Intelligent warehouse safety management method based on artificial intelligence

Abstract

The application relates to the technical field of intelligent warehouse safety control, and discloses an intelligent warehouse safety management method based on artificial intelligence. A safety takeover relationship is established along a sensing area through which a carrying task passes, a time sequence model is trained in a closed takeover track, a takeover interval and an expected observation sequence are determined. A takeover state containing a last trusted physical event, a dynamic occupation range and associated channel resources is generated in a current area; a state is transferred after a local observation in conformity with the sequence is obtained in a next area and the current area confirms driving off, otherwise the two areas jointly maintain occupation. When a control action changes a track or an observable condition, a takeover basis is abandoned and reconstructed according to actual feedback; when a sensing or control feedback fails after takeover, joint maintenance is restored, so that the continuity of identity, occupation and resource control responsibility across the sensing boundary is maintained, and the operation conflict risk caused by early release of resources is reduced.

Description

Technical Field

[0001] This invention relates to the field of mobile operation safety control technology in intelligent warehousing, and in particular to an intelligent warehousing safety management method based on artificial intelligence. Background Technology

[0002] Intelligent warehousing typically employs automated guided vehicles (AGVs), mobile robots, forklifts, stacking equipment, or other handling equipment to transfer goods between inbound workstations, storage aisles, picking workstations, conveyor interface points, and outbound workstations. To obtain the current location and operating status of the handling equipment and its loads, cameras, LiDAR, UWB positioning base stations, RFID gates, photoelectric detectors, or fixed-position detection devices can be deployed within the warehouse. The warehouse control node receives the observation results from the corresponding sensing devices. Due to the influence of warehouse spatial structure, shelving obstructions, equipment communication range, and control node management boundaries, continuous task routes often need to pass through multiple sensing areas sequentially.

[0003] Existing warehouse security management typically determines whether handling equipment has entered a restricted area based on target detection results, positioning results, or equipment operating parameters within a single area, and implements alarms, speed limits, or shutdowns accordingly. When tracking across areas, target identification can be handed over based on the disappearance of a target in the current area and the appearance of a target in the next sensing area. For aisle, storage location, or workstation resources, warehouse management systems generally allocate and release resources based on the progress of handling tasks, equipment arrival signals, or the execution results of control commands.

[0004] When there is overlap in coverage, intermittent observation, or data transmission delays between adjacent sensing areas, the handover time of target identification may not match the actual time when the transport equipment and load cross the area boundary. If the current area releases channel resources only based on the target leaving the field of view, the next sensing area may not have obtained sufficient continuous observations to confirm the same work object; if the next sensing area only obtains control responsibility based on the appearance of a single target, it may incorrectly associate other equipment that reappears after obstruction, erroneous targets formed by reflection, or the same equipment whose load attitude has changed. As a result, the identity of the work object, the actual physical occupancy range, and the control responsibility for channel resources may become disconnected at the area boundary.

[0005] Furthermore, during cross-regional takeover, the arrival time, direction of movement, braking range, and subsequent observable conditions of the handling equipment will change after speed limits, stops, rerouting, or load attitude adjustments are implemented. If the takeover interval and observation sequence established before the control action are still used as the basis for resource release, the original takeover conditions may have become invalid. If a sensing interruption or controller disconnection occurs just after the takeover state has been transferred and before a new credible physical event has been formed in the next sensing area, there may also be a state gap where the previous area has been released from occupation but the next area cannot continue to be confirmed. Therefore, an intelligent warehouse safety management method is needed that can bind and transfer the continuity of targets, dynamic physical occupation, and resource control responsibilities across sensing areas, and rebuild or roll back the takeover state when the observation conditions change due to control actions or faults after takeover. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent warehouse safety management method based on artificial intelligence to solve the problems that when handling equipment and its load cross different sensing areas, the identity of the work object, dynamic physical occupation and channel resource control responsibility cannot be consistently consistent, and the original control basis becomes invalid after the control action changes the subsequent trajectory or observable conditions.

[0007] To achieve the above objectives, the present invention provides an intelligent warehouse security management method based on artificial intelligence, comprising:

[0008] Establish a safe takeover relationship for the work object based on the sensing area traversed by the handling task, and obtain the location of the handling equipment, load attitude, task route, effective coverage of the sensing equipment, and controller feedback;

[0009] Using a time series model trained from a closed historical takeover trajectory, including a sequence state extraction part, a takeover interval output part, and an observation sequence output part, the takeover interval and expected observation sequence of the working object from the current sensing area to the next sensing area are determined.

[0010] Before entering the takeover zone, the current sensing area generates a takeover status including the last trusted physical event, the dynamic occupied range, and the associated channel resources, and sends it to the next sensing area;

[0011] When the next sensing area obtains local observations that conform to the expected observation sequence and the current sensing area confirms that the work object has left, the takeover state is transferred and the associated channel resources that no longer intersect with the dynamic occupancy range are released; otherwise, the two sensing areas jointly maintain the dynamic occupancy range and restrict the entry of subsequent tasks.

[0012] When speed limits, stops, rerouting, or load attitude adjustments alter the subsequent trajectory or observable range, the original takeover section is terminated, and the takeover section and expected observation sequence are reconstructed based on actual controller feedback until the takeover state transition is completed.

[0013] Compared to existing methods, this invention binds the last credible physical event, dynamic occupancy range, and associated channel resources into a takeover state that can be transferred between adjacent sensing areas. This eliminates the need for the identification, actual occupancy, and resource control of the work object to be determined separately by unrelated target tracking results, path scheduling results, and control commands. The takeover condition is formed when the next sensing area obtains local observations conforming to the expected observation sequence and the current sensing area confirms the departure of the work object. This reduces control gaps caused by any sensing area unilaterally releasing resources prematurely.

[0014] In overlapping coverage sections, resources in the previous area are maintained until continuous observation is formed in the next area. In observation discontinuity sections, the dynamic occupancy range is extended to the first observable position, so that resource constraints cross the perception boundary with the actual movement process. When the control action changes the subsequent trajectory or observable range, the original takeover basis is stopped and the takeover interval and expected observation sequence are reconstructed with the actual feedback of the controller, so that the safety control action is consistent with the subsequent takeover conditions.

[0015] When a local observation conflicts with the last credible physical event, the candidate trajectory is retained and its coverage resources are occupied. Furthermore, in the event of a sensing interruption after takeover or loss of control feedback from the handling equipment controller, the dual-region co-maintenance is restored. This ensures that uncertain states are not erroneously compressed into a single path before a new credible physical event is obtained. Therefore, this invention can maintain the continuity and consistency of the identity of the work object, the dynamic occupancy range, and the responsibility for controlling warehouse resources at the distributed sensing and control boundaries. Attached Figure Description

[0016] The above and other technical features, processing relationships and embodiments of the present invention can be further understood in conjunction with the accompanying drawings, which are only used to illustrate the technical relationships of the present invention.

[0017] Figure 1 This is a schematic diagram of the overall process of the intelligent warehouse safety management method provided in the embodiments of the present invention;

[0018] Figure 2 This is a schematic diagram of the sensing area division and security takeover relationship provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the takeover status generation, version transfer, and two-way confirmation process provided in an embodiment of the present invention.

[0020] Figure 4 This is a schematic diagram illustrating the dynamic occupancy range formation and cross-regional retention provided in an embodiment of the present invention;

[0021] Figure 5 A schematic diagram illustrating the process of control actions causing takeover basis failure, version obsolescence, and reconstruction in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the candidate trajectory blocking and takeover state rollback process provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be fully described below with reference to the technical processes in the embodiments of the present invention. The described embodiments are used to explain the implementable path of the technical relationship defined by the claims, and do not limit the present invention to a certain type of handling equipment, a certain type of sensing device, or a certain time series model. Warehouse task generation, equipment basic navigation, communication transmission, and conventional security protection, which are not described in detail, can be implemented using existing technologies in the field.

[0024] The safe takeover and dynamic occupancy control of this invention resides at the warehouse control and resource scheduling layer. It is used to determine the takeover status between sensing areas, channel resource occupancy, and upper-level control requests made to the handling equipment. It does not replace the independently configured safety control loops, emergency stop functions, or onboard safety sensing protection of the handling equipment itself. Equipment-level safety control always operates independently. The dynamic occupancy range and resource restrictions formed by this invention are only used to increase or extend resource occupancy on the warehouse side and are not used as a basis for reducing the equipment-level safety protection area or releasing equipment-level safety braking conditions. When the operating permission given by the warehouse control node is inconsistent with the equipment-level safety control requirements, the more stringent control result is executed. Therefore, even if communication is interrupted, the timing model output is invalid, or the warehouse control node is temporarily unable to complete the takeover judgment, the equipment-level safety protection still executes independently, and the warehouse side maintains or expands the relevant resource occupancy.

[0025] The term "work object" in this invention refers to the combination of a handling device and its current load that requires continuous monitoring and safety control during a handling task. The handling device can be an automated guided vehicle (AGV), a mobile robot, an automated forklift, a manually driven forklift connected to a warehouse control system, stacking equipment, or other equipment capable of moving along a task route. The load can be a pallet, bin, crate, or other goods carried by a pick-and-place mechanism. Because the same handling device has different external boundaries and visible features when unloaded, loaded, turning, lifting, or unloading, safety takeover focuses on the combined state of the handling device and its current load, rather than simply transferring the device number.

[0026] Sensing devices are used to generate local observations that reflect the position, movement, or load status of work objects. These devices may include cameras, lidar, ultra-wideband positioning base stations, RFID gates, geomagnetic detectors, photoelectric switches, weighing detectors, pick-and-place station position detectors, and odometers or position feedback in handling equipment controllers. Warehouse control nodes receive observations from one or more sensing devices and manage the resources of aisles, storage locations, or workstations within their corresponding range. Warehouse control nodes can be located in edge computing devices, handling equipment scheduling servers, or warehouse management platforms, as long as they can distinguish the sources of observations from different sensing areas and perform resource occupancy and release control.

[0027] like Figure 1 As shown, the AI-based intelligent warehouse safety management method provided in this embodiment establishes a safety takeover relationship across sensing areas according to the task route, and sequentially executes the following steps: establishing area relationships, acquiring operational information, determining the takeover interval and expected observation sequence, generating and transmitting takeover status, confirming two-way evidence, maintaining joint anomaly information, reconstructing the takeover basis after control, and updating the model after takeover closure. Each step revolves around the transmission of the same takeover status, and is not a separate, independent target detection, path scheduling, and alarm process.

[0028] In step S101, the sensing area is divided according to the installation location, effective measurement distance, obstruction boundary, communication status, and access relationship of each sensing device in the warehouse site and the warehouse control node. The sensing area is not determined solely by a planar geometric boundary, but refers to the range within which the corresponding warehouse control node can independently confirm the work object based on local observation. Multiple sensing devices received by the same warehouse control node can form a sensing area; the same physical channel can also be divided into multiple adjacent sensing areas due to different control nodes managing it separately.

[0029] The effective coverage area of ​​a sensing device can be determined jointly by installation calibration, warehouse mapping, and the device's online status. For cameras, the effective coverage area is the field of view that can stably obtain the position and shape of the work object, excluding areas permanently obstructed by fixed shelves. For lidar or ultra-wideband positioning devices, the effective coverage area is the measurement area that satisfies positioning continuity and can distinguish adjacent work objects. For RFID gates or fixed detectors, the effective coverage area can be represented as the event confirmation area formed around the fixed detection position. When the device is obstructed, offline, or the measurement quality deteriorates, its effective coverage area can be temporarily removed from the current sensing area. Whether the measurement quality meets the requirements is determined jointly by the effective observation boundary formed by installation calibration and the device's historical normal operation records, with the minimum condition being the ability to stably distinguish work objects and continuously provide position or event results. After the device returns to online status, only when observations that meet this condition and can be continuously interpreted from adjacent sources are obtained again will the corresponding range be restored to the effective coverage area.

[0030] like Figure 2 As shown, overlapping or discontinuous observation sections can form between two adjacent sensing areas along the transport task route. An overlapping section refers to a range where, when the work object passes through this section, both the current sensing area and the next sensing area have areas where effective local observations can be formed. A discontinuous observation section refers to a section where, after the work object leaves the current sensing area, it needs to continue moving for a distance before reaching the first observable position in the next sensing area. If part of the route boundary is overlapping and another part is discontinuous, it can be recorded separately according to the branch routes actually traversed by the work object.

[0031] The safety takeover relationship is recorded sequentially according to the task route, including the current sensing area, the next sensing area, the overlapping coverage or observation discontinuity between them, the corresponding warehouse control node, and the intersecting channels, storage locations, and workstation resources. When a handling task passes through multiple sensing areas, a continuous sequence of takeover relationships can be formed. When the task route is rerouted, only the subsequent takeover relationships affected by the rerouting are reconstructed; previously closed takeover relationships are retained as historical operation records.

[0032] In step S102, the location of the handling equipment, the load attitude, the task route, the effective coverage area of ​​the sensing equipment, and the controller feedback are acquired. The location of the handling equipment can be provided by a positioning system, visual tracking, or controller mileage information; the load attitude at least indicates whether the load exists, the position of the load relative to the handling equipment, and the current state of the pick-and-place mechanism; the task route is provided by the warehouse management system or equipment scheduling system; the controller feedback includes at least information related to the current task, such as the actual speed, actual steering direction, braking or stopping status, and the actual position of the pick-and-place mechanism.

[0033] Each input carries the acquisition time, reception time, data source, and work object identifier. The warehouse control node prioritizes converting different sources to a unified time base based on equipment clock synchronization records. When corrected synchronization deviations cannot be directly obtained, clock deviations are estimated based on the sending time, receiving time, and round-trip time of paired records during synchronization or two-way communication. Residual errors caused by link delay asymmetry, timestamp resolution, and synchronization residuals are retained as time uncertainty ranges. When only one-way message reception records exist, the one-way transmission delay or clock deviation of the message is not directly inferred from the communication round-trip time. The corrected acquisition time is used to determine the sequence of events. When the time uncertainty ranges of two events overlap, departure confirmation or takeover is not only based on time sequence but also requires consideration of positional continuity, controller feedback, and other reliable observations. When the time information of a certain source cannot be corrected, the observation of that source can be used to expand the dynamic occupancy range or trigger joint maintenance, but it is not used alone as a basis for releasing associated channel resources.

[0034] The device clock synchronization can employ hardware synchronization, network clock synchronization based on bidirectional timestamp exchange, or other industrial synchronization mechanisms capable of forming traceable synchronization records. The time uncertainty range is used as a deterministic boundary on the safety side and can be formed by the synchronization time correction residual, timestamp resolution, the upper bound of the link delay asymmetry observed in recent bidirectional communication, and clock drift margin. If any of these boundaries cannot be reliably obtained, a conservative upper bound configured by the storage system for the corresponding device or communication link is used instead, without reducing the time uncertainty range through statistical averaging.

[0035] The warehouse control node writes the input information of the same work object into a time-ordered takeover event sequence according to the corrected acquisition time. For records repeatedly reported by the same source within a acquisition cycle, only those records with complete data and continuous interpretation of changes in position are retained; the remaining records are marked as duplicates. For records with inconsistent object identifiers, movement directions, or load states from different sources, they are not directly overwritten based on the order of arrival at the server; instead, the source, acquisition time, and corresponding state are retained separately. Unresolved conflict records can only trigger the expansion of dynamic occupancy range, entry into shared maintenance, or the formation of candidate trajectories; they cannot be used to confirm departure, complete takeover, or release resources.

[0036] In step S103, a time-series model is trained using historical takeover trajectories, and the takeover interval and expected observation sequence are determined during the current transport task. Historical takeover trajectories consist of task trajectories whose takeover state transitions have been closed. A closed takeover state transition occurs when the current sensing area has confirmed the departure of the work object, the next sensing area has acquired a takeover state, the work object subsequently forms the first credible physical event in the next sensing area, and there are no candidate trajectories that have not yet been eliminated during the takeover process. Only task trajectories that meet the above closure conditions can be used as basic training samples.

[0037] Each historical takeover trajectory is segmented into sample segments based on the takeover process of adjacent sensing areas. Sample inputs include the route segment and direction of travel, the type of handling equipment, the shape and attitude of the load, the operating speed before and after entering the current route segment, the effective coverage of the current sensing area and the next sensing area, and the operating status of each sensing device. For closed takeover processes where control commands change the trajectory, the type of control command, the actual response of the controller, and the dynamic occupancy range after the response are also included as part of the input sequence.

[0038] The sample output includes the start and end positions and times of the takeover interval, as well as the sequence of local observations from different sensing devices in the next sensing area. This sequence is not merely a list of devices; it also records the source of each expected local observation, the earliest expected time of appearance, the latest expected time of appearance, and the corresponding route location range, thus forming the expected occurrence interval. The start position and start time of the takeover interval represent the route location and corresponding time at which the current sensing area should begin preparing for takeover and maintain resources, respectively. The end position and end time represent the route location and corresponding time at which, under normal operating conditions, the next sensing area should have already obtained evidence of takeover, respectively.

[0039] The time-series model can be a sequence model capable of processing time series and simultaneously outputting intervals and event sequences. During training, samples of the same handling equipment category, load shape and attitude, and sensing area combination are divided into training data and validation data according to the task occurrence time, avoiding adjacent records of the same takeover process from entering different datasets simultaneously. The model output is not directly used as a resource release command, but rather to limit the timing of the takeover state generation, the local observation sources required for the next sensing area, and the occurrence interval; whether the resource is released is still determined by the subsequent two-way evidence confirmation process.

[0040] In one executable model implementation, each sample segment is organized into an event sequence arranged according to the corrected acquisition time. Each event record in the sequence includes the source category of the sensing device, the time position relative to the last credible physical event, the mileage position along the task route, the speed and direction of movement of the handling equipment, the existence status of the load and the position of the pick-and-place mechanism, and the online status of the corresponding sensing device. The task route, the category of handling equipment, the dimensions of the load, and the adjacent sensing areas combine to form a static context input along with the event sequence. Continuous values ​​are scaled according to the normal operating boundaries of the corresponding equipment or route segment, and category information is converted into distinguishable category codes. The upper and lower bounds used for scale unification are determined by the rated range of the equipment, the boundaries of the warehouse map, and normal closed records in the training data, and are fixed with the model version. The same set of scale parameters is used for verification and inference. When using upper and lower bounds for scale transformation, the transformation is only performed on continuous fields where the upper bound is greater than the lower bound; fields with the same upper and lower bounds are retained as constants, and no division operations involving the difference between the upper and lower bounds are performed. Inputs that exceed the predetermined boundaries retain their out-of-bounds status according to the same scale relationship and are marked with an out-of-bounds flag, without being truncated to normal values ​​without prompting. When a sensing source has no output within a sample segment, it is recorded with a missing value instead of replacing the actual observation with a zero value. The same model version consistently uses one time unit and one length unit: time location, start and end times, and control response processes are uniformly expressed in milliseconds or seconds; route location, equipment and load dimensions are uniformly expressed in millimeters or meters; and operating speed is expressed in the length unit per second corresponding to the selected length unit. Training, validation, and inference records maintain consistency in unit, scale parameters, and category encoding.

[0041] For continuous numerical fields that require scaling but have different upper and lower bounds, the following scaling transformation can be used:

[0042] ;

[0043] in, This is the original continuous field to be input. and These are the lower and upper bounds of the scale, which are fixed with each model version. This is the input value after scaling. When... When the input exceeds the boundary, the field is retained as a constant and no division is performed; when the input exceeds the boundary, the out-of-bounds state is still retained according to the same scale relationship and an out-of-bounds mark is added, without truncation to disguise it as a normal value.

[0044] The sequence model includes a sequence state extraction part, a takeover interval output part, and an observation sequence output part. The sequence state extraction part forms a sequence state reflecting the operation object's approach to the area boundary, crossing the boundary, and entering the next effective coverage area, according to the order of event occurrence. The takeover interval output part outputs the start and end times relative to the last credible physical event and the start and end positions along the task route based on the sequence state and static context. The observation sequence output part forms ordered observation event items according to the available sensing sources in the next sensing area, and outputs the expected earliest occurrence time, the expected latest occurrence time, and the route location range for each observation event item. During training, the spatiotemporal position where resources are actually maintained and takeover evidence is actually obtained in the closed takeover trajectory is used as the interval supervision result, and the sensing sources and their sequence relationship after confirmation by credible physical events are used as the sequence supervision result. The model parameters are updated so that the interval boundary and the observation event items simultaneously approach the corresponding supervision results. After the model output, a structural constraint check is performed: the start time of the takeover interval must not be later than the end time, and the start point of the route along the mission direction must not be after the end point of the route; the earliest expected occurrence time of each observed event must not be later than the latest expected occurrence time, and the starting position of the route location range must not be after the ending position; observed event items that need to form a sequential order must not have irresolvable duplicate sequential positions. Outputs that do not meet the above relationships are marked as invalid, and a conservative initial takeover interval and observations to be checked are adopted.

[0045] In a specific executable model example, the sequence state extraction part uses a Long Short-Term Memory (LSTM) network to receive event sequences one by one according to the corrected acquisition time. Missing markers, out-of-bounds markers, and the online status of sensing sources are encoded together with the corresponding event fields. The task route, handling equipment type, load dimensions, and adjacent sensing area combinations are combined with the sequence state combination corresponding to the last valid event after context mapping. The takeover interval output part sets up time interval output branches and route position interval output branches, which respectively give the start and end times relative to the last credible physical event and the start and end positions along the task route. The observation sequence output part uses the currently online sensing sources in the next sensing area as a candidate set, and for each candidate source, it gives whether it is an observation to be verified, its position in the observed event items, the expected earliest time of occurrence, the expected latest time of occurrence, and the route position range. During training, the two interval output branches are updated with the spatiotemporal boundary from the actual start of resource maintenance to the actual acquisition of takeover evidence. The observation sequence output part is updated with the sensing sources confirmed by credible physical events in the closed takeover trajectory, their sequence relationship, and the actual occurrence interval. Non-existent sensing sources are masked and do not participate in the corresponding supervision. During inference, unusable candidate sources are first removed based on the current working status of the sensing device, and then the expected observation order is formed according to the order of the outputs. When key observation sources are missing, the output boundary does not meet the aforementioned structural constraints, or the output cannot cover the necessary observation process expanded by the upper bound of the error, the model results are only used to preserve resources in advance, and a conservative initial takeover interval and observations to be verified are used to complete the subsequent takeover.

[0046] When the current input combination of device, payload, and sensing area lacks a closed historical takeover trajectory, continuous fields exceed the validated input boundaries of the model version, key sensing sources are missing, or the model output violates structural constraints such as the takeover interval, expected interval, or observation order, the current input will be marked as exceeding the model's applicability conditions. In this state, the time-series model output cannot be used to shorten the takeover interval, cancel pending observations, or release associated channel resources; it can only be used as a reference for preserving resources, and a conservative initial takeover interval and all necessary pending observations will be used to complete the takeover.

[0047] After the model completes training, it first uses the closed takeover trajectory based on the task occurrence time following the training data to verify the output interval and observation sequence. If the model cannot output a takeover interval with a valid start-end relationship for a certain combination of device, payload, and sensing area, omits a key observation source that is currently online and must be passed through in the closed trajectory, or if the expected occurrence interval output still cannot cover the actual acquisition time recorded on the verification trajectory according to a unified time base after being expanded according to the upper bound of clock synchronization error, then the model output for that combination will not be enabled, and a conservative initial takeover interval will continue to be used. The receiving wait boundary used to determine whether the observation message has timed out and how long the resources need to be held is calculated by adding an upper bound of communication delay to the uncertainty range of the acquisition time, and is checked against the actual receiving time in the verification trajectory. When input is missing during inference, the missing state and its source are input together; when the missing state affects key observation sources, the model results can only be used to hold resources in advance, and cannot be used to shorten the occupation or complete the state transition.

[0048] When a combination of sensing areas lacks sufficient closed historical takeover trajectories, a conservative initial takeover interval can be formed based on the warehouse map, the permissible speed of the handling equipment, the sensing coverage boundary, and the controller's response capability. Observations from all available sensing devices are then used as verification observations. Sufficient closed historical takeover trajectories mean that samples of the same handling equipment type, load shape and attitude, and sensing area combination can generate non-overlapping training and validation data according to the task occurrence time, and that the validation data includes currently online key observation sources; if this condition is not met, it is treated as insufficient samples. The conservative initial takeover interval defines both time and route location boundaries: the starting position is the route location where the work object, approaching the area boundary along the task route, is first likely to enter the boundary resource after extrapolation based on permissible speed, control response process, and safe stop range; the starting time is the earliest possible time of this first possible entry under the corresponding state. The ending position is the farthest route location corresponding to the completion of the current online key observation to be verified along the task route in the next sensing area, and must not be located before the first observable position along the task travel direction; the ending time is the conservative latest time when the work object reaches this position under permissible operating conditions and the corresponding observation can be received, where the communication waiting part uses the measured upper bound of the communication delay. If any boundary cannot be reliably determined, it is extended forward and backward along the route or time to the adjacent credible physical event or the safety boundary already configured in the storage system. This initial takeover interval is only used to maintain resources in advance and is not used to shorten existing safe occupancy. As the closed takeover trajectory accumulates, the time series model then generates an output adapted to the corresponding equipment, load, and area combination.

[0049] In step S104, the current sensing area generates a takeover status before the work object enters the takeover zone. The takeover status includes at least the handling task identifier, work object identifier, current sensing area and next sensing area identifiers, last reliable physical event, dynamic occupancy range, associated channel resources, the takeover zone output by the model, and the expected observation sequence. The takeover status can also record the sensing device status and timing model version used when it is generated, so as to determine whether the original takeover basis is still valid after control actions or changes in device status.

[0050] Each takeover status also has a takeover sequence number and a status version number. The takeover sequence number remains unchanged during a takeover of the same pair of adjacent sensing areas; when the same handling task continues into the next pair of adjacent sensing areas, a new takeover sequence number is generated and associated with the handling task identifier and the previously closed state to distinguish different work objects and different takeover processes. The status version number increments when any of the following changes: the last trusted physical event, dynamic occupancy range, associated channel resources, takeover interval, or expected observation sequence. The warehouse control node only processes messages with the same takeover sequence number and a status version number no earlier than the locally saved version; late older version messages are retained as trace records but must not overwrite the current takeover status.

[0051] During the takeover process, the state has at least the following states: current area maintained, awaiting confirmation, dual-area jointly maintained, control unstable, candidate trajectory locked, transferred and awaiting closure, and closed. After the current area generates a takeover state, it enters the state awaiting confirmation; after the next sensing area completes resource pre-occupation and returns confirmation of the same state version, if the bidirectional takeover conditions have not yet been met, it enters the dual-area jointly maintained state; when the control action changes the takeover basis, it enters the state of control unstable; when the local observation conflicts with the last trusted physical event, it enters the state of candidate trajectory locked; after the bidirectional takeover conditions are met, it enters the state of transferred and awaiting closure; after the next sensing area forms its first trusted physical event, it enters the state of closed. Except in the aforementioned rollback scenarios, no warehouse control node may skip intermediate states and directly mark the awaiting takeover state as closed.

[0052] The last credible physical event is used to determine the last physical node in the current sensing area that can assume responsibility for confirming the true state of the work object. The last credible physical event can be any of the following: the handling equipment passes a fixed detection position, the load-bearing state changes, the pick-and-place mechanism reaches the work position, or the handling equipment enters a stopped state. The fixed detection position can be formed by an RFID gate, photoelectric detector, landmark positioning point, or other location determination device; the change in load-bearing state can be confirmed by feedback from the pick-and-place mechanism, weighing, or visual load observation.

[0053] In a specific implementation, the last credible physical event is jointly confirmed by feedback from the handling equipment controller and observation from at least one warehouse sensing device. The handling equipment controller feedback, as referred to in this invention, means feedback generated by the handling equipment-side control unit or the equipment-side detection unit directly coupled to its operating state, reflecting at least one of the following actual states: the actual position, actual speed, braking or stopping state of the handling equipment, the actual position of the pick-and-place mechanism, and the load-bearing state. It does not include information merely indicating that control commands have been sent or received, or that the task state has been manually modified. For example, the equipment-side feedback indicates that the handling equipment has passed a predetermined route position, while a fixed detector detects the corresponding work object passing by; or the equipment-side feedback indicates that the pick-and-place mechanism has completed lifting, while a camera or weighing detector confirms a change in the load-bearing state. The confirmation results from both sources should point to the same handling task, the same work object, and the same event stage. Equipment-side feedback alone without independent warehouse sensing observation, or only a single visual target appearing briefly, will not be directly used as a credible physical event for state transition, resource release, or takeover closure.

[0054] like Figure 3 As shown, after the current sensing area generates a takeover status, it sends the takeover status to the warehouse control node corresponding to the next sensing area. The next sensing area returns a receipt confirmation after confirming that the takeover status content is complete and that it can access the associated channel resources. If the current sensing area does not receive a receipt confirmation, it does not transfer resource control responsibility to the next sensing area and maintains the current dynamic occupancy range and associated channel resources. This receipt confirmation only indicates that the next sensing area has obtained the takeover information; it does not equate to the takeover status being transferred.

[0055] Before the next sensing area returns a confirmation of receipt, it first establishes a pre-allocation record locally based on the dynamic occupancy range and associated channel resources in the takeover status. The confirmation of receipt includes the takeover sequence number, status version number, and a list of pre-allocated resources. The current sensing area only accepts confirmation of receipt if these three pieces of information match the sending status. If confirmations of receipt arrive repeatedly, resources are not added again. If the status versions are inconsistent, the newer takeover status version will trigger pre-allocation and confirmation again. Therefore, confirmation of receipt and resource pre-allocation use the same version as the current sensing area, avoiding a gap in responsibility when the next sensing area only receives data but has not yet allocated resources.

[0056] Dynamic occupancy range represents the space that a work object may actually occupy or that it needs to safely stop during takeover. The dynamic occupancy range updates with the location, speed, turning status, pick-and-place mechanism status, and load attitude of the handling equipment, and the results are mapped to aisles, intersections, work areas in front of storage locations, and workstation interface areas on the warehouse map. Associated aisle resources are warehouse resources that intersect with the dynamic occupancy range and need to prevent other handling tasks from entering simultaneously; when the dynamic occupancy range is updated, associated aisle resources are added, maintained, or released after the takeover release conditions are met.

[0057] like Figure 4 As shown, the dynamic occupancy range consists of the physical area of ​​the handling equipment and load, the sweeping range of the turning or pick-and-place action, and the range traversed to achieve a safe stop according to the current operating state. The physical area is formed based on the shape of the equipment, the shape of the load, and the load's attitude relative to the equipment; the sweeping range is formed based on the current turning direction, turning path, and the movement process of the pick-and-place mechanism from the current position to the target position; the range traversed to achieve a safe stop is formed based on the controller's current actual speed, permissible braking method, load status, and route direction. These three factors are combined according to the warehouse map coordinates to obtain the dynamic occupancy range at the current moment.

[0058] Based on the composition relationship of the above-mentioned entity range, sweep range, and safe stopping range, the dynamic occupancy range at the current moment can be expressed as:

[0059] ;

[0060] in, This is the calculation time for the current dynamically occupied range. For a moment The dynamic occupancy range, For the physical boundaries of the handling equipment, For the current load entity boundary, The sweeping range formed by turning or picking up / dropping actions. To ensure the coverage area required for a safe shutdown based on the current operating status, each set is converted to the same warehouse map coordinate system before a union operation is performed.

[0061] When forming the dynamic occupancy range, the equipment's external boundary is first transformed to warehouse map coordinates based on the current pose of the handling equipment. Then, the load boundary is transformed based on the load's external dimensions, the position of the pick-and-place mechanism, and the load's attitude relative to the equipment. Each boundary transformation uses the warehouse map coordinate system, map version, and calibration parameters recorded in the control system status. Pose, load, and sensor observations from different times are first aligned to the same calculation time according to a unified time reference. If coordinate transformation or time alignment cannot be reliably completed, the uncertain range of the corresponding observation is retained, and the occupancy boundary is expanded outwards, without shrinking the dynamic occupancy range accordingly. For turning actions, continuous equipment and load boundaries are generated between the current pose and the next route pose according to the actual turning direction fed back by the controller, and the union of the areas traversed by each boundary is taken. For pick-and-place actions, the areas traversed by the load and mechanism are generated according to the movement direction of the pick-and-place mechanism from its actual position to its target position. When control is unstable or there are two possible route branches, the traversed areas of each reachable branch are formed separately and the union is taken; no branch is prematurely deleted before the actual route is confirmed.

[0062] The warehouse map pre-records the spatial boundaries and resource identifiers of each of the channels, intersections, storage area front work areas, and workstation interface areas. The warehouse control node cross-references the current dynamic occupancy range with these spatial boundaries, adding resources that spatially intersect and are within the current takeover period to the associated channel resource list. When the dynamic occupancy range expands, new intersecting resources are added immediately; when the dynamic occupancy range shrinks, only a list of resources to be released is generated. After the takeover status transition is complete, resources that no longer intersect with the updated dynamic occupancy range and are not used by the next takeover relationship are released. For boundary resources maintained by the previous sensing area before this transition, the takeover status retains its resource identifier, control node, and release version until the first trusted physical event occurs in the next sensing area, forming a rollback resource list as the basis for restoring occupancy restrictions when a takeover rollback occurs. When the map or sensing coverage version changes, if the change does not affect the spatial boundaries, resource ownership, or key observation sources involved in this takeover, the takeover can be completed according to the version recorded when the current takeover status was generated. If the change alters the dynamic occupancy range, associated resource boundaries, the next sensing area, or key observation sources, the current takeover basis will be marked as invalid, existing resource occupancy will be maintained, and the takeover status will be rebuilt according to the changed version. The changed version will be directly used for tasks that have not yet entered the takeover area.

[0063] The safe stop range is not a fixed distance for all equipment, but is determined based on the actual speed and braking status, current load status, and route direction provided by the handling equipment controller. When the controller can provide feedback on the actual deceleration or stopping process, it can be determined based on response records from recent closed-loop control processes of the equipment; when the controller cannot provide a complete response, the control boundaries permitted by the equipment manufacturer or the safe stop boundaries already configured in the storage system are used. Regardless of the source, its purpose is to expand or maintain resource occupancy, and the dynamic occupancy range should not be shortened in the absence of reliable feedback.

[0064] In addition to the presence and attitude of the load, the current load status, when the handling equipment or work station can provide relevant information, also includes the load mass or mass range, load center of gravity height, or offset relative to the support surface. For routes where braking capacity is affected by ground conditions, the road surface condition or available braking capacity level corresponding to the current route is also recorded. When matching the safe stop range using closed-loop stop records, the above-mentioned parameters that can be obtained are used as screening criteria for similar states. If the load mass, center of gravity parameters, or road surface condition cannot be reliably obtained, it is not presumed to be a light load, low center of gravity, or high adhesion condition; instead, the most unfavorable load / road surface combination allowed by the equipment or the conservative stop boundary configured by the manufacturer or warehousing system is adopted.

[0065] When determining the safe stopping range using a closed-loop control process, stopping records matching the current state are selected based on the type of handling equipment, load status, route direction, and road conditions, with records whose initial speed is not lower than the current actual speed being given priority. The command issuance position, the position where the controller begins to change speed, and the position where the equipment reaches the stopping state in the stopping records are all corrected by controller feedback and warehouse sensing records. The route from command issuance to the start of speed change is included in the control interface cycle, measured command transmission delay, and execution response delay. When the controller cannot directly provide the command reception or execution time, a conservative upper limit of the delay formed by the closed-loop response records of the same equipment and control actions is used to extend the route. Among the selected matching stopping records, the stopping envelope is formed by the safety side boundary covering the furthest point along the task route from the command issuance position to the actual stopping position, avoiding narrowing the safe stopping range with a single, shorter braking record. When matching records are insufficient, controller feedback is incomplete, or road conditions change, historical records are not used to shorten the range; instead, the stopping envelope formed by the equipment's allowable control boundary is used.

[0066] When a closed-loop stop record exists that matches the current equipment type, load status, route direction, and road conditions, the conservative safe stopping distance along the mission route can be formed according to the matching record:

[0067] ;

[0068] in, To ensure a conservative safe stopping distance along the mission route, This is a set of closed-loop stopping records that match the current state and whose initial velocity is not lower than the current actual velocity. The record index in the closed stop record set, For the first The distance along the route from the location where the stop command was issued to the actual stopping location is recorded in this record. This is the safety-side extension resulting from positioning errors, map marking errors, load boundary measurement errors, and necessary safety margins. Only... This relationship is used when the record is not empty and is valid; when the record is insufficient or the feedback is incomplete, the stopping envelope formed by the device's allowed control boundary is still used.

[0069] After receiving the takeover status, the next sensing area listens to the local observations of each sensing device within its area according to the expected observation sequence. The expected observation sequence includes the source of the sensing device and the expected occurrence interval. For example, when the mission route may sequentially pass through the edge of the ultra-wideband positioning coverage, the effective field of view of the camera, and the fixed detection position, the model outputs the expected occurrence order of the corresponding sources; if the increased load will obscure the device identification, the expected observation can first require obtaining the shape of the load, and then require the formation of device identification at the fixed detection position.

[0070] When determining whether local observations conform to the expected observation sequence, first verify whether the source of the local observation belongs to the current expected source and falls within the corresponding expected occurrence interval; second, verify the identification of the work object corresponding to the local observation; third, verify whether a continuous motion relationship is formed between the local observation location and the last credible physical event; finally, verify whether the payload shape and attitude are consistent with the takeover status. If any of the verification items fails, the takeover status of the next sensing area is not confirmed.

[0071] The next sensing area establishes a verification position for each expected observation event item and writes the time-corrected local observations sequentially according to the acquisition time. If an observation occurs earlier than the corresponding expected occurrence interval, the observation is retained but awaits the establishment of the preceding event item; if an observation occurs later than the expected occurrence interval or the expected source is temporarily offline, subsequent observations are not forcibly moved forward to fill the missing positions, but the incomplete event items are retained and entered into a common retention phase. Repeated observations from the same sensing source are only updated for their corresponding event items if their position and direction of motion are continuous with the previous valid observation; observations with insufficient confidence, conflicting object identifiers, or unexplained payload states are not included in the completion order, but participate in candidate trajectory and occupancy range expansion.

[0072] The continuous motion relationship is determined based on the actual route and controller feedback. The warehouse control node, starting from the position and time of the last credible physical event, establishes the reachable route range of the work object within the current elapsed time along the remaining task route: the front end of the range extends forward based on the actual speed of the controller, the upper limit of allowable acceleration, the actual turning direction, and confirmed control actions; the rear end of the range retains the route positions that may still be occupied in cases of stopping, low speed, or communication delays, and the current dynamically occupied range is superimposed on the reachable route range. The local observation time of the next sensing area, after considering the time uncertainty range, should not be earlier than the last credible physical event; when two time uncertainty ranges overlap, the continuity relationship is not determined solely by time sequence, but its position and direction of motion should also fall within the reachable route range of the corresponding time, and changes in load state should be interpretable by the confirmed actions of the pick-and-place mechanism during the period. If the controller has confirmed a stop, the reachable route range must not continue to be pushed forward at the speed before the stop; if a detour or load attitude adjustment occurs during the period, the re-formed task route and dynamically occupied range are used for verification. This process does not use a fixed distance threshold, but rather uses the actual spatiotemporal range reachable by the current work object for judgment.

[0073] When it is necessary to conservatively extrapolate the reachable route's front end on a one-dimensional mileage coordinate system along a confirmed mission route, the following upper bound relationship can be used:

[0074] ;

[0075] in, Time elapsed since the last credible physical event The location at the beginning of the reachable route range, This represents the route location corresponding to the last credible physical event. This refers to the actual speed along the route confirmed by the controller at that moment. This represents the elapsed time since the last credible physical event. This represents the upper limit of the allowable acceleration along the path under the current control state. This is the safety-side extension amount calculated from the location, map marking, and time uncertainty to the route location. This relationship is only used to expand the reachability front end and not to shorten the dynamic occupancy range when feedback is missing; when there are detours or multiple route branches, the reachability ranges of each branch are formed separately and then the union is taken.

[0076] For the turning process, the reachable attitude range at the corresponding moment can be generated based on the actual turning angle, angular velocity or the turning rate and angular acceleration boundary allowed by the equipment controller. When the controller cannot provide reliable turning dynamics, the reachable route range is not narrowed accordingly, but the union of the sweep range of each route branch and its equipment and load allowed to enter under the current control state is taken.

[0077] When the controller can reliably provide the current angular velocity and the device has traceable permissible angular acceleration boundaries, time elapses. The absolute value of the heading change within the range can be conservatively bounded by the following upper bound:

[0078] ;

[0079] in, The elapsed time used for this reachable attitude extrapolation is... This is the change relative to the current heading. The current actual angular velocity, This is the upper limit of the permissible angular acceleration of the device under the current control state. This is the angle safety margin caused by asynchronous steering feedback, calibration, and sampling. It cannot be reliably obtained. or At this time, instead of using this relationship to narrow down the reachable attitude range, the union of the allowed route branches and sweep range is taken in the manner described above.

[0080] The identification of the work object can be formed by equipment number, task identifier, RFID tag, or cross-device fusion identifier. To avoid taking over based solely on the identifier field, identifier consistency must also be established in conjunction with positional continuity and load shape and attitude. Load shape and attitude consistency includes consistency in no-load or loaded state, continuous interpretation of the relative position of the load on the pick-up / placement mechanism, and consistency between load height changes and the pick-up / placement actions fed back by the controller. If local contour differences occur due to changes in viewing angle, but can be explained by attitude changes and observation direction, it can be considered consistent.

[0081] Confirmation of object departure within the current sensing area means that, after the last credible physical event, the current sensing area determines, based on local observations and controller feedback, that the object has crossed the effective coverage boundary of the area, and there are no credible observations indicating that the object is still within the area. The mere occurrence of temporary loss of visual target, communication data delay, or the controller sending a command to leave the target location does not constitute confirmation of departure.

[0082] The takeover status will only be transferred to the next sensing region when both conditions are met: the next sensing region receives local observations that conform to the expected observation sequence, and the current sensing region confirms that the work object has left. After the transfer, the next sensing region becomes the current holder of the takeover status, and the current sensing region can release associated channel resources that no longer intersect with the updated dynamic occupancy range. Resources that still intersect with the dynamic occupancy range or are continued to be used by the next takeover relationship will remain occupied.

[0083] Once the two-way conditions are met, the next sensing area sends an acknowledgment to the current sensing area. This acknowledgment carries the takeover sequence number, state version number, the local observation source for verification, and its acquisition time. The current sensing area verifies the departure evidence under the same state version and returns a transfer completion acknowledgment. Upon receiving the transfer completion acknowledgment and persisting the transferred pending closure state, the next sensing area becomes the current holder. The current sensing area can only execute the pending release resource list after confirming at the communication layer that the transfer completion acknowledgment has been reliably delivered, or after confirming from the shared state record that the next sensing area has persisted the same state version. If the acknowledgment, transfer completion acknowledgment, or delivery status is lost, both parties retransmit using the same takeover sequence number and state version. Duplicate messages only update the delivery status and do not repeatedly pre-allocate or release resources. Neither party releases resources independently due to communication retry.

[0084] When neither of the two conditions is met simultaneously, the current sensing area and the next sensing area jointly maintain a dynamically occupied range. Joint maintenance means that neither warehouse control node may unilaterally delete the takeover status or release the associated channel resources they can control, and both must report to the warehouse management system that the same work object is in a pending takeover state. When planning subsequent tasks, the warehouse management system treats the jointly maintained range as an occupied area, restricting other handling tasks from entering.

[0085] For overlapping coverage areas, the target object may be observed by two sensing areas simultaneously. After the next sensing area obtains a single local observation, it still needs to obtain subsequent local observations within the expected occurrence interval that can form a continuous relationship with the previous observation, or obtain a fixed detection location event, to be considered as obtaining continuous local observations. Until then, the current sensing area continues to maintain associated channel resources. This process prevents transient false detections, reflecting targets, or adjacent devices in the overlapping field of view from causing the previous area to be prematurely released after a single association.

[0086] In this embodiment, continuous local observation refers to at least two valid observations formed sequentially within the corresponding expected occurrence interval. The location of the latter observation can be interpreted from the location, time difference, task route, and controller feedback of the former observation, and there are no conflicting observations pointing to other work objects between them. When a credible physical event formed at a fixed detection location can simultaneously confirm the identity and direction of the work object, it can be considered as continuous local observation together with the preceding valid local observation. Multiple observations originating from the same instantaneous image, the same original measurement frame, or repeatedly reported by the same device are not considered continuous local observations.

[0087] For observation gaps, the current sensing area cannot relinquish its responsibility even after confirming the departure of the work object. The system extends the dynamically occupied area along the task route to the first observable position in the next sensing area. The extension must at least cover the route segment that the work object can reach under its current operating state and the area required for a safe stop. The current sensing area maintains the resources at the entrance of the observation gap, and the next sensing area maintains the first observable position and its entrance resources. Together, they prohibit other handling tasks from entering the extended dynamically occupied area.

[0088] If a fixed detection location capable of generating a credible physical event is set within an observation discontinuity, the event can update the last credible physical event and the dynamically occupied range, but it does not independently replace the local observation of the next sensing area. When the fixed detection location is managed by an independent control node, its event can be added to the secure takeover relationship as an intermediate takeover node; when managed by the current or next sensing area, it participates in the expected observation sequence as the local observation source of the corresponding area.

[0089] In step S105, the impact of speed limits, stops, rerouting, or load attitude adjustments on the takeover criteria is addressed. The control actions may be triggered by the method detecting that the takeover conditions are not met, or by warehouse scheduling avoidance, personnel entering passageways, equipment safety protection, or workstation requirements. Whenever a control action actually alters the subsequent trajectory, arrival time, braking range, or the load shape observable by the sensing equipment, it is necessary to determine whether the original takeover interval and expected observation sequence have become invalid.

[0090] like Figure 5 As shown, after issuing a control command, the takeover state is first set to an unstable control state. The unstable control state indicates that the expected effect of the control command has not yet been confirmed by the actual operation feedback of the handling equipment, and the original takeover zone is no longer used as the basis for releasing associated channel resources. The current sensing area and the next sensing area continue to maintain the dynamic occupancy range before control, and the resources that need to be maintained are increased according to the possible route range reached by the control action, until actual feedback is obtained from the controller.

[0091] Actual controller feedback includes actual speed changes, actual steering direction, braking or stopping completion status, actual route position after rerouting, and actual position of the pick-and-place mechanism. The system does not consider the sending of control commands or a successful reception response from the controller as actual feedback. For speed limiting and stopping, it is necessary to confirm that the equipment speed has changed according to the control result; for rerouting, it is necessary to confirm the actual steering direction and subsequent route position; for load attitude adjustment, it is necessary to confirm that the actual position of the pick-and-place mechanism and the corresponding changes in load status have occurred.

[0092] After receiving actual feedback from the controller, the dynamic occupancy range is updated based on the actual speed, actual steering direction, or actual position of the pick-and-place mechanism. Speed ​​limiting or stopping will change the safe stopping range and estimated arrival time; rerouting will change the route branch corresponding to the dynamic occupancy range and the next sensing area; load attitude adjustment will change the entity range, the sweep range of the pick-and-place action, and the load shape that the camera or other sensing devices are expected to observe. The updated dynamic occupancy range serves as a new input to the time-series model, rather than simply adding a fixed time to the original model output.

[0093] When determining whether the original takeover criteria have become invalid, the operating status saved when the original takeover status was generated is checked item by item against the actual feedback from the controller. If the actual route branch changes, the actual arrival status exceeds the originally expected range, the actual speed change causes the safe stopping range to exceed the original dynamic occupancy range, the load or pick-up / placement mechanism attitude changes so that the originally expected shape can no longer be used for verification, or the originally expected observation source leaves the effective coverage area, the original takeover range and the corresponding expected observation sequence are marked as invalid. If only a control command is received but the actual state has not changed, the control remains in an unstable state and is not rebuilt based on the command target value.

[0094] The system discards the original takeover zone as the basis for resource release and re-determines the takeover zone and expected observation sequence based on the actual feedback from the controller, the updated dynamic occupancy range, the current effective sensing coverage, and the remaining task routes. If the rerouting causes a change in the next sensing area, the transfer of takeover status to the original next sensing area is terminated, and the resources in the original area that still intersect with the dynamic occupancy range are retained until the work object leaves; an updated takeover status is generated and sent to the new next sensing area.

[0095] During reconstruction, the most recent still valid credible physical event is used as the new sequence starting point. The actual feedback from the controller, the updated dynamic occupancy range, the remaining task routes, the current online sensing sources, and the completed observation events are reassembled into the model input to generate a new takeover interval and expected observation sequence, and the takeover status version number is incremented. Resources already occupied by the old version continue to be retained until the new version is confirmed and pre-occupied by the corresponding next sensing area; resources no longer covered by the new version are only added to the release list after the version switch is confirmed, thus preventing resource gaps from being created during the reconstruction process itself.

[0096] If controller feedback is missing within the predetermined response process or conflicts with local observations, the takeover state remains in an unstable control state. The response process is determined based on the cycle of the transport equipment's control interface, communication delay records, and closed-loop response records of the same control action; no uniform value is set to deviate from the equipment's capabilities. During the unstable control period, time-series model predictions are not used to shorten resource usage, and a stop command or request for manual takeover can be sent to the transport equipment. After manual takeover occurs, the trajectory of that task is not used for subsequent model updates.

[0097] If no actual feedback proving the control action has taken effect is obtained after the scheduled response period, the system will not reuse the same control command indefinitely. Instead, it will maintain or expand the occupation of associated channel resources, send a stop request to the device-level safety control terminal, and transfer the current task to manual handling. Manual handling requires at least the formation of a confirmation record pointing to the current work object and safety location through the vehicle terminal, on-site fixed detection device, or manual confirmation terminal; before the formation of the confirmation record, the takeover status must not be changed to closed.

[0098] In step S106, local observation conflicts, post-takeover failures, and model updates are handled. When the local observations of the next sensing area and the last credible physical event of the current sensing area cannot form a continuous motion relationship, there may be situations such as target misassociation, communication delays, temporary rerouting of the work object, or two adjacent handling devices entering the boundary simultaneously. The system does not immediately select the single trajectory with the higher probability and release the resources occupied by the other trajectory, but instead retains the candidate trajectories formed by the local observations and the last credible physical event respectively.

[0099] like Figure 6 As shown, the first candidate trajectory extends from the last reliable physical event along the actual route and dynamic occupancy range of the controller, indicating that the original work object has not yet been reliably observed by the next sensing area; the second candidate trajectory traces back along the task route from the local observation in the next sensing area, indicating that the local observation may correspond to the work object entering the next sensing area. The system issues a stop command to the current handling equipment and maintains the occupancy of the channel resources covered by the two candidate trajectories. The actual position and stop status formed after the stop command is executed also serve as the observation basis for subsequent exclusion of candidate trajectories.

[0100] If a new credible physical event can establish a continuous motion relationship with one of the candidate trajectories and exclude the other, the candidate trajectory that does not meet the criteria is excluded. For example, if a fixed detection location confirms that the original transport equipment is still within the observation discontinuity, then the conflicting local observation in the next sensing area is not considered evidence of takeover. Similarly, if the fixed detection location and controller feedback in the next sensing area jointly confirm that the original transport equipment has arrived, then the candidate trajectory extending from the last credible physical event but not entering the next sensing area is excluded. Coverage resources for a candidate trajectory are not removed until it has been excluded.

[0101] If no new credible physical event is formed within the candidate trajectory verification time limit determined based on the current control state, the stop command and the resource occupation of the two candidate trajectories are maintained, and manual verification is initiated. The verification time limit is determined jointly based on the candidate trajectory coverage route length, the allowable speed and acceleration boundaries of the equipment under the current control state, the equipment stop response process, the sensing device acquisition cycle, and the measured communication delay. Its purpose is to cover the conservative process required for the work object to develop from the current state to the next verifiable physical event or to complete stop confirmation, and it does not use the route length directly divided by the current instantaneous speed as the sole calculation basis. When the current control state allows for continuous stopping or waiting, the verification process is not extended indefinitely by the route arrival time; after the equipment confirms stop, the process switches to the conservative time window required for stop state confirmation, acquisition of the next credible physical event, and communication completion. Manual verification can only exclude invalid candidate trajectories after confirming the actual location, load status, and corresponding task identifier of the work object; if the verification result still conflicts with any existing automatic observation, automatic takeover is terminated, and relevant resources are maintained until the source of the conflict is recalibrated or restored.

[0102] After the takeover state transition is complete, the next sensing region must generate the first credible physical event to mark the takeover state transition as closed. This first credible physical event serves to prove that the next sensing region has not only obtained takeover observations but has also established a transferable physical confirmation basis within its region. Before the first credible physical event is generated, the previous sensing region retains the last credible physical event, state copy, and rollback resource list related to this takeover, but does not duplicate control over resources already explicitly maintained by the next sensing region.

[0103] If the takeover status has been transferred, but the first credible physical event has not yet occurred in the next sensing area, and the next sensing area experiences a sensing interruption or loses control feedback from the handling equipment controller, the takeover status will be restored to the state jointly maintained by both sensing areas. The party detecting the fault first freezes new task assignments in the rollback resource list using the current takeover sequence number and status version, and sends a rollback request to the other party; the other party restores the boundary resource occupancy limits it can control and returns a recovery confirmation, and the next sensing area continues to maintain its current dynamic occupancy range. No new associated handling tasks are allowed to enter until both parties obtain recovery confirmation for the same status version. This rollback does not require the work object to revert to the previous area, but rather restores shared safety responsibility across boundaries.

[0104] If the rollback resource has been allocated to another handling task before the rollback request is issued, but that task has not yet entered the resource, the allocation is revoked and the task's entry is stopped. If another handling task has already entered the resource, its occupancy record is not directly overwritten. Instead, a stop command is sent to the relevant handling equipment, incorporating its current reachable range and the original work object's dynamic occupancy range into the resource limit, until both work objects form a credible physical event that confirms a safe location. If resource recovery fails or recovery confirmation times out, the fault state is maintained and manual handling is initiated. It is prohibited to end the joint maintenance with a unilateral recovery result.

[0105] After the sensing devices in the next sensing area are restored, it is necessary to reacquire local observations consistent with the takeover status and combine them with the actual feedback from the controller after its recovery to form a new credible physical event. If the controller is disconnected but the independent fixed detection device can still confirm that the work object has stopped in a safe position, resource occupation can be maintained and the system can wait for control to be restored; if a safe position cannot be confirmed, the dynamic occupation range should be expanded to the possible route range, and the relevant channels should be prevented from continuing to assign tasks.

[0106] The recovery termination conditions after takeover rollback are as follows: the next sensing area regains local observations consistent with the current state version; the controller or an independent source capable of reflecting the device's own state recovers feedback; both together form a trusted physical event for the next sensing area; all resources in the rollback resource list have obtained recovery confirmation of the same state version; and the recovered dynamic occupancy range has been rechecked against the list of resources to be released. After the above conditions are met, the state transition is executed again according to the bidirectional confirmation and transfer completion confirmation; if only the communication connection is restored but no trusted physical event is formed, the joint maintenance does not end.

[0107] After a takeover state transition is closed, the actual takeover interval, actual observation sequence, and trajectory changes caused by control commands are recorded to form candidate update samples. These candidate update samples are first merged into a closed sample set corresponding to the handling equipment category, load shape and attitude, and sensing area combination. When the update cycle configured by the warehousing system is reached or a new closed sample condition is added, the candidate update model is trained using the updated closed sample set while retaining the current model version. The update cycle and the new closed sample condition are constrained to the minimum requirement of being able to continue forming training and validation data isolated according to task occurrence time; the current model is not directly replaced by a single trajectory that has not yet been replaced and validated. The actual takeover interval is jointly determined by the actual location and time at which the takeover state begins and the actual location and time at which the next sensing area obtains takeover evidence; the actual observation sequence is formed according to the local observation sources and occurrence intervals after confirmation by credible physical events. Model updates are performed according to the handling equipment category, load shape and attitude, and sensing area combination to avoid indiscriminately changing the output of other combinations during the takeover process of one type of equipment or load.

[0108] Task trajectories involving manual takeover, unresolved sensing equipment malfunctions, or unexcluded candidate trajectories will not be used to update the time series model. Manual takeover may alter normal arrival relationships under automatic control; unresolved sensing malfunctions cannot form a complete observation sequence; unexcluded candidate trajectories indicate that target attribution is not yet closed. These trajectories may be retained for fault tracing, but must not be used as supervisory samples for takeover intervals and expected observation sequences. After the model version is updated, tasks already in the takeover interval will continue to use the model output from when their takeover state was generated, while tasks not yet in the takeover interval will be re-inferred using the new model to avoid inconsistencies in the basis for the same takeover process.

[0109] After forming a candidate update model, a closed takeover trajectory occurring later than the training samples is used for replacement verification. The verification checks whether the takeover interval covers the actual process from the start of resource maintenance to obtaining takeover evidence, whether the expected observation sequence includes the key sources actually used for takeover confirmation, and whether the reconstructed output after control actions is consistent with the actual trajectory. If the candidate update model exhibits any of the following: a start point later than the actual need to maintain resources, an end point earlier than the actual acquisition of takeover evidence, omission of key observation sources, or output of the old route sequence after control, the original model version is retained, and the corresponding samples are transferred for verification. After the new model is enabled, the original model version and applicable combinations are saved. If, during operation, the same device, payload, and sensing area combination continuously outputs that cannot form an effective takeover interval or observation sequence across different takeover processes, the model is reverted to the original model or a conservative initial takeover interval. The continuous anomaly triggering condition is configured based on the recent closed-loop operation records and replacement verification results of the combination. Duplicate messages in the same takeover process are not accumulated as multiple anomalies, and this condition is only used for model rollback, not for resource release.

Claims

1. An intelligent warehouse safety management method based on artificial intelligence, characterized in that, include: Establish a safe takeover relationship for the work object based on the sensing area traversed by the handling task, and obtain the location of the handling equipment, load attitude, task route, effective coverage of the sensing equipment, and controller feedback; Using a time series model trained from a closed historical takeover trajectory, including a sequence state extraction part, a takeover interval output part, and an observation sequence output part, the takeover interval and expected observation sequence of the working object from the current sensing area to the next sensing area are determined. Before entering the takeover zone, the current sensing area generates a takeover status including the last trusted physical event, the dynamic occupied range, and the associated channel resources, and sends it to the next sensing area; When the next sensing area obtains local observations that conform to the expected observation sequence and the current sensing area confirms that the work object has left, the takeover state is transferred and the associated channel resources that no longer intersect with the dynamic occupancy range are released; otherwise, the two sensing areas jointly maintain the dynamic occupancy range and restrict the entry of subsequent tasks. When speed limits, stops, rerouting, or load attitude adjustments alter the subsequent trajectory or observable range, the original takeover section is terminated, and the takeover section and expected observation sequence are reconstructed based on actual controller feedback until the takeover state transition is completed.

2. The intelligent warehouse safety management method based on artificial intelligence according to claim 1, characterized in that, The sensing area is divided according to the effective coverage of the sensing devices received by the same warehouse control node, and overlapping coverage sections or observation discontinuities are formed between adjacent sensing areas. The security takeover relationship is recorded sequentially according to the task route, including each sensing area, the overlapping coverage segment or observation discontinuity segment, and the corresponding warehouse control node.

3. The intelligent warehouse safety management method based on artificial intelligence according to claim 2, characterized in that, The historical takeover trajectory consists of the closed task trajectory of the takeover state transition; the time series model takes the task route, the type of handling equipment, the shape and attitude of the load, the running speed, the effective coverage of the sensing equipment and the working status of the sensing equipment as input, and the sequence state extraction part forms a sequence state arranged in the order of event occurrence. The takeover interval output part outputs the start and end positions and start and end times of the takeover interval, and the observation sequence output part outputs the sequential relationship of the sensing sources to be verified in the next sensing area and the expected occurrence interval. During training, the actual spatiotemporal boundaries of the closed takeover trajectory, as well as the sensing sources, sequence relationships, and actual occurrence intervals confirmed by credible physical events, are used as the supervision results.

4. The intelligent warehouse safety management method based on artificial intelligence according to claim 1, characterized in that, The last credible physical event is any one of the following events: the handling equipment passes through a fixed detection position, the load bearing state changes, the pick-and-place mechanism reaches the working position, or the handling equipment enters a stopped state, and is confirmed by feedback from the handling equipment controller and observation from at least one warehouse sensing device. The dynamic occupancy range covers the physical range of the handling equipment and load, the sweeping range of turning or picking / placing actions, and the safe stop range. The safe stop range is determined based on the actual speed and braking status, current load status, and route direction fed back by the handling equipment controller. When there is a closed stop record that matches the current status, the conservative stop envelope formed by the closed stop record is used. When there are insufficient closed stop records or the controller feedback is incomplete, the equipment allowable control boundary or the safe stop boundary configured by the warehousing system is used.

5. The intelligent warehouse safety management method based on artificial intelligence according to claim 3, characterized in that, The expected observation sequence includes the source of the sensing device for each local observation in the next sensing area and the expected occurrence interval; When it is determined that the local observation conforms to the expected observation sequence, the consistency of the operation object identifier, position continuity, load shape and attitude corresponding to the local observation with the takeover status is checked in turn. If any check item fails to meet the requirements, the takeover status of the next sensing area is not confirmed.

6. The intelligent warehouse safety management method based on artificial intelligence according to claim 5, characterized in that, When the target object is located in the overlapping coverage segment, the associated channel resources of the current sensing area are maintained until continuous local observations are obtained in the next sensing area; When the work object is located in the observation discontinuity segment, the dynamic occupancy range is extended along the task route to the first observable position of the next sensing area, and the current sensing area and the next sensing area jointly prohibit other transport tasks from entering the dynamic occupancy range.

7. The intelligent warehouse safety management method based on artificial intelligence according to claim 6, characterized in that, After issuing the control commands corresponding to the speed limit, stop, rerouting, or load attitude adjustment, the takeover state is set to an unstable control state, and the original takeover zone is no longer used as the basis for resource release. The dynamic occupancy range is updated based on the actual speed of the handling equipment, the actual turning direction, or the actual position of the pick-and-place mechanism, and the update result is used as input for reconstructing the takeover section and the expected observation sequence.

8. The intelligent warehouse safety management method based on artificial intelligence according to claim 7, characterized in that, When the local observation in the next sensing area and the last credible physical event in the current sensing area cannot form a continuous motion relationship, the candidate trajectories formed by the local observation and the last credible physical event are retained respectively. A stop command is issued to the transport equipment, and the occupancy of the channel resources covered by the two candidate trajectories is maintained until a new credible physical event eliminates one of the candidate trajectories.

9. The intelligent warehouse safety management method based on artificial intelligence according to claim 8, characterized in that, After the takeover state has been transferred and before the first credible physical event is formed in the next sensing area, if the next sensing area experiences a sensing interruption or loses control feedback from the handling equipment controller, the takeover state will be restored to a state jointly maintained by the two sensing areas. The occupancy limit of the channel resources between the current sensing area and the next sensing area will be restored, and the handling tasks associated with the channel resources will be stopped from entering.

10. The intelligent warehouse safety management method based on artificial intelligence according to claim 9, characterized in that, After a takeover state transition is closed, an update sample is formed based on the actual takeover interval, the actual observation sequence, and the trajectory changes caused by the control command. The time series model is then updated according to the combination of the handling equipment type, load shape and attitude, and sensing area. Task trajectories that include manual takeover, unrecovered sensing device malfunctions, or candidate trajectories not being excluded are not used to update the time series model.

Images