A multi-sensor fusion meal loss real-time monitoring system

Abstract

The application discloses a kind of multi-sensing fusion's meal loss real-time monitoring system, it is related to food safety monitoring technical field, in edge box collection camera, scale body, bucket, environment and RFID data, complete frame scale alignment, de-identification, exception repair and self zero, generate event record;With confusion correction matrix and Dirichlet prior correction initial probability, and with weight conservation ratio, shunt exception;Introduce inventory and temperature and humidity / ethylene exogenous quantity, sequence fusion smoothing and integration of category meal loss are obtained Batch risk;Carry out online change point and information gain backtracking, locate high waste single product and cause;Generate serving, make up table and disposal work order and write back parameter.The method improves the accuracy of classification and quantification and the stability of time sequence, supports evidence chain compliance retention, reduces meal loss and cost.Improvement strategy can be executed and consistent across teams and can be audited.

Description

Technical Field

[0001] This invention relates to the field of food safety monitoring technology, specifically to a multi-sensor fusion real-time food spoilage monitoring system. Background Technology

[0002] In the uncertain and high-concurrency scenarios of restaurant kitchens, food delivery, collection, and replenishment occur simultaneously. Multiple data acquisition points—video cameras, electronic scales, RFID / barcodes, and environmental sensors—operate independently for extended periods, commonly experiencing clock drift, network latency, occlusion, and jitter. Existing solutions often employ a loosely coupled "camera + scale" statistical approach, using timestamp stitching, linear interpolation, or empirical thresholds for event segmentation and category estimation, with privacy handling primarily relying on simple mosaicking. Under sudden load increases and light / steam disturbances, frame-level classification becomes unstable, and scale values ​​are affected by temperature drift and micro-vibrations, leading to category bias and distorted quality estimation. Furthermore, there is a lack of unified implementation for alignment coupling, robust noise reduction, empty-bucket self-zeroing, and evidence chain storage, and no auditable mapping interface for event granularity representation and quality conservation is established, easily causing inconsistencies in subsequent statistics and traceability. These pain points have been thoroughly observed and recorded on-site, with typical phenomena including "false alignment" during jitter segments and system bias caused by accumulated zero-point drift.

[0003] Based on the aforementioned scenario, the technical problem addressed by this invention is:

[0004] In a multi-source, asynchronous kitchen data stream with privacy constraints, how can we construct a unified, auditable, and online-solvable closed-loop link from observation to decision-making? This link aligns and cleans video frames and scale values ​​at the same time reference and event granularity, forming an event record vector containing time, station location, bin location, weight, image index, and initial class probability. We introduce column randomization for confusion correction and total variation of the class graph within the probability simplex, combining this with a Dirichlet prior for the scene, and use information geometry to obtain a consistent class ratio. Finally, we map this to class-specific meal loss quality according to mass conservation principles. On a unified time grid, a state model coupled with observation fitting, dynamic smoothing, and graph Laplace is combined with inventory incentives and exogenous drivers such as temperature / humidity / ethylene to form time-level sequences and batch-level risks that can be used for diagnosis and scheduling. On the change point detection side, online detection driven by exogenous quantities is established using run length posterior and Bregman divergence to locate category contributions. On the execution and feedback side, scheduling constraints are integrated with weight scale, replenishment rhythm, and disposal weight, and the execution evidence chain is written back to the confusion matrix and prior to maintain parameter consistency. This technical problem is mainly prominent in scenarios such as lunch and evening peak hours, menu switching, inventory anomalies, and cold chain fluctuations. If it cannot be effectively solved, it will lead to event representation distortion, imbalance of category quality allocation, instability of time series inference and anomaly location, and further cause scheduling conflicts, resource misallocation, and evidence chain breakage. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a multi-sensor fusion real-time food waste monitoring system. By collecting data from cameras, scales, bins, the environment, and RFID via an edge box, it performs frame-scale alignment, de-identification, anomaly repair, and zeroing, generating event logs. It uses a confusion correction matrix and Dirichlet prior to correct initial probabilities and adjusts them according to weight conservation ratios to divert anomalies. By incorporating inventory and exogenous temperature / humidity / ethylene levels, it fuses and smooths categorized food waste sequences and integrates them to obtain batch risk. Online change point analysis and information gain backtracking are performed to locate high-waste items and their causes. It generates portion sizes, replenishment orders, and disposal work orders and writes back parameters. This method improves the accuracy and temporal stability of categorized and quantitative analysis, supports compliant evidence chain retention, reduces food waste and costs, and solves the technical problems described in the background art.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A multi-sensor fusion real-time food loss monitoring system includes: acquiring video frames and weighing values ​​at an edge box; employing entropy-normalized optimal transmission-affine time scaling alignment; using robust smoothing and total variation combined with empty bucket self-zeroing to correct the weighing values; generating an event record vector by pooling probabilities according to the alignment weight and the logarithmic opinion pool convergence probability of the quality increment; taking the initial class probability of the event record vector, performing a KL projection overlay on the total variation according to the confusion correction matrix determined by the class, then mixing it with the Dirichlet prior according to the logarithmic geodesy, and combining it with the corrected weighing values ​​to map the class food loss quality according to the quality conservation principle;

[0010] Using a convex quadratic target estimation state, observation fitting, dynamic smoothing, and Laplace coupling are performed on the food spoilage quality. Inventory incentives and temperature, humidity, and ethylene characteristics are introduced to construct a hazard rate and integrate it to obtain smoothed food spoilage and batch risk. Change point detection is performed using the posterior of the running length to map exogenous inventory and environmental characteristics into hazard levels. The evidence is updated with Bregman divergence, and CUSUM is constructed to locate anomalies. Information gain is calculated in the peak window according to the monotonic coefficients of the baseline-extended baseline.

[0011] Using weight scale, replenishment rhythm and batch processing weight as decision quantities, a unified objective is constructed and solved and scheduled under the constraint of upper resource bound; the evidence chain self-calibration is performed: the confusion correction matrix column is updated by convex combination, and the Dirichlet prior is updated by exponential smoothing.

[0012] Furthermore, within the sliding window, an entropy-normalized optimal transmission-affine time scaling joint objective is constructed for the video frame time series and the scale time series. The alignment coupling matrix and time scaling parameters are obtained, and the frame-scale pairing index set is obtained by iteratively solving according to edge constraints and Sinkhorn.

[0013] Furthermore, within the aligned weighing domain, the step structure is recovered by jointly denoising with the pseudo-Huber robust term and the total variation, and the zero-point offset of the bucket position is estimated based on the empty bucket period retrieved from the aligned index set. The weight sequence is then debiased to obtain the corrected weight sequence, which is then bound to the station configuration for storage.

[0014] Furthermore, a detection aggregation map is constructed for each frame at the edge box and a security mask is formed by morphological dilation. Isotropic blurring is applied within the mask area to de-identify the data while preserving the original stitching of non-sensitive areas. This data is then written into the event evidence chain and associated with the frame-scale pairing index set.

[0015] Furthermore, the visual classification probabilities of the aligned frame set are fused using a logarithmic opinion pool. The fusion weight is obtained by normalizing the alignment coupling strength and the quality increment within the corresponding time window, generating an initial event-level category probability and combining it with the timestamp, position, bucket, weight, and image index to form the event record vector.

[0016] Furthermore, the initial class probabilities are mapped to the observation distribution using the confusion correction matrix, and the total variation constraint of the class graph is constructed within the probability simplex to solve for the corrected class proportions. Each column of the confusion correction matrix is ​​Laplace smoothed to be strictly positive and the column sum is one.

[0017] Furthermore, the Dirichlet prior parameters of the scene index and the corrected class proportions are log-geodesically mixed under Fisher-Rao information geometry. The mixing coefficient is adaptively set according to the scene sample size to obtain the prior fused probability vector, which is then used for subsequent mass conservation mapping.

[0018] Furthermore, based on the alignment coupling and mass increment, a consistency index is constructed, and a condensation parameter for power domain sharpening is set accordingly. Element-wise exponentiation and normalization are performed on the prior fused probability vector. Mass conservation mapping is strictly performed according to the event weight, and an abnormal shunting threshold is set with power divergence.

[0019] Furthermore, on a unified time grid, a convex quadratic target estimation state is constructed by observation fitting, dynamic smoothing, and class graph Laplacian coupling. The observation weight matrix is ​​generated by the anomaly metric and scene confidence vector, is diagonally positive definite and adaptively updated over time, and the solution is obtained by banded Cholesky or conjugate gradient.

[0020] Furthermore, an exogenous input model is introduced. The inventory incentive vector is obtained by aligning RFID with inbound and outbound data. The environmental feature vector is composed of temperature, humidity, and ethylene concentration after inverse temperature, logarithmic, and power transformations. These are linearly injected into the state evolution through the inventory coupling matrix and the environmental coupling matrix, respectively, and sparsification is used to suppress overfitting.

[0021] Furthermore, the temperature, humidity, and ethylene characteristics are synthesized into an instantaneous hazard using an exponentially accelerating hazard rate, and integrated in the batch time domain to obtain a batch risk score; the batch risk, category demand pressure, and category size are then combined using a Sigmoid function to form a priority for subsequent scheduling and handling.

[0022] Furthermore, an online change point detection is performed using a runtime posterior framework. The risk level is set as a function of exogenous inventory, environmental features, and the anomaly metric through logical mapping. The evidence item uses the Bregman divergence predicted relative to the intra-segment baseline and is updated in real time within the edge box using truncated recursion.

[0023] Furthermore, a unified objective is constructed using weight scale, replenishment rhythm, and batch processing weight as decision variables and is solved under the constraints of resource and capacity upper bounds. At the same time, work order scheduling is completed by soft maximization of lateness penalty. After execution, the evidence chain is used for the convex combination update of the confusion correction matrix column and the exponential smoothing of the Dirichlet prior.

[0024] (III) Beneficial Effects

[0025] This invention provides a multi-sensor fusion real-time food waste monitoring system, which has the following beneficial effects:

[0026] Within the edge box, entropy-normalized optimal transmission-affine time scaling is used to achieve co-domain alignment between video frames and weighing values. This is combined with robust smoothing and total variation recovery transitions, and weight is corrected by zeroing the empty bucket. Frame-level probabilities are then aggregated using a logarithmic opinion pool to output an event record vector containing key fields. Simultaneously, de-identification and evidence chain binding are completed to form auditable anchor points.

[0027] The initial probability is corrected by the relative entropy projection of the confusion correction matrix and the total variation of the category graph within the probabilistic simplex. Then, it is mixed with the scene Dirichlet prior logarithmic geodesic and sharpened in the power domain by combining the consistency index. Under the mass conservation constraint, it is mapped to the category loss mass. At the same time, the power divergence is used to divert anomalies to maintain consistent connection with upstream and downstream interfaces.

[0028] A state model coupled with observation fitting, dynamic smoothing, and graph Laplace is constructed on a unified time grid. Inventory incentives and exogenous inputs of temperature, humidity, and ethylene are introduced, and batch risk is obtained by integrating the hazard rate. This outputs a time-period-level category food loss sequence and a batch-level risk profile, which are seamlessly connected with the aforementioned quality conservation and event record vectors.

[0029] Online change point detection is performed using runtime posterior, with the risk level set as a logical mapping between exogenous quantities and anomaly measures. The Bregman divergence is used as an evidence term, and anomalies are located in the category dimension using cumulative and statistical methods. Information gain is then calculated around the peak using baseline-extended baseline and monotonic coefficients, and a structured list of causes is output and consistently coupled with the evidence chain.

[0030] A unified objective is constructed using weighted metrics, replenishment cycle time, and batch processing weights, and solved under resource and capacity constraints. Soft maximization of lateness penalties is used to complete work order scheduling. After execution, the evidence chain is used to update the convex combination of the confusion correction matrix columns and the exponential smoothing of the Dirichlet prior. Robust covariance and quantile thresholds are used to adaptively update the anomaly threshold, forming a continuous self-calibration. (See attached figures.)

[0031] Figure 1 This is a schematic diagram of the multi-sensor fusion real-time food loss monitoring system of the present invention. Detailed Implementation

[0032] 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.

[0033] Please see Figure 1 This invention provides a multi-sensor fusion real-time food waste monitoring system, comprising:

[0034] Step 1: Within the edge box, unify multi-source asynchronous observations to the same time base and the same event granularity, output standardized event record vectors, ensure privacy compliance and traceability, and elevate the original observations into aligned and cleaned event representations.

[0035] The various acquisition points in the kitchen experience independent clock drift, network latency, and occlusion; if they are roughly spliced ​​together directly using timestamps, it will lead to category bias and quality estimation distortion. During peak dining hours, video frame sequences often exhibit abrupt changes in activity intensity due to occlusion and jitter, while the weight curve of the scale shows jumps or steps when it is placed and retrieved.

[0036] Linear interpolation alone can lead to false alignment in jitter segments. Therefore, within the edge box, the video frame time series is first formed within a sliding window. With frame activity intensity sequence and weighing time series With the sequence of mass change rate Then, construct the optimal transport-affine time-scaling joint objective with entropy regularization and solve for the alignment coupling matrix. With time scaling parameters , Based on this, a one-to-one pairing relationship between event frames and scale values ​​is established, where:

[0037]

[0038] Where: video frame time series , The video frame timestamps are in seconds and are a real-number increasing sequence; the weighing time series. The unit is seconds; Time-stamped value: Video activity intensity sequence , is a scalar feature of inter-frame pixel difference or optical flow energy, and is a non-negative real number; For frame activity intensity, quality change rate sequence , where is the absolute value of the first-order difference of the weighing signal, expressed in kilograms per second, and is a non-negative real number. This refers to the rate of change in quality.

[0039] Cost Matrix The element is defined as:

[0040]

[0041] in, Time scale (seconds) For the strength scale (consistent with the selected strength definition), the weighting factor Coupling matrix :frame With the scale Matching rights; element field , representing the pairing quality of frame-scale values, satisfying the set of edge constraints:

[0042]

[0043] Among them, edge weights The distributions on the K-dimensional and J-dimensional probability simplexes are usually uniform or weighted according to the activity intensity.

[0044] Entropy regularity coefficient Controlling coupling sparsity; time scaling parameters For scale drift, The clock bias and the target inner product are used to correct the time base inconsistency between the camera and the scale and to resist drift. This represents element-wise multiplication and addition, ensuring the overall cost is differentiable; index and size. Frame rate Number of sampling points for scales All are positive integers;

[0045] Ultimately, it is determined by the coupling matrix. Export the set of indexes mapping frames to scale values, i.e., the alignment index set. It is used for event slicing and subsequent weight allocation.

[0046] In use, by simultaneously estimating the time scaling parameter and coupling, it can still stably provide one-to-one or one-to-many pairings under occlusion or slight frame loss, significantly reducing the magnitude of the loss caused by mismatch; entropy regularization makes the coupling alignment numerically stable and solvable online, satisfying the real-time constraints of edge boxes; the joint cost of activity intensity and mass change rate suppresses irrelevant alignment without action or weight, improving the accuracy of event segmentation.

[0047] To meet compliance requirements, the video side needs to de-identify faces, hands, name tags, readable text, and highly reflective surfaces, while maintaining verifiability with the scale body and RFID / barcode evidence chain.

[0048] Edge boxes construct a detection aggregate map for each frame. (Integrating face detection, hand detection, OCR text boxes, and highlight reflection detection), and employing morphological dilation to protect against edge leakage risks, a content-fidelity blur is then applied within the masked area, which is then stitched together with the non-sensitive area to obtain a de-identified frame. :

[0049]

[0050] Where: original image frame ;No. Frame color image tensor; shape , Example (RGB); Numerical range (Originated from 8-bit image normalization); Unit normalized grayscale / chroma;

[0051] De-identified frames and Same type, de-identified image tensor; shape and Consistency; Domain Directly write back the event evidence chain; detect aggregated graphs. For normalized confidence maps, use a single-channel soft mask with a specific shape. ,domain The S1's detectors for faces, hands, nameplates / text, strong reflections, or uniform badges merge pixels by pixel (e.g., maximum / weighted sum followed by truncation). );

[0052] Morphological dilation operator With structural element radius Expand sensitive areas Take a positive real number, with respect to the radius of the mask as... The expansion; discretely defined as:

[0053]

[0054] Wherein, pixel coordinates The structural element is a disk. The output field remains the same. An anisotropic version is available: using elliptical structural elements. Pointwise multiplication operator Hadamard multiplication; unit tensor and Same type;

[0055] Fuzzy Operator Isotropic nuclear width Smoothing (such as Gaussian kernel) to suppress the risk of reversible reduction. Take positive real numbers. For isotropic Gaussian blurring, each channel undergoes independent convolution:

[0056] Output Index set aligned with scale value Together they form an auditable image index With evidence-collecting slices.

[0057] In use, sensitive areas are morphologically safe and redundant before smoothing, significantly reducing the risk of edge leakage or reverse recovery; non-sensitive areas maintain pixel-level fidelity, ensuring accurate verification of scale changes and RFID / barcodes; mask and smoothing scales can be dynamically configured according to workstation and time period, balancing privacy and readability. This provides an aligned index set for subsequent robust repair and event volume generation. It serves as the core time-series anchor point, while ensuring compliance.

[0058] The original weight sequence of the scale body is affected by temperature drift, micro-vibration, residue adhesion and electromagnetic interference. Direct differential will generate a large number of false peaks when fine-particle materials are added or water is slowly seeping in.

[0059] Therefore, within the aligned weighing domain, the transition structure is first recovered using a joint objective of robust smoothness and total variation. Then, the zero-point offset of each bucket position is estimated using empty bucket segments during the event, ultimately yielding the corrected weight sequence. This provides a reliable baseline for event quality and subsequent allocation, where:

[0060]

[0061] Where: smooth weight sequence : The denoised weighing trajectory of the output; unit: kilogram; obtained from the optimal solution of the following optimization. ; Optimize the variable sequence : Potential smoothed weight to be estimated; unit: kilogram; range: real numbers (optional constraints can be added in engineering). );

[0062] Number of sampling points : The number of points sampled within this time window; a positive integer; the sampling frequency from the scale body and the event window length; the original scale value sequence. : The weighing scale reading collected and time-aligned; unit: kilogram; values ​​are real numbers, which may contain jumps and noise; original weighing sequence For time index Weight observations, in kilograms; smoothed sequences To optimize the variables, the optimal solution is denoted as the smooth weight. ;

[0063] Robust cost function A pseudo-Huber type is used to suppress impulse noise, and its closed-form expression is:

[0064]

[0065] Scale parameters Control the outlier cutoff strength; total variation weight Unnecessary high-frequency fluctuations are suppressed to preserve the steps; all parameters are taken as positive real numbers to ensure that the target is strictly convex and the solution is unique.

[0066] The obtained smooth weight This will be used subsequently for zero-point estimation and event quality calculation. Bucket zeroing will then be performed to align the index set. With barrel location markings Search for suspected empty bucket time periods (e.g., if an empty bucket is identified visually and there is no change in RFID / barcode information during warehousing), and estimate the zero-point offset of the bucket position. Then perform debiasing correction on the entire sequence:

[0067]

[0068] Where: zero-point offset This is the scalar correction value for each bucket location; bucket location number. : The number corresponding to the bin; From station / bucket configuration; bucket mapping Provide index The corresponding bin number; ; Optimization variables; Bucket zero-point offset Bucket position Zero drift of the scale body; unit: kilogram; suspected empty barrel period collection. The number of sampling points was determined by a combination of visual empty bucket confidence, environmental fluctuations, and RFID / barcode status. The number of scale samples within this time window; Corrected weight sequence The result is to remove bias, and the unit is kilogram.

[0069] It should be noted that the first term in the above formula is a robust estimate of the absolute deviation (corresponding to the idea of ​​weighted median), to avoid the mean being dragged down by rare anomalies.

[0070] In use, this integrated processing simultaneously suppresses high-frequency noise and temperature drift, preserving the stepped structure consistent with the delivery action; zero-point estimation based on empty bucket evidence avoids the subjectivity and downtime costs of manual zeroing; the corrected... Maintain quality conservation and cross-bucket consistency even under a large number of concurrent events.

[0071] After obtaining the aligned index set With corrected weight sequence Next, the visual classification results from multiple frames within the event need to be temporally aggregated to output the initial category probability vector. It only reflects visual evidence and will be confused with the correction matrix in step two. With Dirichlet prior parameters Further calibration is required.

[0072] To suppress jitter and redundancy within the same frame, a weighted logarithmic opinion pool is used to fuse the aligned frame set. The weights are determined by both the alignment strength and the quality increment, thus maintaining causal consistency with changes in physical quality.

[0073]

[0074] Where: Initial class probability :event Category Probability; unit is dimensionless; and satisfy Category number Event Intraframe Index Set By alignment index set Determined by the event time window;

[0075] Frame-level visual probability From a lightweight visual classifier in de-identified frames The output is temperature calibrated; aggregate weights With alignment coupling matrix The frame line weight and corresponding weight increment Proportional, and Normalization symbol; proportional normalization symbol Indicates to The dimensional vector is regularized to make .

[0076] The final output is an event record vector. Includes: timestamp (Time of event end, scale value), position (Workstation labeling), bin location (Bottling number), weight (Depend on (Positive transition accumulation within the event window), image index (Frame range and file handle), initial class probability Batch Identification (Analyzed by RFID / barcode).

[0077] In practice, a logarithmic opinion pool is used instead of a simple average to significantly suppress the outlier effects of extreme frames and to increase the weight of keyframes at quality transitions, ensuring that visual evidence is in sync with changes in physical quality; event logging. The field structure and symbol naming are fully aligned with subsequent steps, allowing for seamless transition to the calibration and quantification process in step two; Image index RFID / barcodes enable the chain of evidence to be verifiable and traceable. Aligned continuous observations are compressed into auditable discrete events, forming the minimum sufficient statistics for subsequent inference.

[0078] As a supplementary explanation, the optimal transmission target is given by entropy regularization and edge constraints, and solved online within the CPU / NPU of the edge box using Sinkhorn iteration. Configurable by workstation; morphological expansion operator Using radius Circular structural elements in binarization Performed on; fuzzy operator Isotropic nuclear width Linear smoothness, kernel weights can be pre-generated and cached; robust cost function. The closed-form expression is given and is convex, and each point can be approximated using one-dimensional Newton iteration; the split Bregman / ADMM implementation of the total variation regularization can converge in constant number of iterations within each event window; the zero-point estimation of absolute deviation is equivalent to the weighted median, and the edge box can be solved in linear time using the selection algorithm; the exponential-normalization process of the log opinion pool is given and does not depend on empirical thresholds.

[0079] Step 2: Initialize the class probabilities In the confusion correction matrix With Dirichlet prior parameters Under the common constraints, it is transformed into the posterior class proportion. and with event weight Strictly conserved mapping to categorical food waste quality It also outputs auditable anomaly indicators to update the calibration set and model in a closed loop.

[0080] In a kitchen setting, lightweight visual classifiers often project the observed category as the true category under conditions of strong occlusion, steam reflection, and abrupt changes in local lighting, resulting in a systematic bias related to the category structure. Furthermore, menu changes, time-of-day variations, and customer preferences alter the accessibility probability of categories at the prior level. If these two mechanisms are ignored and directly used... and Multiplication will solidify the model bias into subsequent time series and policy links.

[0081] The class target set records the statistical mapping from the true class to the visual output, and based on this, a column-randomized, strictly positive confusion correction matrix is ​​formed. , of which The column encodes the actual value as At the event level, to avoid the negative components and normalization violations caused by linear pseudo-inverses, the true class proportions are considered as those in the probability simplex. The vector on the vector, and model the observation probability as To this end, a KL projection targeting information geometry is established and a total variation regularization of the category hierarchy diagram is superimposed to solve for the corrected category proportions. :

[0082]

[0083] Where: Total number of categories Confusion correction matrix (Each column is listed in) The sum of the above pairs is 1; the class target set is used for estimation and Laplace smoothing to ensure strict positiveness); initial class probability vector. Event log from step one;

[0084] KL divergence Graph difference operator is used to measure the information difference between the observed distribution and the observed distribution generated from the true scale. It consists of a hierarchy or similarity graph of dish categories, and its first... The row is subjected to first-order difference between adjacent classes; regularization coefficient To suppress unnecessary spikes in adjacent categories in order to maintain interpretability;

[0085] The sum of the absolute values ​​of the elements; optimize variables. The intermediate solution is found within the probabilistic simplex; the unit is dimensionless; in practice, KL-mirror descent can be used: in the dual domain... -Step size update and project back sequentially The edge box can be solved in real time.

[0086] In practice, KL projection on the probability manifold overcomes the numerical instability and physical infeasibility of linear inversion; the total variation constraint of the category graph is introduced to ensure that the correction results conform to the engineering prior of menu structure and the possibility of confusion between adjacent dishes; the resulting corrected proportions... With confusion correction matrix Synchronous evolution can be iteratively updated periodically in step five.

[0087] Visual evidence from a single event is susceptible to chance in small sample sizes, while the prior strength varies significantly across different menus, time periods, and customer groups (collectively referred to as scenarios). To address this, Dirichlet prior parameters are injected without disrupting the simplex structure. Log-geometry blending under Fisher-Rao geometry is used to adjust the corrected scale. with prior average Convex combinations of logarithmic fields:

[0088]

[0089] Where: the prior fusion probability vector :event Category distribution; unit is dimensionless; As the output of S2, it enters the mass conservation law. Number of categories : Positive integer; defined by the menu; scene index function Map events to menu / time period / customer group scenario sets; ;

[0090] Dirichlet Priors For the pseudo-count of categories in the scene, elements are taken as positive real numbers to ensure posterior solvability; log-geodesic mixing coefficient. Automatically set according to the amount of data in the scenario and recent stability (close to 1 when the sample is sufficient, and decrease when the sample is sparse to increase the prior weight); Proportional normalization. Indicates indexed by element Normalization. This mixture is equivalent to approximating the posterior mode of the logarithmic field of the Dirichlet-polynomial model, which is well disclosed and implemented with only vector addition and subtraction and element-wise exponentiation.

[0091] When used, the geodesic method operates within the probability manifold, without introducing negative components or violating normalization; it utilizes scene-adaptive mixing coefficients. Balancing current evidence with steady-state priors significantly mitigates small sample size and short-term drift; the resulting fused prior probability vector Provides a robust and interpretable proportional baseline for subsequent quality allocation.

[0092] Even after demixing and prior injection, the category ratio within an event may still be excessively flat or excessively sharp due to sudden changes in the scene. To couple visual evidence and scale transitions to the same scale, an event consistency index is calculated. It consists of the alignment and coupling matrix from step one. The normalized correlation between the frame line weights and the quality increments within the corresponding time window is given (implemented as a sliding window kernel metric within the edge box). As a supplementary note:

[0093]

[0094] Zero-prevention. Monotonic traction power exponent ;

[0095] The consistency index is calculated by compressing the consistency of a single food waste event across three pieces of evidence—image, weighing, and unmixing probability—into a scalar between 0 and 1. First, the frame set for the event is obtained using frame-weighing alignment. Weights are assigned to each frame based on coupling strength and positive weight increment, and the concentration of weights in a few key frames is observed; higher concentration results in a higher score. Next, the weight changes within the event are examined, and the proportion of positive increments relative to all increments (including positive, negative, and jitter) is calculated; a predominance of increases results in a higher score. Simultaneously, the sharpness (close to a single class) of the unmixed probability is assessed; sharper probabilities result in a higher score. After normalizing each of these three items to the 0-1 range, they are linearly aggregated using three configurable weight sets and scaled overall using the scene confidence level corresponding to the menu, time period, and customer group to obtain the final consistency index. In boundary cases, if there is only one frame, the frame concentration receives full marks; if there is no weight change, the directional consistency receives zero marks. This index adapts to the scene and the strength of the evidence, determining the subsequent power-domain sharpening intensity for the class probability.

[0096] Based on this, the concentration parameters are constructed: ( (For condensed sensitivity), in the logarithmic domain, the probability vector after prior fusion... Perform consistent sharpening and strictly map to category quality according to quality conservation:

[0097]

[0098] Where: the proportion of categories after sharpening , is the normalized result after element-wise exponentiation; the probability after prior fusion. Step 2: The probability obtained from the logarithmic geodesic prior injection; dimensionless. , ;

[0099] Concentration parameters As the consistency metric monotonically increases, it increases the dominant category weight when frame-weight values ​​are strongly consistent, and reverts to a moderate distribution when consistency is weak; event weight Accumulated quality transition after correction from step one; Category of meal loss quality For output, the law of conservation of mass must be satisfied. .

[0100] Both exponentiation and normalization are element-wise differentiable operations, and edge box implementation requires only one logarithmic-exponential and additive normalization operation.

[0101] When using, align the coupling matrix. The provided physical consistency guides visual proportions, avoiding false multimodal assignments caused solely by the camera; power-domain sharpening balances anomaly suppression and principal class enhancement, improving the quality of category-specific food damage. It more closely resembles real-world deployments in strongly consistent scenarios; strict quality conservation ensures seamless interface with inventory and subsequent time-series models.

[0102] To prevent individual anomalous events from contaminating parameter updates, while preserving evidence for retraining, a power-based approach is adopted. - Divergence is used to construct consistent residuals between observations and predictions, and this is used to determine which categories to exclude. First, the normalized class probability vectors are used... (The amount is) Generate predicted observation distribution Then compared with the initial observation Power divergence:

[0103]

[0104] Where: Anomaly measurement For scalars, a larger value indicates a greater deviation between observation and prediction; shape parameter Controlling the heavy-tailed sensitivity of the divergence curve (often taken as...) Increase sensitivity to extreme mismatches); , These are the event index and the category index, respectively.

[0105] Initial class probabilities With predicted observation components Enter element by element; total number of categories Same as before.

[0106] Adaptive thresholds based on pre-built scene Detecting anomalies: When At that time, the event is marked as being removed from the training data but the evidence chain is preserved and archived. and scene tags Write back to the retraining candidate pool for offline recalibration of the confusion correction matrix. With temperature calibration, power divergence at It degenerates into KL divergence.

[0107] When used, power divergence is more sensitive to heavy-tail mismatch and can stably reveal systematic errors (such as bucket occlusion or batch labeling errors) in a few extreme scenarios; the threshold is adaptive according to the scenario to avoid misjudgment of the global threshold during peak hours and off-peak hours; abnormal events are diverted instead of discarded, which not only protects the stability of online parameters, but also provides high-value samples for offline retraining.

[0108] The KL projection objective and constraints are explicitly differentiable; the mirror descent / projection gradient method can solve for the edge box in real time; graph difference operator. The data comes from a preset category graph (nodes represent categories, edges represent similarity or process adjacency), which can be updated by operations and maintenance when the menu changes; the log-geodesic hybrid graph only involves vector addition and element-wise exponentiation;

[0109] Power domain sharpening parameters Consistency Indicators Mapped from this, the consistency metric is implemented as an alignment coupling matrix. Kernel correlation measure of mass increment; power divergence The closed-form expression with differentiability and threshold value have been given. Set offline using the scene quantile method.

[0110] Step 3: Focusing on observational stabilization and physical interpretability, construct a three-layer mapping of observational quantities, state quantities, and risk quantities. This ensures that the categorical meal loss sequence is robustly smoothed over time and coupled with environmental exposure, ultimately outputting a time-period-level categorical meal loss sequence. Batch-level preservation risk score .

[0111] Directly statistically analyzing the event-level category meal loss quality point by point can produce false fluctuations under peak-hour load, short-term occlusion, and inventory anomalies, thus misleading change point detection and strategy triggering. Therefore, this step uniformly records the event-level category meal loss quality output from step two as an observation vector within this step. And introduce state vector Characterizes the quality of the smoothed category of food loss under the same time grid.

[0112] In a real-world kitchen setting, different categories of food waste exhibit gradual changes over adjacent time periods, while categories with similar raw materials or processing methods show similar dynamics over time. To balance temporal smoothness and category structure, a convex quadratic objective is used to govern the observation fitting dynamic consistency cross-class coupling triple constraint; the observation weights follow the anomaly measurement in step two. The scene confidence vector is adaptively adjusted to achieve an adaptive trade-off between peak-hour anomalies and off-peak stability, where:

[0113]

[0114] Where: number of time steps Represents the number of discrete moments within a statistical period; state vector For a moment The category of meal loss is estimated as a non-negative real number in kilograms; observation vector. The amount of events aggregated in step two (in the grid) (cumulative summation), unit: kilogram; observation weight matrix The state transition matrix is ​​a diagonal positive definite matrix used to suppress the influence of outlier observations on the fitting term. To characterize the inertia or gradual change between adjacent time points, a diagonally dominant approximate identity matrix can be used to maintain a conservation trend; process noise covariance It is a symmetric positive definite matrix, which controls the smoothing intensity of the dynamic terms;

[0115] Category graph Laplace matrix From the similarity matrix degree matrix Composition, positive semidefinite, used to penalize unnecessary divergence in similar categories of raw materials / processes; cross-class coupling weights. The value is a real number, used to adjust the coupling strength.

[0116] The above objective is a strictly convex optimization with a unique solution, which can be achieved using striped Cholesky gradients or conjugate gradients within a marginal box; the operator form is fully disclosed, satisfying the feasibility requirement. To ensure the observation weight matrix... As the anomaly metric adapts to the scene, its diagonal term is defined as the product of exponential decay and quality percentage:

[0117]

[0118] Among them: anomaly measurement Power divergence output from step two; attenuation coefficient For real numbers, The larger the value, the more sensitive it is to anomalies; Scene confidence vector Prior confidence levels at the menu / time / customer group level, with elements located in the range of positive real numbers and calibrated offline; quality percentage vector. Pick , where small constant To prevent the denominator from being zero; element-wise Hadamard multiplication is denoted as Diagonalization operator Map the vector to a diagonal matrix.

[0119] In practice, this objective function incorporates cross-class similarity of observation confidence time dynamics within an integrated framework, thereby maintaining sequence stability even under anomalous pulses during the evening peak; through the observation weight matrix The exponential decay of the anomalous events means that the impact of anomalous events is softly suppressed rather than hard deleted, ensuring the continuity of the evidence chain; the category graph Laplace synchronizes the meal loss elasticity of similar dishes, which helps in the structural explanation of subsequent root cause backtracking.

[0120] Changes in the kitchen's inventory and environmental disturbances are the core external factors causing short-term fluctuations in food waste. Without explicit driving modeling, the smoother will be forced to operate with a large process noise covariance. External factors can be absorbed, leading to over-smoothing. To address this, exogenous stimuli are explicitly added to the process model, making state evolution sensitive to inventory and environment without overfitting.

[0121]

[0122] Where: state vector With state transition matrix Same as before; Inventory incentive vector To aggregate the net outbound quantity by category (inbound is recorded as negative, damage is recorded as positive, etc., in kilograms), the quantity is aligned with the inventory system via RFID / barcode in step one and then displayed in the grid. Accumulation; Environmental feature vector The temperature, humidity, and ethylene levels at the workstation / cold storage are transformed by taking quantiles / means / peak values ​​within the grid and then concatenated. Dimensions... Inventory Coupling Matrix It is a sparse, approximately diagonal matrix, allowing for a small amount of cross-level coupling within categories sharing raw materials; the environment coupling matrix... Mapping environmental exposure to sensitivity to incremental food loss; process noise and Same as before. Discretized exogenous inputs are linearly injected through a fixed design matrix, and parameters can be obtained from historical maximum likelihood or Bayesian ridge regression.

[0123] When used, the explicit inventory incentive vector Environmental feature vectors This enables the model to respond quickly and directionally when replenishment, returns, temporary chain disruptions, and cold storage fluctuations occur, reducing lag caused by unmodeled external factors; sparsification of the parameter matrix suppresses overfitting, making the state sensitive to key drivers and robust to noise.

[0124] The loss of freshness of perishable food is dominated by the coupled acceleration effect of temperature, humidity and ethylene concentration. In order to project the continuous exposure trajectory into batch risk, the exponential acceleration hazard rate is used as the core. The three exposures are incorporated into the instantaneous hazard through physical inspiration transformation, and then the risk is obtained by integration in the batch time domain.

[0125]

[0126] Among them: instantaneous hazard rate For batch At any moment Freshness loss intensity; baseline hazard rate The hazard per unit time under reference conditions; acceleration factor. These are learnable parameters;

[0127] Temperature transformation The reciprocal scale representing the Eyring-like reaction rate, with a reference temperature. (Celsius) constant; humidity changing Capturing the logarithmic effect of water activity on microbial growth, with reference humidity. Percentage; Ethylene transformation Characterizing half-power acceleration under diffusion-limited conditions, with reference concentration Batch temperature ,humidity ethylene Obtained by mapping the environment and RFID path in step one;

[0128] Once the instantaneous hazard is known, the batch cumulative risk is defined as the complement of the survival function:

[0129]

[0130] Among them: preservation risk score For batch A measure of freshness loss probability; batch time domain The continuous interval set formed by arrival time → storage time → outbound use / disposal; the integral is achieved in the edge box by piecewise constant / linear interpolation Riemann summation, and the step size is adaptively selected by the sensor sampling frequency and RFID event boundary.

[0131] In practice, the triple transformation of the hazard rate avoids the scale inconsistency problem caused by direct linear superposition; the inverse temperature difference is particularly sensitive to high temperatures; the logarithmic transformation of humidity suppresses the percentage boundary effect; and the square root transformation of ethylene fits the diffusion-limited mechanism. The integral obtained... Monotonic growth under prolonged residence and fluctuating exposure faithfully reflects the combined effect of time and intensity. Furthermore, batches with high risk but low quality may not be the highest priority, while batches with high quality but moderate risk should not be ignored. Therefore, a risk-scale-demand composite priority is designed to ensure that the ordering of FIFO and emergency response adheres to food safety principles while minimizing costs.

[0132]

[0133] Among them: priority For batch The urgency of execution; Sigmoid mapping Linear composites are converted into comparable scores; risk scoring From the previous formula; demand pressure The POS short-term forecast and table rollout are mapped to a near-term demand / gap indicator for the category to which the batch belongs; category mapping The batch will be categorized into its dish type; time window For strategy assessment windows (e.g., the most recent shift); time-based category meal loss quality ; scale constant For reference quality; weight Historical operational KPIs (meal wastage rate) can be used to determine the meal wastage rate. Unit meal loss cost The multi-objective calibration was obtained.

[0134] When used, priority is synthesized by harmonizing security risk, inventory level, and demand pressure, avoiding extreme decisions based on a single indicator; the Sigmoid function suppresses the explosive impact of abnormally large values, making the sorting smoother and more executable; the output... Directly connect to the work order generation in step five (FIFO / emergency handling / quantity adjustment).

[0135] As a supplementary explanation, the category diagram Laplace It can be constructed based on the similarity of raw material components / historical confusion rate: Let the similarity be... ,in A symmetric measure of compositional differences or historical confusion For bandwidth; based on this, and Exogenous matrix It can be obtained through ridge regression or expectation maximization over a historical period, with process noise. It can be estimated by residual variance decomposition; the hazard integral is approximated by discrete summation. It is implemented with constant complexity within the edge box; the synthetic mapping form is explicit, and the weights can be obtained through Bayesian optimization or multi-objective grid search.

[0136] Step 4: Taking risk perception change point detection as the entry point and information gain-driven causal backtracking as the main line, the time-period categorized meal loss sequence is analyzed. Unified modeling is performed with exogenous drivers, batch risks, and operational tags to ultimately output anomaly peak sets and Top List of causes and a list of actionable corrective measures.

[0137] If anomalies are identified solely by thresholds or single statistics, peak hour surges or replenishment operations may be misreported as a surge in waste; while if the causes are traced back based solely on empirical rules, it is difficult to guarantee consistency across stores and time periods.

[0138] In a real kitchen, inventory inflows and outflows, along with environmental fluctuations, alter the prior frequency of mutations. To avoid false alarms triggered by fixed priors, the risk of change points is adaptively modeled based on exogenous inputs, and a divergence-based evidence term replaces the fragile density assumption, forming an online posterior update for the runtime. This allows for robust change point localization without sacrificing real-time performance, as detailed below:

[0139]

[0140] Among them: running length posterior , indicating time The length of the segment since the last change point is The posterior probability;

[0141] Danger level Given by the logical mapping, take , where logical functions For Sigmoid, parameters Vector parameters , Scalar parameters The learnable coefficient;

[0142] Normalized exogenous inventory vector With normalized environment vector From step three Linear scaling; anomaly measurement From step two;

[0143] Evidence Items Smooth the mass vector Deviation from the baseline prediction within the segment is converted into likelihood weights: take ; where category weight The normalized allocation of batch risk in step three across categories and time periods is combined with the confidence decay in step two.

[0144] Intra-segment baseline prediction From the dynamics of step three In length The most recent historical revaluation yielded the Bregman divergence. Selecting the potential function To fit the logarithmic elasticity of nonnegative mass; initial prior The rest are 0;

[0145] Proportion symbol It indicates that subsequently in Normalization makes the sum equal to 1. When the project is implemented, it is truncated within the edge box. The online recursive calculation is computationally complex, but its computational complexity is limited by the maximum segment length threshold.

[0146] When in use, the risk level adapts to exogenous inputs and anomaly metrics, preventing platform replenishment / chain breakage / environmental transitions from being misjudged as wasteful change points; the evidence terms are expressed in divergence, which is weakly coupled with the distribution hypothesis, resulting in numerical stability and robustness to heavy-tailed noise; the posterior a posteriori of the runtime directly produces the change point confidence level, facilitating integration with policy thresholds, and ensuring low and reliable detection latency.

[0147] To assign anomalies at the category level and suppress noise, a generalized CUSUM with risk weights is constructed to track the persistent deviation of actual quality from baseline predictions. A forgetting coefficient is used to control the memory length, thereby providing segment-level change points in the runtime posterior while simultaneously locating category-level anomaly contributions. Specifically:

[0148]

[0149] Among them: cumulative statistics Indicate category At any moment abnormal cumulative intensity, Corresponding to the previous moment;

[0150] Forgetting coefficient To control the decay of historical contributions, memories are longer when the value is close to 1; class weights This is used to emphasize observations with high batch risk and high reliability; Bregman divergence With potential function Same as before; baseline prediction The output is directly derived from the dynamics of step three at the current moment; drift compensation. This is a category threshold term, which can be set according to the historical steady-state divergence quantile.

[0151] Furthermore, upon reaching the threshold This is recorded as a class anomaly and compared with the posterior probability of the change point in the runtime posterior. Commonly triggered peak time The determination.

[0152] When used, compared to single-point deviation, CUSUM characterizes structural anomalies with the intensity of cumulative deviation and is not sensitive to short-term pulses; after introducing risk weights, categories with high preservation risk and high confidence are more easily highlighted, forming a unified scale of detection, risk and category; the linkage triggering with the posterior of the running length reduces the early and late reporting rates.

[0153] Correlation coefficients alone cannot protect against collinearity and confounding. To form an explanation with quasi-causal semantics within the detection window, the peak time needs to be considered. Construct a before-and-after comparison window The improvement of the baseline-extended baseline under the Bregman quasi-likelihood is used as the label causal contribution. For each candidate label feature... (Covering POS sales volume, customer traffic intensity, recipe switching instructions, team dummy variables, and table layout coding), in the extended baseline, using monotonic coefficients Inject it into the category Baseline prediction, then calculate contrast information gain:

[0154]

[0155] Among them: information gain The larger the value, the higher the label value. Category The stronger the explanatory power of the anomaly; the higher the category quality. Compared with baseline prediction From step three;

[0156] Tag features The information system and on-site coding are uniformly sampled and time-aligned; monotonic coefficient By using convex optimization within the window (minimizing the sum of subsequent terms), we can ensure that label enhancement does not overfit and that the direction is consistent; weights To mitigate the leakage of explanations related to multicollinearity, the historical collinearity of the labels and window stability are scaled.

[0157] Among them, information gain By accumulating the SKUs according to the recipe mapping, we obtain the causal contribution score for each product. Then, we summarize the contribution scores and category size to produce the Top. A list of high-waste items and their corresponding causes; this metric relies only on explicit functions and the convexity problem within a window.

[0158] When used, the Bregman information gain of the baseline-extended baseline is used instead of the correlation coefficient, which naturally resists the effects of scale and skewness; the monotonic constraint stabilizes the interpretation direction and avoids reverse attribution; the confounding penalty limits the inflated contribution of collinear labels, and the output causal list is more robust when migrating across stores.

[0159] Explanation does not equate to actionability. In the action library, candidate actions are bound to each cause (such as table setup adjustments, replenishment rhythm adjustments, portion reduction, rapid first-in-first-out, cutting / pickled / donation / damage reporting), and their expected benefits are evaluated using a multi-objective indicator of waste rate-cost. The minimum action set is then solved under the constraints of team and warehouse capacity and budget.

[0160]

[0161] Among them: action set For optimal selection; action library Stores executable actions; action elements An index for a specific intervention command that is a candidate and can be executed; a set of actions. At the level of optimizing variables, it refers to the subset of actions to be performed; waste rate. Unit meal loss cost For platform KPIs; expected improvement amount , Through contribution score-effect mapping and preservation risk Priority Joint inference; trade-off coefficients Calibration based on historical strategy replay; cost With budget Financially feasible;

[0162] Human load vector With upper limit of production capacity Depict the multi-dimensional resources of the work team (duration, workstations, cold storage occupancy, etc.). This indicates that each component does not exceed a certain value. This optimization is an extension of the 0-1 knapsack problem, and can be achieved using a greedy approximation or Lagrange relaxation to provide a suboptimal solution within the edge box.

[0163] When in use, action selection is completed in a multi-objective, cross-constraint quantitative framework to avoid unilateral decision-making that only considers risks and not costs; causal contribution and batch priority are combined into the model to match the order of use / reduction / disposal with actual execution capabilities; the output minimum action set naturally connects to the work order arrangement and feedback collection in step five.

[0164] Given the explicit recursion of the posterior of the run length and the logical mapping of the risk, the evidence terms are explicitly constructed using Bregman divergence, which can be implemented in edge boxes with truncated run length and exponential weighting; CUSUM only requires constant-level state storage for online updates, and the threshold and drift terms can be determined by historical quantiles; the monotonic coefficients are obtained using windowed convex optimization, and the information gain is presented in explicit difference form, all of which can be solved in real time on CPU / NPU;

[0165] Step 5: Construct a unified closed loop encompassing strategy generation, work order scheduling, execution feedback, and parameter self-calibration. This ensures consistent coupling of anomalies, risks, and scale within the decision domain, and quantifies the execution effects to write back to the model parameters, driving the system's self-stable evolution in real-world kitchen scenarios. If fixed rules are triggered solely by a list of causes and risk ranking, it will lead to excessive or insufficient intervention during cross-time and cross-store migrations.

[0166] In the high-concurrency environment of lunch and evening peak hours, the impact of portion size adjustment (smaller / half-portion), replenishment frequency, and batch processing path on food waste is interdependent: reducing portion size changes the required replenishment frequency, and the choice of processing path, in turn, affects the available scale. To avoid fragmented decision-making, these three factors are solved within the same solvable framework with a unified objective.

[0167] Define decision quantity: weight scale (Ratio relative to standard portion), beat of the counter-attack (Time interval between two replenishment sessions), Batch processing weight vector (In first-in-first-out, cutting and preparation, marinating, donation, and loss reporting, etc.) Based on the probability or share along each path, and with risk, scale, and service as multiple objectives, the following strategy optimization objectives are established (the larger the objective, the better):

[0168]

[0169]

[0170] Where: strategic objective Time index set For the current scheduling window discrete time; category: meal loss quality The smooth output from step three;

[0171] portion size This represents the proportion relative to the standard serving size; a smaller value indicates a smaller serving size. (Serving sensitivity coefficient) Flexibility in controlling portion size reduction and minimizing food waste; replenishment rhythm Reference beat Based on experience; beat index To suppress excessively frequent platform replenishment; observation weight It is obtained by weighting the information gain from step four and the exogenous strength from step three;

[0172] Batch set , for time In-stock batch index; priority From step three; processing the weight vector for Simplex assignment of path probabilities; payoff vector Components are defined as ,in For batch preservation risk scoring, For path Unit risk-return coefficient; cost vector Cost per unit of path execution.

[0173] The constraints include: the resource-workstation-human capacity of each type and time period shall not exceed the upper limit; the quantity and cycle time shall be within the process-feasible range; and the batch processing weight shall be normalized moment by moment.

[0174] When used, the unified objective groups the revenue from portion reduction, the revenue from replenishment rhythm, and the net revenue from disposal onto the same axis, avoiding local optima caused by single-item optimization; the weighted terms give higher decision-making attention to high-risk, high-scale categories and batches; the logarithmic and exponential terms are more sensitive at small scales, matching the amplifying effect of peak-period fine-tuning on meal loss, thereby steadily reducing total meal loss and disposal costs.

[0175] The strategy variables need to be implemented as specific work orders in the warehouse, food preparation, hot kitchen, cold dishes, and front-of-house operations, and must satisfy multi-dimensional capacity constraints and due date priorities within the time window. Therefore, a soft-maximization penalty function is used to characterize the cost of lateness, and a logarithmic-exponential smoothed upper bound is used to approximate the maximum value, achieving differentiability optimization within the edge box.

[0176]

[0177] Among them: scheduling objectives Smaller is better; action set It includes three types of tasks: adjusting the solved quantities, replenishing the platform, and handling the problem; batch mapping. Work order Associated with its batch; priority Same as before;

[0178] Start time With expiration time Derived from upstream peak time and cold chain window; smoothing coefficient Controlling the steepness of the soft lateness curve; number of capacity dimensions Work orders in terms of capacity resource consumption (Including manpower minutes, workstation occupancy, cold storage time slots, etc.); Upper limit of production capacity Penalty weight .

[0179] Given the feasible time window and processing sequence constraints of a work order, near-optimal scheduling can be obtained online using greedy-local exchange or coordinate descent.

[0180] When used, the late cost of soft maximization avoids the non-differentiability and scheduling oscillations caused by hard thresholds, and gives higher weight to urgent batches; multi-dimensional capacity penalty enables work orders to automatically stagger peak times between cold / hot / cutting and packing stages, improving execution feasibility; after being linked with technical point A, the strategy naturally transitions from variable values ​​to task time, reducing implementation deviations.

[0181] The executed image + weight + batch evidence chain and randomized annotation provide new samples of the true category—visual observation. Meanwhile, adjustments to weight and tempo alter the prior strength distribution for different scenarios. To avoid the lag caused by offline one-time recalibration, a sliding window Bayesian-exponential smoothing update strategy is adopted.

[0182] Firstly, regarding the confusion correction matrix... Column (real category is) Perform the following convex component update:

[0183]

[0184] Where: column conditional distribution vector The updated real class is A probability column that is visually classified into various classes; dimensionless; its components are non-negative and sum to 1; it can be used as a matrix. The Column replacement write-back; old column conditional distribution vector : Existing quantity before update; dimensionless; same as above;

[0185] Confusion correction matrix column vector Normalize to 1; Update step size ; Counting vector The frequency of observations within the window that have been randomly checked and confirmed to fall into each observation category; Laplace smoothing. Preventing zero counting; unit vector Same type.

[0186] Secondly, indexing scene menus / time periods / customer groups. Dirichlet prior parameters are subjected to weighted exponential smoothing:

[0187]

[0188] Where: Dirichlet prior pseudo-counting vector Scene Updated prior; component The unit is dimensionless (equivalent sample size); dimension Number of categories; old pseudo-count vector Prior information before the update; same dimension and unit as above; source is the stock from the previous period;

[0189] Prior parameters Smoothing coefficient ;Category Cumulative Quantity From the time period level within the window It is aggregated according to scenario, and the components are: Priority aggregation Batch priority Mapping by category of affiliation Import category dimension; weight Prior reinforcement driven by regulatory strategies.

[0190] When using, The convex component update rapidly incorporates the latest sampling results while maintaining column normalization and non-negativity, avoiding numerical instability; for prior parameters... Exponential smoothing allows gradual changes in menu or customer structure to enter the inference layer in a continuous manner, reducing overcorrection caused by prior mutations; both share the same source of evidence chain, improving consistency and accountability.

[0191] As work orders are executed, the statistical properties of the observation residuals change with shift load and scene disturbances. To maintain the stability of the fusion unit and detector, robust covariance estimation with Cauchy weights is adopted, and the anomaly metric and CUSUM threshold are adaptively updated using the quantile threshold method. Firstly, scene indexing... Observation noise covariance In the window Updated to :

[0192]

[0193] Wherein: observation noise covariance Symmetric positive definite; observation vector With state vector From step three; weight Cauchy weights are used to suppress heavy-tailed residuals; robust scaling. The robust multiple of the absolute deviation of the historical median is used; Euclidean norm. The denominator is the sum of weights.

[0194] Secondly, based on the latest residuals and statistics, update the outlier measures and CUSUM threshold to obtain... :

[0195]

[0196] Among them: anomaly detection threshold Power divergence used in step two Traffic splitting; CUSUM threshold For step four Quantitative functions Take probability Empirical quantiles, often used as upper quantiles to constrain false alarm rates; sets For the scene window.

[0197] When used, Cauchy weighted covariance remains convergent under sudden peaks and heavy-tailed residuals, preventing excessive smoothing or oversensitivity caused by unstable observations; quantile thresholds make the false alarm rate and false negative rate controllable and reproducible across different scenarios, and the thresholds are automatically adjusted according to load and season; the two work together to maintain the long-term stability of the fusion-detection-strategy chain.

[0198] By transforming traceable execution evidence into learnable parameter updates, the feedback channel from strategy to model is closed, enabling system parameters to oscillate in sync with on-site dynamics during continuous operation rather than drifting with lag.

[0199] Here, a clear objective function and feasible region (weight interval, cycle interval, disposal weight simplex and multidimensional capacity upper bound) are given. The objective term adopts an explicit form of exponential and power type, which can be approximately solved in the edge box by coordinate descent or sequential quadratic programming. Update to convex combination, column normalization ensures interpretability. The exponential smoothing is a linear closed form; the covariance update and threshold adaptation are given specific expressions for the weight function and quantile, respectively.

[0200] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0201] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0202] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0203] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0204] 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-sensor fusion real-time food waste monitoring system, characterized in that: include, Video frames and weighing values ​​are acquired at the edge box. Within a sliding window, an entropy-normalized optimal transmission-affine time-scaling joint objective is constructed for the video frame time series and weighing value time series. The alignment coupling matrix and time-scaling parameters are obtained, and the frame-weighing pairing index set is obtained by iteratively solving according to edge constraints and Sinkhorn. The visual classification probabilities of the aligned frame set are fused by logarithmic opinion pooling. The fusion weight is obtained by normalizing the alignment coupling strength and the quality increment within the corresponding time window. The alignment coupling strength is the coupling weight between the aligned frame and the corresponding weighing value time in the alignment coupling matrix. Generate initial category probabilities at the event level and combine them with timestamp, station, bucket, event weight, and image index to form an event record vector. The event weight is the weight accumulated from the positive transitions of the corrected weight sequence within the event time window. The initial class probability of the event record vector is used to make a KL projection overlay diagram of the total variation according to the confusion correction matrix determined by the class, and then mixed with the Dirichlet prior according to the logarithmic geodesy, and combined with the correction scale value according to the mass conservation to map the class meal loss mass. Using a convex quadratic target estimation state, observation fitting, dynamic smoothing, and Laplace coupling are performed on the food spoilage quality. Inventory incentives and temperature, humidity, and ethylene characteristics are introduced to construct a hazard rate and integrate it to obtain smoothed food spoilage and batch risk. Change point detection is performed using the posterior of the running length to map exogenous inventory and environmental characteristics into hazard levels. The evidence is updated with Bregman divergence, and CUSUM is constructed to locate anomalies. Information gain is calculated in the peak window according to the monotonic coefficients of the baseline-extended baseline. Using weight scale, replenishment rhythm and batch processing weight as decision variables, a unified objective is constructed and solved and scheduled under the upper bound constraint of resources; the evidence chain self-calibration is performed: the column vector of the confusion correction matrix is ​​updated by convex combination, and the Dirichlet prior is updated exponentially.

2. The multi-sensor fusion real-time food waste monitoring system according to claim 1, characterized in that: Within the aligned weighing domain, the step structure is recovered by denoising with a pseudo-Huber robust term and total variation. The zero-point offset of the bucket position is estimated by retrieving the empty bucket period based on the frame-scale pairing index set. The original weight sequence of the scale body is debiased to obtain the corrected weight sequence, which is then bound to the station configuration for storage.

3. The multi-sensor fusion real-time food waste monitoring system according to claim 2, characterized in that: For each frame, a detection aggregation map is constructed at the edge box and a security mask is formed by morphological dilation. Isotropic blurring is applied within the mask area to de-identify the data while preserving the original stitching of non-sensitive areas. This data is then written into the event evidence chain and associated with the frame-scale pairing index set.

4. The multi-sensor fusion real-time food waste monitoring system according to claim 3, characterized in that: The initial class probability is mapped to the observation distribution using the confusion correction matrix, and the total variation constraint of the KL projection superimposed on the class map is constructed in the probability simplex to solve for the corrected class proportion. Each column of the confusion correction matrix is ​​Laplace smoothed to be strictly positive and the column sum is one.

5. The multi-sensor fusion real-time food waste monitoring system according to claim 4, characterized in that: The Dirichlet prior parameters of the scene index are mixed with the corrected class proportions in Fisher-Rao information geometry using log-geodesic mixing. The mixing coefficient is adaptively set according to the scene sample size to obtain the prior fused probability vector, which is then used for subsequent mass conservation mapping.

6. The multi-sensor fusion real-time food waste monitoring system according to claim 5, characterized in that: Based on the alignment coupling and mass increment, a consistency index is constructed, and a condensation parameter for power domain sharpening is set accordingly. Element-wise exponentiation and normalization are performed on the prior fused probability vector. Mass conservation mapping is strictly performed according to the event weight, and anomaly measurement is calculated using power divergence, and anomaly shunting threshold is set accordingly.

7. A multi-sensor fusion real-time food waste monitoring system according to claim 6, characterized in that: The convex quadratic target estimation state is constructed on a unified time grid by observation fitting, dynamic smoothing, and class graph Laplacian coupling. The observation weight matrix is ​​generated by the anomaly metric and scene confidence vector, is diagonally positive definite and adaptively updated over time, and the solution is obtained by banded Cholesky or conjugate gradient.

8. The multi-sensor fusion real-time food waste monitoring system according to claim 7, characterized in that: An exogenous input model is introduced. The inventory incentive vector is obtained by aligning RFID with inbound and outbound data. The environmental feature vector is composed of temperature, humidity and ethylene concentration after inverse temperature, logarithmic and power transformations. The state evolution is linearly injected through the inventory coupling matrix and the environmental coupling matrix respectively, and sparsification is used to suppress overfitting.

9. A multi-sensor fusion real-time food waste monitoring system according to claim 8, characterized in that: The temperature, humidity, and ethylene characteristics are combined into an instantaneous hazard using an exponentially accelerating hazard rate, and the batch risk score is obtained by integrating in the batch time domain. The batch risk, category demand pressure, and category size are then combined using a sigmoid function to form a priority for subsequent scheduling and handling.

10. A multi-sensor fusion real-time food waste monitoring system according to claim 9, characterized in that: Online change point detection is performed using a runtime posterior framework. The risk level is set as a function of exogenous inventory, environmental features and the anomaly measure through logical mapping. The evidence item uses the Bregman divergence predicted relative to the intra-segment baseline and is updated in real time within the edge box using truncated recursion.

11. A multi-sensor fusion real-time food waste monitoring system according to claim 10, characterized in that: A unified objective is constructed using weight scale, replenishment rhythm, and batch processing weight as decision variables and is solved under the constraints of resource and capacity upper bounds. At the same time, work order scheduling is completed by soft maximization of lateness penalty. After execution, the evidence chain is used for the convex combination update of the column vectors of the confusion correction matrix and the exponential smoothing of the Dirichlet prior.

Images