A multi-specification material storage and automatic dispensing control system

Abstract

The application discloses a multi-specification material storage and automatic distribution control system, and relates to the technical field of discrete article distribution control. The system comprises the following steps: reading the category and the number of requests at the distribution time, calling the corresponding specification image, performing specification conditioning processing on the weight, vision and force perception quantity, forming multi-source residual and source quality marks; then fusing the multi-source residual to generate three-state reliable records of over-distribution, correct distribution and missed distribution, and performing baseline write-back according to the protection pass marks; subsequently, the three-state reliable records are used to control the discharge gate, the review tray, the separation mechanism and the supplement execution unit to switch between the direct distribution path, the over-distribution review path and the missed distribution review path; after the review, the remaining intermediate state is output, and when the boundary is crossed, the model parameter, the path threshold and the new specification isolation state are updated and locked according to the cross-source consistency and the drift degree. The application can improve the verification adaptability and path stability of continuous distribution of multi-specification materials.

Description

Technical Field

[0001] This invention relates to the field of discrete item distribution control technology, specifically a multi-specification material storage and automatic distribution control system. Background Technology

[0002] Automated warehousing and discrete item distribution control technologies are frequently used in applications such as factory spare parts cabinets, tool requisition cabinets, medical consumables dispensing cabinets, and unmanned retail kiosks. In actual engineering, the same cabinet or the same retrieval mechanism needs to handle materials with large differences in specifications, such as gaskets, clips, connectors, blades, and small packages of consumables, and complete the material retrieval under the guidance of personnel scanning, work order requisition, or production line replenishment instructions. The determination of whether to take the goods is based on the detection of changes in the weight of the goods, camera recognition, photoelectric detection at the discharge port, or manual verification. Existing equipment mostly consists of fixed channels, independent storage compartments, weight change detection, camera recognition, photoelectric detection at the discharge port, and manual verification. The main focus is on cabinet door control, product recognition, inventory statistics, and settlement management, which can meet the basic needs of single channel or stable product category scenarios.

[0003] Chinese patent document CN110443950A discloses a smart vending machine. This vending machine includes a vending machine cabinet for holding beverages, a control terminal installed inside the cabinet, a QR code on the surface of the cabinet, a visual recognition module installed inside the vending machine, and a smart lock. First, after a user accesses the control terminal corresponding to the QR code via their mobile phone, the control terminal controls the smart lock to open the cabinet door, allowing the user to directly select items from inside. Second, the visual recognition module detects whether there are sufficient items in the vending machine and sends the result to the control terminal. The control terminal analyzes the data collected from the visual recognition module and controls the smart lock and the visual recognition module based on the analysis results. The document also discloses a gravity recognition module, which includes a gravity detection sensor connected to the algorithm host. The gravity detection sensor is designed to be installed at the bottom of each shelf inside the vending machine cabinet, used in conjunction with the visual recognition module to count the number of items inside the vending machine cabinet.

[0004] Therefore, the technical solution presented in this paper utilizes the cabinet, QR code, visual recognition camera, gravity detection sensor, algorithm host and smart lock to form the recognition and cabinet door control link in the unmanned vending scenario, focusing on the changes of goods inside the cabinet and the user's selection when opening the cabinet.

[0005] The aforementioned existing technologies are applicable to scenarios where users open the cabinet to select items or to count the number of items in the cabinet, but they still have certain limitations when using a shared single storage and distribution mechanism to automatically distribute multiple specifications of materials one by one. For example, the visual recognition module and the gravity recognition module work around the quantity of goods in the cabinet or the weight of the shelf, and the detection object is biased towards the overall state of the compartment or shelf. However, when entering the automatic distribution scenario, it is necessary to determine whether the output of a single material in a certain picking and placing action conforms to the instruction. At the same time, considering that there are differences in the weight, shape, contact area, surface friction and stacking posture between different specifications of materials, small-sized thin sheets are prone to multiple pieces sticking together due to static electricity, oil film or edge adhesion, and large-sized connectors or tool heads are prone to problems such as incomplete material dropping due to support point offset, unstable clamping or release posture deviation. At this time, if only the overall weight change or image quantity statistics are used, it is easy to mix sensor fluctuations, posture occlusion and real mis-dispensing, and it is impossible to determine whether the abnormality comes from more or less picking or sensor status change. In addition, due to continuous use and peak picking, the zero point drift of weighing, changes in visual lighting, mechanical vibration and verification queue will overlap. If the verification boundary relies on a fixed threshold for a long time, it is easy to cause repeated verification, increased manual intervention and unstable output cycle.

[0006] Therefore, the key technical problem to be solved in this field is: how to improve the adaptability of single distribution verification to specification differences, mis-distribution direction and operational disturbances when a shared storage and distribution mechanism continuously distributes materials of multiple specifications. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a multi-specification material storage and automatic distribution control system. Based on protection, a calibration baseline is written back using markers. Subsequently, based on a three-state reliable record, the system controls the discharge gate, verification tray, re-separation mechanism, and replenishment execution unit to switch between direct distribution, over-received verification, and missed distribution verification paths. After verification, a residual intermediate state is output, and when boundaries are exceeded, the system latches and updates model parameters, path thresholds, and new specification isolation states based on cross-source consistency and drift. This invention improves the verification adaptability and path stability of continuous distribution of multi-specification materials and solves the technical problems described in the background art.

[0009] (II) Technical Solution

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A multi-specification material storage and automatic distribution control system, executed by a storage-distribution controller of a shared storage-distribution mechanism, includes: at the time of a distribution event, reading the category and requested quantity, calling the specification profile of the category, performing specification conditional processing on at least two of the weight, visual and force perception quantities, and forming multi-source residual, source quality mark and protection pass mark;

[0012] Based on the source quality markers, multi-source residuals are fused to generate three-state reliable records that distinguish between over-taken, correct, and missed records, and the baseline writing status is determined by the protection through the markers to determine the specification profile.

[0013] Based on the three-state trusted record and protection through the tag, the discharge gate, the verification tray, the re-separation mechanism and the supplementary execution unit are controlled to switch between the direct distribution path, the over-the-count verification path and the missed distribution verification path.

[0014] After the review is completed, a legacy intermediate state containing drift trend, non-convergence marker, and congestion metric is generated. When the legacy intermediate state goes out of bounds, cross-source consistency and drift are calculated based on sliding window residual statistics. When the cross-source consistency is not lower than the meta threshold and the drift is not higher than the drift threshold, the model parameters and path threshold of the three-state trusted record are latched and updated. When the above conditions are not met, the current version is maintained and the candidate update is discarded.

[0015] Furthermore, the formation of source quality markers includes: acquiring the weighing stable state, visual occlusion state, and force sampling continuous state at the time of the distribution event; writing the stability gap, out-of-bounds state, and timing deviation of each source into the source quality marker; continuing to generate a three-state reliable record when at least two types of sources are in a usable state, and setting the protection through the marker to a conservative state when fewer than two types of sources are in a usable state.

[0016] Furthermore, generating a three-state credible record includes: selecting available multi-source residuals according to source quality labels, and fusing the available multi-source residuals into a directional evidence quantity;

[0017] When the directional evidence quantity is in the positive interval, an over-dominating state is generated; when the directional evidence quantity is in the negative interval, a missed-issuance dominant state is generated; when the directional evidence quantity is in the neutral interval and the source quality mark is not in a conservative state, a correct dominant state is generated; when there is a directional conflict between available multi-source residuals, the confidence of continuous distribution is reduced and written into the conservative review entry.

[0018] Furthermore, the model parameters for the three-state trusted record include source weights, over-scanning discrimination bands, and under-scanning discrimination bands; the storage-distribution controller generates candidate updates based on the verification test deviations after the verification is completed;

[0019] When the cross-source consistency of the candidate update is not lower than the meta threshold and the drift is not higher than the drift threshold, the candidate update is latched; when the above conditions are not met at the same time, the previous stable parameter version is maintained and the candidate update is discarded.

[0020] Furthermore, when one of the weighing, visual, or force perception quantities is missing within the time window corresponding to the time of the assigned event, the source of the missing quantity is written into the source quality tag, and a multi-source residual is formed with the other at least two types of perception quantities.

[0021] When only one type of sensing quantity arrives within the time window, the direct distribution path is prohibited, and the corresponding material is imported into the verification tray and a conservative verification record is generated.

[0022] Furthermore, after the storage-distribution controller issues path commands to the discharge gate, verification tray, re-separation mechanism and replenishment execution unit, it listens for execution receipts within the command receipt time window;

[0023] When the execution receipt arrives, continue executing the path command; when the execution receipt has not arrived, keep the discharge gate closed and resend the path command; when the number of resends reaches the resend limit, keep the corresponding material in the verification pallet and generate a manual transfer status.

[0024] Furthermore, when multiple distribution events simultaneously request the same executable in the review tray, re-separation mechanism, and replenishment execution unit, the storage-distribution controller generates a queue order based on the event timestamp, continuous distribution reliability, and recycle count.

[0025] The first-team-assigned event locks its path command while it is in the same execution state; non-first-team-assigned events remain in a pending review state and are prohibited from opening the discharge gate.

[0026] Furthermore, the multi-source sensing data is obtained from at least two of the following: the weighing support plate, the visual acquisition unit at the discharge port, and the force sensing sampling unit at the pick-up and drop-off end; the weighing support plate is located in the bearing area in front of the discharge gate.

[0027] The visual acquisition unit at the discharge port covers the bearing area; the force sensing sampling unit at the picking and placing end generates a force sensing quantity using at least one of the wrist force sensor and the motor current sampling circuit.

[0028] Furthermore, under low computing power conditions at the edge, the storage-distribution controller generates a three-state trusted record based on the direction interval of the multi-source residual and the source quality mark calling rule discrimination table;

[0029] When the edge computing power is in normal condition, a three-state posterior estimate is generated based on the multi-source residual, source quality label and the discrimination band corresponding to the category, and the continuous distribution confidence is formed by the three-state posterior estimate.

[0030] Furthermore, the storage-distribution controller includes a local control unit and a remote computing unit;

[0031] When the local control unit is connected to the network, the remote computing unit generates candidate model parameters based on the legacy intermediate state, and the local control unit latches the candidate model parameters based on cross-source consistency and drift.

[0032] When the local control unit is offline, it uses local sliding window residual statistics to maintain the current version and caches the legacy intermediate state.

[0033] Furthermore, the storage-distribution controller generates a distribution control message, which includes a command number, storage location number, path field, forced mode field, event timestamp, and control verification field. The path field records one of the following: direct distribution path, over-the-top verification path, missed distribution verification path, and conservative verification path. The storage-distribution controller drives the corresponding discharge gate, verification pallet, re-separation mechanism, and replenishment execution unit based on the path field.

[0034] Furthermore, the storage-distribution controller generates a model update message after the legacy intermediate state goes out of bounds. The model update message includes update number, scope of application, target object, parameter increment, gating pass flag, drift degree, rollback version, parameter version, update timestamp, and update verification field.

[0035] When the gate passes and is marked as passed, the parameter increment is written; when the gate passes and is marked as failed, the previous stable parameter version is maintained according to the rollback version.

[0036] Furthermore, the storage-distribution controller drives the retrieval interface to display the current distribution events according to the status sequence of pending distribution, verification in progress, direct distribution, review, re-separation, replenishment, success, and transfer to manual processing;

[0037] The retrieval interface displays the storage location number, path field, continuous allocation confidence range, and parameter version for each status. The storage-retrieval-allocation controller synchronously outputs local logs, which record status sequences, allocation control messages, and model update messages.

[0038] (III) Beneficial Effects

[0039] This invention provides a multi-specification material storage, retrieval, and automatic distribution control system, which has the following beneficial effects:

[0040] By introducing specification profiles at the moment of distribution, and processing weighing, visual, and force perception quantities into specification conditions, weight deviations, quantity deviations, and handling resistance deviations of materials with different specifications can be explained under the same category. This avoids misjudgments caused by the use of a unified verification strategy in shared storage and distribution institutions, such as over-dispensing of small items due to adhesion or under-dispensing of large items due to slippage. Through multi-source residual fusion, a three-state reliable record is formed that distinguishes between over-dispensing, correct dispensing, and under-dispensing. This solution no longer merely judges whether the material is abnormal, but further identifies the direction of the mis-dispensing, which is beneficial for subsequent distribution actions to enter the re-separation path for over-dispensing, the replenishment path for under-dispensing, and the direct distribution or light verification path for correct dispensing.

[0041] By setting source quality markers and protection pass markers, this solution reduces the impact of abnormal sources on the results when there is visual occlusion, unstable weighing, interrupted force sampling, or single source absence. It also imports distribution events with insufficient sources into conservative review, reducing the impact of a single sensor failure on the entire distribution process.

[0042] By calibrating and compensating for variations in symmetrical weight zero drift, visual interference, and mechanical resistance, and writing back the specification profile with a calibration baseline under protective conditions, the transmission of sensor state changes to the three-state discrimination results can be suppressed, ensuring the stability of the verification boundary during continuous distribution. By controlling the discharge gate, verification tray, re-separation mechanism, and replenishment execution unit through markers based on the three-state reliable record and protection, this solution can transform the algorithm's judgment results into an executable physical path, forming a closed-loop relationship between verification, review, re-separation, replenishment, and release, reducing repeated material handling and undirected retries.

[0043] By using legacy intermediate threshold drift, new specification non-convergence, and review congestion for subsequent gating latches, type change isolation, and queue scheduling, we prevent suspicious review results from directly polluting model parameters. At the same time, we ensure that the system does not change the allocation criteria and execution path in type change and high-concurrency application scenarios. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall structure of the multi-specification material storage and automatic distribution control system of the present invention;

[0045] Figure 2 This is a schematic diagram of the overall process of the multi-specification material storage, retrieval and automatic distribution control method of the present invention;

[0046] Figure 3 This is a schematic diagram of the process for generating event logs and specification-conditional multi-source residuals in this invention;

[0047] Figure 4 This is a schematic diagram of the three-state trusted record generation process in this invention;

[0048] Figure 5 This is a schematic diagram of the adaptive calibration and anti-interference protection boundary control process in this invention;

[0049] Figure 6 This is a schematic diagram of the differentiated allocation path based on three-state trusted records in this invention;

[0050] Figure 7 This is a schematic diagram of the high-order feedback closed loop driven by the legacy intermediate state in this invention. Detailed Implementation

[0051] 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.

[0052] Please see Figures 1-7 This invention provides a multi-specification material storage and automatic distribution control system, comprising:

[0053] Step 1: Before the material is released in the distribution action, align the specifications and quantity that should be retrieved in this request with what the field sensors actually observe in the same distribution event record, and output the specification-conditional multi-source residuals and data quality tags that can be directly called in Step 2.

[0054] In a shared single-set storage-distribution mechanism, small-sized materials are prone to sticking and being over-distributed due to surface adhesion, electrostatic adsorption, or end-clamping inertia, while large-sized materials are prone to being missed due to insufficient edge support, tilting, or slippage. Using only a fixed weight threshold or a fixed visual threshold at the discharge port essentially applies the same standard to all categories, causing over-distribution of light items to be buried within the normal fluctuations of larger items, and misjudging missed distribution of heavy items as normal small-item distributing. Therefore, step one first reads the category identifier and requested quantity from the distribution instruction, then retrieves the corresponding specification profile for that category, ensuring that each sensor reading is based on the desired physical characteristics of that category, rather than on a general equipment threshold. The specification profile should at least include the desired single-item weight. Weight dispersion Geometric outline range Baseline of pick-up and drop forces and allowable attitude band Category labeling Corresponding to the materials to be distributed this time, the mechanism's operating condition indicator. The corresponding combination of pick-up and drop-off end, storage location level, and discharge path.

[0055] Upon receiving a distribution instruction, the storage-distribution controller does not generate verification data during the inventory check upon receipt. Instead, it generates a distribution event record at the moment the material is picked up at the retrieval end, passing over the weighing support plate or about to enter the discharge port. This distribution event record includes the category identifier, requested quantity, event timestamp, measured distribution weight, visually recognized quantity, material posture description, force response or motor current response, source quality mask, and degradation flag.

[0056] Preferably, the weighing support plate is arranged on the transition bearing surface in front of the discharge port, and its effective bearing range covers the landing area of ​​the material with the largest outer contour; the vision acquisition unit is set above or to the side of the discharge port so that the field of view covers the bearing surface and the end release path; the force sampling unit is set at the wrist of the picking and placing end, or the picking and placing resistance is equivalently characterized by the current curve of the drive motor.

[0057] Tension / compression sensors, torque sensors, light curtain counters, or structured light cameras with equivalent functions can all be used as alternative sensing sources. However, any implementation should retain at least two types of sensing quantities from different physical sources to avoid losing the directional basis for subsequent three-state discrimination when a single source fails.

[0058] The storage-distribution controller first parses the category identifier in this distribution instruction. and number of requests Then, the corresponding expected single-piece weight is loaded from the local specification profile library or the specification profile library issued by the upper-level scheduling terminal. Weight dispersion Geometric outline range Baseline of pick-up and drop forces With Permissible Attitude Band Among them, weight dispersion Instead of directly providing ordinary second-order fluctuations, the profile parameters are solidified by the quantile width of qualified samples, batch tolerance, and packaging adsorption margin. When materials of various specifications undergo compression, adhesion, or minor surface adhesion in the storage cell, extreme samples directly widen the ordinary statistics, thus masking over-receiving of light items and under-receiving of heavy items with the same wide threshold. After the specification profile is loaded, the storage-distribution controller writes the specification profile version number into the distribution event record, ensuring that both the calibration baseline write-back in step three and the gated latch write-back in step five can locate the same profile source.

[0059] During the specification profile establishment phase, the storage-distribution controller performs weighing, visual acquisition, and pick-and-place response acquisition on qualified samples of the same category, and aggregates the qualified weight samples. The outer contour sample set and the pick-and-place response sample set are written into the specification profile library; during runtime, the profile is called according to the specification profile version number to avoid mixing newly created samples and official profiles. Expected single-piece weight. From qualified weight sample set The median center is determined, and the weight dispersion is... From qualified weight sample set The quantile band is determined, and the geometric outline range is defined. Determined by the outer contour envelope, the allowable attitude band The baseline for taking and placing forces is determined by the attitude distribution on the support plate. Institutional operating condition identification The grab, lift, and release response envelopes are determined.

[0060]

[0061] Weight dispersion Category identifier The qualified weight sample quantile band, with positive values, is used to normalize the weighing deviation; the qualified weight sample set This indicates the weight record of a qualified single item that has been manually confirmed or verified within the filing or isolation learning window; upper quantile operator. The lower quantile operator represents the 0.75th quantile of the sample set. Represents the 0.25th quantile of the sample set; both are obtained through sample sorting; stable micro-quantities. It is a positive number, used to avoid the normalization denominator becoming invalid due to the narrow quantile band.

[0062] In an exemplary field embodiment, the storage-distribution mechanism serves a maintenance spare parts cabinet, which simultaneously stores washers, clips, connectors, and small tool heads. After the operator selects a connector and enters the requested quantity at the requisition terminal, the storage-distribution controller first retrieves the connector's specification profile and then drives the retrieval terminal to the corresponding storage location. After the terminal grips the material, it does not drop it directly but first passes it over a weighing support plate. The vision acquisition unit simultaneously captures the outline and quantity on the support plate, while the wrist force sampling unit records the resistance change before and after gripping and releasing. During this process, slight adhesion of washers, jamming of connector pins, and slippage of tool heads are not treated as the same weight change but are instead linked to their respective specification profiles and entered into the same distribution event record.

[0063] After the specification profile is loaded, the storage-distribution controller conditions the perceived quantities of weighing, vision, and force. For the weighing source, firstly, the measured distribution weight is subtracted from the product of the requested quantity and the expected single-item weight. Then, the weighing baseline bias that varies with time and storage location is deducted. Finally, the weighing conditional residual is obtained by normalizing according to the effective scale formed by the weight dispersion and the source mass gap during the weighing process. :

[0064]

[0065] Where: weighing conditional residual Indicates distribution event Regarding the direction and magnitude of deviation from the current specifications in the weight dimension, there is no fixed upper limit on the value. A positive deviation corresponds to an over-selection tendency, while a negative deviation corresponds to an under-selection tendency; actual distributed weight. The weight reading of the weighing support plate after it has stabilized within the event time window is non-negative; Request quantity The target number of pieces given in the distribution instruction is a positive integer; the expected weight per piece is... Category Identification The corresponding single-item weight profile parameter, with positive values, is used to establish the weight benchmark for this product category; weighing baseline offset. Category Identification In time The zero drift, temperature drift, and residual compensation of the support plate can be positive, zero, or negative;

[0066] Weight dispersion The image scale parameter for category identification, with a positive value; weighing source quality penalty coefficient. The non-negative adjustment is used to amplify the normalized denominator when the weighing source is unstable; the mass gap of the weighing source. The range of values The higher the value, the less reliable the weighing source was in this event; stable trace amounts It is a positive number, used to avoid the failure of the normalization denominator when the weight dispersion is too narrow or the source mass gap is small.

[0067] For visual sources, the storage-distribution controller not only reads the visually recognized quantity but also matches the material outline with the permissible orientation range. The visually recognized quantity and the requested quantity create a visual quantity discrepancy. The orientation of the outer center and edges, or their overlapping and occlusion, creates a pose offset. Together, they constitute the visual conditioned residual. :

[0068]

[0069] Where: visual conditional residual Indicates the joint deviation of the visual source from the quantity and posture of this distribution; visual effective gating. The value is either 0 or 1. It is set to 0 when occlusion, reflection, or overlap exceeds the allowable boundary to prevent incorrect counts from entering subsequent fusion; visual quantity deviation. The result obtained by subtracting the number of requests from the number of visual recognitions can be positive, zero, or negative; pose weights These are non-negative coefficients used to control the effect of attitude offset on the visual conditional residual; attitude offset amount The material outline relative to the allowable attitude zone The deviation is non-negative. This formula ensures that events where the quantity appears correct but the posture shows overlapping edges are not directly classified as correct, but rather leave a questionable direction for step two.

[0070] For force-sensing or motor current sources, the storage-distribution controller uses the response curves of the pick-and-place end during the gripping, lifting, moving, and releasing phases as the object, extracting the difference between the clamping resistance, release resistance, or motor current envelope and the pick-and-place force baseline. Force-sensing conditional residuals are then used. The following formula is used to obtain:

[0071]

[0072] Where: Force-conditional residual Indicates the deviation of pick-and-place resistance relative to the current product category and mechanism operating conditions; force response characteristics. The resistance or current envelope features extracted from the capture to release phase of this event are taken as non-negative values; the release force baseline is taken. Category Identification With the machine's operating condition label The baseline response is non-negative; force scale This represents the permissible response scale under the current product category and institutional operating conditions, and is a positive number.

[0073] Force Source Quality Penalty Coefficient Non-negative adjustment quantity; force source mass gap The range of values ​​is Stable trace amounts It is a positive number, used to avoid the normalization denominator becoming invalid due to an excessively narrow force perception scale.

[0074] When multiple lightweight materials are clamped at the end, the weight difference may not be significant, and the force-conditioning residual will still provide an over-grabbing tendency in terms of gripping resistance or release resistance; when large materials slip, the force-conditioning residual will, together with the weighing-conditioning residual, point to a tendency to miss.

[0075] After completing the three types of conditional residuals, the storage-distribution controller sets a source quality mask for each sensing source.

[0076] The quality of the weighing source is determined by the weighing stabilization time, range boundaries, support plate residue, and vibration interference; the quality of the visual source is determined by the occlusion ratio, reflective area, outline integrity, and image sharpness; the quality of the force source is determined by the continuity of current sampling, the integrity of the clamping phase, and the impact degree of the release phase. The source quality mask does not replace the residual itself, but rather varies with the weighing conditions of the residual. Visual conditioning residuals Conditioned residuals of force perception Outputting them together allows step two to read both the deviation direction and the reliable boundary of the deviation from the source during fusion.

[0077] To ensure that the source quality mask has a computationally feasible source, the storage-distribution controller generates a source validity score for each sensed source. Source validity Residual without changing weighing conditions Visual conditioning residuals Conditioned residuals of force perception The directional meaning is not used as a source trusted boundary during the second step of fusion, so that occlusion, overrange, and timing misalignment will not enter the three-state trusted record in an equal weight manner.

[0078]

[0079] Where: source validity Indicates source index In the distribution event The availability of the data is within a certain range. Source Index Pick , or These represent the sources of weight, visual perception, and force perception, respectively; source mass gap. Represents the stability gap, with values ​​ranging from 0 to 1. ; out-of-bounds quantity Indicates the degree of over-range, glare obstruction, or overload; value ranges from [value missing]. Delay This indicates the alignment deviation between the source timestamp and the distributed event timestamp, taken as... Gap weight Outbound weight and delay weight All values ​​are non-negative and initialized by factory calibration or historical pass / fail events. Trim function. This means that when the input is less than 0, the output is 0; when the input is greater than 1, the output is 1; and when the input is within the range, the output is the input itself.

[0080] When a sensing source malfunctions, becomes blocked, or exceeds its range, the storage-distribution controller does not delete the distribution event record. Instead, it marks the corresponding source as unavailable and writes the degradation mark to the same record.

[0081] For example, when the camera's field of view is briefly obstructed by the operator's hand, visual effective gating... Set to 0, and adjust the residual based on the weighing condition. Conditioned residuals of force perception It is still retained; when the weighing support plate is in a state of vibration instability, the mass gap of the weighing source is still present. Improved, visually conditioned residual With force-conditioning residual Continue to construct a subset of available sources. If only a single source remains available, the storage-distribution controller will set the degradation flag and transfer the event to the conservative verification entry point, allowing step four to select a review path, instead of providing a final decision in step one.

[0082] The distribution event log output in step one includes the following: Category identifier Number of requests Event timestamp, weighing conditional residual Visual conditioning residuals Force-conditioned residuals Source quality mask, degradation flag, and specification profile version number.

[0083] Preferably, a time-aligned window The stable section before and after release at the end of the pick-and-place mechanism is covered, and the value is configured based on the mechanism cycle time, sensor sampling period, and material settling time; in high-cycle scenarios, a time alignment window is used. Employing shorter windows and preserving force feedback segments before release, time-aligned windows are used in scenarios involving easily scrolling or bouncing materials. Use a longer window and wait for the weighing to stabilize. Time-aligned window. Its purpose is to ensure that the three types of residuals describe the same event, rather than mistakenly including the previous residual, the next pre-fetch, or the vibration of the storage location in the current record.

[0084] In vertical containers where it is inconvenient to install weighing support plates at the end, the visual acquisition unit and force sampling unit serve as the primary sources, while the weighing source is indirectly provided by the receiving pallet downstream of the discharge port. In scenarios where small parts are difficult to weigh stably, the force sampling unit or motor current sampling unit serves as the primary source, while the visual acquisition unit is used to confirm quantity and adhesion edges. In dusty, oily, or low-light environments, the visual acquisition unit is reduced to an auxiliary source, and the weighing support plate and force sampling unit constitute a usable subset of sources. In the cloud-based centralized implementation, the specification image library is shared and distributed by multiple storage-distribution mechanisms.

[0085] In the edge implementation, the specification profile library is stored in the local controller, and the event log generation in step one can still be completed even when the network is disconnected. All the above alternative implementations maintain the same output object, namely the specification-conditionalized multi-source residuals and data quality markers, and therefore do not change the input interface in step two.

[0086] By loading specification profiles, Step 1 interprets the same weight variation as an over-taking tendency, correct taking tendency, or under-taking tendency under the corresponding category, avoiding blindness to materials with large size spans by fixed thresholds. Through conditional residuals of weighing, vision, and force perception, Step 1 retains the event chain even in the event of occlusion, vibration, over-range, or single-source failure, and guides events that are not suitable for direct judgment to subsequent conservative verification paths.

[0087] Step 2: Fuse the standardized multi-source residuals output from Step 1 into a three-state posterior that can distinguish between over-distribution, correct distribution, and under-distribution, and generate continuous distribution confidence from the three-state posterior. This provides a clear control basis for the direct transmission, light verification, review and separation in step four.

[0088] When multiple specifications of materials share the same storage and distribution mechanism, the same weight deviation does not correspond to the same distribution outcome. When small, thin sheets adhere together, the weighing condition residual... It only shows a slight positive deviation, but the visual edges will show overlapping outlines; when large-size connectors slip off, the weighing conditional residuals are... Negative deviations may occur, resulting in force-conditional residuals. Then, the release resistance will suddenly decrease. If step two still uses a single threshold to classify events as pass or fail, then over-fetched events and missed events will be processed by the same review strategy in subsequent actions, and cannot correspond to the two different physical treatments of reseparation and re-fetching.

[0089] Therefore, step two interprets the conditional residuals from step one as directional evidence, and also incorporates the source quality mask into the fusion process, so that each allocated material has traceable tri-state semantics before being released.

[0090] The storage-distribution controller first reads the three types of conditional residuals from the distribution event record, and then determines the appropriate category based on the category identifier. Retrieve the specified conditional weights; after removing sources marked as unusable by the source quality mask, fuse the remaining sources according to their validity to obtain the directional evidence quantity. Then, the directional evidence quantity Mapped to overtake posterior Correct posterior and post-missed testing Finally, based on the concentration of the three-state posterior and the source deficiency penalty, the continuous allocation confidence level is generated. .

[0091] Furthermore, the output of step two includes both which state is dominant and how reliable that state is, thus avoiding step four only seeing a vague anomaly marker.

[0092] In a preferred embodiment, the storage-distribution controller is deployed within the edge control box of the automated dispensing cabinet. Its input terminals are connected to the weighing support plate, the visual acquisition unit at the discharge port, and the force sampling unit at the pick-and-place end. Its output terminals are connected to the sorting gate, the recovery tray, the re-separation mechanism, and the replenishment execution unit. The calculation cycle for step two is triggered by the timestamp of the distribution event, rather than by a fixed inspection cycle. This ensures that each three-state post-hoc corresponds to an actual pick-and-place action, avoiding the mixing of previous item residue, subsequent item pre-grabbing, or manual intervention actions into the current event.

[0093] In this processing method, the storage-distribution controller does not initially determine whether the distribution was successful. Instead, it transforms the residuals from weighing, vision, and force into evidence in the same direction. Weighing sources tend to describe actual quantity and quality differences, visual sources tend to describe outlines and stacking conditions, and force sources tend to describe resistance changes during the picking and placing phases. Since these three sources have different physical meanings, directly averaging them would mask the differences in their origins. The source quality mask and source quality gap are combined to form the source validity, thus reducing the influence of masked visual sources, over-range weighing sources, or force sources with sample breakage in this event, rather than simply deleting them or allowing them to participate with equal weight.

[0094]

[0095] Among them, the amount of directional evidence Indicates distribution event The overall deviation direction indicates that a positive deviation suggests over-taking, a negative deviation suggests under-taking, and a deviation close to the neutral range indicates correctness; weighing validity. Visual effectiveness and the effectiveness of force perception The range of values ​​is all The weights are determined by the corresponding source quality mask and the corresponding source quality gap, respectively, and are used to indicate whether the source is available and to what extent in this event; Visual weight and force weight All are non-negative numbers, identified by category. Corresponding specification profiles are provided to illustrate the degree of dependence of different product categories on different sources; weighing condition residuals. Visual conditioning residuals Conditioned residuals of force perception All outputs from step one are used to retain deviation information from the three types of physical sources; stable trace amounts A positive number is used to avoid denominator invalidation when available sources are extremely limited. Weighing weight. Visual weight and force weight The initial value is determined by the separation degree of the corresponding source in historical review events. The storage-distribution controller separately calculates the separability of weighing source, visual source, and force source for over-collection and under-collection events, and then normalizes them into category weights. Thus, different categories can form different source dependency structures.

[0096]

[0097] Where: Source weight Indicates source index Category labeling The fusion contribution is a non-negative number; source separation degree Reference source index Category labeling The degree of separability between over-fetched and under-fetched events, with a non-negative value; Source Index Traversing the sources of weighing Visual source Helijue Source Stable trace amounts The number of values ​​is set to a positive number to prevent the normalization denominator from becoming invalid when there are insufficient historical samples.

[0098] For example, in a maintenance spare parts cabinet, when an operator requests the distribution of several thin washers, if a positive weight deviation occurs on the weighing support plate and overlapping edges appear simultaneously in the visual image, the storage-distribution controller will adjust the directional evidence. Move towards the over-picking side; when the operator requests a heavier connector, if the weight on the weighing support plate is negatively biased and the release resistance at the picking / placing end is lower than the corresponding picking / placing force baseline, then the directional evidence quantity... The process moves towards the side that has been missed. This process does not alter the sensor hardware layout, but it compresses different fault modes caused by different materials onto the same directional axis, facilitating the subsequent generation of a stable three-state posterior. Furthermore, signals from sources with poor quality will not dominate the judgment, and the residuals of materials of different specifications are retained as comparable directional evidence. Step four can directly read the control tendency of the side that has been missed or the side that has been overtaken.

[0099] In obtaining the amount of directional evidence Subsequently, the retrieval-distribution controller does not immediately compare with a fixed threshold, but instead constructs over-retrieval utility, correct utility, and under-retrieval utility. The boundaries of the three-state utility are determined by category identifiers. Corresponding over-scan and under-scan discrimination bands are defined, therefore, thin sheets, clips, connectors, and tool heads will not share the same discrimination band. Meanwhile, mask missing degree... The three-state utility is used to reduce the dominance of the correct state when the source is insufficient, and to prevent uncertain events from being mistaken for correct events due to the absence of a key source.

[0100]

[0101] Among them, over-utilization Indicates distribution event The strength of evidence tending to be excessive; understated utility Indicates distribution event Strength of evidence tending to be omitted; correct utility Indicates distribution event The strength of evidence tends to be correct; over-sensitivity coefficient Missed transmission sensitivity coefficient and correct sensitivity coefficient All are positive numbers, indicated by the product category identifier. Corresponding specification profiles or verification samples are generated; over-selection discrimination bands are obtained. A positive number is used to define the overshoot-side trigger boundary; the missed detection band. A positive number is used to define the trigger boundary on the missed transmission side;

[0102] Missing source correctness penalty coefficient A non-negative number used to reduce correctness utility when the source is insufficient; mask missing value. The range of values ​​is The number of unavailable sources and their category weights are jointly determined. This setting ensures that a lack of sources only reduces the chance of correct dominance, without directly creating over-fetching or under-fetching directions.

[0103]

[0104] Among them, the three-state posterior Representing state In the distribution event The posterior proportion in the range is: State Index Pick , or These correspond to over-fetching, correct, and missed sending, respectively; State Index To calculate the summation index, iterate through three states: over-sent, correct, and missed; exponential function. This is a commonly used positive value mapping function, used to convert three-state utilities into normalizable positive numbers. In practice, it can be implemented by the controller's floating-point arithmetic unit, lookup table interpolation, or fixed-point approximation; (over-taking posterior) Correct posterior and post-missed testing The sum of the three is 1.

[0105] In terms of on-site actions, the display interface of the storage-distribution controller does not need to show the formula results. It only needs to mark the current event as over-retrieval dominant, correct dominance, or missed distribution dominant in the background.

[0106] For example, when thin washers are stacked together and fall onto the support plate, a double-edged visual outline appears, resulting in a weighing conditional residual. and visual conditional residuals Same direction support over-take posterior When the connector is stuck at the edge of the storage location and has not completely fallen, the force-conditional residual is... With weighing conditional residuals Jointly support post-missed detection .

[0107] Furthermore, the output of step two is no longer an isolated abnormal score, but a state probability that corresponds one-to-one with the subsequent physical action, which can provide different entry points for further separation, supplementation and direct output.

[0108] After the three-state post-hoc is generated, the storage-distribution controller will also calculate the continuous distribution confidence level. The reliability of continuous distribution mentioned here is not the correct posterior. Instead of copying, it considers the gap between the maximum a posteriori and the second-largest a posteriori, as well as the mask missing degree. This setting is because a slightly higher posterior in a certain state does not necessarily mean that the discrimination is stable; only when the dominant state and the competing state are spaced far apart and there is no significant missing key source is the event suitable to enter the low-intervention path of step four.

[0109]

[0110] Among them, the credibility of continuous distribution The value of , used to indicate the distribution event The three-state discrimination stability; maximum a posteriori It is the posterior. Correct posterior and post-missed testing The maximum value in; the second largest posterior. The second largest value after the maximum a posteriori, used to represent the degree to which the competing state interferes with the dominant state; the source deficiency penalty coefficient. A non-negative value used to adjust the strength of confidence in continuous allocation by deducting the mask missing degree; mask missing degree Following the above rules, the clipping function... Representative input Output 0 when less than 0, input When the value is greater than 1, the output is 1. Otherwise, the output itself is not calculated. The controller can be implemented using upper and lower limit comparators.

[0111] Step two ultimately outputs a three-state trusted record, which includes a category identifier. Number of requests Quantity of directional evidence Post-test Correct posterior Missing items will be checked later. Continuous distribution credibility The steps include: dominant state, available source set, degradation flag, and credibility model version number. Step four involves reading the dominant state and continuous allocation credibility. Subsequently, highly reliable and correctly prioritized events enter the direct distribution path; over-selected low-reliability events enter the review and separation path; missed-selected low-reliability events enter the supplementary or re-selection path; and events with insufficient sources but continuous distribution reliability... Lower-risk events are entered into the conservative review path. Thus, step two connects the algorithm's judgment results with the mechanical actions of the storage-distribution mechanism, rather than remaining at the level of pure mathematical classification.

[0112] In low-computing-power implementations, the storage-distribution controller uses directional evidence quantity. , over-discrimination band Missing data detection band and mask missing degree The decision table for calling the rule, when the directional evidence quantity... Less than the over-band When the output is over-dominant, the directional evidence quantity is... When the negative miss detection band is greater than the output miss detection band, the output miss detection is dominant, and when the directional evidence amount is greater than the output miss detection band, the output miss detection is dominant. It is in the neutral range and the mask is missing. When the source is not exceeded, correct output is preferred; when the mask is missing degree When the source deficiency boundary is exceeded, a conservative verification flag is output.

[0113] When implemented in parallel, if it is inconvenient to install a weighing support plate at the end, the storage-distribution controller should use visual conditional residuals. Conditioned residuals of force perception As a set of available sources, and weighing validity Lower the setting; if the weighing of small parts is unstable, the force perception conditional residual will be lowered. As primary evidence, visually conditioned residuals Used to determine adhesion edges; if the controller's computing power is limited, the three-state utility can be replaced by a segmented rule tree, with the rule tree still based on directional evidence. , over-discrimination band Missing data detection band and mask missing degree As input, output the three-state posterior approximation with the same name and the confidence level of continuous distribution. .

[0114] When deployed centrally in the cloud, the three-state posterior calculation is completed on the host server after the residual is uploaded by the edge controller; when deployed autonomously at the edge, the controller uses a fixed-point lookup table method to complete the exponential function approximation, and the output fields remain unchanged.

[0115] To facilitate on-site verification, the three-state reliable record also carries a status reason code, which consists of the dominant source, the missing source, and the trigger direction of the discrimination band. What operators see on the front-end interface is not an abstract score, but rather executable prompts such as visual material stacking support over-weighing, negative bias support omission, or insufficient source requiring verification. On the mechanical side, the same record drives gate holding, pallet retrieval, end-point replenishment, or re-separation actions. This field design ensures that the algorithm output, interface prompts, and actuators maintain the same event source, reducing secondary guesswork in the distribution process during manual verification.

[0116] In practice, the effects of specification-conditional weights and source validity eliminate the need for a uniform threshold in multi-source residual fusion, maintaining consistent interpretation across different product categories. The tri-state posterior separates over-allocation and under-allocation into different states, facilitating subsequent physical allocation actions to adapt to the direction of misallocation. Continuous allocation reliability... By simultaneously expressing the stability of the dominant state and the lack of sources, an executable hierarchical entry point is formed between conservative review, direct allocation, and re-separation.

[0117] Step 3: During the continuous occurrence of distribution events, adaptive dynamic calibration and anti-interference boundary control are performed on the multi-source sensing quantities. The specification profile is written back through the calibration baseline of the protection conditions so that the subsequent three-state reliability is still judged based on the actual material specifications rather than sensing drift.

[0118] In a shared storage and distribution system for multiple material specifications, the weighing support plate may experience zero-point offset due to temperature, installation stress, and residual powder. The visual acquisition unit may experience missing counting edges due to cabinet door opening and closing, external light spots, and reflections from transparent packaging. Furthermore, the force response at the pick-and-place end may shift after prolonged operation due to guide rail resistance and gripper wear. If these changes directly affect the weighing condition residual in step one... Visual conditioning residuals Conditioned residuals of force perception Step two interprets changes in device status as over-sampling or under-sampling, causing step four to frequently enter the verification and re-separation path.

[0119] Step three therefore does not aim to correct a single sensor, but rather to simultaneously observe the sources of offsets in weighing, vision, and force sensing along the event chain. It writes compensable slow offsets back to the specification profile and isolates uncompensable sudden interferences outside the conservative mode, thereby breaking the transmission chain of sensor drift - residual offset - three-state misjudgment - error writeback.

[0120] After receiving the tri-state trusted record, the storage-distribution controller does not immediately rewrite the specification profile. Instead, it reads the source quality mask, degradation flag, and weighing conditional residual from the distribution event record. Visual conditioning residuals Force-conditioned residuals And the reliability of continuous allocation output in step two For distribution events with high confidence and correct dominance, the controller uses them as baseline observation events; for distribution events with over-dominance, under-dominance, or insufficient sources, the controller only records the interference state and does not use the event for baseline correction. The calibration object comes from a relatively stable real distribution process, rather than from distribution results that have been marked as suspicious.

[0121] Preferably, the weighing support plate is fixed on a rigid support seat in front of the discharge port, the vision acquisition unit is set above or to the side of the support plate, and the force sensing sampling unit at the picking and placing end is set at the gripper drive end; a torque sensor, motor current sampling circuit, light curtain counter, or structured light camera with equivalent functions can also replace the corresponding sensing source. The communication message uses the distribution event record field and adds calibration status field, protection status field, and baseline version field, so that step one can read the updated specification profile in the next distribution event.

[0122] Step 3: Identify the nature of changes in multi-source sensing quantities within the sliding event window. If the weighing readings deviate in the same direction after no-load, material dropping, and release, and the visual count and force response cannot synchronously support over-taking or under-taking, this value is interpreted as weighing baseline bias. If the visual outline shows continuous edge breaks in a category, while the weighing and force responses still revolve around the specification image, this value is interpreted as being affected by visual illumination or occlusion interference. If the force response increases simultaneously in a category, but the weighing and visual responses do not show any quantitative anomalies, this value is interpreted as a change in mechanical resistance.

[0123] The storage-distribution controller generates baseline candidates based on this, instead of writing back individual event residuals directly.

[0124]

[0125] Among them, the event filtering window Indicates a sliding event window Simultaneously satisfying correct dominant gating Weighing can be gated The event set for distribution is used to exclude over-collection, under-collection, and weighing unavailability events; weighing baseline candidate offset. Indicates time The offset of the weighing baseline can be positive, zero, or negative; truncated mean operator. Indicates the removal of the filter event window The average value is calculated after considering the abnormal residuals at both ends, and the elimination ratio is preset by the controller according to the sliding event window capacity; the actual distribution weight is measured. Number of requests Expected unit weight Weighing baseline offset and sliding event window The aforementioned definitions are used for all cases.

[0126] In an exemplary implementation of the maintenance spare parts cabinet, when operators repeatedly retrieved the same type of gaskets, a small amount of oil film remaining on the support plate caused the no-load reading to rise slowly. The storage-distribution controller observed that the visual acquisition unit could still see the outline of a single piece, and the force sampling unit did not exhibit clamping resistance due to multiple pieces sticking together. Therefore, this change was categorized as a candidate offset of the weighing baseline. This process, without increasing the gasket size, limits weighing drift to the weighing baseline offset level, preventing step two from mistakenly interpreting normal single-piece distribution as over-dispensing.

[0127] After obtaining the candidate offset of the weighing baseline, the storage-distribution controller further determines whether the offset has the characteristics of being slow, continuous, and consistent across events. For offsets that meet the protection conditions, the controller generates the baseline correction amount by proportionally compensating for zero drift, integrally compensating for slow drift, and differentially suppressing abrupt changes, and writes it back to the weighing baseline offset in the specification profile; for offsets that do not meet the protection conditions, the controller maintains the original baseline version and puts the corresponding storage location into conservative mode. This compensation is not a simple single-point zeroing, but rather a correction that superimposes the correct dominant gating given in step two with the source quality mask in step one, so that suspicious verification events will not contaminate the specification profile in reverse.

[0128]

[0129] Among them, weighing mutation amount The signed difference between the current candidate offset of the weighing baseline and the candidate offset of the weighing baseline at the previous calibration time, which can be positive, zero, or negative, is used to compensate for short-term directional jumps; weighing baseline offset. It is a category identifier At that time The effective weighing baseline compensation can be positive, zero, or negative; the clipping function... Limit the input to between the lower and upper bounds. Scale-compensated weights. Integral Compensation Weight and differential suppression weights All are non-negative; cumulative offset of weighing It is the cumulative magnitude of the weighing baseline candidate offset within the filtering event window; the lower bound of the weighing offset. and weighing bias upper bound This limits the range of the weighing baseline offset write-back.

[0130] When implementing the preferred method, the proportional compensation weight Greater than the integral compensation weight Differential suppression weights Increased efficiency during cabinet vibration, material drop impact, or after the support plate has just started; when the equipment is in a low-frequency nighttime use state, the storage-distribution controller uses periodic batch calibration and caches the baseline correction in the local controller, and synchronizes it to the upper end after communication is restored; when multiple storage-distribution mechanisms serve the same batch of materials, each node uploads baseline candidates that have passed the protection conditions, which are then aggregated by the cloud to form a shared calibration baseline, and then distributed to the corresponding nodes.

[0131] Regardless of whether local edge implementation or centralized cloud implementation is adopted, the write-back field points to the weighing baseline offset. Without changing the expected unit weight and weight dispersion The meaning of material properties. Therefore, the calibration process is confined to the sensing baseline layer, preventing the equipment status from being mistakenly fixed as material specifications.

[0132] After completing the baseline compensation calculation, the storage-distribution controller also needs to determine whether the current sensing source has exceeded the compensable boundary. If the weighing support plate exceeds the range, the visual occlusion ratio exceeds the image usability boundary, the force response exceeds the end overload boundary, or the same storage location experiences consecutive source unavailability events, the controller will no longer perform shallow write-back, but will instead write the conservative mode flag into the distribution event record and lock the storage location as pending review or manual confirmation.

[0133] The purpose of this protective boundary is not to increase sensitivity, but to provide a clear physical security entry point for step four, preventing anomalous sources from continuing to participate in the generation of three-state credibility.

[0134]

[0135] Among them, protection is achieved through marking. The value can be 0 or 1, with 1 indicating a category identifier. In time The multi-source sensing state is still within the compensable range; setting it to 0 indicates entering conservative mode; indicator function The value is 1 if the condition within the parentheses is true and 0 if it is false; (Weighing out of bounds) For the weighing source in time The comprehensive characterization of range approximation, residual anomalies, and insufficient stabilization time is taken as a non-negative value; visual out-of-bounds quantity. For the visual acquisition unit in time The comprehensive representation of the occlusion ratio, reflective area, and outline loss, with a non-negative value; force perception boundary crossing amount. For picking up and putting down the end in time The comprehensive characterization of overload, impact, and sampling fracture is taken as a non-negative value; weighing protection threshold. Visual protection threshold and force protection threshold All are category identifiers The corresponding non-negative boundary is used to limit the compensable range of the corresponding source.

[0136] For example, during connector distribution, the reflective properties of transparent packaging prevent the visual acquisition unit from fully capturing the pin edges, but the weighing support plate remains stable, and the force response is consistent with the pick-and-place force baseline. The storage-distribution controller degrades the visual effectiveness and allows step two to continue fusing the available subset of weighing and force feedback sources. If subsequent weighing instability, visual obstruction, and force overload occur simultaneously at the same storage location, protection is activated via marking. If the value is 0, the controller pauses the automatic direct delivery for that warehouse location, and step four is redirected to verification or manual processing. For parallel implementations where small parts are difficult to weigh stably, the weighing source can be used only as an auxiliary source, with the visual acquisition unit and motor current sampling unit as the main sources; for dusty or low-light environments, the visual acquisition unit is reduced to an auxiliary source, with the weighing support plate and force sampling unit as the main sources.

[0137] The above parallel embodiments all retain protection through marking. Weighing baseline offset The output has the same name as the conservative mode tag, so it does not change the interface between steps four and five.

[0138] By stripping away spatiotemporal interference, step three separates compensable zero drift, temperature drift, and mechanical drift from actual over-dispensing and under-dispensing events, ensuring that the specification-conditional residuals from step one are still generated around the current product category specification profile. Through proportional-integral-differential baseline compensation, step three limits the shallow calibration write-back to sensor baseline parameters such as weighing baseline offset and pick-and-place force baseline, preventing suspicious dispensing results from being directly written as new material specifications. This is achieved through protection and marking. With the conservative mode flag, step three stops unreliable writeback in scenarios of over-range, occlusion, overload, and continuous anomalies, and provides a verification entry point for step four.

[0139] Step 4: Based on the credibility of continuous distribution The three-state posterior and conservative pattern labels are used to select differentiated distribution paths and form legacy intermediate states that can be processed in step five.

[0140] In a shared storage-distribution mechanism, materials retrieved from their storage locations typically pass through a support plate, a visual field, a verification pallet, a discharge gate, and a recycling channel. If the controller only considers the result of step two as pass or fail, over-dispensing caused by sheet adhesion, under-dispensing caused by connector slippage, and insufficient supply caused by camera obstruction will all enter the same retry action, easily leading to repeated material retrieval, repeated backtracking, and queue accumulation. Step four maps the correct-dominant event, over-dispensing dominant event, and under-dispensing dominant event to direct distribution, re-separation verification, and supplementary distribution verification, respectively, establishing a one-to-one correspondence between the algorithm's three-state semantics and the mechanical gate, pallet, gripper, and recycling channel. The engineering intent is to improve the reliability of continuous distribution obtained in the previous step. It is no longer a value displayed on the screen, but a control quantity jointly controlled by the discharge path, verification strength, and number of cycles.

[0141] The storage-distribution controller first reads the over-fetch post-hoc from the three-state trusted record. Correct posterior Missing items will be checked later. Continuous distribution credibility Dominant state, available source set, and degradation flag; then read the protection flag output from step three. Weighing baseline offset And conservative mode marking. If protection is achieved through marking. For it to be effective, the controller determines the dynamic threshold. With light verification threshold Determine the distribution path; if the conservative mode flag is already set, the controller will not execute high-reliability direct transmission, but will instead send the event to the review tray or manual confirmation entry.

[0142] In the preferred embodiment, the verification tray is located in front of the discharge gate, and the mechanical re-separation mechanism is located between the verification tray and the recycling channel. The replenishment execution unit is still completed by the same pick-up and drop-off end. A flap gate, paddle, or vibration separation trough with the same function can be used as a substitute, but the output action is still classified into three categories: direct distribution path, light verification path, and verification and re-separation path.

[0143] In the preferred embodiment, the support plate is constructed using a rigid metal plate and a buffer pad, installed in layers. The front end of the support plate connects to the discharge gate, and the side end connects to the verification tray. A recycling channel is located below the verification tray. The pick-and-place end operates along a fixed trajectory between the storage location, the support plate, the verification tray, and the discharge bin. The storage-distribution controller first closes the discharge gate during each path change, then drives the levers or short-stroke conveyor belt to change the material's destination. When material enters the verification tray, the weighing support plate releases its previous load, and the pick-and-place end returns to its standby position to prevent residual mechanical vibration from continuing to affect the verification reading. This arrangement allows the same mechanism to perform both rapid direct delivery and temporary bypass in the event of low-reliability events, reducing the need for separate discharge subsystems for different specifications.

[0144] Dynamic threshold In step four, the threshold is determined by the category basis, along with the source missingness and protection boundary. The dynamic threshold does not fluctuate significantly with a single suspicious event, nor does it lower the direct allocation standard due to congestion in the review queue; the review queue length... This priority value is reserved for use in step five, so that the security boundary of the direct allocation entry will not be weakened when performing throughput scheduling.

[0145]

[0146] Among them, dynamic through threshold For the range of values Control the direct distribution entry point; the basic category is controlled through the threshold. Category Identification Corresponding initial discharge boundary; source deficiency penalty coefficient A non-negative number is used to indicate the mask missing degree. Increase the direct discharge threshold when increasing the size; protect the penalty coefficient. A non-negative number, used for protection via a marker. When ineffective, raise the direct discharge threshold; pass the lower threshold. and through the upper threshold Each defines the adjustable range of the direct allocation boundary; pruning function. Same as above.

[0147] The storage-distribution controller first checks the reliability of continuous distribution. With dynamic threshold Light verification threshold Perform grading, and superimpose correct dominance conditions and protection through marking. Generate path selection results Path selection results Instead of relying solely on maximum a posteriori, it requires that correct dominance, source boundary, and protection boundary be satisfied simultaneously to avoid direct discharge of materials after low-quality sources accidentally give high scores.

[0148]

[0149] Where: Path selection result Indicates distribution event The execution path, and its value. Corresponding direct allocation path and value Corresponding light check path, value Corresponding to the verification path and value Corresponding missed review path and value Corresponding conservative review path; continuous distribution credibility Dynamic threshold Light verification threshold Correct posterior Post-test Missing items will be checked later. Maximum a posteriori and protection through marking Same as before.

[0150] Generate path selection results Subsequently, the storage-distribution controller encapsulates a control message, which includes at least a command number, a storage location number, a path field, a forced mode field, and an event timestamp.

[0151] For example, when the material to be dispatched in the spare parts cabinet is a gasket, if it is correct, it will be inspected later. Dominant and continuous distribution credibility Achieving dynamic threshold The discharge gate opens, and the gasket enters the receiving bin from the support plate; if continuous distribution is reliable... If the material falls into the light verification zone, the discharge gate remains temporarily closed. The vision acquisition unit re-images the edge of the support plate, and the weighing support plate waits for the material to settle before giving another reading. If the material exceeds the verification range, the weighing process will proceed. Dominance or protection through marking If invalid, the paddle directs the material into the verification tray, where the re-separation mechanism separates the adhering material and returns it to the verification position. This process directly binds the path action to the three-state result of step two, preventing over-fetch events from being mistakenly sent to the supplementary fetch action.

[0152] Events entering the light verification path or the verification-separation path will generate verification measurement residuals. The storage-distribution controller does not directly replace the posterior result of step two with the verification result; instead, it first calculates the verification deviation. Then judge the verification deviation. Does it fall within the acceptable boundaries? If the review deviation is found... In the review deviation threshold Inside, the controller releases the material and writes the verified measurement deviation into the residual statistics; if the verification deviation... If the boundary is exceeded, the controller will perform re-separation or re-fetching according to the over-fetch dominance or under-fetch dominance direction respectively, and limit the number of re-loops to prevent a single low-reliability event from occupying the mechanism for a long time.

[0153]

[0154]

[0155] Where: verification deviation The range of values ​​is Used to indicate the degree of constraint on the change in residuals before and after verification; verification weighing residuals Verify visual residuals and the residual of the verification force These are the residuals obtained by re-collecting and conditionally adjusting the data according to the rules in step one, respectively, in the verification path;

[0156] Weighing conditional residuals Visual conditioning residuals Conditioned residuals of force perception The residual of the original event in step one; called the repeated kernel weight. Visual review weight Helijue Review Weight All are non-negative, reflecting the reliability of different sources during the verification stage; constrained mapping function Compress any real number input to Used to mitigate single-point abrupt changes caused by material drop impact or short-term shielding; input variables This is to verify the difference between the original residual and the original residual.

[0157]

[0158] Where: Verification direction quantity This indicates the change in direction of the residual after verification relative to the original residual, and can be positive, zero, or negative; verification of the weighing validity. Verify visual effectiveness and the effectiveness of review and verification All values ​​are taken from The weight of the repeated kernel is obtained from the source quality mask in the verification path. Visual review weight Helijue Review Weight All are non-negative; verify the weighing residuals. Verify visual residuals and the residual of the verification force The verification direction quantity was obtained by re-collecting data from the verification path and following the rules in step one. Used for reviewing deviations When crossing the boundary, distinguish between the separation direction and the replenishment direction.

[0159] In an exemplary implementation of the connector, after the pick-and-place end places the connector on the support plate, the vision acquisition unit sees that the edge of the pin is obscured, and the continuous distribution reliability is determined. Entering the light calibration zone. The controller closes the discharge gate, the grippers retract to the safe position, and the vision acquisition unit re-acquires the image from another angle and weighs the support plate, waiting for the vibration to subside before rereading the value.

[0160] If the review deviation Falling into the review deviation threshold The controller opens the discharge gate; if the verification still shows that the missing item is dominant, the pick-up and drop-off end returns to the same storage location to retrieve one more item, and regenerates the distribution event record after the re-retrieval. The above actions ensure that the basic correction comes from the verified measured deviation, rather than from the initial suspicious reading, thus protecting the calibration baseline of step three and the three-state discrimination band of step two.

[0161] Step four, after completing the actions of the current event path, suspends the state that the basic closed loop cannot resolve on its own as a legacy intermediate state. This legacy intermediate state is not a regular runtime log, but rather the trigger input for step five, originating from review deviations, threshold changes, conservative mode counts, and review queue status. The storage-distribution controller encapsulates threshold drift trends, low-confidence write-back flags, new specification non-convergence, conservative mode counts, review queue length, recycle exceeding limits, and high-bit latency in the same intermediate state vector. This allows step five to identify three different deep states: the threshold is being pulled by suspicious samples, the new specification is not yet stable, or low-confidence events are blocking the channel.

[0162]

[0163] Where: Residual intermediate state vector Indicates time Next, the output of step four is fed into the set of suspended states in step five; threshold drift trend. Indicates dynamic passage threshold Offset direction and intensity in consecutive events, with non-negative values; write back low-confidence flags. A value of 0 or 1 is used to verify whether the measured deviation is insufficient to support the basic write-back; new specification non-convergence flag. Use 0 or 1 to indicate whether the product category is still under observation for the new specifications; conservative counting method. A non-negative integer is used for consecutive and cumulative entries into conservative mode for category identification; check queue length. A non-negative integer representing the number of events awaiting review; recycle overrun flag. Use 0 or 1 to check if the re-separation path has reached the re-loop limit; high-order delay It is a non-negative number used to characterize the waiting time level at higher positions in the review queue.

[0164] When implemented in parallel, the multi-outlet storage-distribution mechanism can assign the direct distribution path to the main discharge port and the verification and separation path to the lateral verification port; in weak network scenarios, the storage-distribution controller locally caches legacy intermediate state vectors. Upload will resume once communication is restored; in low-computing-power scenarios, path selection results... This can be implemented using a rule tree, which still reads the continuous allocation trustworthiness. Dynamic threshold Light verification threshold and protection through marking The technical advantage of this processing method is that step four can complete a single material release, and can also hand over the drift, transformation and congestion states that are difficult to handle in the basic closed loop to step five through a unified interface, thus avoiding the shallow write-back from having to undertake all the correction tasks alone during long-term operation.

[0165] By employing credibility grading and path command encapsulation, step four transforms the dominance of correctness, over-taking, and missed transmission into direct release of verification actions, followed by separate verification and supplementary verification actions. Under this mechanical approach, the verification results possess a certain degree of mechanisticness; through verification deviations... With restricted mapping function The verification results are used for basic calibration and residual statistics, preventing material drop impacts or sudden shading changes from directly affecting the baseline; by using the residual intermediate state vector Threshold drift, non-convergence of new specifications, and explicit suspension of verification congestion are used for gating latching, type change isolation, and starvation prevention scheduling in step five.

[0166] Step 5: Based on the remaining intermediate state suspended in Step 4, the gating latch credibility model and dynamic threshold are used to isolate the new specification samples and schedule the verification queue and storage location migration to prevent the basic write-back from drifting, contaminating or becoming congested during long-tail operation.

[0167] Steps one through four can handle over-distribution, correct distribution, and missed distribution in a single distribution, but their write-back primarily serves to calibrate the baseline and residual statistics. When reflective transparent packaging, tray residue, gripper wear, or short-term congestion occur consecutively, low-confidence events will repeatedly enter the verification path, and the verification results will then dynamically guide the threshold through. And credibility model parameters Slow offset. If a new specification material is added to the storage location and its sample has not yet stabilized before being merged into the existing category model, the source weights and discrimination bands of the original category will be contaminated by the residuals of the new specification. Therefore, step five regards the remaining intermediate state in step four as a lesion indication of the basic closed loop, first determines whether the conditions for reliable write-back are met, and then decides on latching, freezing, rolling back, isolation, or scheduling.

[0168] The storage-distribution controller reads the legacy intermediate state vector. This includes threshold drift trends. Write back low confidence markers New specification non-convergence marker Conservative mode counting Length of the verification queue Recycle Exceedance Flag and high latency .

[0169] When any of the above states crosses the corresponding boundary, the controller retrieves the weighing conditional residual from the sliding event window in the distribution event record. Visual conditioning residuals Force-conditioned residuals The three-state posterior and verification residuals are used to form sliding window residual statistics. When the remaining intermediate state does not trigger the boundary, step five goes into hibernation, while steps one through four continue to complete the normal distribution. This design avoids rewriting the model for every event and reduces self-traction caused by occasional suspicious events.

[0170] The storage-distribution controller first compresses the drift, cold start, and queue states given in step four into the trigger flags in step five. Upon triggering, the controller does not directly update the model. Instead, it compares whether the directions of the weighing, visual, and force residuals within the same sliding event window support each other, and compares the deviations between the recent tri-state posterior distribution and the stable baseline distribution. Subsequent latching only has an input basis when the deviation directions from different sources can be mutually interpreted and the recent distribution has not yet crossed the drift boundary.

[0171] The storage-distribution controller triggers a flag in calculation step five. First, the length of the review queue is checked. Recycle Exceedance Flag and high latency Forming queue congestion This ensures that congestion triggers have a clear data source.

[0172]

[0173] Where: queue congestion degree This represents the pressure of the review queue, with a value range of [value range missing]. Queue weight Recirculating weights and delay weight All are non-negative numbers; check queue length Number of events awaiting review; Recycle limit exceeded flag. The value can be 0 or 1; high-order delay To verify the high-order statistics of queue waiting time; pruning function. The aforementioned definition will be used.

[0174]

[0175] In the formula: Step 5 trigger flag The value is either 0 or 1, used to determine whether step five should be intervened; threshold drift trend It is a non-negative quantity used to characterize the dynamic passage threshold. Continuous offset; drift trigger threshold A non-negative value used to define acceptable drift boundaries; write back low-confidence flags. The value is 0 or 1, used to characterize that the verification test deviation is insufficient to support the basic write-back; new specification non-convergence flag. The value can be 0 or 1, and is used to represent the category identifier. Currently under quarantine observation;

[0176] Queue congestion The range of values ​​is The congestion trigger threshold is obtained from the review queue length, the recycle overrun flag, and high-order delay compression. The range of values ​​is This is used to define the boundary for the review queue to enter the scheduling intervention;

[0177] Indicator Function This indicates that the output is 1 if the condition within the parentheses is true, and 0 otherwise; logical symbol. This means that the event can be triggered if any one of the conditions is met.

[0178]

[0179] Where: cross-source consistency The range of values ​​is This is used to characterize whether different sources support the same state in the same direction and whether the magnitude of the residuals before and after verification is explainable; drift degree It is a non-negative quantity used to characterize the deviation of the recent three-state confidence distribution from the stable baseline distribution;

[0180] Source Index Pick , or These correspond to the sources of weighing, vision, and force, respectively; sign function Indicates positive, zero, and negative directions; consistent direction The main direction of valid sources within the sliding event window; bin collection It is composed of the three-state dominant state and the confidence interval of continuous allocation; recent distribution Derived from the normalized result of the number of events in each bin within the current sliding event window; stable baseline distribution It is derived from the normalized results of confirmed allocation events under the previous stable parameter version;

[0181] Source weight The value is non-negative, and the sum of the weights of the three source categories is 1, determined by the source quality mask and the category weight; the mean of the sliding window residuals is verified. Source Index The constrained mean in the review path; the mean of the original sliding window residuals. Source Index The constrained mean in the original events of step one; drift. It is a non-negative quantity used to characterize the deviation of the recent tri-state posterior distribution from the stable baseline distribution;

[0182] binning index This is the discrete interval index of the tri-state posterior distribution; recent distribution. and stable baseline distribution All are non-negative normalized distributions; smoothing amount It is a positive number, used to prevent the division from failing when the distribution interval is empty;

[0183] Exponential function Sum and logarithmic functions All of these are standard mappings that the controller can implement using table lookup or floating-point operations;

[0184] For example, when the spare parts cabinet collects gaskets in the morning, the verification trays continuously queue, and some gaskets enter the re-separation mechanism due to oil film adhesion. The controller, seeing the increase in the verification queue and the recirculation exceeding the limit, sets the trigger flag in step five. Then, the residuals of the weighing, visual, and force sensing methods are checked; if the visual stacking edge is consistent with the positive bias direction of the weighing, and the force sensing release resistance also shows adhesion, the cross-source consistency is considered good. An increase indicates that the state can be used as a basis for write-back. If only camera reflection causes a visual jump, and the weighing and force perception do not support the same direction, then subsequent latching is blocked. Step five only incorporates the accumulated state supported by the actual physical link; occasional occlusion or single-source failures will not trigger parameter write-back.

[0185] After the cross-source trusted screening is completed, the storage-distribution controller will construct a write-back gating tag. At this point, the controller can read different cross-source consistency levels. and drift Furthermore, latching and updating only proceeds when multiple sources mutually support each other and the overall distribution remains within control. If the drift exceeds the boundary, the current running parameters are maintained, and forward learning ceases; the credibility model parameters are then updated. Repeat the process for the previous stable parameter version. Here, latching is used to write the update result to the candidate area first, and then replace the running area after the gating is passed, without directly rewriting the discharge criteria for suspicious samples.

[0186]

[0187] Where: Credibility model parameters For time The effective parameter package includes at least the source weights from step two, the three-state discrimination band, and the dynamic pass threshold from step four.

[0188] Projection operator This means performing boundary pruning on each parameter in the candidate parameter set to ensure that the candidate parameter does not exceed the center of the stable parameter set. Maximum allowable offset on both sides When the candidate parameter is a threshold parameter, it is also constrained by the physical boundary of that parameter; stable parameter center This is the parameter package corresponding to the previous stable parameter version;

[0189] Maximum Allowable Offset The non-negative boundary is used to prevent a single update from exceeding the interpretable range; the step size function. If the number is positive, it is preferable to use a form that decreases with the number of triggers. The initial step size A positive number indicates the attenuation coefficient. The number is non-negative; Step 5 triggers the flag. Using the aforementioned definition; write back the gating flag. The value is 0 or 1, used to control whether candidate parameters can be latched; candidate write-back value. The parameter increments, calculated from sliding window residual statistics, verification bias, and queue status, can be obtained by recursive Bayesian update, recursive least squares update, or bounded grid search.

[0190] Meta threshold The range of values ​​is Used to define the lower bound of cross-source consistency; drift boundary A non-negative quantity used to limit the deviation of recent distributions from the upper bound; logical symbol This means that both conditions are true;

[0191] At the execution level, the controller writes the reverse write-back message to the local parameter area. The message fields include update number, scope, target object, parameter increment, gating pass flag, drift degree, rollback version, parameter version, timestamp, and verification field. The target object is limited to the weighing baseline offset, pick-and-place force baseline, candidate profile version, confidence model, calibration compensation weight, and dynamic threshold in the specification profile, without changing the allocation event record itself. For the expected single-piece weight and weight dispersion, the candidate profile is replaced with the official profile only when the category is in the new specification isolation learning window and the isolation merging flag is 1. For existing official profiles, the expected single-piece weight and weight dispersion are not directly rewritten by a single verification event or low-confidence event.

[0192] For example, when the gripping resistance of the connector slowly increases with the wear of the grippers, step three first corrects the baseline of the gripping and releasing force. If the verification path still has low-confidence flags written back, the force source weight and the missing detection band can be slightly adjusted according to the gating latch method. If the recent distribution suddenly deviates from the stable baseline, the candidate parameters will be frozen and rolled back to the previous stable parameter version, and the discharge gate will still operate according to the stable criteria. Step five changes the suspicious output from the direct learning object to the gated candidate evidence to reduce the risk of threshold self-poisoning, so that the parameter write-back in steps one to four maintains the same version chain.

[0193] When the remaining intermediate state displays a new specification non-convergence flag. When setting a category, the storage-distribution controller does not immediately incorporate the sliding window residuals of that category into the global model. Instead, it establishes an isolated learning window for that category. Within the isolated learning window, a conservative prior and a higher verification entry are used. All direct delivery results, light verification results, and verification results only update the candidate profile for that category.

[0194] Once the sample size, residual amplitude changes, and cross-source consistency meet the merging boundary, the candidate profile is replaced by the official profile. For peak material handling scenarios, the controller synchronously calls the digital twin pre-simulation, inputting the storage location map, material geometry, verification queue status, and pick-up / placement end trajectory, and outputting the expected recirculation risk, collision risk, and suggested relocation storage location; subsequently, the controller adjusts the path priority and relocates high-frequency and difficult-to-separate materials to shallow storage locations.

[0195]

[0196] In the formula: Isolation and merging mark The value can be 0 or 1, and is used to determine the category identifier. Can candidate portraits be incorporated into official portraits; number of isolated samples Category Identification The number of valid assigned events accumulated within the isolation learning window, with values ​​being non-negative integers;

[0197] Minimum number of samples A positive integer, used to prevent a small number of events from being directly incorporated; residual amplitude variation. To isolate the intensity of variation in the residual quantile width within the learning window, a non-negative value is used to replace ordinary variance in describing the degree of metric convergence.

[0198] Band boundary A non-negative value used to define the stable boundary of candidate profiles; category cross-source consistency. Category Identification The corresponding cross-source consistency score has a range of values. Meta threshold The aforementioned definition will be used;

[0199]

[0200] Scheduling priority value Used to determine the events assigned to the review queue. Processing order; urgency The range of values ​​is The amount is determined by the task type and waiting level; aging quantity. It is a non-negative value that increases with the event waiting time, used to prevent low-priority events from being delayed for extended periods; continuous allocation reliability. Following the output from step two, the value range is... This is used to direct low-confidence events to more adequate review resources; number of recycles. A non-negative integer, used to restrict repeated entry of the re-separation mechanism for the same event; emergency weight. aging weight Low credibility weight and recurrent suppression weights All are non-negative quantities, configured by the on-site distribution strategy;

[0201] In one optimal implementation, the multi-specification spare parts cabinet includes a front row of shallow storage locations, a rear row of deep storage locations, a weighing support plate, an upper vision acquisition unit, a wrist force sampling unit, a verification tray, and a lateral reseparation slot. Newly added elastic clips are first placed in an isolation learning window, and the controller uses a conservative verification entry for these clips. The receiving end can still issue distribution requests, but the candidate profile of the clip does not affect the final profile of the gasket and connector. During the midday peak traffic, digital twin simulations revealed that thin gaskets in the deep storage locations tend to be repeatedly reseparated in the verification queue. The controller issues a storage location migration command, causing the retrieval end to move the thin gaskets to the shallow storage locations during idle intervals and reducing the recurrence limit for such events. Once the limit is exceeded, the process switches to manual confirmation. The new specification does not contaminate the existing model, congestion events do not livelock in the verification tray, and storage location migration directly affects retrieval distance, separation actions, and queue waiting.

[0202] Based on the triggering mark in step five and write-back gated tags This sets deep anomalies in the basic closed loop as restricted write-back conditions, preventing the credibility model parameters from being affected by questionable verification results. By isolating and incorporating markers The new specifications will not be included in the global model until the candidate profiles are stable, reducing the impact of the transformation phase on existing categories; and scheduling priority values ​​will be used. By integrating the review queue, re-separation mechanism, and storage location migration into the same execution chain, the throughput pressure is transformed into bounded path scheduling.

[0203] To ensure consistent parameter sources across steps one through five, the storage-distribution controller stores specification profile versions, reliability model versions, calibration baseline versions, and path threshold versions in its local parameter area. The specification profile version includes at least the expected single-piece weight, weight dispersion, pick-and-place force baseline, and permissible attitude band; the reliability model version includes at least the source weight, over-picking discrimination band, and under-picking discrimination band; the calibration baseline version includes at least the weighing baseline offset, proportional compensation weight, integral compensation weight, differential suppression weight, and protection threshold; and the path threshold version includes at least the dynamic throughput threshold, light verification threshold, and recycle upper limit. The initial values ​​of these parameters are provided by equipment calibration, qualified samples, historical verification records, or conservative default boundaries, and are only updated during operation after the protection conditions in step three or the write-back gating conditions in step five. For new specification materials, the storage-distribution controller first establishes a candidate profile, without directly overwriting the existing formal profile; the candidate profile only becomes the formal profile after the isolation and merging mark is valid.

[0204] In summary, the distribution event record output in step one provides standardized input for the three-state trusted record, while the three-state trusted record output in step two provides directional guidance for path selection in step four. The protection output in step three defines the basic write-back boundary through the marking and calibration baseline version, and the legacy intermediate state vector output in step four provides a trigger entry point for step five. Step five then uses write-back gating markers, isolation and incorporation markers, and scheduling priority values ​​to inversely constrain the aforementioned parameters. In this way, sensor acquisition, three-state discrimination, calibration compensation, path execution, and higher-order write-back all run along the same version chain, avoiding the drift of the meaning of the same parameter in different steps.

[0205] 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 multi-specification material storage and automatic distribution control system, executed by a storage-distribution controller sharing a storage-distribution mechanism, characterized in that, include: At the moment of distribution, read the category and the quantity requested, call the specification profile of the category, and perform specification conditional processing on at least two of the weight, visual and force perception quantities to form multi-source residuals, source quality marks and protection pass marks; Based on the source quality markers, multi-source residuals are fused to generate three-state reliable records that distinguish between over-taken, correct, and missed records, and the baseline writing status is determined by the protection through the markers to determine the specification profile. Based on the three-state trusted record and protection through the tag, the discharge gate, the verification tray, the re-separation mechanism and the supplementary execution unit are controlled to switch between the direct distribution path, the over-the-count verification path and the missed distribution verification path. After the review is completed, a legacy intermediate state containing drift trend, non-convergence marker, and congestion metric is generated. When the legacy intermediate state goes out of bounds, cross-source consistency and drift are calculated based on sliding window residual statistics. When the cross-source consistency is not lower than the meta threshold and the drift is not higher than the drift threshold, the model parameters and path threshold of the three-state trusted record are latched and updated. When the above conditions are not met, the current version is maintained and the candidate update is discarded.

2. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 1, characterized in that: The formation of source quality markers includes: acquiring the weighing stable state, visual occlusion state, and force sampling continuous state at the time of the distribution event; Write the stability gaps, out-of-bounds states, and timing deviations of each source into the source quality flag; continue to generate three-state trusted records when at least two types of sources are available, and set the protection to a conservative state by the flag when fewer than two types of sources are available.

3. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 1, characterized in that: Generating a three-state credible record includes: selecting available multi-source residuals according to source quality labels, and fusing the available multi-source residuals into a directional evidence quantity; When the directional evidence quantity is in the positive interval, an over-dominating state is generated; when the directional evidence quantity is in the negative interval, a missed-issuance dominant state is generated; when the directional evidence quantity is in the neutral interval and the source quality mark is not in a conservative state, a correct dominant state is generated; when there is a directional conflict between available multi-source residuals, the confidence of continuous distribution is reduced and written into the conservative review entry.

4. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 3, characterized in that: The model parameters for the three-state trusted record include source weights, over-scanning discrimination bands, and under-scanning discrimination bands; the storage-distribution controller generates candidate updates based on the verification test deviations after the verification is completed; When the cross-source consistency of the candidate update is not lower than the meta threshold and the drift is not higher than the drift threshold, the candidate update is latched; when the above conditions are not met at the same time, the previous stable parameter version is maintained and the candidate update is discarded.

5. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 1, characterized in that: When one of the weighing, visual, or force perception quantities is missing within the time window corresponding to the time of the assigned event, the source of the missing quantity is written into the source quality tag, and a multi-source residual is formed with the other at least two types of perception quantities. When only one type of sensing quantity arrives within the time window, the direct distribution path is prohibited, and the corresponding material is imported into the verification tray and a conservative verification record is generated.

6. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 5, characterized in that: After the storage-distribution controller sends path commands to the discharge gate, verification tray, re-separation mechanism and replenishment execution unit, it listens for execution receipts within the command receipt time window. When the execution receipt arrives, continue executing the path command; when the execution receipt has not arrived, keep the discharge gate closed and resend the path command; when the number of resends reaches the resend limit, keep the corresponding material in the verification pallet and generate a manual transfer status.

7. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 6, characterized in that: When multiple distribution events simultaneously request the same executable in the review tray, re-separation mechanism, or replenishment execution unit, the storage-distribution controller generates a queue order based on the event timestamp, continuous distribution reliability, and recycle count. The first-team-assigned event locks its path command while it is in the same execution state; non-first-team-assigned events remain in a pending review state and are prohibited from opening the discharge gate.

8. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 7, characterized in that: The multi-source sensing data is obtained from at least two of the following: the weighing support plate, the visual acquisition unit at the discharge port, and the force sensing sampling unit at the pick-and-place end; the weighing support plate is set in the bearing area in front of the discharge gate; The visual acquisition unit at the discharge port covers the bearing area; the force sensing sampling unit at the picking and placing end generates a force sensing quantity using at least one of the wrist force sensor and the motor current sampling circuit.

9. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 8, characterized in that: In low computing power conditions at the edge, the storage-distribution controller generates a three-state trusted record based on the direction interval of the multi-source residual and the source quality mark calling rule discrimination table. When the edge computing power is in normal condition, a three-state posterior estimate is generated based on the multi-source residual, source quality label and the discrimination band corresponding to the category, and the continuous distribution confidence is formed by the three-state posterior estimate.

10. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 9, characterized in that: The storage-distribution controller includes a local control unit and a remote computing unit; When the local control unit is connected to the network, the remote computing unit generates candidate model parameters based on the legacy intermediate state, and the local control unit latches the candidate model parameters based on cross-source consistency and drift. When the local control unit is offline, it uses local sliding window residual statistics to maintain the current version and caches the legacy intermediate state.

11. The multi-specification material storage and automatic distribution control system according to claim 1, characterized in that: The storage-distribution controller generates distribution control messages, which include command number, storage location number, path field, forced mode field, event timestamp, and control verification field. The path field records one of the following: direct allocation path, bypass review path, missed allocation review path, and conservative review path. The storage-distribution controller drives the corresponding discharge gate, verification tray, reseparation mechanism and replenishment execution unit based on the path field.

12. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 11, characterized in that: After the legacy intermediate state goes out of bounds, the storage-distribution controller generates a model update message. The model update message includes update number, scope, target object, parameter increment, gating pass flag, drift degree, rollback version, parameter version, update timestamp, and update verification field. When the gate passes and is marked as passed, the parameter increment is written; when the gate passes and is marked as failed, the previous stable parameter version is maintained according to the rollback version.

13. The multi-specification material storage, retrieval, and automatic distribution control system according to claim 12, characterized in that: The storage and distribution controller drives the retrieval interface to display the current distribution events in the following status sequence: pending distribution, verification in progress, direct distribution, review, re-separation, replenishment, success, and transfer to manual processing. The retrieval interface displays the storage location number, path field, continuous allocation reliability range, and parameter version for each status. The storage-distribution controller synchronously outputs local logs, which record status sequences, distribution control messages, and model update messages.

Images