Intelligent warehouse intelligent shelf precise perception system and method based on multi-sensor fusion of wind power

Abstract

The application discloses a wind power intelligent warehouse intelligent shelf precise perception system and method based on multi-sensor fusion, and particularly relates to the technical field of intelligent warehouse perception, and the method comprises the following steps: establishing slot identification, and forming an initial reference model of a multi-sensor of a shelf; taking the change of a door magnet state as a priority trigger basis, combining the change of weighing, the change of vision and the change of an RFID tag set to determine a candidate operation event, and determining an operation event confirmation value based on time sequence coincidence degree, slot consistency degree and quality interpretation degree, and outputting the confirmed operation event; forming radio frequency deviation degree, visual deviation degree and weighing residual degree based on the confirmed operation event, and determining a long-term drift discrimination value; and when the conditions are met, performing a constrained and gradual updating on the layer domain read-write reference, the slot image mapping reference, the weighing baseline and the standard quality of the material unit. The application can improve the accuracy, stability and long-term operation reliability of inventory perception.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehousing sensing technology, and more specifically, to a precise sensing system and method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion. Background Technology

[0002] Wind farms typically have spare parts warehouses to store various materials such as fasteners, electrical components, sensor assemblies, tool accessories, and storage boxes. Due to the large variety of materials, significant differences in specifications, frequent material requisitions and returns, and the fact that similar materials may be scattered across different shelves, shelf levels, and storage locations, accurate sensing of the location status, material identification, and quantity changes of materials in each storage location is fundamental to ensuring efficient inventory management, traceability of material requisitions, and accuracy of spare parts management. Current technologies typically employ electronic tag identification, weighing detection, image acquisition, and door status detection to identify the storage status of materials on intelligent shelves. This information is then combined with preset shelf codes, storage location codes, and inventory ledger information to complete material entry / exit records or update inventory status.

[0003] However, in the actual operation of spare parts warehouses at wind farms, a single material requisition or return operation typically involves a short, continuous process of opening the door, taking out and placing turnover boxes, taking out or placing parts, and closing the door. Different sensors have inconsistent sampling times, sensing ranges, and response methods, which can easily lead to issues such as tag cross-reading, visual obstruction, short-term weighing disturbances, and brief door openings. Especially under long-term operating conditions, fundamental parameters such as shelf reading range, slot image mapping relationships, weighing benchmarks, and material standard quality may gradually deviate from initial calibration results due to changes in installation status, wear and tear, environmental fluctuations, and material turnover. Existing solutions often rely on initial calibration information, fixed judgment rules, or single observation results for inventory identification. When there are inconsistencies in timing, slot correspondence, or quality interpretation among multi-source observation results, it can easily lead to slot judgment bias, material identification bias, and inaccurate quantity change judgment, thus affecting the accuracy and long-term stability of inventory perception results. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a precise sensing system and method for wind power smart warehouse intelligent racking based on multi-sensor fusion to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion includes the following steps: Establish slot identification and form an initial baseline model for the multi-sensor shelving; The system prioritizes door magnet status changes as the triggering basis and determines candidate operation events by combining weighing changes, visual changes, and RFID tag set changes within a preset triggering period. It constructs an event window and organizes multi-source data within the window, extracting the action center time, action center location, stability quality difference, and tag status change results. Based on the timing consistency, slot consistency, and quality interpretability, it determines the operation event confirmation value. When the operation event confirmation value reaches the preset confirmation threshold, it outputs the confirmed operation event. Only confirmed operational events are written into the self-calibration sample pool. Based on the confirmed operational events corresponding to the analysis object, radio frequency deviation, visual deviation and weighing residual are generated, and then the long-term drift discrimination value is determined. When the long-term drift discrimination value is greater than or equal to the preset drift threshold and holds for multiple consecutive statistical periods, or when at least two deviations constituting the long-term drift discrimination value exceed their respective sub-thresholds, the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass are updated in a constrained and progressive manner. When the parallel determination of the parameters before and after the update shows that the updated parameters meet the preset improvement requirements, the updated parameters are written into the formal operation configuration.

[0006] In a preferred embodiment, the initial reference model of the multi-sensor shelving includes at least a layer domain read / write reference, a slot image mapping reference, a weighing baseline, an environmental reference baseline, slot geometric information, and material unit standard mass information; the establishment of slot identification includes: establishing a slot identification for each actual storage location according to the unified coding rules of shelving group, shelving side, shelving layer, and slot, and associating each slot with its corresponding material category, allowable storage range, turnover box identification, electronic tag identification, and weighing acquisition location.

[0007] In a preferred embodiment, determining the candidate operation event includes: when the door magnetic detection unit detects that the door has switched from a closed state to an open state, if, within the preset trigger time period, the weighing output of the target slot or adjacent slot changes beyond a preset threshold relative to the stable quality benchmark, or a hand, turnover box, or material outline appears in the visual image entering the shelf slot area, or the RFID reading and writing unit detects an increase or decrease in the tag set within the corresponding layer, then the operation process corresponding to this door magnetic opening is defined as a candidate operation event; constructing the event window and organizing the multi-source data within the window includes: extracting data for a preset pre-trigger time centered on the first trigger time, and continuously receiving data for a preset post-trigger time; when the door is still in the open state before the preset post-trigger time ends, or the visual branch continuously detects that the hand or turnover box is still near the target slot, or the weighing output has not yet recovered to the stable range, the event window is extended until the door closes and the multi-source observation results re-enter a stable state, or the preset longest window duration is reached.

[0008] In a preferred embodiment, the operation event confirmation value is formed by timing consistency, slot consistency, and quality interpretability. The timing consistency is determined based on the time interval between the action center moment of the visual branch, the moment when the stable quality change of the weighing branch reaches the judgment condition, and the moment when the tag status change of the RFID branch first occurs, combined with a preset timing tolerance. The slot consistency is determined based on the consistency between the target slot pointed to by the visual branch, the target slot corresponding to the weighing change, and the target slot corresponding to the RFID branch. The quality interpretability is determined based on the correspondence between the stable quality difference and the unit standard quality of the candidate material. When the operation event confirmation value reaches a preset confirmation threshold, the target slot, candidate material, operation type, quantity change result, door magnetic opening and closing moment, action center moment, stable quality difference, and corresponding layer tag change result are recorded.

[0009] In a preferred embodiment, only confirmed operation events whose confirmation values ​​reach a preset confirmation threshold are written into the self-calibration sample pool, and an index record is established for each confirmed operation event. The index record includes at least the event number, target slot, belonging layer, corresponding material, door magnetic opening time, door magnetic closing time, visual motion center position, visual motion center time, stable quality value before operation, stable quality value after operation, RFID tag status change result, environmental status, and the final confirmed operation type and quantity change result. When the number of confirmed operation events corresponding to the analysis object is not less than the preset minimum sample size, the calculation of the long-term drift discrimination value is initiated.

[0010] In a preferred embodiment, the analysis object is a slot, layer, or material; the long-term drift discrimination value is formed by radio frequency deviation, visual deviation, and weighing residual; wherein, the radio frequency deviation is determined based on the correspondence between candidate slots obtained by RFID mapping and the finally confirmed slots, the visual deviation is determined based on the offset between the visual motion center position and the center position of the finally confirmed slot, combined with the characteristic length of the slot, and the weighing residual is determined based on the difference between the stable mass difference and the theoretical mass difference corresponding to the number of pieces and the standard mass of the material unit based on the quantity change result.

[0011] In a preferred embodiment, when the candidate slot obtained by RFID mapping is consistent with the final confirmed slot, the corresponding event is recorded as a consistent event; when the candidate slot obtained by RFID mapping is inconsistent with the final confirmed slot but is located in adjacent positions within the same layer domain, the corresponding event is recorded as a proximity deviation event; when the candidate slot obtained by RFID mapping is significantly inconsistent with the final confirmed slot, the corresponding event is recorded as a significant deviation event, and the radio frequency deviation degree is determined based on the distribution of consistent events, proximity deviation events, and significant deviation events in the confirmed operation events; when multiple confirmed operation events show that the visual motion center of the same slot continuously shifts in the same direction, the visual deviation degree is increased; when multiple confirmed operation events continuously show that there is a difference deviation in the same direction between the stable quality difference and the theoretical quality difference, the weighing residual degree is increased.

[0012] In a preferred embodiment, the constrained progressive update includes: updating the weight distribution of tag reading and writing results to specific slot assignments within the layer domain based on the correspondence between candidate slots and final confirmed slots given by the RFID branch in confirmed operation events related to the target layer domain; smoothing the slot image mapping boundary or center position based on the long-term offset direction and offset amount between the visual action center position and the final confirmed slot center position; updating the weighing baseline based on the weighing value in the empty slot confirmation event, and progressively adjusting the standard mass of the material unit based on the quality difference samples in multiple confirmed material requisition and return events; and adjusting the weight coefficients in the operation event confirmation value and long-term drift discrimination value according to the consistency between each sensor branch and the final confirmation result within a preset time period.

[0013] In a preferred embodiment, the parallel determination of the pre-update parameters and post-update parameters includes: performing dual-path parallel determination on newly generated material requisition and return operation events using the pre-update parameters and post-update parameters respectively within a preset verification period, and comparing the changes of the two sets of parameters in target slot identification accuracy, quantity change determination accuracy, proportion of events awaiting review, and false alarm rate; only when the post-update parameters are consistently better than the pre-update parameters in at least a preset number of items among the target slot identification accuracy, quantity change determination accuracy, proportion of events awaiting review, and false alarm rate, and do not cause significant deterioration of other indicators, are the post-update parameters written into the formal operation configuration; otherwise, the settings revert to the pre-update parameters.

[0014] In a preferred embodiment, the following modules are included: The baseline modeling module is used to establish slot identification and form an initial baseline model of the multi-sensor rack. The event confirmation module is used to identify candidate operation events, construct an event window and organize multi-source data within the window, extract the action center time, action center location, stability quality difference and label status change results, determine the operation event confirmation value and output the confirmed operation event; The drift discrimination module is used to write confirmed operation events into the self-calibration sample pool, and to generate radio frequency deviation, visual deviation and weighing residual based on the confirmed operation events corresponding to the analysis object, thereby determining the long-term drift discrimination value; The update verification module is used to perform constrained incremental updates to the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass when the long-term drift discrimination value meets the update conditions. When the updated parameters meet the preset improvement requirements, the updated parameters are written into the formal operation configuration.

[0015] The technical effects and advantages of this invention are as follows: This invention establishes slot identification and forms an initial benchmark model of multi-sensor racking. It uses changes in door magnetic states as the primary trigger, and combines changes in weighing, visual, and RFID tag sets to determine candidate operation events. The operation events are then confirmed based on temporal consistency, slot consistency, and quality interpretability. This unifies and comprehensively judges multi-source observation results during the same material requisition or return process. Consequently, it effectively reduces identification deviations caused by tag cross-reading, visual obstruction, weighing disturbances, and asynchronous multi-source signals, improving the accuracy of target slot identification, material identification, and quantity change determination. It also reduces false alarm rates and the proportion of events requiring verification, thereby enhancing the accuracy and stability of inventory perception results in wind power smart warehousing scenarios.

[0016] This invention further writes only confirmed operational events into the self-calibration sample pool, and generates radio frequency deviation, visual deviation, and weighing residual based on the confirmed operational events to determine long-term drift discrimination values. When conditions are met, the layer domain read / write baseline, slot image mapping baseline, weighing baseline, and material unit standard mass are updated in a constrained, progressive manner. Whether to write the updated parameters into the formal operation configuration is determined through parallel evaluation of the parameters before and after the update. Therefore, it is possible to continuously correct multi-sensor baseline mismatch during long-term operation, avoid interference from occasional anomalies on model updates, ensure the controllability and effectiveness of the parameter update process, and thus improve the robustness, reliability, and continuous sensing capability of the system under long-term operating conditions. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart illustrating the precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to the present invention. Figure 2This is a timing diagram illustrating the triggering of candidate operation events and the construction of the event window in this invention. Figure 3 This is a schematic diagram of the closed-loop operation of long-term drift detection and linkage update of the present invention; Figure 4 This is a schematic diagram of the structure of the wind power smart warehouse intelligent shelf precision sensing system based on multi-sensor fusion of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: The present invention provides a precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion, such as... Figure 1 As shown, it includes the following steps: Step 1: Establish slot markings to create an initial baseline model for the multi-sensor shelving system; Step 2: Using changes in the door magnetic sensor status as the primary triggering criterion, candidate operation events are determined by combining changes in weighing, visual changes, and changes in the RFID tag set within a preset triggering period. An event window is constructed, and multi-source data within the window is organized. The time of the action center, the location of the action center, the stability quality difference, and the tag status change results are extracted. The operation event confirmation value is determined based on the temporal consistency, slot consistency, and quality interpretability. When the operation event confirmation value reaches the preset confirmation threshold, the confirmed operation event is output. Step 3: Only confirmed operation events are written into the self-calibration sample pool. Based on the confirmed operation events corresponding to the analysis object, radio frequency deviation, visual deviation and weighing residual are generated, and then the long-term drift discrimination value is determined. Step 4: When the long-term drift discrimination value is greater than or equal to the preset drift threshold and this holds true for multiple consecutive statistical periods, or when at least two deviations constituting the long-term drift discrimination value exceed their respective sub-thresholds, the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass are updated in a constrained and progressive manner. When the parallel determination of the parameters before and after the update shows that the updated parameters meet the preset improvement requirements, the updated parameters are written into the formal operation configuration.

[0020] Specifically: Step 1: After the system is deployed in the wind farm spare parts warehouse, the edge control unit first reads the shelf layout data, slot code data, material master data, turnover box binding data, electronic tag binding data, material unit standard quality data, allowed storage relationships, and environmental threshold parameters. Following a unified coding rule for shelf groups, shelf sides, shelf layers, and slots, a unique slot identifier is established for each actual storage location. The shelf group distinguishes different shelf bodies, the shelf side distinguishes the orientation or access surface of the same shelf, the shelf layer distinguishes vertical levels, and the slot distinguishes the specific storage unit within the same layer. After coding, the edge control unit associates each slot with its corresponding material category, allowed storage range, turnover box identifier, electronic tag identifier, and weighing acquisition location, establishing a slot master data table and an inventory status table. This ensures that subsequent observations collected by different sensors can all be traced back to the same slot object.

[0021] After modeling the shelf objects, the edge control unit performs access and time synchronization processing for each sensor. Specifically, RFID reading / writing units located in each shelf layer, weighing acquisition units located at the bottom of slots or shelves, visual acquisition units located on the front or top of the shelf, door magnetic detection units located at door positions, and temperature and humidity detection units located inside the shelf complete device registration, communication link establishment, and clock synchronization, respectively. The layer refers to the reading / writing area within the same shelf layer covered by the same RFID reading / writing unit. The edge control unit assigns a unified time reference to each sensor, uniformly marking the visual frame acquisition time, RFID tag reading time, weighing data sampling time, door magnetic status change time, and temperature and humidity sampling time. This ensures that in subsequent material requisition or return operations, the observation results generated by different sensors can be aligned and compared along the same timeline.

[0022] After all sensors have been connected, the edge control unit collects initial baseline data under unattended conditions and when the shelf is stationary. For the RFID read / write unit, the edge control unit continuously collects the electronic tag read / write results for a preset duration in each layer, statistically analyzes the stable occurrence of each tag in its corresponding layer, and writes the set of stable tags, their frequency of occurrence, and the layer correspondence into the layer read / write baseline. For the vision acquisition unit, the edge control unit calibrates the projection range of each slot in the image based on the shelf installation location and slot structure boundaries, determines the slot boundaries, slot center, and the boundary relationship between adjacent slots, and forms a slot image mapping baseline. For the weighing acquisition unit, the edge control unit collects the output values ​​of each slot under no-load and current stable load conditions, forming the no-load baseline value and stable mass baseline value for each slot. For the temperature and humidity detection unit, the edge control unit simultaneously records the initial temperature and humidity conditions inside the shelf, forming an environmental reference baseline.

[0023] Considering that the main technical problem of this implementation is the gradual mismatch of multi-sensor baselines after long-term operation, this step also records the key basic quantities required for subsequent judgment. Specifically, the edge control unit stores the geometric center position of each slot, the slot feature length, the layer domain number to which it belongs, and the unit standard mass of the corresponding material for each slot. The slot feature length is used to characterize the representative size of the slot in the image plane or shelf plane, and is used to normalize the distance between the action center and the confirmed slot center when calculating the visual deviation. The unit standard mass is used to characterize the standard mass of a single piece or a group of corresponding materials, and is used to convert the stable mass difference into an interpretable change in the number of pieces when calculating the mass interpretability and the weighing residual.

[0024] To prevent the initial baseline from being affected by occasional disturbances, the edge control unit removes abnormal samples during the initial baseline acquisition process. If, during the acquisition period, a door magnet is detected to be open, personnel are seen entering the visual image, there is a sudden change in the slot quality value, there is an abnormal increase or decrease in the RFID tag set, or there is a drastic fluctuation in temperature and humidity within a short period of time, the data for the corresponding time period is determined to be invalid baseline data and is re-acquired. Only when the outputs of each sensor remain stable for multiple consecutive sampling periods, and there is no significant conflict between the visual position, weighing result, and RFID layer domain relationship of the same slot, is that segment of data written into the shelf multi-sensor initial baseline model. Through the above processing, the system forms a shelf multi-sensor initial baseline model, which includes at least a layer domain read / write baseline, a slot image mapping baseline, a weighing baseline, an environmental reference baseline, slot geometric information, and material unit standard mass information.

[0025] Step Two: After forming the initial baseline model of the multi-sensor shelving in Step One, as follows... Figure 2 As shown, the edge control unit enters continuous monitoring mode, cyclically receiving and buffering the real-time outputs of the door magnetic detection unit, weighing acquisition unit, vision acquisition unit, and RFID reading and writing unit. Since the material requisition and return operations in the wind farm spare parts warehouse typically involve a short, continuous process of opening the door, taking out the box, taking out the part, putting it back, and closing the door, and the door magnetic status, vision trajectory, weighing changes, and RFID tag status changes are not synchronized during this process, this step does not directly determine inventory based on data from a single moment. Instead, it constructs an event window based on a continuous operation process, so that the same physical action corresponds to the same processing unit in time.

[0026] In practice, the edge control unit first uses the change in the door magnetic sensor status as the priority trigger signal. When the door magnetic sensor detection unit detects that the door has switched from a closed state to an open state, the edge control unit records the moment the door magnetic sensor opens. If any auxiliary triggering condition is further detected within a preset triggering period after that moment, the operation process corresponding to this door magnetic sensor opening is defined as a candidate operation event. The auxiliary triggering conditions include: a change in the weighing output of the target slot or adjacent slots relative to the stable quality benchmark in step one exceeding a preset threshold; the appearance of a hand, turnover box, or material outline entering the shelf slot area in the visual image; or the RFID reader / writer unit detecting an increase or decrease in the tag set within the corresponding layer. This design is intended to avoid initiating the complete identification process solely due to a short-term door opening, inspection, or environmental disturbance.

[0027] After a candidate operation event is established, the edge control unit, centered on the initial trigger time, extracts data for a preset pre-trigger duration and continuously receives data for a preset post-trigger duration, forming the event window corresponding to the current operation. The pre-trigger duration is used to retain static reference data before triggering, in order to extract stable quality values, static label sets, and visual idle states before the operation; the post-trigger duration is used to cover the stable recovery phase after operation execution, turnover box return, and door closure, in order to obtain stable quality values ​​and the final state after the operation. If the door is still open before the preset post-trigger duration ends, or the visual branch continuously detects that the hand or turnover box is still near the target slot, or the weighing output has not yet recovered to the stable range, the edge control unit extends the event window until the door is closed and the multi-source observation results re-enter a stable state, or until the preset maximum window duration is reached.

[0028] After the event window is determined, the edge control unit organizes the multi-source data within the window. For the image sequence output by the vision acquisition unit, the edge control unit detects the hand, turnover box, and material outline in each frame of the image based on the slot image mapping reference formed in step one, and maps their positions in the image to the actual slot area, thereby forming a motion trajectory sequence. In this sequence, the edge control unit further extracts the motion center moment and motion center position, whereby the motion center moment is used to characterize the representative moment when the hand or turnover box enters the target slot area and generates substantial interaction, and the motion center position is used to characterize the slot area where the interaction mainly falls. For the data output by the weighing acquisition unit, the edge control unit uses the static interval before the start of the event window as the pre-operation reference and the re-stabilized interval before the end of the event window as the post-operation reference, and calculates the stable mass value of the target slot before operation and the stable mass value after operation, respectively. The difference between the two is recorded as the stable mass difference corresponding to the event window. For the data output by the RFID read / write unit, the edge control unit compares the appearance, disappearance, and continuous presence status of the tag set within the event window based on the layer domain read / write benchmark formed in step one. It extracts the tag state change result corresponding to the current event and constrains it within the layer domain range corresponding to the slot where the visual motion trajectory is located, thereby forming the candidate material identification result. For the data output by the door magnetic detection unit, the edge control unit extracts the door opening and closing times to define the effective time boundary of this operation.

[0029] After the multi-source evidence extraction is completed, the edge control unit calculates the operation event confirmation value of the current event window to determine whether the operation can be directly used as the basis for inventory change. The operation event confirmation value is used to characterize the degree of support of multi-source evidence for the same inventory change conclusion within the current event window, and is formed by weighting the temporal consistency, slot consistency, and quality interpretability. In specific implementation, the edge control unit extracts the key event moments corresponding to the visual branch, weighing branch, and RFID branch respectively. The key event moment of the visual branch is the moment of the action center, the key event moment of the weighing branch is the moment when the stable quality change reaches the judgment condition, and the key event moment of the RFID branch is the moment when the tag status change first occurs. Then, the time interval between the visual branch and the weighing branch, the time interval between the visual branch and the RFID branch, and the time interval between the weighing branch and the RFID branch are calculated respectively, and the time intervals are compared with the preset temporal tolerance. When none of the above time intervals are greater than the preset timing tolerance, the timing consistency is determined to be at a high level; when only one time interval exceeds the preset timing tolerance and the excess is not greater than the preset relaxation range, the timing consistency is reduced according to the preset reduction rule; when two or more time intervals exceed the preset timing tolerance, or any time interval exceeds the preset relaxation range, the timing consistency is determined to be at a low level.

[0030] After determining the timing consistency, the edge control unit further determines the slot consistency. Specifically, the edge control unit extracts the target slot pointed to by the visual branch, the target slot corresponding to the weight change, and the target slot corresponding to the RFID branch, and compares the consistency among the three. When all three correspond to the same slot, the slot consistency is determined to be at a high level. When two of them correspond to the same slot, and the other corresponds to an adjacent slot within the same layer, the slot consistency is reduced according to a preset reduction rule. When the correspondence among the three is inconsistent, or the result of the RFID branch exceeds the layer range of the visual motion trajectory, the slot consistency is determined to be at a low level. To avoid occasional crosstalk between adjacent slots affecting the final determination, the edge control unit preferably only allows the adjacent slot results of the RFID branch to participate in the fusion according to the reduction method when at least one of the visual branch and the weighing branch is consistent with the final target slot. When neither the visual branch nor the weighing branch can support the RFID branch, the result of the RFID branch is not used as a high-confidence slot basis.

[0031] After determining the slot consistency, the edge control unit determines the quality explanatory power based on the correspondence between the stable quality difference and the unit standard quality of the candidate materials. Specifically, the edge control unit uses the stable quality difference corresponding to the event window as a basis, combines it with the unit standard quality of the candidate materials to form the number of candidate items, and obtains the corresponding theoretical quality difference based on the number of candidate items and the unit standard quality of the candidate materials; then, it compares the degree of deviation between the stable quality difference and the theoretical quality difference. When the deviation between the stable quality difference and the theoretical quality difference is not greater than the preset quality tolerance, the quality explanatory power is determined to be at a high level; when the deviation is greater than the preset quality tolerance but not greater than the preset relaxation range, and the direction of the stable quality difference is consistent with the direction of inventory change corresponding to the number of candidate items, the quality explanatory power is reduced according to the preset reduction rule; when the deviation exceeds the preset relaxation range, or the direction of the stable quality difference is inconsistent with the direction of inventory change corresponding to the number of candidate items, the quality explanatory power is determined to be at a low level. In the case of multiple candidate materials, the edge control unit preferably selects the candidate material with the smallest deviation from the stable quality difference and whose corresponding layer tag change result can be supported by the RFID branch as the candidate material for the current event.

[0032] To ensure the comparability of judgment results between different events, the edge control unit preferably processes the timing consistency, slot consistency, and quality interpretability into the same value range before performing a weighted calculation to obtain the operation event confirmation value for the current event window. The preset confirmation threshold can be determined based on calibration samples, historical review samples, or confirmed operation events generated during the trial operation phase. Furthermore, the preset confirmation threshold is preferably determined based on a comprehensive balance of the proportion of events awaiting review, the false alarm rate, and the accuracy of quantity change judgment. This ensures that confirmed operation events maintain high credibility while preventing a large number of real events from being permanently retained as events awaiting review due to an excessively high threshold setting.

[0033] After calculating the operation event confirmation value, the edge control unit compares it with a preset confirmation threshold. When the operation event confirmation value reaches the preset confirmation threshold, the system outputs the event as a confirmed operation event and simultaneously records the target slot, candidate material, operation type, quantity change result, door magnet opening and closing time, action center time, stable quality difference, and corresponding layer label change result. The operation type is determined comprehensively based on the positive and negative direction of the stable quality difference, the visual trajectory direction, and the label status change direction. Specifically, when the stable quality value after the operation is less than the stable quality value before the operation and the visual trajectory shows movement from inside the slot to outside the slot, it can be determined as a material requisition event; when the stable quality value after the operation is greater than the stable quality value before the operation and the visual trajectory shows movement from outside the slot to inside the slot, it can be determined as a return event. If the operation event confirmation value does not reach the preset confirmation threshold, the system marks the event as a pending review event, only saving the corresponding original evidence and intermediate processing results, without directly modifying the inventory status. Step two outputs confirmed operation events and pending review events, with the confirmed operation events serving as input for subsequent long-term drift discrimination.

[0034] Step 3: After obtaining the confirmed operation events in Step 2, the edge control unit does not directly use all historical observation data for model correction. Instead, it only writes confirmed operation events whose confirmation values ​​meet the preset confirmation threshold into the self-calibration sample pool. The reason for this is that in the daily operations of the wind farm spare parts warehouse, visual obstruction, RFID cross-reading, short-term door opening, temporary shelving of turnover boxes, and short-term weighing disturbances may all lead to incomplete or unreliable observation results. If such events are used as correction criteria, sporadic errors are easily accumulated into the direction of model correction. On the contrary, confirmed operation events have been constrained by the time sequence consistency, slot consistency, and quality interpretability in Step 2, and can stably represent a real material requisition or return process. Therefore, they are suitable as a reliable sample source for subsequent long-term drift judgment. Based on this, the edge control unit establishes an index record for each confirmed operation event. The index record includes at least the event number, target slot, belonging layer, corresponding material, door magnetic opening time, door magnetic closing time, visual motion center position, visual motion center time, stable quality value before operation, stable quality value after operation, RFID tag status change result, environmental status, and finally confirmed operation type and quantity change result.

[0035] To identify persistent mismatches between the multi-sensor model and the actual operating state during long-term operation, the edge control unit summarizes and analyzes the index records in the self-calibration sample pool according to a preset statistical period or a preset cumulative number of events. Preferably, the slot is used as the basic analysis object; when layer-level read / write correction or material-level quality correction is required, the layer or material can also be used as the analysis object, respectively. Let the current analysis object be k, the edge control unit extracts the set of confirmed operation events corresponding to object k from the self-calibration sample pool. The set of confirmed operation events refers to the set of historical operation events whose operation event confirmation value reaches a preset confirmation threshold and whose target slot, its layer, or its corresponding material is related to object k. To avoid unstable statistical results due to an insufficient sample size, the edge control unit preferably initiates long-term drift calculation only when the number of events in the set of confirmed operation events is not less than a preset minimum sample size; when the number of samples is less than the preset minimum sample size, the current model parameters of object k remain unchanged, and only the samples continue to accumulate.

[0036] After obtaining the set of confirmed operation events for object k, the edge control unit extracts long-term deviation features from three directions: RF read / write results, visual action position results, and weighing interpretation results, and constructs a long-term drift discriminant value. This long-term drift discriminant value characterizes whether object k has experienced persistent mismatch within the current statistical period, and is formed by a weighted average of RF deviation, visual deviation, and weighing residual. In specific implementation, the edge control unit performs item-by-item statistical analysis on the set of confirmed operation events corresponding to object k according to the order of event occurrence, and generates RF deviation, visual deviation, and weighing residual values ​​respectively.

[0037] For radio frequency (RF) deviation, the edge control unit compares the correspondence between candidate slots mapped by RFID and the final confirmed slots in each confirmed operation event. When the candidate slots mapped by RFID and the final confirmed slots are consistent, the corresponding event is recorded as a consistent event. When the candidate slots mapped by RFID and the final confirmed slots are inconsistent but located in adjacent positions within the same layer domain, the corresponding event is recorded as a neighboring deviation event. When the candidate slots mapped by RFID and the final confirmed slots are significantly inconsistent, the corresponding event is recorded as a significant deviation event. The edge control unit determines the RF deviation based on the distribution of consistent events, neighboring deviation events, and significant deviation events in the set of confirmed operation events corresponding to object k, such that the higher the proportion of significant deviation events, the higher the RF deviation. When multiple confirmed operation events continuously exhibit slot deviation in the same direction, the impact of the corresponding deviation on the RF deviation further increases.

[0038] For visual deviation, the edge control unit compares the offset between the visual motion center position in each confirmed operation event and the final confirmed slot center position, and normalizes the offset by combining it with the slot feature length recorded in step one, thereby obtaining the visual deviation of object k in the current statistical period. When multiple confirmed operation events show that the visual motion center of the same slot continuously shifts in the same direction, it indicates that there is a stable deviation between the visual mapping relationship and the actual shelf space relationship, and the corresponding visual deviation increases; when the visual motion center only shows a discrete distribution or random fluctuation, the corresponding visual deviation remains at a low level.

[0039] For weighing residuals, the edge control unit compares the stable mass difference in each confirmed operational event with the theoretical mass difference corresponding to the number of pieces and the standard mass per unit of material based on the quantity change results. It then determines the weighing residual based on the direction and magnitude of this difference. When multiple confirmed operational events consistently show a deviation in the same direction, it indicates a stable mismatch between the weighing baseline or the standard mass per unit of material and the current actual state, resulting in an increase in the corresponding weighing residual. When the difference only occurs briefly in individual events, it is not directly used as the primary basis for determining long-term drift.

[0040] After obtaining the radio frequency deviation, visual deviation, and weighing residual, the edge control unit processes these three deviations into the same value range and then performs a weighted calculation to obtain the long-term drift discrimination value of object k. To improve discrimination stability, the edge control unit preferably requires that object k meets the preset drift threshold for the long-term drift discrimination value in multiple consecutive statistical periods, or requires that at least two types of deviations constituting the long-term drift discrimination value simultaneously exceed their respective sub-thresholds before it is recorded as a valid long-term drift result. If the sample reliability screening in step two is missing, RFID cross-reading, visual occlusion, weighing disturbance, and event fragmentation results will simultaneously enter the long-term statistical process, thus causing the radio frequency deviation, visual deviation, and weighing residual to lose a reliable reference basis.

[0041] Step 3 outputs the set of confirmed operation events, radio frequency deviation, visual deviation, weighing residual, long-term drift discrimination value, and drift discrimination conclusion corresponding to object k. Further, the preset minimum sample size, preset drift threshold, and sub-thresholds corresponding to each deviation can be determined based on the statistical results of the system calibration phase, trial operation phase, or historical review samples. When determining the above parameters, the edge control unit preferably comprehensively considers the false alarm rate, false negative rate, proportion of events awaiting review, and changes in recognition accuracy after updates, so that long-term drift discrimination can reflect continuous mismatch while avoiding frequent model updates triggered by a small number of occasional abnormal events. For different layers, different slots, or different material categories, the edge control unit can also set corresponding local correction parameters under a unified discrimination framework to adapt to the differences in read / write characteristics, visual occlusion degree, and weighing fluctuation characteristics of different objects.

[0042] Step 4: After obtaining the long-term drift discrimination value of object k in Step 3, as follows... Figure 3 As shown, the edge control unit first determines whether the object meets the linkage update conditions. Preferably, when the long-term drift discrimination value of object k is greater than or equal to a preset drift threshold, and this state remains true for multiple consecutive statistical periods, or when at least two of the radio frequency deviation, visual deviation, and weighing residual that constitute the long-term drift discrimination value simultaneously exceed their respective sub-thresholds, the edge control unit marks the object as an object to be corrected and generates a corresponding linkage update task. For objects that do not meet the linkage update conditions, the system keeps the current parameters unchanged and only continues to accumulate subsequent confirmed operation events.

[0043] After the linked update task is generated, the edge control unit uses the RF deviation, visual deviation, and weighing residual output in step three as correction criteria to perform a constrained, incremental update of the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass. This constrained, incremental update means that all parameters are gradually corrected based on the current operating parameters, taking into account the stable deviations reflected by recently confirmed operational events, rather than directly replacing the original parameters based on a single abnormal event.

[0044] To correct the layer-domain read / write benchmark, the edge control unit extracts confirmed operation events related to the target layer-domain, calculates the correspondence between candidate slots given by the RFID branch in each event and the final confirmed slot, and updates the weight distribution of tag read / write results assigned to specific slots within the layer-domain based on the correspondence. For correspondences that appear repeatedly in multiple confirmed operation events and remain consistent with the final confirmed slot, their weight in the layer-domain read / write benchmark is increased; for correspondences that have not been supported by confirmed operation events for a long time, their weight in the layer-domain read / write benchmark is decreased. After the update, the edge control unit performs unified constraint processing on the assignment weights corresponding to each slot within the same layer-domain to maintain the stability of the read / write assignment relationship within the layer-domain.

[0045] For the correction of the slot image mapping baseline, the edge control unit extracts the visual action center position and the final confirmed slot center position from the confirmed operation events related to object k, and statistically analyzes the long-term offset direction and offset amount of the visual action center relative to the confirmed slot center. When multiple confirmed operation events show a continuous offset in the same direction for the same slot, the image mapping boundary or center position of that slot is smoothly corrected. During the correction process, the edge control unit maintains the boundary relationship between adjacent slots and ensures that the corrected slot center position remains within the effective range of the corresponding slot, so as to avoid mapping overlap between adjacent slots due to over-correction.

[0046] For the correction of weighing baseline and material unit standard mass, the edge control unit handles weighing zero-point drift and mass interpretation mismatch respectively. When the same storage slot consistently shows an empty value deviating from the original weighing baseline in empty slot confirmation events or stable inventory events, the edge control unit uses the weighing value collected in the most recent empty slot confirmation event as the correction basis to update the empty slot baseline value. When the same material consistently shows a fixed directional deviation between stable mass difference and theoretical mass difference in multiple confirmed material requisition and return events, the edge control unit extracts mass difference samples from the corresponding events and prioritizes events with clear quantity changes and high mass interpretation to progressively adjust the unit standard mass of the material. Discrete outliers that only appear in individual events are not directly used as the basis for correcting the weighing baseline or material unit standard mass.

[0047] After completing the above three branch model updates, the edge control unit further adjusts the weight coefficients in the operation event confirmation value and the long-term drift discrimination value to make the fusion determination result adapt to the actual operating state of the current system. Specifically, the edge control unit determines the recent stability of the radio frequency branch, the vision branch, and the weighing branch respectively according to the consistency between each sensor branch and the final confirmation result in the recent period, and adjusts the corresponding weight coefficients based on the recent stability. For branches with higher recent stability and smaller fluctuations, increase their weights in the fusion determination; for branches with a decline in recent stability or obvious deviation, reduce their weights in the fusion determination; for branches in the correction process, perform restricted adjustment while retaining their participation in the fusion. After adjustment, the weight coefficients are preferably uniformly constrained to maintain the continuity and stability of the fusion determination result.

[0048] The above linkage update does not immediately replace the original operation configuration in full amount after parameter generation, but first enters the online verification stage. Specifically, in the subsequent preset verification period, the edge control unit performs dual-path parallel determination on the newly generated material requisition and return operation events using the pre-update parameters and the post-update parameters respectively. Among them, the pre-update parameters are used to form the original determination result, and the post-update parameters are used to form the corrected determination result. For the same event, the system calculates the target slot, candidate materials, quantity change results, and operation event confirmation values obtained under the two sets of parameters respectively, and compares the two sets of determination results with the final inventory confirmation result of the event. The final inventory confirmation result can be derived from the delayed confirmation result of subsequent confirmed operation events, the manual review result, the subsequent static inventory result, or the actual inventory change result registered in the system.

[0049] To evaluate the correction effect, the edge control unit statistically analyzes the changes in indicators such as the target slot recognition accuracy, quantity change determination accuracy, proportion of events to be reviewed, and false alarm rate before and after the update. The preset improvement requirement means that the post-update parameters are continuously better than the pre-update parameters in at least a preset number of items among the indicators, and the remaining key indicators do not show a decrease or increase exceeding the preset deterioration threshold; among them, for the target slot recognition accuracy and quantity change determination accuracy, an increase in the value is used as the determination basis for being better than the pre-update parameters; for the proportion of events to be reviewed and the false alarm rate, a decrease in the value is used as the determination basis for being better than the pre-update parameters. The continuous superiority preferably means that it is satisfied in consecutive multiple verification results formed by statistical batches within the preset verification period; the obvious deterioration preferably means that any key indicator exceeds the preset deterioration threshold relative to the pre-update parameters and appears continuously.

[0050] The system will only write the updated parameters into the running configuration and replace the original parameters as the new current benchmark if the updated parameters meet the above-mentioned preset improvement requirements. If the updated parameters do not meet the preset improvement requirements, or although some indicators are improved, the proportion of events to be reviewed increases, the false alarm rate increases, or the accuracy of quantity change judgment decreases beyond the preset deterioration threshold, the system will revert to the parameters before the update and record this correction process as an invalid update sample.

[0051] After this step, the system outputs the updated layer read / write baseline, the updated slot image mapping baseline, the updated weighing baseline, the updated material unit standard mass, the updated fusion weight coefficient, the online verification results, and a decision on whether to write the updated parameters into the formal operation configuration. Thus, the system forms a long-term closed-loop operation mechanism encompassing operation triggering, operation event confirmation value calculation, confirmed operation event pooling, long-term drift discrimination value calculation, linked updates, and online verification. This ensures that the smart shelving system for wind power smart warehousing can continuously maintain the accuracy and stability of inventory perception results during long-term operation.

[0052] Example 2: The design of the wind power smart warehouse intelligent shelf precision sensing system based on multi-sensor fusion is based on the method in Example 1, specifically as follows... Figure 4 As shown, it includes a baseline modeling module, an event confirmation module, a drift detection module, and an update verification module.

[0053] The baseline modeling module, after system deployment, is used by the edge control unit to read shelf layout data, slot code data, material master data, turnover box binding data, electronic tag binding data, material unit standard quality data, allowed storage relationships, and environmental threshold parameters. It establishes unique slot identifiers according to unified coding rules for shelf groups, shelf sides, shelf layers, and slots, and completes the association between slots and material categories, allowed storage ranges, turnover box identifiers, electronic tag identifiers, and weighing acquisition locations, forming a slot master data table and an inventory status table. Simultaneously, it completes the access and time synchronization processing of RFID reading / writing units, weighing acquisition units, visual acquisition units, door magnetic detection units, and temperature and humidity detection units, and collects initial baseline data under static shelf conditions to form a multi-sensor initial baseline model for the shelf. This multi-sensor initial baseline model includes at least layer-domain reading / writing baselines, slot image mapping baselines, weighing baselines, environmental reference baselines, slot geometric information, and material unit standard quality information.

[0054] The event confirmation module receives real-time outputs from the door magnetic detection unit, weighing acquisition unit, vision acquisition unit, and RFID reading and writing unit under continuous monitoring. It prioritizes door magnetic status changes as the triggering criterion, combining weighing changes, changes in the outline of hands, turnover boxes, or materials in the visual image, and changes in the RFID tag set to determine candidate operation events. After a candidate operation event is established, an event window is constructed. The visual data, weighing data, RFID data, and door magnetic data within the window are processed, extracting the action center time, action center position, stability quality difference, tag status change results, and door opening and closing times. The operation event confirmation value is calculated based on timing consistency, slot consistency, and quality interpretability. When the operation event confirmation value reaches a preset confirmation threshold, a confirmed operation event is output; when the operation event confirmation value does not reach the preset confirmation threshold, a pending review event is output.

[0055] The drift discrimination module is used to write confirmed operation events into the self-calibration sample pool, and to summarize and analyze the historical events in the self-calibration sample pool according to a preset statistical period or a preset cumulative number of events. For the analysis object corresponding to the slot, layer or material, it extracts the set of confirmed operation events related to it, and forms radio frequency deviation, visual deviation and weighing residual respectively, and calculates the long-term drift discrimination value accordingly to obtain the corresponding drift discrimination conclusion, so as to determine whether the current object has experienced continuous mismatch.

[0056] The update verification module is used to mark the corresponding object as an object to be corrected and generate a linkage update task when the long-term drift discrimination value reaches the preset drift threshold and meets the conditions of multiple consecutive statistical periods or at least two types of deviations exceeding the corresponding sub-thresholds simultaneously. Subsequently, based on the radio frequency deviation, visual deviation, and weighing residual, the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass are updated in a constrained and progressive manner, and the weight coefficients in the operation event confirmation value and the long-term drift discrimination value are adjusted. After the update is completed, the online verification stage is entered, and the parameters before and after the update are judged respectively. The changes in the target slot identification accuracy, quantity change judgment accuracy, proportion of events to be reviewed, and false alarm rate are compared. When the updated parameters meet the preset improvement requirements, they are written into the formal operation configuration.

[0057] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0058] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0059] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0060] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0061] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for precise perception of intelligent shelves in wind power intelligent warehousing based on multi-sensor fusion, characterized in that, Includes the following steps: Establish slot identification and form an initial baseline model for the multi-sensor shelving; The system prioritizes door magnet status changes as the triggering basis and determines candidate operation events by combining weighing changes, visual changes, and RFID tag set changes within a preset triggering period. It constructs an event window and organizes multi-source data within the window, extracting the action center time, action center location, stability quality difference, and tag status change results. Based on the timing consistency, slot consistency, and quality interpretability, it determines the operation event confirmation value. When the operation event confirmation value reaches the preset confirmation threshold, it outputs the confirmed operation event. Only confirmed operational events are written into the self-calibration sample pool. Based on the confirmed operational events corresponding to the analysis object, radio frequency deviation, visual deviation and weighing residual are generated, and then the long-term drift discrimination value is determined. When the long-term drift discrimination value is greater than or equal to the preset drift threshold and holds for multiple consecutive statistical periods, or when at least two deviations constituting the long-term drift discrimination value exceed their respective sub-thresholds, the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass are updated in a constrained and progressive manner. When the parallel determination of the parameters before and after the update shows that the updated parameters meet the preset improvement requirements, the updated parameters are written into the formal operation configuration.

2. The multi-sensor fusion-based intelligent perception method for wind power intelligent warehouse intelligent shelves according to claim 1, characterized in that: The initial reference model of the multi-sensor rack includes at least the layer domain reading and writing reference, slot image mapping reference, weighing baseline, environmental reference baseline, slot geometric information, and material unit standard mass information. The establishment of slot identification includes: establishing a slot identification for each actual storage location according to the unified coding rules of shelf group, shelf side, shelf layer and slot, and associating each slot with its corresponding material category, allowable storage range, turnover box identification, electronic tag identification and weighing collection location.

3. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion as described in claim 1, characterized in that: The process of determining candidate operation events includes: when the door magnetic detection unit detects that the door has switched from a closed state to an open state, if the weighing output of the target slot or adjacent slot changes beyond a preset threshold relative to the stable quality benchmark within the preset trigger time period, or a hand, turnover box, or material outline appears in the visual image entering the shelf slot area, or the RFID reading and writing unit detects an increase or decrease in the tag set in the corresponding layer, then the operation process corresponding to the opening of the door magnetic sensor is defined as a candidate operation event; the process of constructing an event window and organizing the multi-source data within the window includes: taking the first trigger time as the center and extracting data for a preset pre-trigger time, and continuously receiving data for a preset post-trigger time; when the door is still in the open state before the preset post-trigger time ends, or the visual branch continuously detects that the hand or turnover box is still near the target slot, or the weighing output has not yet recovered to the stable range, the event window is extended until the door is closed and the multi-source observation results re-enter a stable state, or the preset longest window duration is reached.

4. The precise sensing method for intelligent wind power storage racks based on multi-sensor fusion as described in claim 3, characterized in that: The operation event confirmation value is formed by timing consistency, slot consistency, and quality interpretability. The timing consistency is determined based on the time interval between the action center moment of the visual branch, the moment when the stable quality change of the weighing branch reaches the judgment condition, and the moment when the tag status change of the RFID branch first occurs, combined with a preset timing tolerance. The slot consistency is determined based on the consistency between the target slot pointed to by the visual branch, the target slot corresponding to the weight change, and the target slot corresponding to the RFID branch. The quality interpretability is determined based on the correspondence between the stable quality difference and the unit standard quality of the candidate material. When the operation event confirmation value reaches a preset confirmation threshold, the target slot, candidate material, operation type, quantity change result, door magnetic opening and closing moment, action center moment, stable quality difference, and corresponding layer tag change result are recorded.

5. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to claim 4, characterized in that: Only confirmed operation events whose confirmation values ​​reach a preset confirmation threshold are written into the self-calibration sample pool, and an index record is created for each confirmed operation event. The index record includes at least the event number, target slot, layer, corresponding material, door magnetic opening time, door magnetic closing time, visual motion center position, visual motion center time, stable quality value before operation, stable quality value after operation, RFID tag status change result, environmental status, and the final confirmed operation type and quantity change result. When the number of confirmed operation events corresponding to the analysis object is not less than the preset minimum sample size, the calculation of the long-term drift discrimination value is initiated.

6. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to claim 5, characterized in that: The analysis object is a slot, layer, or material; the long-term drift discrimination value is formed by radio frequency deviation, visual deviation, and weighing residual; wherein, the radio frequency deviation is determined based on the correspondence between the candidate slot obtained by RFID mapping and the final confirmed slot, the visual deviation is determined based on the offset between the visual motion center position and the center position of the final confirmed slot and combined with the characteristic length of the slot, and the weighing residual is determined based on the difference between the stable mass difference and the theoretical mass difference corresponding to the number of pieces and the standard mass of the material unit based on the quantity change result.

7. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to claim 6, characterized in that: When the candidate slot obtained by RFID mapping is consistent with the final confirmed slot, the corresponding event is recorded as a consistency event; when the candidate slot obtained by RFID mapping is inconsistent with the final confirmed slot but is located in adjacent positions within the same layer domain, the corresponding event is recorded as a proximity deviation event. When the candidate slot obtained by RFID mapping is significantly inconsistent with the final confirmed slot, the corresponding event is recorded as a significant deviation event, and the radio frequency deviation degree is determined according to the distribution of consistent events, adjacent deviation events and significant deviation events in the confirmed operation events; when multiple confirmed operation events show that the visual motion center of the same slot continues to shift in the same direction, the visual deviation degree is increased; when multiple confirmed operation events continuously show that there is a difference deviation in the same direction between the stable quality difference and the theoretical quality difference, the weighing residual degree is increased.

8. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to claim 6, characterized in that: The constrained progressive update includes: updating the weight distribution of tag reading and writing results to specific slot assignments within the layer domain based on the correspondence between candidate slots and final confirmed slots given by the RFID branch in confirmed operation events related to the target layer domain; smoothing the slot image mapping boundary or center position based on the long-term offset direction and offset amount between the visual action center position and the final confirmed slot center position; updating the weighing baseline based on the weighing value in the empty slot confirmation event, and progressively adjusting the standard mass of the material unit based on the quality difference samples in multiple confirmed material requisition and return events; and adjusting the weight coefficients in the operation event confirmation value and long-term drift discrimination value based on the consistency between each sensor branch and the final confirmation result within a preset time period.

9. The precise sensing method for intelligent shelving in wind power smart warehousing based on multi-sensor fusion according to claim 8, characterized in that: The parallel determination of the parameters before and after the update includes: performing dual-path parallel determination on newly generated material requisition and return operation events using the parameters before and after the update within a preset verification period, and comparing the changes of the two sets of parameters in target slot identification accuracy, quantity change determination accuracy, proportion of events awaiting review, and false alarm rate; only when the updated parameters are consistently better than the parameters before the update in at least a preset number of items among the target slot identification accuracy, quantity change determination accuracy, proportion of events awaiting review, and false alarm rate, and do not cause significant deterioration in other indicators, are the updated parameters written into the formal operation configuration; otherwise, the settings are reverted to the parameters before the update.

10. A precise sensing system for intelligent wind power warehouse shelving based on multi-sensor fusion, characterized in that, The sensing system is used to implement the method according to any one of claims 1-9, and includes the following modules: The baseline modeling module is used to establish slot identification and form an initial baseline model of the multi-sensor rack. The event confirmation module is used to identify candidate operation events, construct an event window and organize multi-source data within the window, extract the action center time, action center location, stability quality difference and label status change results, determine the operation event confirmation value and output the confirmed operation event; The drift discrimination module is used to write confirmed operation events into the self-calibration sample pool, and to generate radio frequency deviation, visual deviation and weighing residual based on the confirmed operation events corresponding to the analysis object, thereby determining the long-term drift discrimination value; The update verification module is used to perform constrained incremental updates to the layer domain read / write benchmark, slot image mapping benchmark, weighing baseline, and material unit standard mass when the long-term drift discrimination value meets the update conditions. When the updated parameters meet the preset improvement requirements, the updated parameters are written into the formal operation configuration.

Images