Corrugated carton production management system and method

Abstract

The invention discloses a corrugated carton production management system and method, and relates to the technical field of corrugated carton production, the system comprises a production scheduling module, a material tracing module, an equipment cooperation module, a multi-mode detection module and a cost control module, and solves the problem that supply and demand relationships, production efficiency and quality management and control are not coordinated; according to the technical key points, an order ID is coded into an integer chromosome, sequential crossover and reverse mutation operation are applied, a fuzzy comprehensive evaluation model is constructed by combining Hash chain table and local hill climbing method optimization, and factors are quantified through a trapezoidal membership function, so that the order delivery punctuality rate and the equipment utilization rate are improved, and the mold replacement cost is reduced; the real-time production cost is collected and calculated through Kafka and Flink, an operation cost method model is constructed to maintain a standard cost database, abnormity is detected through a CEP engine, the cost hyper-branched risk is predicted through an LSTM-Transform hybrid neural network, a thermodynamic diagram is generated through a cache result to assist decision making, the cost accounting delay time is shortened, and the hyper-branched processing efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of corrugated paper box production, and in particular to a production management system and method for corrugated paper boxes. Background Art

[0002] As an important product in the packaging industry, the production management of corrugated boxes involves many complex links, including order scheduling, material control, equipment scheduling, quality inspection, and cost accounting. Traditional production management systems have the following shortcomings:

[0003] 1. The production scheduling algorithm has difficulty balancing multiple optimization objectives such as order delivery time, equipment utilization, and mold change costs;

[0004] 2. Data traceability throughout the material lifecycle relies on a centralized database, which poses a risk of data tampering and reduces supply chain collaboration efficiency.

[0005] 3. Equipment status monitoring and coordinated control of energy consumption and production capacity lack real-time and accuracy;

[0006] 4. Single-modality quality inspection technology has problems such as high missed detection rate and poor adaptability;

[0007] 5. Cost accounting is lagging behind and the ability to predict overspending risks is insufficient.

[0008] Currently, a Chinese patent application with the patent number "CN202410047133.8" discloses a production management system and method for colored corrugated boxes. The system involves mixing, coating, and drying pulp and pigment according to a preset formula to produce colored corrugated cardboard. The colored corrugated cardboard is then cut, folded, and bonded together to form colored corrugated boxes. The colored corrugated boxes undergo quality inspections. Upon confirming that the colored corrugated boxes meet predetermined quality standards, the colored corrugated boxes are packaged. Furthermore, production data for the colored corrugated boxes is collected, processed, stored, analyzed, and displayed. While this system can control the production process and ensure the stability and consistency of product quality, it lacks adaptive adjustment capabilities, leading to frequent imbalances between supply and demand, production efficiency, and quality control.

[0009] To address the above problems, the present invention provides an intelligent production management system and method that integrates multiple technology modules. By integrating technologies such as genetic algorithms, blockchain, digital twins, multimodal detection and deep learning, it achieves efficient collaboration and precise control of the entire corrugated cardboard production process. Summary of the Invention

[0010] (1) Technical problems solved

[0011] In view of the deficiencies in the prior art, the present invention provides a production management system and method for corrugated paper boxes, which solve the problems raised in the background art.

[0012] (2) Technical solution

[0013] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0014] A production management system for corrugated paper boxes, the system comprising:

[0015] Production Scheduling Module: Based on order information, material inventory, and equipment capacity, a hybrid optimization algorithm combining genetic algorithms and taboo search is used to generate production schedules. The hybrid optimization algorithm uses chromosome encoding, sequential crossover (OX) operations, and a fuzzy comprehensive evaluation model to perform multi-objective adjustments.

[0016] Material traceability module: This module uses the Hyperledger Fabric consortium chain architecture to record the entire material life cycle data, including procurement batches, production process parameters, and quality inspection results, and uses zero-knowledge proof for supply chain traceability.

[0017] Equipment collaboration module: Builds a 3D dynamic model based on point cloud data, maps physical equipment status in real time through the OPC UA protocol, and combines the particle swarm optimization algorithm to coordinate energy consumption and production capacity control;

[0018] Multimodal inspection module: Integrates machine vision, spectral analysis, and laser ranging technology, and uses an improved YOLOv5 model for defect detection;

[0019] Cost Control Module: Based on the Apache Flink stream computing framework, it updates production costs through the activity-based costing (ABC) model and predicts overspending risks based on the updated production costs.

[0020] Furthermore, the hybrid optimization algorithm sets a dynamic weight adjustment mechanism, which automatically adjusts the weight coefficients of each objective function according to changes in the production environment through an adaptive fuzzy controller. The construction of the hybrid optimization algorithm includes the following steps:

[0021] S101, encode the production order sequence into an integer chromosome, where each gene bit represents an order ID;

[0022] S102, using sequential crossover (OX) and reverse mutation operations to extract gene fragments;

[0023] S103. Record the solutions of the most recent N searches, remove duplicate paths, and use a hash list structure to store the solution’s feature vectors. The retrieval time complexity is O(1).

[0024] S104. Based on the population generated by the genetic algorithm, candidate solutions are generated by exchange and insertion, and a local hill climbing method is introduced to iterate the candidate solutions for 5-10 times;

[0025] S105. Construct a fuzzy comprehensive evaluation model based on multiple objectives, where the multiple objectives include order delivery time, equipment utilization rate, and mold change cost; terminate the algorithm when there is no change in the optimal solution of consecutive K generations, and output the optimal production schedule.

[0026] Furthermore, the fuzzy comprehensive evaluation model in the evaluation function includes:

[0027] Determine the evaluation factor set: {on-time delivery rate, equipment utilization rate, mold change cost};

[0028] Constructing the fuzzy membership matrix: using trapezoidal membership function to quantify each factor;

[0029] Calculate the comprehensive evaluation value: synthesize the weights of each factor through the weighted average operator.

[0030] Furthermore, the construction process of the three-dimensional dynamic model includes:

[0031] S201. Obtain equipment geometric parameters through laser scanning and deploy sensors to collect real-time operation data of production equipment;

[0032] S202. Based on the acquired device geometric parameters, a simplified model of the device is generated using a triangular mesh simplification algorithm. The model accuracy level is switched according to the user interaction distance using level of detail (LOD) technology, with a maximum support of 5 levels of detail.

[0033] S203. Establish a real-time communication channel between the physical device and the virtual model through the OPC UA protocol, with a data update frequency of ≥100 Hz, and set a heartbeat detection mechanism. When the communication interruption exceeds 500ms, it automatically switches to the local cache data to continue operation;

[0034] S204, introducing a Kalman filter algorithm to eliminate sensor noise and align the virtual model with the physical device;

[0035] S205. Set up a model version management mechanism to automatically trigger the model reconstruction process when the device structure changes.

[0036] Furthermore, the three-dimensional dynamic model also includes a multi-physics field coupling analysis module. The multi-physics field coupling analysis module uses spatiotemporal data cube technology to integrate equipment operation data, environmental parameters and product process parameters to construct a four-dimensional dynamic simulation model and perform multi-physics field coupling analysis. The multi-physics field coupling analysis module includes:

[0037] Physical field modeling: Construct temperature fields, stress fields, and flow fields, and set corresponding abnormal warning thresholds. When the parameters in the physical field exceed the set abnormal warning thresholds, an alarm will be issued;

[0038] Data mapping mechanism: This mechanism uses the OPC UA protocol to acquire real-time device usage data, and uses Kriging interpolation to map discrete sensor data into continuous field variables. Device usage data includes device vibration acceleration and motor current.

[0039] Coupling solution: Use weak coupling method to iteratively calculate the multi-physics field of temperature-stress-fluid;

[0040] Real-time optimization control: Adjust process parameters based on the multi-objective particle swarm optimization (MOPSO) algorithm and output optimized equipment operating parameters.

[0041] Furthermore, in the multimodal quality detection module, machine vision uses a line scan camera to obtain the surface image of the carton; spectral analysis detects the ink composition by deploying a near-infrared spectrometer; and laser ranging uses a triangulated laser sensor to measure the size of the carton.

[0042] Furthermore, the fusion method of the multimodal quality detection module includes:

[0043] Feature fusion layer: Feature pyramid network (FPN) is used to perform cross-modal feature fusion of image texture features (HOG), spectral absorption peak features, and size parameters;

[0044] Decision fusion layer: The Dempster-Shafer evidence theory is applied to synthesize multi-source detection results, and a dynamic adjustment mechanism for the confidence threshold is set to adaptively adjust the confidence threshold according to the production speed.

[0045] Furthermore, the improved YOLOv5 model includes:

[0046] Backbone network: Introduces EfficientNet-B3 as a feature extractor and embeds the CBAM attention module;

[0047] Detection head: Uses depthwise separable convolution to replace traditional convolution, introduces an adaptive anchor frame optimization algorithm, and regenerates anchor frames based on the corrugated cardboard defect dataset;

[0048] Training strategy: Focal Loss is used to adjust positive and negative samples, and knowledge distillation technology is implemented to transfer the knowledge of the teacher model (YOLOv5s) to the student model.

[0049] Furthermore, the stream computing framework receives production equipment sensor data, material consumption data, and order information in real time through the Kafka message queue, applies the Flink framework to implement sliding window calculations, and uses the CEP complex event processing engine to detect abnormal cost fluctuations;

[0050] Establish a real-time cost calculation model based on activity-based costing (ABC) and dynamically maintain a standard cost database, including standard labor hours and material consumption for corrugated box specifications;

[0051] The cost calculation results are cached in Redis to generate a cost change heat map.

[0052] A production management method for corrugated paper boxes comprises the following steps:

[0053] Based on order information, material inventory, and equipment capacity, a hybrid optimization algorithm combining genetic algorithm and taboo search is used to generate production schedules. The hybrid optimization algorithm uses chromosome encoding, sequential crossover (OX) operation, and fuzzy comprehensive evaluation model to perform multi-objective adjustments.

[0054] The Hyperledger Fabric consortium chain architecture records the entire material life cycle data, including procurement batches, production process parameters and quality inspection results, and uses zero-knowledge proof for supply chain traceability;

[0055] Build a 3D dynamic model based on point cloud data, map the physical equipment status in real time through the OPC UA protocol, and combine the particle swarm optimization algorithm to coordinate the control of energy consumption and production capacity;

[0056] Integrating machine vision, spectral analysis, and laser ranging technology, it uses an improved YOLOv5 model for defect detection;

[0057] Based on the Apache Flink streaming computing framework, production costs are updated through the activity-based costing (ABC) model, and overspending risks are predicted based on the updated production costs.

[0058] (3) Beneficial effects

[0059] The present invention provides a production management system and method for corrugated paper boxes, which has the following beneficial effects:

[0060] First, the order ID is encoded as an integer chromosome (for example, the order set {O1, O2, O3} is encoded as [1, 2, 3]), and the parent gene fragments are recombined by sequential crossover operation, and the reverse mutation is randomly reversed to ensure the diversity of the solution space; the feature vectors of the most recent N search solutions are stored in a hash list (the retrieval time complexity is O(1)), and the candidate solutions are optimized for 5-10 iterations in combination with the local hill climbing method to avoid falling into the local optimum. A fuzzy comprehensive evaluation model including "delivery punctuality, equipment utilization, and mold change cost" is constructed, and each factor is quantified by a trapezoidal membership function (for example, the delivery punctuality adopts an ascending semi-trapezoidal function, and the mold change cost adopts a descending semi-trapezoidal function). The target weight coefficient is dynamically adjusted by an adaptive fuzzy controller. The algorithm is terminated when there is no change in the optimal solution for consecutive K generations. The output schedule can significantly optimize the order delivery punctuality, equipment utilization, and mold change cost.

[0061] Second, in the material procurement process, the purchase batch number and supplier information are uploaded to the chain for evidence storage; in the production process, process parameters (such as corrugated roller temperature and pressure) and quality inspection results (such as burst strength and edge pressure strength) are automatically collected and recorded through smart contracts; when tracing the supply chain, zero-knowledge proof technology is used to allow the data owner to prove the authenticity of the data to the verification party without leaking the original information (such as when a supplier proves the compliance of a material to a customer, it only needs to prove the correctness of the knowledge without disclosing specific parameters). The distributed ledger characteristics of the alliance chain ensure that data tampering can be traced, and the on-chain data storage time accuracy reaches milliseconds, which improves the efficiency of material traceability; while protecting the privacy of the company's core data (such as formulas and process parameters), zero-knowledge proof technology meets the transparency needs of multi-party collaboration in the supply chain and reduces data sharing costs.

[0062] Third, laser scanning is used to obtain the geometric parameters of the equipment (accuracy ±0.5mm), and the number of model faces can be reduced through triangulated mesh simplification. The level of detail (LOD) technology is used to support switching to the highest level 5 accuracy (displaying complete details within 10 meters of the equipment and automatic simplification at a long distance); a real-time communication channel is established through the OPCUA protocol (data update frequency ≥100Hz), and a heartbeat detection mechanism is configured (automatically switching to local cached data when communication interruption exceeds 500ms) to ensure that the virtual model is synchronized with the physical equipment status; the Kalman filter algorithm is introduced to eliminate sensor noise and achieve a position alignment error of ≤0.3mm. Discrete data such as equipment vibration acceleration and motor current are integrated and mapped into continuous field variables through the Kriging interpolation method to construct a four-dimensional dynamic simulation model of temperature field, stress field, and flow field; when the physical field parameters exceed the warning threshold (such as bearing temperature >80℃), the multi-objective particle swarm optimization algorithm is triggered to automatically adjust process parameters such as motor speed and pressure.

[0063] Fourth, the line scan camera captures carton surface images at a frequency of 2000Hz (resolution ≥1200dpi), the near-infrared spectrometer scans the 900-1700nm band to detect ink components, and the triangulated laser sensor measures the carton size with an accuracy of ±0.1mm; in the feature fusion layer, the image texture features, spectral absorption peak features (such as the characteristic peaks of harmful components in ink), and size parameters are fused through the feature pyramid network; the decision fusion layer applies the Dempster-Shafer evidence theory to synthesize multi-source detection results, and dynamically adjusts the detection strategy according to the production speed (such as automatically increasing the confidence threshold to 0.95 when the high-speed line is greater than 300 boxes / minute). The backbone network introduces EfficientNet-B3 as a feature extractor and embeds the CBAM attention module to focus on the defect area; the detection head uses depthwise separable convolution to reduce the amount of calculation, and regenerates the anchor frame based on a data set of 100,000 corrugated cardboard defects; during training, the caustic loss is used to balance positive and negative samples (the weight of difficult samples is increased by 2 times), combined with knowledge distillation technology (the teacher model is ResNet-50).

[0064] Fifth, through the Kafka message queue, equipment sensor data (such as power consumption and consumables consumption), material in-and-out data, and order information are collected in real time, and the real-time production cost is calculated using the Flink sliding window (window size of 5 minutes); an activity-based costing model is constructed to dynamically maintain a standard cost database containing 200+ corrugated cardboard box specifications (standard working hour accuracy of ±2 minutes, material consumption accuracy of ±0.5%), and abnormal cost fluctuations are detected through the CEP complex event processing engine (such as triggering an early warning when the material cost of a single box suddenly increases by 10%). The LSTM-Transformer hybrid neural network is used. The LSTM layer captures the short-term fluctuation characteristics of the time series, and the Transformer layer models the long-distance dependencies of the cost data (such as the impact of seasonal factors on raw material prices), realizing the prediction of overspending risks in the next 72 hours; the calculation results are cached in Redis and a cost change heat map is generated to visualize the cost proportion and fluctuation trend of each process, assisting management in real-time decision-making and shortening the cost accounting lag time from 24 hours to minutes. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the overall structure of the present invention. DETAILED DESCRIPTION

[0066] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0067] Market research:

[0068] Corrugated cardboard boxes are core products in the packaging industry. Their production management involves complex processes such as order scheduling, material control, equipment scheduling, quality inspection, and cost accounting. Traditional production systems have the following pain points:

[0069] Single scheduling algorithm: It is difficult to balance multiple optimization objectives such as order delivery time, equipment utilization, and mold change costs;

[0070] Insufficient traceability: Centralized databases are subject to data tampering risks, resulting in low supply chain collaboration efficiency;

[0071] Extensive equipment control: Relying on manual inspections, lacking real-time performance and accuracy;

[0072] Quality inspection limitations: Single-modality technology has a high missed detection rate and cannot meet the needs of high-end packaging inspection;

[0073] Cost accounting is lagging: there is a lack of real-time data support and weak warning capabilities for overspending risks.

[0074] To address the above problems, this application proposes a production management system that integrates intelligent scheduling, blockchain traceability, digital twin equipment collaboration, multimodal detection and stream computing cost control. It realizes multi-objective scheduling through a hybrid optimization algorithm of genetic algorithm and taboo search, builds a material life cycle traceability system based on Hyperledger Fabric alliance chain and zero-knowledge proof, and uses a three-dimensional digital twin model combined with multi-physical field coupling analysis to realize real-time equipment management and control. It integrates machine vision, spectral analysis and laser ranging technology to improve the YOLOv5 model for defect detection. It realizes real-time production cost accounting and overspending risk prediction based on the Apache Flink stream computing framework and LSTM-Transformer neural network. The system significantly improves production efficiency, quality stability and cost control capabilities, and provides an effective solution for the intelligent transformation of the corrugated cardboard industry.

[0075] R&D concept:

[0076] For examples, see Figure 1 This embodiment provides a corrugated cardboard production management system. This system is mainly applicable to the full-process intelligent management of corrugated cardboard production enterprises. It covers five core links: order scheduling, material control, equipment scheduling, quality inspection, and cost accounting. It also integrates technologies such as the Industrial Internet, blockchain, digital twins, multimodal AI, and stream computing to build a closed-loop system of "data collection - intelligent decision-making - real-time control". The specific structure is as follows:

[0077] 1. Genetic-Tabu Search Hybrid Algorithm

[0078] 1.1, genetic operation:

[0079] Map the order ID to an integer sequence (e.g., the order set {O1, O2, O3} is encoded as [1, 2, 3]), each gene bit represents the production order, and encode the production order sequence as an integer chromosome, each gene bit corresponds to a unique order ID. Suppose the order set is O = {O1, O2, ..., O3} n}, then the chromosome is represented as Chromosome = {ID1, ID2, ..., ID n}, where ID i For order O i The unique identifier of the gene bit corresponds to the order priority and process constraints (such as adjacent production of orders with the same flute type).

[0080] For example, the order set O = {O1 (specification 3-layer B type, quantity 1000, delivery date April 10), O2 (5-layer AB type, quantity 500, delivery date April 9)}; equipment capacity: printing machine A is available from 14:00 to 24:00 on April 8, and die-cutting machine B is available from 16:00 to 24:00 on April 8. Material inventory: 2000 sheets of 3-layer corrugated paper and 1000 sheets of 5-layer corrugated paper.

[0081] Order Crossover (OX): Randomly select gene segments from parent 1 and parent 2, retain the segment order of parent 1, and fill in the unselected genes from parent 2 in sequence to avoid duplication and conflict. For example: parent 1 = [A, B, C, D], parent 2 = [C, D, A, B], select segment [B, C], and generate offspring [A, B, D, C] after crossover.

[0082] Reverse Insertion (RI): Randomly select two sites on the chromosome and reverse the gene order between the two sites to enhance local search capabilities. [1, 2, 3, 4, 5] becomes [1, 4, 3, 2, 5] after reversing sites 2-4.

[0083] 1.2, Tabu Search (TS):

[0084] The solutions of the most recent N=20 searches are recorded (taboo table), and the feature vectors are stored in a hash list (MD5 hash algorithm is used, and the feature vector length is 128 bits, such as the hash value of the order sequence). The retrieval time complexity is O(1), which avoids repeated searches. In addition, in order to improve the quality of the local optimal solution, the local hill climbing method is introduced, and the candidate solutions are iterated 5-10 times. Each time, a new solution is generated by exchange or insertion operation. If the new solution is better, it is retained, otherwise the iteration is terminated.

[0085] 2. Multi-objective fuzzy evaluation model

[0086] 2.1. Evaluation factors:

[0087] U = {u1, u2, u3} = {delivery timeliness, equipment utilization, mold change cost}, specifically as follows:

[0088] On-time delivery rate: target ≥ 95%. The trapezoidal membership function defines that the membership degree rises to 1 when the on-time rate is [85%, 95%] and decreases when it is below 85%;

[0089] Equipment utilization rate: target ≥80%, ascending semi-trapezoidal function (membership degree = 1 when ≥80%, linear increase when <80%);

[0090] Mold change cost: target ≤ 500 yuan / time, descending semi-trapezoidal function (membership degree = 1 when ≤ 500 yuan, linearly decreasing when > 500 yuan).

[0091] Membership function: Trapezoidal membership function is used to quantify each factor:

[0092] Trapezoidal membership function (taking delivery on-time rate u1 as an example): Among them, x is the actual on-time delivery rate (percentage), [a1, b1, c1, d1] are the trapezoidal interval parameters, which define the range of membership variation. a1 = 80% is the low boundary, and the membership is 0 when the on-time rate is lower than this value; b1 = 85% is the rising boundary, and the membership begins to rise linearly after the on-time rate exceeds this value; c1 = 95% is the stable boundary, and the membership of the on-time rate within this interval is 1; d1 = 100% is the high boundary, and the membership decreases linearly after the on-time rate exceeds this value (the highest in practice is 100%).

[0093] Equipment utilization u2: ascending semi-trapezoidal function

[0094] Mold change cost u3: descending semi-trapezoidal function

[0095] Calculation of comprehensive evaluation value: Among them, S is the comprehensive evaluation value (the value range is 0-1, the higher the value, the better the solution), ω i is the weight coefficient of the i-th evaluation factor (dynamically adjusted by the adaptive fuzzy controller, ω1+ω2+ω3=1), is the membership value of the i-th factor (e.g. u2 is the equipment utilization rate, using the ascending semi-trapezoidal function, the membership is 1 when x ≥ 80%).

[0096] 2.2 Dynamic Weight Adjustment:

[0097] Through the adaptive fuzzy controller, the weights are automatically adjusted according to production environment parameters (such as the proportion of urgent orders and the equipment failure rate) (for example, the weight of on-time delivery rate is increased to 0.6 for urgent orders). The system terminates when the optimal solution remains unchanged after K consecutive generations (such as K = 20), and outputs a scheduling plan that includes order sequence, equipment allocation, and time window.

[0098] For example, assuming that the order set is {O1, O2, O3}, the chromosome is initialized to [1, 2, 3], and after crossover, it generates [1, 3, 2], and after reversal mutation, it becomes [1, 2, 3] (no change). In the fuzzy evaluation, if the on-time delivery rate is 90% (membership degree 0.8), the equipment utilization rate is 85% (membership degree 1), and the mold change cost is 600 yuan (membership degree 0.6), the weight ω = [0.4, 0.3, 0.3], then the comprehensive evaluation value S = 0.4×0.8+0.3×1+0.3×0.6=0.8.

[0099] 3. Material traceability module - blockchain-driven full-link traceability

[0100] 3.1、Hyperledger Fabric consortium chain:

[0101] Deployment nodes include alliance members such as manufacturers, suppliers, and logistics providers. Each node maintains the same ledger. Data on the chain includes the procurement stage, production stage, and circulation stage, as follows:

[0102] Procurement stage: batch number, supplier, purchase date, material composition (such as corrugated paper weight and bursting strength); production stage: process parameters (fluting type, glue formula, pressing pressure), quality inspection results (compressive strength, water resistance); circulation stage: logistics nodes, storage environment (temperature and humidity), dealer information.

[0103] Among them, when purchasing materials, a block containing the purchase batch number BID, supplier L, and purchase date T is generated; in the production stage, process parameters P = {p1, p2, ..., p m} and the quality inspection result Q are written into the blockchain through smart contracts.

[0104] 3.2 Zero-Knowledge Proofs (ZK-SNARKs):

[0105] When downstream customers verify the compliance of materials, a proof function f(BID, L, T, P, Q)→True / False is constructed. Customers can verify compliance without obtaining specific data. Suppliers do not need to disclose specific data (such as ingredient ratios), but only need to mathematically prove that "the data meets the preset conditions" (such as "burst resistance ≥ 800kPa"). This method can protect commercial secrets (such as formula parameters) while ensuring that the data cannot be tampered with (the blockchain hash value is unique).

[0106] For example, when purchasing a batch of corrugated paper, the block data is {BID=20250301001, L=Company A, T=2025-03-01}. During production, process parameters (such as corrugated flute type B, glue viscosity 1500cP) and quality inspection results (qualified) are added. Through zero-knowledge proof, downstream customers can verify that the batch of materials meets environmental protection standards without obtaining the specific viscosity value.

[0107] 3.3 Smart Contract Implementation:

[0108] When materials enter the production process, sensors automatically collect process parameters, triggering smart contracts to encrypt the data and upload it to the chain. The block contains the hash value of the previous block, forming a chain structure.

[0109] The user enters the batch number, and the contract returns the full process node information of the batch of materials (including timestamp and operator signature), supporting traceability in seconds.

[0110] 4. Equipment collaboration module - real-time control driven by digital twins

[0111] 4.1. 3D dynamic model construction:

[0112] (1) Data acquisition: Laser scanner acquires device point cloud data (Accuracy ±0.5mm, point cloud density 100 points / cm 2 ), where (x i ,y i , z i ) is the coordinate of the i-th point in three-dimensional space, and N is the number of points in the point cloud;

[0113] The geometric parameters of the equipment (such as roller radius and conveyor belt length) are obtained. The edge folding algorithm is used to convert the point cloud data into a triangular mesh model. The QEM (quadratic error metric) is used to simplify the algorithm and retain sharp features (such as bolt connections). Five levels of accuracy are defined. The relationship between the level of detail and the number of faces is shown in the following table:

[0114] Level of Detail Face number range Application Scenario 5 More than 100,000 Close inspection 3 10,000-100,000 Central control room monitoring 1 1,000-10,000 Remote Overview

[0115] (2) Operation data: vibration sensor (range ±10g, frequency response 0-10kHz), current sensor (accuracy 0.5% FS), temperature sensor (accuracy ±1°C), establish a communication channel through the OPCUA protocol, define the device status data model (such as "device ID. Operation status. Temperature"), the data update frequency is 100Hz, and support TLS1.2 encryption; the heartbeat detection mechanism sends a heartbeat packet every 50ms. If no response is received within 500ms, it switches to the local cache (caching the data of the last 10 seconds). The deployment density is ≥5 devices / device. When the server does not respond for 10 timeouts (i.e. 500ms), it switches to the local cache (cache depth 1000 data, storage time 10 minutes).

[0116] (3) Level of Detail (LOD) technology: Automatically switches model accuracy (up to 5 levels) based on the user interaction distance. When the user is d≤2m away from the device, the highest accuracy model (level of detail 5) is loaded, and when d>10m, the lowest accuracy model (level of detail 1) is loaded. Bolt-level details are displayed at the near end (<2m) and the device outline is displayed at the far end (>10m), thereby improving rendering efficiency.

[0117] (4) Kalman filter: eliminates sensor noise and ensures that the alignment error between the virtual model and the physical device is less than 0.1 mm;

[0118] Equation of state: x k =Ax k-1 +Bu k +w k ;

[0119] Observation equation: z k =H g x k +g k ; where x k is the device state vector at time k (including motion parameters such as position and velocity), A is the state transfer matrix (describing the change of device state over time, such as A = 1 for uniform motion), x k-1 is the device state vector at time k-1, B is the control input matrix (if there is an external control signal, such as motor speed regulation, B is the input influence coefficient), u k is the control input at time k (such as motor speed command), w k is the process noise (obeying normal distribution N(0, Q), Q is the covariance matrix), z k is the sensor observation value at time k (such as vibration acceleration, position coordinates), H g is the observation matrix (mapping the state vector to the observation space, such as H when the position is observed g =[1,0]), g kis the observation noise (which obeys the normal distribution N(0, R), where R is the covariance matrix), and the optimal state is estimated by recursive update;

[0120] (5) Version management: When the device structure changes, the model reconstruction process is triggered, a new version number V = major version. minor version. revision number (such as 1.0.1) is generated, and a change log is recorded.

[0121] 4.2 Multi-physics coupling analysis:

[0122] Four-dimensional simulation model: The spatiotemporal data cube structure is DataCube(x, y, z, t), (X, Y, Z) is the spatial coordinate, and the equipment operation status spatiotemporal cube is constructed (data granularity: 1 second / time), and t is the time (resolution 1 second).

[0123] Physical field modeling includes temperature field (bearing temperature > 80°C triggers an early warning, and the lubrication frequency is adjusted through the particle swarm algorithm), stress field (early warning when the stress of key components is > 200MPa or when the pressure of the corrugated roller is > 300kN, and the pressing speed is automatically optimized) and flow field (early warning when the cooling fan flow rate is < 0.5m / s, and the fan power is adjusted in a linked manner).

[0124] Data mapping (Kriging interpolation): Let the discrete sensor locations be {s1, s2, ..., s m}, the measured values are {z(s1), z(s2), ..., z(s m )}, the interpolation formula is: in, is the predicted value of the continuous field variable (such as temperature, stress) at position s0, s i is the position (three-dimensional coordinates) of the i-th sensor, z(s i ) is the measurement value of the i-th sensor (such as vibration acceleration, motor current), λ i is the weight of the i-th sensor measurement value (determined by minimizing the prediction variance, satisfying ∑λ i =1), assuming that the sensor data obeys the second-order stationary hypothesis, the variation function The semivariogram parameters are determined by least squares fitting to achieve the mapping from discrete points to continuous fields.

[0125] Coupling solution: We adopt the weak coupling method, first calculate the temperature field, then use the temperature as the load input stress field calculation, and finally couple the flow field. The iterative convergence condition is that the change rate of each field parameter is less than 1%, and the equipment operation parameter optimization plan (such as motor speed, pressure value) is output every 5 seconds. We adopt the weak coupling method to iteratively calculate the temperature-stress-flow field, and the objective function is: min(E h +0.8C h ), where E h is energy consumption, Ch is the capacity loss; in the particle swarm optimization algorithm, the particle speed and position are updated: in, is the velocity of the i-th particle at the k-th iteration, is the velocity of the i-th particle at the k-1th iteration, ω g is the inertia weight (controls the degree of retention of the particle's historical velocity, usually 0.8-1.2), c1, c2 are learning factors (adjust the step size of the particle moving towards the individual optimum and the global optimum, usually 2), r1, r2 are random numbers in the interval [0, 1] (to increase the randomness of the search), pbest i is the historical optimal position of the i-th particle (corresponding to the minimum energy consumption or the highest production capacity), gbest is the global optimal position (the optimal solution of all particles), is the position of the i-th particle at the k-th iteration (representing the equipment process parameter combination, such as motor speed and pressure value), is the position of the i-th particle at the k-1th iteration.

[0126] 5. Multimodal Detection Module—Fusion Perception Defect Detection

[0127] 5.1 Multi-source data collection:

[0128] Machine vision: LMI Gocator 2420 line scan camera (12k pixels, 2000Hz) captures the carton surface image I∈R W×H×3 , W is the image width (in the case of line scanning, it corresponds to the number of pixels per line, that is, W = 12000), H is the image height corresponding to the number of scanned lines. Assuming the scanning time is t seconds, then H = 2000t, because 2000 lines are scanned per second), 3 is the number of channels of the image, and R is the set of real numbers. Detect scratches, stains, and print offsets.

[0129] Spectral analysis: Bruker MPA near-infrared spectrometer (900-1700 nm) was used to detect the ink composition, with a spectral resolution of 10 nm and an output absorption peak vector S = [s1, s2, ..., s 80 ]; Identify the volatile organic compound (VOC) content in ink (such as excessive solvent residue).

[0130] Laser ranging: The Keyence LK-G80 triangulation laser sensor (accuracy ±0.1mm) measures the length, width, height, and flute height of cartons, detecting dimensional deviations (e.g., greater than ±1mm). The measurement range is 0-2000mm, and the output dimensional parameters D = (L, W, H), where L, W, and H are length, width, and height, respectively.

[0131] 5.2 Improved YOLOv5 model:

[0132] Backbone network: EfficientNet-B3 is introduced as the feature extractor, and the CBAM attention module is embedded to weight the channel and spatial dimensions separately: Among them, F′ input feature map (size H×W×C), M c is the channel attention, M c (F) is the channel attention weight (dimension 1×1×C, generated by global average pooling and fully connected layers), M s is spatial attention, M s (F) is the spatial attention weight (dimension H×W×1, generated by channel dimension compression and convolution), By performing element-by-element multiplication (enhancing the feature response of key channels and spatial areas), focusing on defect areas (such as texture features at edge damage), the detection speed is increased by 20% and the number of parameters is reduced by 30%.

[0133] Detection head: Depthwise Convolution + Pointwise Convolution is used to replace traditional convolution, reducing the number of parameters by 40%. The adaptive anchor box clusters the corrugated cardboard defect dataset (containing 100,000 samples and 12 defect types) based on the K-means++ algorithm to generate 6 groups of anchor boxes: (12×16), (24×32), (48×64), (80×100), (120×150), and (200×250).

[0134] Adaptive anchor frames: Regenerate anchor frames based on a dataset of 100,000 defects. The sizes of the anchor frames are {(16×16), (32×32), ..., (128×128)} (e.g., a 16×16 anchor frame for a small target with edge damage, and a 128×128 anchor frame for a large target with box deformation).

[0135] Training strategy: Reduce the weight of easy-to-classify samples (such as defect-free areas) and focus on difficult samples (defect edges), reducing the missed detection rate from 8% to 5%; the teacher model (ResNet-50) guides the student model (YOLOv5) to learn deep features, and the detection accuracy of the small model is improved by 5%. The loss function is L = L CE +λL KD , where L CE is the cross entropy loss, L KD is the teacher-student output distribution difference (KL divergence), λ = 0.5.

[0136] FocalLoss FL(p t )=-αt(1-p t ) γlog(p t ), where p t is the model's predicted probability for the sample (positive sample p t =p, negative sample p t =1-p), αt is the category balance factor (to solve the imbalance of positive and negative samples, such as αt = 0.25 when there are too many negative samples), γ is the focusing parameter (to adjust the weight of easy-to-classify samples, and when γ = 2, the loss contribution of high-confidence samples is reduced), which reduces the weight of easy-to-classify samples.

[0137] 5.3 Multimodal Fusion Strategy

[0138] Feature fusion: The image texture (2048 dimensions), spectral absorption peak (80 dimensions), and size parameter (3 dimensions) are fused through the feature pyramid network (FPN) to generate a 512-dimensional multimodal feature vector and output the multimodal feature vector F∈R 512 .

[0139] Decision fusion: Dempster-Shafer evidence theory is used to synthesize the three-modal detection results. The confidence threshold is dynamically adjusted with the production speed (for example, the threshold for high-speed lines (>200 pieces / minute) is reduced from 0.8 to 0.75 to balance speed and accuracy). The detection results are {D1, D2, D3} corresponding to machine vision, spectroscopy, and laser detection respectively), the confidence is {m1, m2, m3}, m1∈[0,1], and the conflict coefficient is defined. When k < 0.7, Dempster's synthesis rule is used: Among them, m(A) is the confidence of proposition A (such as "carton defect") after synthesis, A1, A2, A3 are the detection results of machine vision, spectral analysis, and laser ranging (such as "qualified" and "defective") respectively, and m i (A i ) is the result A of the i-th detection technology i The basic probability assignment is based on the machine vision (e.g. the confidence level of a defect detected by machine vision is 0.7), excluding cases where the detection results are completely conflicting (e.g. two technologies consider it qualified and one considers it defective).

[0140] Dynamic adjustment of confidence threshold: Assume that the production speed is v (pieces / minute), the threshold θ = 0.7 + 0.1 × tanh (v / 50), and lower the threshold during high-speed production to reduce missed detections.

[0141] 6. Cost control module - real-time dynamic cost control

[0142] 6.1 Stream Computing Architecture (Apache Flink + Kafka):

[0143] Data pipeline: Kafka message queues receive real-time data on device energy consumption (50Hz), material consumption (per batch), and order output value (in real time), with a throughput of 100,000 messages per second; Flink sliding windows (5-minute windows, 1-minute steps) calculate unit product costs with a latency of less than 200ms.

[0144] Activity-based costing (ABC) model: Direct costs: material costs (corrugated paper, ink, glue, inventory deducted in real time according to the order BOM) + labor costs (standard working hours allocated by process, such as 0.1 hours / piece for printing process, 0.15 hours / piece for die-cutting process), manufacturing expenses (equipment depreciation: allocated by usage time, energy consumption: calculated by power × running time); indirect costs: equipment depreciation, energy consumption, and management expenses, allocated through cost drivers (such as machine hours, order quantity).

[0145] For example: Indirect cost of printing process = machine hours of this process × (total indirect cost / total machine hours of the whole factory). Cost allocation formula: Cost driver rate r = total manufacturing overhead / total standard working hours.

[0146] Standard cost database fields: {carton specifications, standard working hours, standard material consumption, standard energy consumption}. For example, the standard working hours for a 3-layer B-type carton is 0.2 hours / piece, and the standard paper consumption is 1.5 kg / piece. Among them, direct material cost: direct material costs such as base paper and ink for producing cartons (yuan / batch); direct labor cost: direct labor wages in the production process (yuan / batch); indirect costs: equipment depreciation, factory rent and other shared costs (yuan); cost driver rate: basis for allocating indirect costs (such as allocation by machine working hours, driver rate = indirect costs / total working hours); output: number of cartons produced in the current period (pieces).

[0147] Real-time cost formula: Among them, M i is the cost of the i-th material, L i is the working hours of the i-th process, w is the labor unit price, O i is the i-th manufacturing cost, Q t is the output at time t.

[0148] Anomaly detection (CEP engine defined rules): If the unit cost fluctuation for three consecutive windows is greater than 15%, an anomaly warning is triggered.

[0149] 6.2. Overspending Risk Prediction (LSTM-Transformer Hybrid Network):

[0150] LSTM layer: captures short-term cost fluctuations (such as an increase in material consumption over three consecutive windows). The hidden layer has 256 dimensions, a memory period of 12 hours, and is stacked in two layers with 128 neurons per layer. This layer captures short-term dependencies in time series (such as cost fluctuations over the last eight hours).

[0151] Transformer layer: The multi-head attention mechanism (8 heads) analyzes long-range dependencies (such as the relationship between quarterly order volume and raw material purchase price). The position encoding uses a sine function to capture long-term dependencies (such as weekly cost trends). The formula is Among them, Q, K, V are query, key, and value matrices, d k =64 as a dimension), identifying abnormal cost patterns (such as the chain reaction caused by a sudden price increase by a supplier).

[0152] Output layer: Predicts the probability of overspending in the next four hours with an accuracy of ≥85%, triggering procurement warnings in advance (such as switching to alternative suppliers). The loss function is binary cross entropy. The overspending probability is output through a fully connected layer + Sigmoid function. The loss function is FocalLoss (balancing positive and negative samples).

[0153] 6.3 Visualization and Decision Support

[0154] Cost heat map: Displays cost distribution by process, equipment, and time dimensions (e.g., the corrugated forming process accounts for 35% of the total cost, and the energy consumption cost late at night is 10% higher than that during the day).

[0155] Standard cost database: Maintains standard working hours for each specification of cartons (e.g. 0.15 hours / piece for a 3-layer A-flute carton) and standard material consumption (1.2kg / piece), and compares the deviation between actual cost and standard value in real time (warning when >5%).

[0156] 6.4 System implementation process and key nodes:

[0157] Hardware: laser scanner (accuracy ±0.5mm), industrial sensors (vibration / temperature / current), line array camera (12k pixels), near-infrared spectrometer (900-1700nm).

[0158] Software: Hyperledger Fabric node (≥4 cores and 8GB memory), Flink cluster (3 nodes, 16 cores and 32GB per node), deep learning server (NVIDIA A100 GPU).

[0159] 6.5 Implementation steps:

[0160] G001, Data Acquisition and Modeling (1-2 weeks): Use laser scanning equipment to generate point cloud data and build a 3D digital twin model (supporting 5 levels of detail). Deploy a sensor network and access device operation data via the OPCUA protocol (update frequency 100Hz).

[0161] G002, Algorithm Training and Optimization (2-3 weeks): Use historical order data to train the hybrid optimization algorithm and tune the fuzzy evaluation weights (e.g., a delivery weight of 0.6 for urgent orders and a weight of 0.5 for equipment utilization for regular orders); annotate 100,000 defect images, train and improve the YOLOv5 model, and compress the model size through knowledge distillation (from 200MB to 120MB). For knowledge distillation, the teacher model is ResNet-50 (pre-trained on ImageNet), the student model is YOLOv5s, and the distillation loss is L KD =KL(p teacher ||p student ), where p teacher Temperature scaled (T=10);

[0162] G003, System Integration and Testing (1 week): Jointly debug the interfaces of each module (e.g., scheduling results drive parameter adjustments in the equipment collaboration module, and test results trigger abnormality warnings in the cost module); Stress testing: simulate 2,000 orders / day and 30 devices running concurrently to verify that scheduling calculation time is less than 10 minutes and test latency is less than 100ms;

[0163] G004, Online and Iteration (Continuous): Establish model version management (e.g., automatically trigger digital twin model reconstruction after equipment modification, version number 1.0→1.1); regularly update the defect dataset (add new defect samples), and fine-tune the YOLOv5 anchor frame parameters.

[0164] Example of a standard cost database:

[0165] Carton specifications Standard working hours (h / piece) Standard paper (kg / piece) Standard energy consumption (kW·h / unit) 3-layer Type B 0.2 1.5 0.15 5-layer AB type 0.3 2.5 0.25

[0166] Real-time cost calculation:

[0167] Assume that from 15:00 to 15:05 on April 8, 500 3-layer B-type cartons were produced, consuming 760 kg of corrugated paper (unit price 3 yuan / kg), labor cost 200 yuan, and equipment energy consumption 80 kW·h (unit price 1 yuan / kW·h), then The standard cost is 1.5×3+0.2×20+0.15×1=4.5+4+0.15=8.65 yuan / piece (Note: The example data here is only for illustration of the calculation logic; the actual standard cost needs to include more indirect costs).

[0168] A production management method for corrugated paper boxes, the method steps are as follows:

[0169] Data input: The order management system imports order information (including specifications, quantity, and delivery date), the ERP system synchronizes material inventory (corrugated paper inventory, ink balance), and the equipment management system obtains equipment capacity (available time and maximum capacity of each production line);

[0170] Intelligent production scheduling: The production scheduling module runs a hybrid optimization algorithm to generate a production schedule that includes order sequence, equipment allocation, and time nodes, and outputs it to the MES system;

[0171] Material control: Purchase batch information is uploaded to the blockchain, process parameters (such as printing pressure and die-cutting temperature) are recorded when materials are collected for production, and quality inspection results (pass / fail) are written to the blockchain in real time. All parties in the supply chain verify material compliance through zero-knowledge proof without exposing commercially sensitive data.

[0172] Equipment collaboration: A 3D dynamic model maps equipment status in real time. When bearing temperature exceeds 80°C, an early warning is triggered. The multi-physics field analysis module uses a particle swarm optimization algorithm to adjust the motor speed (target: energy consumption down by 10%, production capacity up by 5%). After equipment parameters are optimized, the new operating parameters are fed back to the production scheduling module to dynamically adjust subsequent production plans.

[0173] Quality inspection: A line scan camera captures carton surface images (resolution 12,000 × 1,000), a spectrometer scans the ink area (each scan takes 50 milliseconds), and a laser sensor measures dimensions (taking the average of three inspections per carton). Multimodal data is fused using a feature pyramid network, and an improved YOLOv5 model is used to detect defects. Decision results are synthesized using evidence theory, and defective products trigger rework.

[0174] Cost control: Flink calculates unit costs in real time. When cost fluctuations exceed 15% over three consecutive 15-minute windows, the CEP engine triggers an exception alert and notifies the production manager. The LSTM-Transformer model predicts the probability of overspending in the next two hours. If the probability is ≥70%, adjustments to the production schedule are recommended (for example, prioritizing high-profit orders).

[0175] The weight coefficient is determined using the coefficient of variation method, which is a method of assigning weights to each indicator based on the degree of variation between the current value of each evaluation indicator and the target value. If the numerical difference of an indicator is large, it can clearly distinguish the evaluated objects, indicating that the indicator has rich discrimination information, and thus the indicator should be given a larger weight. On the contrary, if the numerical difference of each evaluated object on a certain indicator is small, then the ability of this indicator to distinguish the evaluation objects is weak, and thus the indicator should be given a smaller weight. This method directly uses the information contained in each indicator to obtain the weight of the indicator through calculation, and therefore is objective.

[0176] In the application, the several formulas involved are all calculated by taking their numerical values after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they are not elaborated here.

[0177] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate 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 performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0178] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0179] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A production management system for corrugated paper boxes, characterized in that: The system includes: Production Scheduling Module: Based on order information, material inventory, and equipment capacity, a hybrid optimization algorithm combining genetic algorithms and taboo search is used to generate production schedules. The hybrid optimization algorithm uses chromosome encoding, sequential crossover operations, and a fuzzy comprehensive evaluation model to perform multi-objective adjustments. Material traceability module: This module uses the Hyperledger Fabric consortium chain architecture to record the entire material life cycle data, including procurement batches, production process parameters, and quality inspection results, and uses zero-knowledge proof for supply chain traceability. Equipment collaboration module: Builds a 3D dynamic model based on point cloud data, maps physical equipment status in real time through the OPC UA protocol, and combines the particle swarm optimization algorithm to coordinate energy consumption and production capacity control; Multimodal inspection module: Integrates machine vision, spectral analysis, and laser ranging technology, and uses an improved YOLOv5 model for defect detection; Cost Control Module: Based on the Apache Flink stream computing framework, this module updates production costs through an activity-based costing model and predicts overspending risks based on the updated production costs.

2. The production management system for corrugated paper boxes according to claim 1, characterized in that: The hybrid optimization algorithm sets a dynamic weight adjustment mechanism, which automatically adjusts the weight coefficients of each objective function according to changes in the production environment through an adaptive fuzzy controller. The construction of the hybrid optimization algorithm includes the following steps: S101, encode the production order sequence into an integer chromosome, where each gene bit represents an order ID; S102, extracting gene fragments using sequential crossover and reverse mutation operations; S103. Record the solutions of the most recent N searches, remove duplicate paths, and use a hash list structure to store the solution’s feature vectors. The retrieval time complexity is O(1). S104. Based on the population generated by the genetic algorithm, candidate solutions are generated by exchange and insertion, and a local hill climbing method is introduced to iterate the candidate solutions for 5-10 times; S105. Construct a fuzzy comprehensive evaluation model based on multiple objectives, where the multiple objectives include order delivery time, equipment utilization rate, and mold change cost; terminate the algorithm when there is no change in the optimal solution of consecutive K generations, and output the optimal production schedule.

3. The production management system for corrugated paper boxes according to claim 2, characterized in that: The fuzzy comprehensive evaluation model in the evaluation function includes: Determine the evaluation factor set: {on-time delivery rate, equipment utilization rate, mold change cost}; Constructing the fuzzy membership matrix: using trapezoidal membership function to quantify each factor; Calculate the comprehensive evaluation value: synthesize the weights of each factor through the weighted average operator.

4. The production management system for corrugated paper boxes according to claim 1, characterized in that: The construction process of the three-dimensional dynamic model includes: S201. Obtain equipment geometric parameters through laser scanning and deploy sensors to collect real-time operation data of production equipment; S202. Based on the acquired device geometric parameters, a simplified model of the device is generated using a triangular mesh simplification algorithm. The model accuracy level is switched based on the user interaction distance using a level of detail technique, with a maximum support of 5 levels of detail. S203. Establish a real-time communication channel between the physical device and the virtual model through the OPC UA protocol, with a data update frequency of ≥100 Hz, and set a heartbeat detection mechanism. When the communication interruption exceeds 500ms, it automatically switches to the local cache data to continue operation; S204, introducing a Kalman filter algorithm to eliminate sensor noise and align the virtual model with the physical device; S205. Set up a model version management mechanism to automatically trigger the model reconstruction process when the device structure changes.

5. The production management system for corrugated paper boxes according to claim 4, characterized in that: The three-dimensional dynamic model also includes a multi-physics coupling analysis module. The multi-physics coupling analysis module uses spatiotemporal data cube technology to integrate equipment operation data, environmental parameters and product process parameters to build a four-dimensional dynamic simulation model and perform multi-physics coupling analysis. The multi-physics coupling analysis module includes: Physical field modeling: Construct temperature fields, stress fields, and flow fields, and set corresponding abnormal warning thresholds. When the parameters in the physical field exceed the set abnormal warning thresholds, an alarm will be issued; Data mapping mechanism: This mechanism uses the OPC UA protocol to acquire real-time device usage data, and uses Kriging interpolation to map discrete sensor data into continuous field variables. Device usage data includes device vibration acceleration and motor current. Coupling solution: Use weak coupling method to iteratively calculate the multi-physics field of temperature-stress-fluid; Real-time optimization control: Adjust process parameters based on the multi-objective particle swarm optimization algorithm and output optimized equipment operating parameters.

6. The production management system for corrugated paper boxes according to claim 1, characterized in that: In the multimodal quality detection module, machine vision uses a line scan camera to obtain the surface image of the carton; spectral analysis uses a near-infrared spectrometer to detect the ink composition; and laser ranging uses a triangulated laser sensor to measure the size of the carton.

7. The production management system for corrugated paper boxes according to claim 6, characterized in that: The fusion method of the multimodal quality detection module includes: Feature fusion layer: A feature pyramid network is used to perform cross-modal feature fusion on image texture features, spectral absorption peak features, and size parameters; Decision fusion layer: The Dempster-Shafer evidence theory is applied to synthesize multi-source detection results, and a dynamic adjustment mechanism for the confidence threshold is set to adaptively adjust the confidence threshold according to the production speed.

8. The production management system for corrugated paper boxes according to claim 7, characterized in that: The improved YOLOv5 model includes: Backbone network: Introduces EfficientNet-B3 as a feature extractor and embeds the CBAM attention module; Detection head: Uses depthwise separable convolution to replace traditional convolution, introduces an adaptive anchor frame optimization algorithm, and regenerates anchor frames based on the corrugated cardboard defect dataset; Training strategy: Use caustic loss to adjust positive and negative samples, implement knowledge distillation technology, and transfer the knowledge of the teacher model to the student model.

9. The production management system for corrugated paper boxes according to claim 8, characterized in that: The stream computing framework receives production equipment sensor data, material consumption data, and order information in real time through the Kafka message queue, applies the Flink framework to implement sliding window calculations, and uses the CEP complex event processing engine to detect abnormal cost fluctuations. Establish a real-time cost calculation model based on activity-based costing and dynamically maintain a standard cost database, including standard labor hours and material consumption for corrugated box specifications; The cost calculation results are cached in Redis to generate a cost change heat map.

10. A method for production management of corrugated paper boxes, using the system according to any one of claims 1 to 9, characterized in that: The steps include: Based on order information, material inventory, and equipment capacity, a hybrid optimization algorithm combining genetic algorithm and taboo search is used to generate production schedules. The hybrid optimization algorithm uses chromosome encoding, sequential crossover operations, and a fuzzy comprehensive evaluation model to perform multi-objective adjustments. The Hyperledger Fabric consortium chain architecture records the entire material life cycle data, including procurement batches, production process parameters and quality inspection results, and uses zero-knowledge proof for supply chain traceability; Build a 3D dynamic model based on point cloud data, map the physical equipment status in real time through the OPC UA protocol, and combine the particle swarm optimization algorithm to coordinate the control of energy consumption and production capacity; Integrating machine vision, spectral analysis, and laser ranging technology, it uses an improved YOLOv5 model for defect detection; Based on the Apache Flink streaming computing framework, production costs are updated through the activity-based costing model, and overspending risks are predicted based on the updated production costs.

Images