in the same workstation within a preset time period. This allows each robot to no longer plan its own path in isolation, but to make comprehensive judgments based on short-term motion information from the global perspective (i.e., all robots in the same workstation), thus laying the foundation for generating collision-free cooperative trajectories.

[0040] After acquiring the predicted paths of all relevant robots, each robot utilizes its locally deployed path planning algorithm. Based on its own predicted path and combined with predicted paths exchanged from other robots as dynamic obstacle constraints, it calculates and generates a new collision-free interpolation trajectory in real time. When generating the trajectory, the path planning algorithm comprehensively considers factors such as avoiding spatial collisions and optimizing energy consumption or time, ensuring that the generated trajectory neither conflicts with the predicted paths of other robots nor fails to meet kinematic smoothness requirements. Ultimately, each robot is controlled to execute actions along its respective collision-free interpolation trajectory, thereby achieving safe and efficient collaborative movement within a shared workspace.

[0041] By generating and exchanging predicted paths within a preset time period in real time in a multi-robot scenario at the same workstation, and using a path planning algorithm to dynamically generate collision-free interpolation trajectories based on the exchanged path information, collaborative obstacle avoidance and path optimization among robots are achieved. This mechanism effectively replaces the traditional fixed teaching program, eliminates the waiting and collision risks caused by trajectory conflicts, and reduces motion energy consumption and improves overall work efficiency while ensuring safe collaborative operation of multiple robots.

[0042] In some embodiments of this application, before generating motion control commands including motion speed and acceleration / deceleration curves according to the above motion adjustment strategy, the method further includes: predicting whether the target workstation will experience a processing timeout based on the above load state data; if the processing timeout is predicted to occur at the target workstation, transmitting adjustment data to the adjacent workstations of the target workstation in advance for a preset time period, so that the adjacent workstations of the target workstation adjust the operating state of their respective workstations according to the adjustment data; wherein the adjustment data includes a gripping offset and / or a conveyor belt speed adjustment coefficient.

[0043] Specifically, before generating motion control commands, the currently collected load status data is analyzed and its trends are predicted in real time. Load status data reflects the real-time operating load of the workstation, such as visual recognition time, joint torque fluctuations, or workpiece arrival frequency. By monitoring the changing trends of these key indicators, it is determined whether the current processing speed lags behind the production cycle requirements, thereby predicting whether the target workstation will experience processing timeouts (i.e., inability to complete the current process within the specified cycle) in the near future. This predictive step aims to identify potential bottlenecks or anomalies in advance, providing a basis for subsequent proactive intervention decisions.

[0044] When a processing timeout is predicted at the target station (e.g., a predicted increase of several milliseconds in vision processing), to prevent this delay from propagating throughout the production line and causing blockages or idle runs, adjustment data is transmitted to the upstream and downstream stations of the target station within a specific time window (i.e., a preset time period, such as 0.5 seconds) before the timeout occurs. The preset time period typically corresponds to the physical time window for material transfer between stations, ensuring that the adjustment actions of adjacent stations are synchronized with the actual disturbance time of the target station. Through this feedforward control method, adjacent stations can respond in advance, achieving a smooth, uninterrupted transition of the production line. Adjustment data consists of parameters used to guide the specific adjustment actions of adjacent stations. For upstream stations, adjustment data may include a gripping offset, used to fine-tune the gripping position or timing to accommodate possible delays or positional changes at the target station; for downstream stations, adjustment data may include a conveyor belt speed adjustment coefficient, used to fine-tune the conveyor belt's running speed, thereby extending or shortening the workpiece's dwell time at the downstream station, providing buffer time for processing at the upstream station. By sending these specific adjustment parameters, adjacent workstations can quickly and accurately adjust the motion control signals or actuator states of their robots without the need for complex replanning, ensuring the rhythm matching and coordinated operation of the entire production line.

[0045] By predicting the processing timeout risk of the target station based on load status data, and transmitting adjustment data, including the grab offset and / or conveyor belt speed adjustment coefficient, to adjacent stations before the timeout occurs using a preset time window, proactive feedforward pre-compensation for local disturbances is achieved. This mechanism enables upstream and downstream stations to synchronize and fine-tune their operating status in advance, ensuring that the adjustment action is precisely matched with the arrival time of the disturbance, thereby effectively eliminating the impact of local delays on the overall production cycle time and achieving uninterrupted smooth transition and millisecond-level rapid response of the production line.

[0046] In some specific embodiments of this application, the load status data includes visual recognition time. Based on the load status data, predicting whether the target workstation will experience a processing timeout includes: obtaining the visual recognition time of the target workstation, where the visual recognition time represents the duration of processing the current frame image; predicting the time required to process the next frame image based on the visual recognition time; and determining that the target workstation will experience a processing timeout when the predicted time required to process the next frame image exceeds a preset time threshold.

[0047] Specifically, visual recognition time refers to the time consumed from receiving the current frame of image data to completing image analysis, feature extraction, and result output. This data directly reflects the computational load and operating status of the current vision processing module and is a key indicator for evaluating the processing efficiency of the workstation. This time consumption data is recorded at a specific sampling period (e.g., per workpiece or per frame of image) as the baseline input for subsequent predictive analysis. Using the real-time acquired current visual recognition time consumption data, combined with historical data trends or a preset prediction model, the time required for processing the next frame of image (i.e., the next workpiece) is estimated. Since visual recognition time may dynamically change due to factors such as workpiece surface reflection, positional deviation, lighting changes, or system load fluctuations, prediction based on the current time consumption can predict in advance whether a performance bottleneck will occur in the next frame processing process. This prediction mechanism transforms discrete instantaneous time consumption into a continuous estimate of future processing cycles, improving the sensitivity to potential delays. The predicted time required to process the next frame of image is compared with the nominal cycle time specified by the production line (i.e., the preset time threshold). If the predicted processing time exceeds the threshold, it means that the target workstation cannot complete the visual recognition and subsequent actions of the current workpiece within the specified time, thus determining that the target workstation is about to experience a processing timeout. This determination triggers a subsequent anomaly pre-compensation mechanism, which allows adjustment data to be sent to adjacent workstations in advance within a preset time period before the actual timeout occurs, in order to avoid production line mismatch caused by visual processing delays.

[0048] By acquiring the visual recognition time in real time and predicting the time required to process the next frame of image, it is possible to accurately identify potential timeout risks caused by fluctuations in visual processing performance. When the predicted time exceeds a preset threshold, an abnormal pre-compensation mechanism can be triggered in advance, allowing upstream and downstream workstations to make synchronous fine adjustments before the disturbance actually occurs. This effectively avoids production line blockage or cycle mismatch caused by visual processing delays, achieving refined prediction and uninterrupted smooth transition for visual load fluctuations.

[0049] In some embodiments of this application, determining the motion adjustment strategy for the target workstation based on the shared state includes: calculating the comprehensive load coefficient of the target workstation based on the load state data; determining the target action using a pre-trained reinforcement learning model based on the comprehensive load coefficient and the shared state; and generating an execution instruction for the motion adjustment strategy based on the target action. The pre-trained reinforcement learning model is trained using multiple sets of data through machine learning. Each set of data includes: a state input and a label corresponding to the state input. The state input includes the comprehensive load coefficient of the target workstation and the shared state. The label represents at least one of the following control behaviors: adjusting the motion speed of the robot at the target workstation; issuing a speed curve to control the acceleration and deceleration process of the robot at the target workstation; and generating a control signal to be sent to adjacent workstations of the target workstation. The control signal includes a deceleration factor or a waiting compensation value.

[0050] In some specific embodiments of this application, calculating the comprehensive load coefficient of the target workstation based on the aforementioned load status data includes: acquiring the real-time sampled values ​​of each sensor index and the corresponding nominal reference values ​​in the aforementioned load status data, wherein the aforementioned load status data includes the workpiece arrival frequency, joint torque, and visual recognition time collected by each sensor; calculating the ratio between the real-time sampled values ​​of each of the aforementioned sensor indices and the corresponding nominal reference values ​​to obtain the normalized load ratio of each of the aforementioned sensor indices; inputting the normalized load ratio of each of the aforementioned sensor indices into the saturation limiting function to obtain the limiting value of each of the aforementioned sensor indices; and weighting and summing the limiting values ​​according to the dynamic weights of each of the aforementioned sensor indices to obtain the aforementioned comprehensive load coefficient.

[0051] Specifically, the load status data includes the following key indicators: first, the workpiece arrival frequency detected by photoelectric sensors, reflecting the production cycle rhythm; second, the joint torque data collected by robot joint torque sensors, reflecting the load pressure of the mechanical actuator; and third, the visual recognition time recorded by the vision processing board, reflecting the computational load of the computing unit. Simultaneously, the system stores nominal reference values ​​for each indicator under ideal or standard operating conditions (i.e., normal baseline values, such as nominal visual recognition time of 20ms, nominal joint torque threshold, etc.), serving as a benchmark for calculating load deviations.

[0052] To eliminate the influence of different physical dimensions on load assessment, the real-time sampled value of each sensor indicator is compared with its corresponding nominal reference value. For example, if the current visual recognition time is 30ms and the nominal reference value is 20ms, the ratio is 1.5. This ratio is the normalized load ratio, which characterizes the degree of deviation of the current indicator from the standard state. When the ratio is greater than 1, it indicates that the load of the indicator is higher than the standard state; when it is equal to 1, it indicates that it is in the standard state; and when it is less than 1, it indicates that the load is lower than the standard state. This step achieves the standardization of multi-source heterogeneous data, making it comparable.

[0053] Considering the potential for significant load fluctuations under extreme and abnormal operating conditions, direct accumulation could distort the overall load factor or lead to overly aggressive control strategies. Therefore, a saturation limiting function is introduced to process the normalized load ratio. This function sets upper and lower thresholds (e.g., limiting the load ratio between 0.5 and 2.0). Regardless of the actual value of the normalized load ratio, the output limit value is confined within a preset reasonable range after processing by the saturation limiting function. This step prevents extreme outliers of a single indicator from dominating the overall load assessment, ensuring that the load factor remains within a controllable range.

[0054] After obtaining the limiting values ​​for each sensor indicator, the limiting values ​​are weighted and summed according to a preset dynamic weighting coefficient to calculate the final comprehensive load coefficient. Specifically, the formula for calculating the comprehensive load coefficient is as follows: ,in, For the comprehensive load factor, The dynamic weight of the i-th sensor index, It is a saturation limiting function. Let i be the real-time sampled value of the i-th sensor indicator. Let be the nominal reference value of the i-th sensor indicator. This is the normalized load ratio.

[0055] Dynamic weights can be adjusted based on the current production task type or historical fault data. For example, in the gluing process, joint torque is given a higher weight. Through weighted summation, load information from multiple dimensions, such as workpiece arrival frequency, joint torque, and visual recognition time, is integrated into a single, quantified comprehensive load coefficient. This coefficient can comprehensively and objectively reflect the real-time comprehensive load status of the target workstation, providing an accurate basis for the state input of the subsequent reinforcement learning model.

[0056] By acquiring real-time sampled values ​​and nominal reference values ​​of multi-dimensional sensor indicators such as workpiece positioning frequency, joint torque, and visual recognition time, a normalized load ratio is calculated and subjected to saturation limiting processing. This effectively eliminates the influence of different physical dimensions and suppresses the interference of extreme outliers on the evaluation results. A weighted summation based on dynamic weights yields the comprehensive load coefficient. This mechanism achieves the quantification and standardized representation of local load fluctuations at the workstation, enabling the comprehensive load coefficient to more sensitively and accurately reflect the actual operating load state of the workstation, providing a stable and highly discriminative state input for subsequent reinforcement learning models.

[0057] In some other specific embodiments of this application, based on the aforementioned comprehensive load coefficient and the aforementioned shared state, a pre-trained reinforcement learning model is used to determine the target action, including: discretizing the aforementioned comprehensive load coefficient and the aforementioned shared state to obtain the current state; obtaining an action value mapping table, wherein the aforementioned action value mapping table is established by the reinforcement learning model through trial and error learning during the training phase, and is used to store the mapping relationship between the discretized state and the action value of the corresponding candidate action; querying the aforementioned action value mapping table, filtering the action values ​​of all the aforementioned candidate actions corresponding to the aforementioned current state to obtain multiple candidate action values; and determining the candidate action corresponding to the maximum value among the multiple aforementioned candidate action values ​​as the aforementioned target action.

[0058] Specifically, to accommodate table-based reinforcement learning methods (such as Q-learning), the calculated overall load coefficient and shared states obtained from adjacent workstations (such as load levels and waiting signals of upstream and downstream workstations) are mapped to a predefined discrete state space. For example, the overall load coefficient can be divided into several intervals such as low load, normal load, and high load, and the shared states can be divided into categories such as upstream idle, upstream backlog, downstream idle, and downstream busy. Through this discretization, continuous real-time operating data is transformed into a finite number of discrete state indices, thereby constructing the current state that the reinforcement learning model can identify and process. This step reduces the dimensionality of the state space, keeping the size of the action value mapping table within a manageable range, facilitating rapid retrieval and learning.

[0059] The aforementioned action-value mapping table, commonly referred to as the Q-Table in Q-learning-based implementations, is the core data structure of the reinforcement learning model. The Q-Table is a two-dimensional data structure where rows correspond to the discretized state space (i.e., the current state) and columns correspond to the action space (i.e., all possible candidate actions). Each element in the Q-Table (i.e., the Q-value) stores the long-term cumulative reward expected to be obtained by performing a specific action in a specific state. This table is established and dynamically updated through a trial-and-error process involving the agent and the environment during offline training or online learning. It reflects the merits of different control strategies under various operating conditions, ensuring that the decision-making process no longer relies on human experience or fixed rules, but rather on more suitable strategies learned from history.

[0060] Based on the current state, a lookup is performed in the action value mapping table. Specifically, all entries in the Q table corresponding to the row containing the current state are located. These entries represent the action value (Q value) corresponding to performing each predefined candidate action (such as maintaining the current speed, decelerating by a preset ratio, pausing and waiting, etc.) in this state. By reading these Q values, a list of expected benefits for all available control actions under the current operating condition is obtained, i.e., the values ​​of multiple candidate actions.

[0061] To make the optimal decision under the current conditions, the values ​​of multiple candidate actions are compared. In one specific embodiment, the candidate action with the highest value is selected as the target action. For example, if the current state shows high load and downstream busyness, and the action "decelerate and send a waiting signal" in the Q table has the highest Q value, then this action is determined as the target action. This target action is then translated into specific control instructions (such as adjusting the robot speed curve, generating a collaborative broadcast signal, etc.), and executed to adjust the production cycle. In this way, the optimal control strategy can be dynamically selected based on real-time load fluctuations and upstream and downstream collaborative status, achieving adaptive optimization of the production line.

[0062] By discretizing the comprehensive load coefficient and shared state into the current state, and querying the action value mapping table generated by the pre-trained reinforcement learning model for all candidate action values ​​in the corresponding state, the action with the highest value is selected as the target action. Leveraging decision-making experience accumulated through historical trial-and-error learning, rapid and deterministic optimal decision-making based on multidimensional load characteristics and upstream / downstream collaborative states is achieved within millisecond-level real-time control cycles. This avoids the latency caused by online search or complex optimization calculations, ensuring the production line's immediate response capability and control stability under dynamic disturbances.

[0063] The pre-trained reinforcement learning model is trained on a large amount of historical production data. Its training process uses multiple sets of data samples, each containing specific state inputs (i.e., the combined load coefficient and shared state) and corresponding optimal control labels. The labels represent a set of validated optimal control actions under specific states, aiming to balance local efficiency and global coordination. Specifically, the control behaviors represented by the labels include: adjusting the robot's speed at the target workstation to adapt to the current load level; issuing smooth speed curves (such as S-curves) to control the robot's acceleration and deceleration, avoiding sudden stops and starts that could impact the mechanical structure and reduce energy consumption; and generating cooperative control signals to be sent to adjacent workstations, such as sending a deceleration factor upstream to slow down the feeding speed, or sending a waiting compensation value downstream to extend the dwell time. Through learning and optimization by the reinforcement learning model, these control behaviors enable the target workstation to make appropriate decisions that balance its own operational stability and upstream / downstream coordination efficiency without human intervention.

[0064] By calculating the comprehensive load coefficient reflecting the multidimensional operating load of the workstation in real time and combining the shared state of upstream and downstream, the system uses a reinforcement learning model trained with multiple sets of state and label data to accurately predict and output the optimal target action. This generates motion adjustment strategies, including robot speed adjustment, acceleration and deceleration curve distribution, and collaborative signals between adjacent workstations. This mechanism realizes closed-loop decision-making from local load perception to global collaborative control. It not only eliminates the response lag of fixed cycle time to dynamic fluctuations, but also enables millisecond-level uninterrupted smooth transition of each workstation on the production line through feedforward collaborative instructions, thereby improving the overall efficiency, flexibility, and energy consumption level in mixed-flow production environments.

[0065] In some embodiments of this application, synchronizing the load status data to the adjacent workstations of the target workstation to establish a shared state between the target workstation and the adjacent workstations includes: sending the load status data to the adjacent workstations of the target workstation; receiving the load status data of the adjacent workstations from the adjacent workstations; constructing the shared state by the load status data of the adjacent workstations and the load status data of the target workstations; and storing the shared state in the local storage unit of the target workstation.

[0066] Specifically, to achieve distributed collaborative control on the production line, after calculating the load status data of the current workstation, this data is broadcast or unicast to adjacent workstations upstream and downstream via a low-latency communication network. In one specific embodiment, direct machine-to-machine communication links are established between workstations via Wi-Fi 6 Mesh self-organizing network modules or 5G URLLC (Ultra-Reliable Low-Latency Communication) modules. This mechanism breaks the limitations of traditional centralized scheduling that relies on periodic polling from a central supervisory unit, enabling adjacent workstations to obtain near real-time information about the target workstation's operating load status (e.g., determining whether material accumulation is imminent upstream or whether downstream has receiving capacity), thereby providing the necessary data foundation for subsequent collaborative decision-making.

[0067] Meanwhile, the target workstation also receives load status data from its upstream and downstream neighboring workstations via the same low-latency communication link. For example, when the target workstation is workstation B, it will receive real-time load information from upstream workstation A (such as the loading workstation) and downstream workstation C (such as the pressing workstation). The received data also includes key indicators such as the workpiece arrival frequency, joint torque, visual recognition time, and overall load coefficient of neighboring workstations. Through bidirectional data exchange, the target workstation not only understands its own local state but also grasps the global context closely related to it, namely the operating rhythm, potential bottlenecks, and load pressure of neighboring workstations.

[0068] The received load status data from adjacent workstations are fused and integrated with the load status data of the current workstation to construct a shared state. This shared state is a comprehensive dataset that includes not only the real-time load characteristics of the target workstation itself, but also the load characteristics and predictive information (such as short-term trajectory prediction points or load trends of adjacent workstations) of its upstream and downstream neighboring workstations. After construction, this shared state data is temporarily stored in local storage units (such as edge controller memory or cache). The purpose of storing the shared state is to provide complete and multi-dimensional state input for subsequent steps to determine the target action based on the reinforcement learning model. Based on this shared state with a global perspective, the reinforcement learning model can more accurately evaluate the impact of the current decision on the entire production line, thereby selecting the optimal control strategy that optimizes the efficiency of the current workstation while taking into account upstream and downstream coordination.

[0069] By achieving real-time bidirectional synchronization of load status data between workstations, interconnection and interoperability of operational information between adjacent workstations are realized. This allows the target workstation to monitor the load status of upstream and downstream processes in real time, thereby constructing a shared state that includes both local and global information. This provides comprehensive state input for the reinforcement learning model, enabling control decisions to take into account upstream and downstream collaboration. It effectively avoids line congestion or waiting problems caused by information asymmetry, and improves the overall collaborative efficiency and stability of the production line in a dynamically fluctuating environment.

[0070] To enable those skilled in the art to better understand the technical solution of this application, the implementation process of the production line scheduling and control method of this application will be described in detail below with reference to specific embodiments.

[0071] This embodiment relates to a specific production line scheduling and control method, taking an air conditioner compressor housing assembly line as an example. Specifically, this embodiment uses a mixed-flow production line with workstations A (loading robot), B (gluing robot), and C (pressing robot) arranged sequentially as an example. The edge controller, communication module, actuator, sensing system, and upper-level MES of this production line are described below. Regarding the edge controller, each workstation is equipped with a lightweight edge computing unit based on ARM architecture, running a Linux operating system, and has a built-in Q-learning inference framework and RRT-Connect path planning library. Regarding the communication module, the workstations are interconnected through a Wi-Fi 6 Mesh self-organizing network module, supporting direct machine-to-machine (M2M) communication with microsecond-level latency. The actuator consists of a six-axis industrial robot, a conveyor belt, and a vision recognition system (including an industrial camera and image processing board). The sensing system consists of robot joint torque sensors and workpiece positioning photoelectric sensors. The upper-level MES is only used for issuing production plans and offline statistics, and does not participate in millisecond-level real-time closed-loop control. The robot communicates with the edge controller via an EtherCAT wired motion bus (hard real-time), while the edge controllers communicate with each other via a Wi-Fi 6 Mesh wireless direct connection (soft real-time collaboration). Vision / sensors are connected to the edge controller at this workstation via CANopen. All devices are selected around the "ARM+Linux+ROS2 / Docker" ecosystem to ensure seamless deployment of Q-learning and RRT-Connect algorithms.

[0072] To further illustrate the specific hardware deployment in this embodiment, the key hardware equipment selection and configuration details for workstations A, B, and C are shown in Table 1.

[0073] Table 1

[0074]

[0075] Q-learning (reinforcement learning) enables each workstation to autonomously determine acceleration / deceleration coefficients and compensation times through trial and error, adapting to varying load levels and upstream / downstream conditions, eliminating the need for manual parameter tuning. RRT-Connect (bidirectional fast expanding random tree), acting as a "cooperative avoidance engine," rapidly searches for collision-free paths bidirectionally when multiple robots share a space, incorporating energy consumption weights to replace fixed teaching programs. The S-shaped velocity curve, acting as a "smoothing actuator," transforms abrupt cycle commands into continuously accelerating / decelerating trajectories, preventing the impact on the mechanical structure from sudden stops and starts.

[0076] To support the operation of the aforementioned hardware, this embodiment constructs a production line scheduling and control system based on real-time load awareness. For example... Figure 2 As shown, the system adopts a distributed edge computing architecture. The MES / upper-level management system (upper-level monitoring system) is responsible for the macro-management and decision-making of the entire production process. Its main function is to formulate and issue production plans / statistical instructions. These instructions are sent to the edge controllers at each workstation on the production line. Workstations A, B, and C are each equipped with an edge controller, namely Edge Controller A, Edge Controller B, and Edge Controller C. These edge controllers are local intelligent units at the workstations, receiving production instructions from the upper-level monitoring system. Simultaneously, they directly interact with the specific equipment at the workstations, driving the equipment (such as pressing robots, gluing robots, and loading robots) and reading real-time data from sensor / vision systems (such as equipment status, product quality, etc.). While executing tasks, each edge controller at each workstation feeds back its own load / status information (i.e., equipment operating load, fault status, completion progress, etc.) to the upper-level monitoring system. This feedback information is crucial for the upper-level system to perform production scheduling, resource optimization, and fault early warning. The communication and data transmission of the entire system rely on a communication network consisting of a "local low-latency bus / Wi-Fi 6Mesh". This network provides high-speed, low-latency connectivity for data exchange between the upper-level monitoring system, edge controllers, and various devices. For example, the edge controller uploads collected sensor data to the upper-level system, and the upper-level system sends new production plans to the edge controller; all of this depends on the stable operation of this communication network. This network ensures real-time information delivery, enabling the entire production process to proceed efficiently and collaboratively.

[0077] The following are the specific implementation steps:

[0078] Step S1: Initial Parameters and Nominal Cycle Time Setting: All workstations start operating at a nominal cycle time of 3 seconds per piece, fixing the initial movement speed (e.g., 100mm / s) and providing a benchmark reference point for production line operation. When there is no disturbance, the system operates efficiently at the predetermined cycle time, reducing unnecessary computational overhead.

[0079] Step S2: Real-time load perception and feature extraction: Each edge controller collects sensor data (workpiece arrival frequency, joint torque, visual recognition time) at a 10ms cycle, and calculates the comprehensive load coefficient, which is a quantitative indicator of "whether the line is blocked". For example, if the visual recognition time increases from 20ms to 30ms, the load coefficient will rise from 1.0 to 1.2. This is more sensitive to micro-fluctuations than simply looking at the binary state of "equipment running / stopping", providing accurate state input for Q-learning.

[0080] ;

[0081] in, For the comprehensive load factor, The dynamic weight of the i-th sensor index, It is a saturation limiting function. Let i be the real-time sampled value of the i-th sensor indicator. Let be the nominal reference value of the i-th sensor indicator. This is the normalized load ratio.

[0082] Step S3: Local Dynamic Cycle Adjustment Based on Q-learning: When the visual recognition time at station B increases and the load factor rises to 1.2, the local Q-learning strategy is triggered: Local adjustment: The optimal action is obtained by looking up the Q table—reducing the glue application speed from 100mm / s to 85mm / s, and issuing an S-shaped speed curve for smooth transition. Collaborative broadcasting: A deceleration factor of 0.9 (slowing down material feeding) is sent to station A, and a waiting compensation of +80ms (extending the dwell time) is sent to station C. The existence of Q-learning makes station B no longer a fixed actuator. When its own load is too high (corresponding to state S), an appropriate speed reduction is found by retrieving the Q value and used as the optimal action (A).

[0083] Step S4: Multi-robot Cooperative Avoidance and Shared Trajectory Planning (RRT-Connect): When two robots need to simultaneously enter the same work area to flip the shell: each edge node exchanges predicted trajectory points for the next 0.3 seconds. Each node independently runs the improved RRT-Connect algorithm, aiming for an energy consumption weight of 0.6 and a time weight of 0.4, to generate collision-free interpolation trajectories in real time. Exchanging predicted trajectories for 0.3 seconds allows the robot to anticipate its movement trajectory in advance. The RRT-Connect bidirectional fast exploration random tree connection algorithm prioritizes routes with low joint acceleration and smooth path curvature. This solves the waiting and energy waste caused by fixed trajectories.

[0084] Step S5: Anomaly Pre-compensation and Uninterrupted Smooth Transition: If the vision system at station B predicts that the processing time for the next frame will increase by 10ms, it immediately transmits fine-tuning data (grabbing the offset or conveyor belt speed fine-tuning coefficient) to stations A and C 0.5 seconds in advance. This step is a feedforward enhancement of Q-learning. By sending a signal to the equipment 0.5 seconds in advance and utilizing the physical transmission time window of the conveyor belt, the adjustments of upstream and downstream are synchronized with the arrival of the disturbance, achieving a seamless smooth transition where the production line does not stop for milliseconds despite the disturbance.

[0085] The specific operational logic is as follows: first, the perception is performed in step S2; then, the decision is made in step S3; next, the execution and coordination are carried out in step S4; then, the feedforward pre-compensation is performed in step S5, and so on in a continuous loop. The upper-layer MES only periodically reads the statistical data of each edge node and does not interfere with real-time control, thereby reducing the network load and computing power bottleneck of the central server.

[0086] This embodiment employs a dynamic, adaptive cycle adjustment mechanism, which aims to intelligently adjust the robot's movement cycle based on the real-time load of each workstation, thereby ensuring the stability and efficiency of the production process. Figure 3 This is a flowchart of the dynamic beat adjustment process in this embodiment, as follows: Figure 3 As shown, this process is executed cyclically in fixed-cycle units. The process begins at the start of each cycle. First, the load factor of the current workstation is detected. This factor is a quantitative value that integrates key indicators such as the workstation's arrival frequency, running torque, and visual processing time, used to accurately reflect the real-time workload of the workstation. After obtaining the load factor, an update amount of the Q value is calculated according to a preset calculation model (the formula is ΔQ=α×(r+r×max(Q')-Q)). This value reflects the deviation between the current load and the target load. The current load factor is compared with a preset critical threshold of 1. If the load factor is greater than 1, it is determined to be in an "overload" state. In this case, a deceleration command is generated, instructing the robot to reduce its movement speed by 5% to 15% to alleviate the pressure on the current workstation. At the same time, a "deceleration factor" and "delay compensation" command are broadcast to all relevant upstream and downstream equipment to ensure coordinated adjustment of the entire production chain. Conversely, if the load factor is less than 1, it is determined to be in a "light load" state. At this point, a speed-up command will be generated, increasing the robot's speed by 5% to 15% to improve production efficiency. Correspondingly, a "speed-up factor" and "advance compensation" command will be broadcast. If the load coefficient is exactly equal to or very close to 1, it is determined to be in a "normal" state, and the robot will maintain its current operating speed while broadcasting a "normal cycle time" signal.

[0087] Regardless of whether the robot is in an "overload," "light load," or "normal" state, after all branch instruction generation and broadcast operations are completed, an S-shaped speed curve smoothing filter will be performed. This step ensures that any changes in robot speed are smooth and continuous, avoiding mechanical shocks and positioning errors caused by sudden speed changes, thus guaranteeing stable robot operation and processing accuracy. Finally, based on the processed speed instructions, the robot is updated and sent for execution, then enters a "waiting for the next cycle" state, preparing for the next dynamic adjustment. Through this periodic detection, decision-making, and execution, the production rhythm can be continuously optimized, achieving stable and efficient automated production.

[0088] In the formula ΔQ = α × (r + γ × max(Q') - Q), ΔQ represents the increment or update of the Q value, reflecting the expected change in long-term cumulative reward after performing an action under the current strategy; α represents the learning rate, ranging from 0 to 1, which determines the degree to which new information covers old information. In production line scheduling, a larger α allows for rapid adaptation to sudden disturbances, while a smaller α makes the strategy more stable; r represents the immediate reward, which is positive when line congestion is successfully avoided and energy consumption is low, and negative if an emergency stop occurs or the waiting time is too long; γ represents the discount factor, ranging from 0 to 1, used to measure the importance of future rewards. The higher the γ, the more the algorithm focuses on long-term production efficiency and energy consumption optimization, while the lower the γ, the more it focuses on the current immediate load balance; Q' represents the Q value in the next state; max(Q') represents the maximum Q value among all possible actions in the new state after taking the current action. This reflects the assumption that the action that maximizes long-term reward will be chosen in the next step. Q represents the original Q value of the current state-action relationship.

[0089] Figure 4 This is a schematic diagram of a multi-robot collaborative trajectory, such as... Figure 4 As shown, Robot 1 and Robot 2 each have their own predicted trajectories, which exchange information through "shared short-term paths." Based on the shared short-term path information, the robot enters the "path fusion optimization" stage. This stage uses algorithms to collaboratively optimize the trajectories of the two robots, aiming to eliminate potential path conflicts, avoid collisions, and achieve efficient path fusion. This optimization process ensures that the robots can safely and efficiently enter the "non-waiting overlapping collaborative work area" without waiting or overlapping, thereby improving the efficiency and smoothness of collaborative work. However, in the traditional work mode, Robot 1 and Robot 2 each follow a preset fixed path. Due to the fixed paths and lack of dynamic coordination, when the paths of the two robots intersect in a certain area (such as the "waiting area / idling"), they must avoid conflict by "entering in sequence." This means that one robot must wait before the other robot enters the work area, or idle in a specific area until the other completes its task. This mode inevitably leads to wasted work time and reduced resource utilization. Therefore, the dynamic collaborative trajectory scheme in this embodiment fundamentally solves the problems of waiting, idling, and low efficiency in the traditional fixed trajectory mode by introducing real-time path sharing and fusion optimization, and realizes more intelligent and efficient multi-robot collaborative operation.

[0090] Figure 5 This is a timing diagram for anomaly pre-compensation in this embodiment, mainly describing the proactive pre-compensation mechanism implemented in the production line to ensure continuous, uninterrupted, or unstalled material flow when a potential anomaly is detected at workstation B (the machine). For example... Figure 5As shown, workstation B, through its vision processing system, pre-determines that if the current operating speed remains unchanged, its downstream workstation C (downstream) will face a risk of timeout within the next 0.5 seconds due to insufficient processing capacity. To address this potential problem, workstation B does not wait for actual congestion to occur but instead takes proactive intervention measures to achieve anomaly pre-compensation. When workstation B predicts the timeout risk, its internal vision processing time increases by 10 milliseconds, indicating that the system is performing more complex analysis to make an accurate judgment. Based on this judgment, workstation B immediately broadcasts a key instruction to its upstream workstation A (upstream): "Deceleration factor 0.9." This instruction is explicitly marked as being issued "0.5 seconds in advance," reflecting the accuracy of the prediction. This instruction means that upstream workstation A is required to reduce the operating speed of its placement pusher to 90% of its original speed. To ensure that this deceleration instruction takes effect, workstation A will correspondingly increase the delay of its placement pusher by 0.1 times the cycle time, thereby physically decelerating and freeing up processing time for downstream workstation C. Simultaneously, station B broadcasts a command to its downstream station C: "Wait for compensation +80ms." This command informs station C that the pressing head operation will begin after a 80-millisecond delay. Because station B has already reduced its speed and smoothed the S-curve transition, its output rhythm has slowed, causing the pressing head at station C to "linger" for 80 milliseconds. Upon receiving this command, station C executes the action of "pressing head lingering for 80ms before retreating," effectively preventing material accumulation or overload due to excessively rapid material input. Throughout this process, station B successfully mitigates a potential production bottleneck by reducing its own speed, coordinating upstream deceleration, and guiding downstream waiting. Through this series of pre-compensation actions, the entire production system achieves continuous material flow without interruptions or accumulation. This not only eliminates the risk of downtime caused by downstream anomalies but also improves the overall efficiency and stability of the production line.

[0091] This application also provides a production line scheduling and control device. It should be noted that the production line scheduling and control device of this application can be used to execute the production line scheduling and control method provided in this application. This device is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0092] The following describes the production line scheduling and control device provided in the embodiments of this application.

[0093] Figure 6 This is a schematic diagram of a production line scheduling and control device according to an embodiment of this application. Figure 6 As shown, the device includes a data acquisition unit 61, a determination unit 62, a first generation unit 63, and a transmission unit 64. The data acquisition unit acquires load status data of the target workstation and synchronizes this data to adjacent workstations to establish a shared state between the target workstation and its adjacent workstations. The determination unit determines a motion adjustment strategy for the target workstation based on this shared state. The first generation unit generates motion control commands, including motion speed and acceleration / deceleration curves, according to the motion adjustment strategy and controls the robot at the target workstation to execute these commands. The transmission unit sends coordination commands to adjacent workstations, enabling them to adjust the motion control signals of their robots based on these coordination commands.

[0094] In this embodiment, by collecting the load status data of the target workstation and synchronizing it with its adjacent workstations to establish a shared state, the motion adjustment strategy is dynamically determined based on the shared state and corresponding motion control commands are generated. At the same time, collaborative commands are sent to adjacent workstations to adjust their running signals in a coordinated manner, thereby realizing the collaboration and adaptive adjustment between workstations. This effectively eliminates the global cycle mismatch caused by local load fluctuations, reduces the waiting time and energy consumption in the multi-robot collaboration process, and solves the problem of low scheduling and control efficiency in existing production lines.

[0095] In some embodiments of this application, the above-mentioned apparatus includes a second generation unit, an exchange unit, and a third generation unit. The second generation unit is used to generate a predicted path for each of the robots based on the motion adjustment strategy when multiple robots are located in the working area of ​​the same target workstation. The predicted path represents the motion trajectory within a preset time period in the future. The exchange unit is used to exchange the predicted paths of the robots with other robots within the target workstation. The third generation unit is used to generate a collision-free interpolation trajectory for each robot based on the exchanged predicted paths using a path planning algorithm, and control each robot to move along the collision-free interpolation trajectory.

[0096] By generating and exchanging predicted paths within a preset time period in real time in a multi-robot scenario at the same workstation, and using a path planning algorithm to dynamically generate collision-free interpolation trajectories based on the exchanged path information, collaborative obstacle avoidance and path optimization among robots are achieved. This mechanism effectively replaces the traditional fixed teaching program, eliminates the waiting and collision risks caused by trajectory conflicts, and reduces motion energy consumption and improves overall work efficiency while ensuring safe collaborative operation of multiple robots.

[0097] In some embodiments of this application, the above-described apparatus further includes a prediction unit and a transmission unit. The prediction unit is used to predict, based on the load state data, whether a processing timeout will occur at the target workstation before generating motion control commands including motion speed and acceleration / deceleration curves according to the motion adjustment strategy. The transmission unit is used to transmit adjustment data to adjacent workstations of the target workstation in advance for a preset time period if the processing timeout is predicted, so that the adjacent workstations of the target workstation adjust their respective operating states according to the adjustment data. The adjustment data includes a gripping offset and / or a conveyor belt speed adjustment coefficient.

[0098] By predicting the processing timeout risk of the target station based on load status data, and transmitting adjustment data, including the grab offset and / or conveyor belt speed adjustment coefficient, to adjacent stations before the timeout occurs using a preset time window, proactive feedforward pre-compensation for local disturbances is achieved. This mechanism enables upstream and downstream stations to synchronize and fine-tune their operating status in advance, ensuring that the adjustment action is precisely matched with the arrival time of the disturbance, thereby effectively eliminating the impact of local delays on the overall production cycle time and achieving uninterrupted smooth transition and millisecond-level rapid response of the production line.

[0099] In some specific embodiments of this application, the aforementioned load status data includes visual recognition time consumption, and the aforementioned prediction unit includes an acquisition module, a prediction module, and a determination module. The acquisition module is used to acquire the aforementioned visual recognition time consumption of the aforementioned target workstation, wherein the aforementioned visual recognition time consumption represents the duration of processing the current frame image; the prediction module is used to predict the time consumption required for processing the next frame image based on the aforementioned visual recognition time consumption; the determination module is used to determine that the aforementioned target workstation will experience the aforementioned processing timeout when it is predicted that the time consumption required for processing the next frame image exceeds a preset time threshold.

[0100] By acquiring the visual recognition time in real time and predicting the time required to process the next frame of image, it is possible to accurately identify potential timeout risks caused by fluctuations in visual processing performance. When the predicted time exceeds a preset threshold, an abnormal pre-compensation mechanism can be triggered in advance, allowing upstream and downstream workstations to make synchronous fine adjustments before the disturbance actually occurs. This effectively avoids production line blockage or cycle mismatch caused by visual processing delays, achieving refined prediction and uninterrupted smooth transition for visual load fluctuations.

[0101] In some embodiments of this application, the determining unit includes a calculation module and a generation module. The calculation module is used to calculate the comprehensive load coefficient of the target workstation based on the load state data. The generation module is used to determine the target action based on the comprehensive load coefficient and the shared state using a pre-trained reinforcement learning model, and to generate the execution instruction of the motion adjustment strategy based on the target action. The pre-trained reinforcement learning model is trained using multiple sets of data through machine learning. Each set of data includes: a state input and a label corresponding to the state input. The state input includes the comprehensive load coefficient of the target workstation and the shared state. The label represents at least one of the following control behaviors: adjusting the movement speed of the robot at the target workstation; issuing a speed curve to control the acceleration and deceleration process of the robot at the target workstation; generating a control signal to be sent to adjacent workstations of the target workstation, the control signal including a deceleration factor or a waiting compensation value.

[0102] In some specific embodiments of this application, the above-mentioned calculation module includes a first acquisition submodule, a calculation submodule, a first processing submodule, and a second processing submodule. The first acquisition submodule is used to acquire the real-time sampled values ​​of each sensor index and the corresponding nominal reference values ​​in the above-mentioned load state data. The above-mentioned load state data includes the workpiece arrival frequency, joint torque, and visual recognition time collected by each sensor. The calculation submodule is used to calculate the ratio between the real-time sampled values ​​of each of the above-mentioned sensor indexes and the corresponding nominal reference values ​​to obtain the normalized load ratio of each of the above-mentioned sensor indexes. The first processing submodule is used to input the normalized load ratio of each of the above-mentioned sensor indexes into the saturation limiting function to obtain the limiting value of each of the above-mentioned sensor indexes. The second processing submodule is used to perform a weighted summation of the above-mentioned limiting values ​​according to the dynamic weights of each of the above-mentioned sensor indexes to obtain the above-mentioned comprehensive load coefficient.

[0103] By acquiring real-time sampled values ​​and nominal reference values ​​of multi-dimensional sensor indicators such as workpiece positioning frequency, joint torque, and visual recognition time, a normalized load ratio is calculated and subjected to saturation limiting processing. This effectively eliminates the influence of different physical dimensions and suppresses the interference of extreme outliers on the evaluation results. A weighted summation based on dynamic weights yields the comprehensive load coefficient. This mechanism achieves the quantification and standardized representation of local load fluctuations at the workstation, enabling the comprehensive load coefficient to more sensitively and accurately reflect the actual operating load state of the workstation, providing a stable and highly discriminative state input for subsequent reinforcement learning models.

[0104] In some other specific embodiments of this application, the generation module includes a third processing submodule, a second acquisition submodule, a filtering submodule, and a determination submodule. The third processing submodule is used to discretize the comprehensive load coefficient and the shared state to obtain the current state; the second acquisition submodule is used to acquire an action value mapping table, which is established by the reinforcement learning model during the training phase through trial and error learning, and is used to store the mapping relationship between the discretized state and the action value of the corresponding candidate action; the filtering submodule is used to query the action value mapping table and filter the action values ​​of all the candidate actions corresponding to the current state to obtain multiple candidate action values; the determination submodule is used to determine the candidate action corresponding to the maximum value among the multiple candidate action values ​​as the target action.

[0105] By discretizing the comprehensive load coefficient and shared state into the current state, and querying the action value mapping table generated by the pre-trained reinforcement learning model for all candidate action values ​​in the corresponding state, the action with the highest value is selected as the target action. Leveraging decision-making experience accumulated through historical trial-and-error learning, rapid and deterministic optimal decision-making based on multidimensional load characteristics and upstream / downstream collaborative states is achieved within millisecond-level real-time control cycles. This avoids the latency caused by online search or complex optimization calculations, ensuring the production line's immediate response capability and control stability under dynamic disturbances.

[0106] By calculating the comprehensive load coefficient reflecting the multidimensional operating load of the workstation in real time and combining the shared state of upstream and downstream, the system uses a reinforcement learning model trained with multiple sets of state and label data to accurately predict and output the optimal target action. This generates motion adjustment strategies, including robot speed adjustment, acceleration and deceleration curve distribution, and collaborative signals between adjacent workstations. This mechanism realizes closed-loop decision-making from local load perception to global collaborative control. It not only eliminates the response lag of fixed cycle time to dynamic fluctuations, but also enables millisecond-level uninterrupted smooth transition of each workstation on the production line through feedforward collaborative instructions, thereby improving the overall efficiency, flexibility, and energy consumption level in mixed-flow production environments.

[0107] In some embodiments of this application, the acquisition unit includes a sending module, a receiving module, and a construction module. The sending module is used to send the load status data to the adjacent workstations of the target workstation; the receiving module is used to receive the load status data of the adjacent workstations from the adjacent workstations; the construction module is used to construct the shared state by combining the load status data of the adjacent workstations and the load status data of the target workstation, and store the shared state in the local storage unit of the target workstation.

[0108] By achieving real-time bidirectional synchronization of load status data between workstations, interconnection and interoperability of operational information between adjacent workstations are realized. This allows the target workstation to monitor the load status of upstream and downstream processes in real time, thereby constructing a shared state that includes both local and global information. This provides comprehensive state input for the reinforcement learning model, enabling control decisions to take into account upstream and downstream collaboration. It effectively avoids line congestion or waiting problems caused by information asymmetry, and improves the overall collaborative efficiency and stability of the production line in a dynamically fluctuating environment.

[0109] The scheduling and control device for the aforementioned production line includes a processor and a memory. The acquisition unit, determination unit, first generation unit, and sending unit are all stored as program units in the memory. The processor executes these program units stored in the memory to achieve the corresponding functions. All of the above modules reside in the same processor; alternatively, the modules may be located in different processors in any combination.

[0110] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0111] This invention provides a production line scheduling and control system, comprising: multiple edge controllers deployed at each workstation, the edge controllers being used to execute any of the above-described production line scheduling and control methods; and multiple robots disposed at each of the aforementioned workstations, the robots being communicatively connected to the edge controllers at their respective workstations and being used to execute motion control signals issued by the edge controllers.

[0112] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the scheduling and control method for the production line.

[0113] This invention provides a processor for running a program, wherein the program executes the production line scheduling and control method during runtime.

[0114] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the aforementioned production line scheduling and control method. The device described herein can be a server, PC, PAD, mobile phone, etc.

[0115] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform the steps of initializing the scheduling and control method for the aforementioned production line.

[0116] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0117] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0122] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0123] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0124] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0126] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for scheduling and controlling a production line, characterized in that, The method is applied to an automated production line comprising multiple workstations and multiple robots, and the method includes: Collect load status data of the target workstation and synchronize the load status data to the adjacent workstations of the target workstation to establish a shared status between the target workstation and the adjacent workstations; Based on the load status data, calculate the comprehensive load coefficient of the target workstation; Based on the comprehensive load coefficient and the shared state, a pre-trained reinforcement learning model is used to determine the target action, and a motion adjustment strategy for the target workstation is generated according to the target action; wherein, the label of the pre-trained reinforcement learning model represents at least one of the following control behaviors: adjusting the motion speed of the robot at the target workstation; issuing speed curves to control the acceleration and deceleration process of the robot at the target workstation; generating control signals to be sent to adjacent workstations of the target workstation; Based on the motion adjustment strategy, motion control commands including motion speed and acceleration / deceleration curves are generated, and the robot at the target workstation is controlled to execute the motion control commands. A coordination command is sent to the adjacent workstations of the target workstation, so that the adjacent workstations adjust the motion control signals of their robots based on the coordination command.

2. The method according to claim 1, characterized in that, When multiple robots are located within the same target workstation's work area, the method includes: Based on the motion adjustment strategy, a predicted path is generated for each of the robots, and the predicted path represents the motion trajectory within a preset time period in the future. The predicted paths of each robot are exchanged with those of other robots within the target workstation; Based on the predicted paths after the exchange, a path planning algorithm is used to generate collision-free interpolation trajectories for each robot, and each robot is controlled to move along the collision-free interpolation trajectories.

3. The method according to claim 1, characterized in that, Before generating motion control commands including motion speed and acceleration / deceleration curves according to the motion adjustment strategy, the method further includes: Based on the load status data, predict whether the target workstation will experience a processing timeout; If it is predicted that the processing timeout will occur at the target workstation, adjustment data is transmitted to the adjacent workstations of the target workstation in advance within a preset time period, so that the adjacent workstations of the target workstation can adjust the operating status of their respective workstations according to the adjustment data. The adjustment data includes the grab offset and / or the conveyor belt speed adjustment coefficient.

4. The method according to claim 3, characterized in that, The load status data includes visual recognition time. Based on the load status data, predicting whether the target workstation will experience a processing timeout includes: The visual recognition time of the target workstation is obtained, and the visual recognition time represents the time taken to process the current frame image; Based on the visual recognition time consumption, predict the time required to process the next frame of image; When it is predicted that the processing time required for the next frame of image exceeds a preset time threshold, it is determined that the processing timeout will occur at the target workstation.

5. The method according to claim 1, characterized in that, Based on the comprehensive load coefficient and the shared state, a pre-trained reinforcement learning model is used to determine the target action, including: Discretize the overall load coefficient and the shared state to obtain the current state; Obtain an action value mapping table, which is established by the reinforcement learning model through trial and error learning during the training phase, and is used to store the mapping relationship between the discretized state and the action value of the corresponding candidate action; Query the action value mapping table, filter the action values ​​of all candidate actions corresponding to the current state, and obtain multiple candidate action values; Among the multiple candidate action values, the candidate action corresponding to the maximum value is determined as the target action.

6. The method according to claim 1, characterized in that, Based on the load status data, the comprehensive load coefficient of the target workstation is calculated, including: The real-time sampled values ​​of each sensor index and the corresponding nominal reference values ​​in the load status data are obtained. The load status data includes the workpiece arrival frequency, joint torque and visual recognition time collected by each sensor. The normalized load ratio of each sensor index is obtained by calculating the ratio between the real-time sampled value of each sensor index and the corresponding nominal reference value. The normalized load ratio of each sensor index is input into the saturation limiting function to obtain the limiting value of each sensor index. The amplitude limit value is weighted and summed according to the dynamic weights of each sensor index to obtain the comprehensive load coefficient.

7. The method according to claim 1, characterized in that, Synchronizing the load status data to adjacent workstations of the target workstation to establish a shared status between the target workstation and the adjacent workstations includes: The load status data is sent to the adjacent workstations of the target workstation; Receive load status data of the adjacent workstations; The load status data of the adjacent workstations and the load status data of the target workstation are used to construct the shared state, and the shared state is stored in the local storage unit of the target workstation.

8. A scheduling and control device for a production line, characterized in that, include: The data acquisition unit is used to acquire the load status data of the target workstation and synchronize the load status data to the adjacent workstations of the target workstation to establish a shared status between the target workstation and the adjacent workstations. A determining unit is configured to determine the motion adjustment strategy for the target workstation based on the shared state. The first generation unit is used to generate motion control instructions including motion speed and acceleration / deceleration curves according to the motion adjustment strategy, and control the robot at the target workstation to execute the motion control instructions. The sending unit is used to send a coordination instruction to the adjacent workstations of the target workstation, so that the adjacent workstations of the target workstation adjust the motion control signals of the robots of the adjacent workstations of the target workstation based on the coordination instruction.

9. A production line scheduling and control system, characterized in that, The system includes: Multiple edge controllers are deployed at each workstation, and the edge controllers are used to execute the production line scheduling and control method as described in any one of claims 1 to 7; Multiple robots are respectively set up at each of the workstations. The robots are communicatively connected to the edge controllers of the corresponding workstations and are used to execute motion control signals issued by the edge controllers.

Images