Full-chain food safety monitoring method and system based on multi-modal data fusion and intelligent algorithm

Abstract

The invention discloses a full-chain food safety monitoring method and system based on multi-modal data fusion and an intelligent algorithm. The method comprises the steps of S1, sensing layer data acquisition; s2, edge calculation data processing; s3, block chain evidence storage; s4, intelligent decision analysis is carried out; and S5, an application layer responds, so that a closed-loop system of'full-chain data real-time acquisition-dynamic credible evidence storage-cross-modal intelligent analysis' is successfully constructed, and the problems of'data islands, fragile evidence storage and lagged decision 'of traditional monitoring are solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of food safety and information technology, and specifically relates to a full-chain food safety monitoring method and system based on multi-modal data fusion and intelligent algorithm, which is fused with a multi-modal sensor array and a dynamic blockchain storage technology, and is suitable for the full life cycle digital safety management of agricultural products and processed foods from planting / production to terminal sales. BACKGROUND

[0002] The existing food safety monitoring technology has the following major defects, resulting in serious lack of regulatory effectiveness and data credibility. The main problems are as follows:

[0003] 1. Fragmented monitoring coverage:

[0004] The coverage rate of manual sampling inspection is less than 5%, and the risk of missing detection is high; single-link isolated monitoring (such as only production end pesticide residue detection), lack of cross-link risk transmission analysis (such as warehouse humidity exceeding standard without early warning, leading to 18% of terminal sales mold exceeding standard).

[0005] 2. Lack of data credibility:

[0006] Centralized database has the security risk of being easily tampered with, such as some large e-commerce platforms in 2022 being exposed for fabricating detection reports by some stores within the platform, highlighting the lack of credibility in the data storage link;

[0007] In addition, multi-source data formats are not unified (such as production end spectrum data and transportation end temperature and humidity data cannot be directly associated), and the average time spent on abnormal traceability is more than 72 hours.

[0008] 3. Risk response lag:

[0009] In the existing conventional process, manual review of detection reports leads to abnormal response delay of more than 48 hours, missing the best disposal window (such as no intervention within 24 hours before fresh food spoilage).

[0010] In addition, single model misjudgment rate is high, because the pesticide residue detection relying only on visual recognition has a misjudgment rate of 15% for unripe fruits, and traditional rule engine cannot handle multi-parameter coupled risks (such as temperature and humidity abnormalities + transportation jolts). SUMMARY

[0011] Therefore, one of the purposes of the present application is to provide a full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm, to build a closed-loop system of "full-chain data real-time collection-dynamic credible storage-cross-modal intelligent analysis", and to solve the problems of "data island, fragile storage, and decision lag" in traditional monitoring.

[0012] One of the purposes of the present application is achieved by the following technical solutions:

[0013] The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm comprises the following steps:

[0014] Step S1: Perception layer data acquisition: using a multi-modal sensor array and related information acquisition equipment to synchronously acquire heterogeneous data of the physical state, environmental parameters and identity of the food in the circulation link;

[0015] Step S2: Edge computing data processing: constructing an edge computing layer, using an NVIDIA Jetson AGX Orin node to perform real-time aggregation on the heterogeneous data of step S1, wherein image data is encrypted using an AE-256-GCM algorithm, and is subjected to protocol conversion and data standardization through an abnormal initial diagnosis screening module;

[0016] Step S3: Blockchain storage: uploading the standardized data to a blockchain network, verifying by enterprise nodes, regulatory nodes and third-party nodes, and generating a time-stamped storage log chain through an intelligent contract engine;

[0017] Step S4: Intelligent decision analysis: using a random forest model to perform abnormal detection on the storage data, and simultaneously using a Bayesian network to dynamically predict the risk probability, and generating and outputting intelligent decision analysis data including risk scores;

[0018] Step S5: Application layer response: in the consumer client, obtaining desensitized traceability data by scanning the product two-dimensional code, and calculating the risk score in real time by the Bayesian network; the enterprise management end is synchronously connected with the intelligent decision analysis data, and automatically intercepts the goods with a risk score less than a set threshold, and synchronously generates a recall work order to the manufacturer system; the regulatory platform is synchronously connected with the intelligent decision analysis data, and triggers the corresponding disposal process according to the risk score, and leaves digital traces in the whole process.

[0019] As an optional technical solution, in step S1, the multi-modal sensor array comprises a near-infrared spectrometer, a temperature and humidity sensor, a vibration sensor and a pressure sensor, and the related information acquisition equipment comprises a blockchain electronic lock, a Beidou / GPS dual-mode terminal and a two-dimensional code scanning device.

[0020] As an optional technical solution, in step S3, the specific steps comprise:

[0021] (1) Data on-chain preprocessing: separating and storing operation event data and detection data, the operation event data comprising data collected by the related information acquisition equipment, and the detection data comprising data collected by the multi-modal sensor array, and calculating SHA-256 hash values for the two types of data;

[0022] (2) Block construction: generate a block body containing double hash values, calculate the Merkle root of the block body, and construct the block header;

[0023] (3) Weighted consensus verification: a consensus network composed of enterprise nodes, regulatory nodes and third-party nodes is used to verify using an improved Byzantine fault-tolerant algorithm;

[0024] (4) Evidence storage execution: when the consensus is passed, the new block is linked to the evidence storage chain within ≤2 seconds, and the data is tamper-proof.

[0025] As an optional technical solution, the step S4 specifically includes:

[0026] (1) Data preprocessing: divide the original data into spectral data and temperature and humidity / vibration data, perform Savitxky-Golay filtering on the spectral data to eliminate high-frequency noise; use the 3σ rule to denoise the temperature and humidity / vibration raw data, and delete the data points deviating from the mean value ±3σ;

[0027] (2) Cross-modal feature fusion: compress the spectral data from 1024 dimensions to 10 dimensions through PCA dimension reduction to obtain spectral principal component data; perform sliding window statistics on sensor data to calculate mean, variance and maximum to obtain environmental statistical features; fuse spectral principal component data, environmental statistical features, time features and spatial features to generate a 21-dimensional input vector:

[0028] (3) Attention-weighted model collaboration: based on the 21-dimensional input vector, calculate the feature weight of each link through the attention mechanism, input the weighted features into the random forest model, and output the abnormal classification result and confidence;

[0029] (4) Dynamic risk prediction: when the confidence is greater than 0.8, activate the Bayesian network to calculate the 72-hour risk probability;

[0030] (5) Decision output and synchronous push.

[0031] As an optional technical solution, the Bayesian network includes a three-level causal model trained based on at least 3000 historical accident data, and the hierarchical structure of the three-level causal model includes:

[0032] First-level nodes: production links;

[0033] Second-level nodes: storage links;

[0034] Third-level nodes: transportation links.

[0035] As an optional technical solution, the random forest model uses the SMOTE oversampling technique to handle the sample imbalance problem during training, and the node splitting criterion is that the Gini index variation is greater than 0.25.

[0036] The second object of the present application is to provide a full-chain food safety monitoring system based on multi-modal data fusion and intelligent algorithm, the system comprises

[0037] The perception layer unit: contains a multi-modal sensor array and related information acquisition equipment, which is used for synchronous acquisition of heterogeneous data of physical state, environmental parameters and identity of food in the circulation link, and each device is connected through an industrial bus;

[0038] The edge computing unit: an embedded device equipped with AGX Orin chip, built-in data binning module, AE-256-GCM encryption module and abnormal initial diagnosis screening module, the data binning module is used to convert continuous numerical features into discrete features, which is convenient for model processing and understanding, the AE-256-GCM encryption module is used for real-time encryption of image data, and the abnormal initial diagnosis screening module is used to realize multi-source data protocol conversion;

[0039] The blockchain storage unit: including a consensus network composed of enterprise node servers, regulatory node hosts and third-party node clusters, and using an improved Byzantine fault tolerance algorithm for verification, the enterprise node server is integrated with an intelligent contract engine;

[0040] The intelligent decision unit: the server includes a random forest anomaly detection module, a Bayesian network risk prediction module and a cross-modal data fusion module, which realizes PCA dimension reduction and attention weight allocation;

[0041] The application layer interaction unit: including enterprise management end, regulatory platform and consumer client server, which are connected with the server of the intelligent decision unit through the API gateway.

[0042] As an optional technical solution, in the blockchain storage unit, the enterprise node server is configured with a double hash generator; and the third-party node cluster is deployed with a Merkle tree verification FPGA.

[0043] The beneficial effects of the present application are:

[0044] (1) Realize the full-chain digital coverage of food safety: through the multi-modal sensor array, the real-time synchronous acquisition of multi-dimensional data in the production, storage, transportation and sales links is realized, and the coverage rate is improved to more than 99%;

[0045] (2) Build a dynamic and reliable storage system: build a three-dimensional blockchain storage model of "operation track + detection data + environmental parameters", ensure that the data cannot be tampered with and can be traced back to the second operation node;

[0046] (3) Cross-modal intelligent decision: develop a risk assessment model that integrates time and space features, compress the abnormal response time to within 10 minutes, and reduce the misjudgment rate in complex scenarios to less than 5%.

[0047] (4) Flexible adaptation capability: design standardized hardware interface (support 20+ mainstream sensor protocols) and algorithm plugin library, realize plug-and-play deployment of different categories (fresh / processed food), scenes (normal temperature / cold chain).

[0048] Additional advantages, objects, and features of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following specification or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to make the objects, technical solutions and advantages of the present application clearer, the following will further describe the present application in combination with the drawings, in which:

[0050] Figure 1 It is a full-chain food safety monitoring system architecture diagram;

[0051] Figure 2 It is a cross-modal data fusion and risk assessment flow chart:

[0052] Figure 3 It is a dynamic block chain storage diagram;

[0053] Figure 4 It is an intelligent contract execution diagram. DETAILED DESCRIPTION

[0054] It should be understood that the preferred embodiments are only for illustrating the present application, and are not intended to limit the protection scope of the present application. Similar reference numerals are used to represent similar elements throughout the drawings, and the illustrated structures and devices are not necessarily drawn to scale. The terms "component", "system", "interface", and the like as appearing herein are intended to represent a computer-related entity, hardware, software (e.g., in execution), and / or firmware. For example, a component can be a processor, a process operated on the processor, a controller, an object, an executable, a program, a storage device, and / or a computer with a processing device. In an illustrative manner, an application operating on a server and the server can also be a component. One or more components can reside within a process, and the components can be localized in one computer and / or distributed between two or more computers. Groups of elements or other components can be described herein, and the term "group" can be interpreted as "one or more".

[0055] The specific embodiments of the present application will be further described below, and the main problems existing in the prior food safety monitoring include:

[0056] (1) Data collection dimension limitation: existing systems rely on a single sensor (such as deploying only temperature and humidity sensors) or manual input, lacking the ability to synchronously collect multi-dimensional data such as composition, time, and space.

[0057] (2) Static storage mechanism: traditional blockchain storage only records result data (such as detection reports), and does not realize dynamic chaining of production-transportation-sale operation track, leaving data tampering blind spots.

[0058] (3) Linear risk assessment: traditional algorithms (such as logistic regression) cannot model cross-link causal relationships (such as the three-level transmission path of "transportation vibration → packaging damage → microbial growth"), and lack real-time risk prediction capability.

[0059] Therefore, in order to solve the above problems of the prior art, as shown in Figure 1 The present application discloses a full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm, which comprises the following steps:

[0060] Step S1: Perception layer data collection: use a multi-modal sensor array and related information collection equipment to synchronously acquire heterogeneous data of the physical state, environmental parameters, and identity of the food in the circulation link;

[0061] Step S2: Edge computing data processing: build an edge computing layer, use an NVIDIA Jetson AGX Orin node to perform real-time aggregation on the heterogeneous data of step S1, wherein image data is encrypted using an AE-256-GCM algorithm, and is subjected to protocol conversion and data standardization through an abnormal initial diagnosis screening module;

[0062] Step S3: Blockchain storage: upload the standardized data to a blockchain network, verify by enterprise nodes, regulatory nodes, and third-party nodes, and generate a time-stamped storage log chain through an intelligent contract engine;

[0063] Step S4: Intelligent decision analysis: use a random forest model to detect abnormalities in the storage data, and simultaneously use a Bayesian network to dynamically predict risk probability, generating decision analysis data including risk score;

[0064] Step S5: Application layer response: on the consumer client, obtain desensitized traceability data by scanning the product two-dimensional code, and the risk score is calculated in real time by the Bayesian network; the enterprise management end is simultaneously connected to the intelligent decision analysis data, automatically intercepts the goods with a risk score less than the set threshold, and simultaneously generates a recall work order to the manufacturer system; the regulatory platform is simultaneously connected to the intelligent decision analysis data, triggers the corresponding disposal process according to the risk score, and digitizes the whole process.

[0065] The above steps will be further described through a plurality of specific embodiments.

[0066] Embodiment one

[0067] In step S1, the multi-modal sensor array and related information collection device adopts a full-chain data collection network deployment, specifically including:

[0068] Production end: deploy near-infrared spectrometer (Bruker MPA, accuracy ± 0.01%) to detect pesticide residues / moisture, and deploy blockchain electronic lock (military-grade AES-256 encryption, supporting LoRa / Wi-Fi dual communication) to record production line ID and operation timestamp (accuracy ± 50ms).

[0069] Warehouse end: integrate pressure sensor (accuracy ± 0.5kg) and temperature and humidity sensor (SHT35, error ± 0.3℃ / ± 0.8% RH) into the shelf to collect inventory turnover rate and environmental data every 2 minutes.

[0070] Transportation end: install Beidou / GPS dual-mode terminal (positioning accuracy ± 2m) and vibration sensor (ADXL345, resolution ± 0.1g) on the cold chain vehicle to upload location, temperature and humidity (± 0.5℃ / ± 1% RH) and vibration acceleration data every 10 minutes.

[0071] Sales end: deploy two-dimensional code scanning device (supporting ISO 18000-6C protocol) to enable consumers to obtain desensitized traceability data and real-time risk score by scanning the code.

[0072] The working principle of the multi-modal data collection layer is as follows:

[0073] Space-time alignment technology: realize cross-link data timestamp unification through device clock synchronization module (NTP protocol, error ± 10ms), and realize spatial coordinate mapping of transportation track and warehouse location based on GPS / Beidou positioning (accuracy ± 5m).

[0074] Multi-source heterogeneous fusion: production end spectrum data (400-2500nm waveband, 1024-dimensional features) and transportation end vibration data (three-axis acceleration, 128-dimensional features) are reduced to 32-dimensional features through PCA, and are spliced with temperature and humidity (4-dimensional) and time (3-dimensional) features into a 160-dimensional input vector.

[0075] Embodiment two

[0076] This embodiment further describes step S2 above, and the data processing and dynamic storage process is as follows:

[0077] The edge computing node (NVIDIA Jetson AGX Orin, 200TOPS) cleans the data in real time: the spectral data is filtered and denoised by Savitzky-Golay, and the temperature and humidity data is filtered by the 3σ rule to remove outliers, and the standardized data stream is generated.

[0078] The "key event trigger + periodic packaging" uplink mechanism is adopted: the core data (component detection value, abnormal warning record) is uplinked in real time, and the regular data is packaged every 5 minutes (including timestamp, device ID, data hash value), and the PBFT consensus algorithm (consensus node ≤ 30) is used to ensure data consistency.

[0079] Embodiment three

[0080] As shown in Figure 2 , the embodiment further illustrates the above step S3, and the specific steps of step S3 include:

[0081] (1) Data uplink preprocessing: separate storage operation event data and detection data, operation event data includes data collected by related information collection equipment, and the detection data includes data collected by a multi-modal sensor array, and SHA-256 hash values are calculated for the two types of data respectively;

[0082] (2) Block construction: generate a block body containing double hash values, calculate the Merkle root of the block body and construct the block header;

[0083] (3) Weighted consensus verification: the consensus network is composed of enterprise nodes, regulatory nodes and third-party nodes, and the improved Byzantine fault tolerance algorithm is used for verification;

[0084] (4) Evidence storage execution: when the consensus passes, the new block is linked to the evidence chain within ≤2 seconds, and the data is tamper-proof.

[0085] Embodiment four

[0086] As shown in Figure 3 , the embodiment further illustrates the above step S3, and the specific steps include:

[0087] (1) Data preprocessing: the original data is divided into spectral data and temperature and humidity / vibration data, Savitxky-Golay filtering is performed on the spectral data to eliminate high-frequency noise; the 3σ rule is used to denoise the temperature and humidity / vibration original data, and the data points deviating from the mean value ±3σ are deleted;

[0088] (2) Cross-modal feature fusion: The spectral data is compressed from 1024 dimensions to 10 dimensions by PCA dimension reduction to obtain spectral principal component data, solving the problem of high-dimensional data redundancy; the sensor data is subjected to sliding window statistics to calculate the mean, variance and maximum to obtain environmental statistical features; the spectral principal component data, environmental statistical features, time features and spatial features are fused to generate a 21-dimensional input vector, and the 21-dimensional input vector is used to realize cross-modal feature standardization fusion;

[0089] (3) Attention-weighted model collaboration: based on the 21-dimensional input vector, the feature weight of each link is calculated through the attention mechanism, the weighted features are input into the random forest model, and the abnormal classification result and confidence are output;

[0090] (4) Dynamic risk prediction: when the confidence is greater than 0.8, the Bayesian network is activated to calculate the 72-hour risk probability;

[0091] (5) Decision output and synchronous push.

[0092] The cross-modal risk decision mechanism is as follows:

[0093] The random forest model (200 decision trees are used in this embodiment, Gini index splitting, and SMOTE oversampling is used to handle unbalanced data) identifies abnormal features (such as excessive pesticide residues, temperature and humidity exceeding the limit, and vibration amplitude > 2g) in real time, and the identification accuracy is ≥95%, and SMOTE oversampling overcomes the model bias caused by the scarcity of abnormal samples;

[0094] The Bayesian network constructs a three-level causal model (production-warehouse-transportation) to realize the quantification of cross-link risk transmission; the cross-link features (such as storage time 10 days + transportation vibration frequency > 50 times / minute) are input, and the 72-hour risk probability (such as mold probability ≥30% triggering red warning) is output, which is synchronously pushed to the manufacturer, distributor and regulatory department. Through the coupling of the random forest model and the Bayesian network, the detection-prediction decision-making closed loop is realized.

[0095] The working principle of the intelligent decision layer is as follows:

[0096] Cross-link risk transmission modeling: based on 3000+ historical accident data to train the Bayesian network conditional probability table, define “transportation vibration → packaging damage” (conditional probability 0.65), “packaging damage → microorganism breeding” (conditional probability 0.72) and other transmission rules to realize dynamic deduction of compound risks.

[0097] Dynamic threshold self-adaptation: independent risk assessment models are established for different categories (such as strawberries / milk), and online learning algorithms (model parameters are updated every 24 hours) are used to adapt to seasonal and regional differences, and the model generalization error is ≤8%.

[0098] Example Five

[0099] This embodiment will further illustrate the specific implementation process of the method:

[0100] 1. Production and storage stage:

[0101] The raw material pesticide residues are detected by a near-infrared spectrometer, and the data is bound with the blockchain electronic lock information (production batch, operator, equipment number) after edge node cleaning, to generate a traceability ID containing a 256-bit SHA-256 hash value (such as "PRO-20250418-001").

[0102] The warehouse shelf pressure sensor calculates the inventory turnover rate in real time (turnover rate = out-of-stock weight / shelf carrying capacity), and when the humidity > 85% and the turnover rate < 30% for 6 hours, it is automatically marked as a "high mold risk area", triggering the warehouse dehumidification equipment to start (response time ≤ 5 minutes).

[0103] 2. Transportation stage:

[0104] The cold chain vehicle terminal uploads data in real time, and if the temperature fluctuation exceeds ±2℃ and lasts for 15 minutes, the blockchain electronic lock will automatically lock the goods box (locking time ≤ 10 seconds), and at the same time, the Gaode map API will be called to obtain the real-time congestion index, and the Dijkstra algorithm will be used to generate the optimal detour route (path cost increase ≤ 15%).

[0105] The vibration sensor detects 3 consecutive impacts > 5g (such as sudden braking scenarios), and automatically marks "packaging risk events" and is associated with the subsequent sales end unboxing quality inspection link (triggering the probability of manual sampling to increase to 50%).

[0106] 3. Sales stage:

[0107] The consumer scans the product two-dimensional code, obtains desensitized traceability data through zero-knowledge proof technology, and the risk score (1-10 points) is calculated in real time by Bayesian network (in this embodiment, the risk score < 6 points is displayed as a yellow warning, and < 4 points is displayed as a red warning).

[0108] The supermarket POS system interfaces with the monitoring platform, automatically intercepts goods with a risk score < 4 points, and synchronously generates a recall work order (including problem link positioning, risk type, and recommended treatment measures) to the manufacturer's ERP system, and the work order response time is ≤ 30 minutes.

[0109] It should be noted that the reverse scoring logic (low score high risk) is used at the sales end (consumer perspective) to facilitate user intuitive understanding (such as "4 points or less" is more easily perceived as dangerous than "8 points or more").

[0110] 4. Closed-loop disposal process:

[0111] The regulatory department monitors in real time through the platform, and three abnormal events (risk score ≥ 8) automatically trigger the process of "enterprise self-check (report submitted within 2 hours) - third-party re-inspection (completed within 4 hours) - market withdrawal (executed within 6 hours)", the whole process is digitized and the disposal time is compressed by 80% compared with the traditional mode.

[0112] It should be noted that, unlike the sales stage, the regulatory end (enterprise / regulatory perspective) uses a positive scoring logic (high score, high risk), which is consistent with the risk level quantification management habit (such as 8 points corresponding to the highest risk level). The difference between the two is shown in the following table:

[0113] Scenario Score Range Risk vs Score Corresponding Action Sales End (Consumer) 1-10 Lower score = higher risk <4 Interdict product, trigger recall Regulatory End (Tier 3 Anomaly) 1-10 Higher score = higher risk ≥8 Trigger enterprise self-audit + 3rd party re-inspection

[0114] Example Six

[0115] Based on the design idea of the foregoing method, the present application also provides a full-chain food safety monitoring system based on multi-modal data fusion and intelligent algorithm, which comprises

[0116] The perception layer unit includes a multi-modal sensor array and related information acquisition equipment, which is used to synchronously acquire heterogeneous data of the physical state, environmental parameters and identity of the food in the circulation link, and each device is connected through an industrial bus;

[0117] The edge computing unit is an embedded device equipped with an AGX Orin chip, which has a data binning module, an AE-256-GCM encryption module and an abnormal initial diagnosis screening module, the data binning module is used to convert continuous numerical features into discrete features for model processing and understanding, the AE-256-GCM encryption module is used for real-time encryption of image data, and the abnormal initial diagnosis screening module is used for multi-source data protocol conversion;

[0118] The blockchain storage unit includes a consensus network composed of enterprise node servers, regulatory node hosts and third-party node clusters, and uses an improved Byzantine fault tolerance algorithm for verification, and the enterprise node server integrates an intelligent contract engine;

[0119] The intelligent decision unit includes a random forest anomaly detection module, a Bayesian network risk prediction module and a cross-modal data fusion module, which realizes PCA dimension reduction and attention weight allocation;

[0120] The application layer interaction unit includes an enterprise management end, a regulatory platform and a consumer client server, which are connected with the server of the intelligent decision unit through an API gateway.

[0121] In this embodiment, the hardware components of the multi-modal data acquisition of the perception layer unit are set as follows:

[0122] (1) Production end

[0123] Near-infrared spectrometer: Bruker MPA, supporting diffuse reflection collection, wavelength range 400-2500 nm, scanning speed 10 times / sec;

[0124] Blockchain electronic lock: Model SECURE-LOCK V2.0, supporting LoRa (transmission distance 1 km) / Wi-Fi dual mode, battery endurance 180 days (low power consumption mode).

[0125] (2) Transportation end

[0126] Vehicle-mounted terminal: integrated Ublox NEO-M8U positioning module (supports Beidou + GPS + GLONASS three satellite positioning), SHT30 temperature and humidity sensor (-40℃ ~ 85℃ working temperature), ADXL345 vibration sensor (±16g range, 13-bit resolution);

[0127] 5G communication module: supports NSA / SA dual mode, upload bandwidth ≥10Mbps, meets real-time data transmission requirements.

[0128] The edge computing architecture of the edge computing unit is set as follows:

[0129] (1) NVIDIA Jetson AGX Orin, 200TOPS computing power, supporting 50+ sensor concurrent access, localized data processing delay ≤100ms, supporting Docker containerized deployment of algorithm model.

[0130] (2) Communication protocol

[0131] Device level: Modbus RTU (RS485 interface) is used in production / warehousing link, supporting multi-master station communication, error checking rate ≤0.1%;

[0132] Wide area network: MQTT 5.0 protocol (QoS=2 to ensure reliable transmission) is used in transportation / sales link, data encryption uses AES-256-GCM mode, transmission error rate ≤10^-9.

[0133] In this embodiment, the dynamic blockchain storage key technology of the blockchain storage unit is as follows:

[0134] (1) Data on-chain strategy:

[0135] Core data (component detection value, abnormal early warning record, operation log): real-time on-chain, using event triggering mechanism (such as completing on-chain within 1 second when the detection value exceeds the standard);

[0136] Routine data (temperature and humidity, location trajectory): packed into chunks every 5 minutes (single chunk data volume ≤ 2 MB), including chunk header (previous hash + timestamp + Merkle root) and chunk body (data list), on-chain storage cost reduced by 40%.

[0137] (2) Consensus and security mechanism:

[0138] Improved PBFT algorithm (iPBFT) is adopted: consensus nodes are limited to enterprises (40%), regulatory authorities (30%), and third-party certification centers (30%), transaction confirmation time ≤ 2 seconds, Byzantine fault tolerance rate ≤ 33%;

[0139] Figure 4 For the execution graph of the smart contract, as shown in Figure 4 Data access control is achieved through hierarchical permissions (enterprise readable and writable, regulatory authority read-only audit, and consumer read-only desensitized data) by smart contract, and attribute-based encryption (ABE) technology is used to protect enterprise sensitive data (such as formula information).

[0140] The module architecture and functions in the intelligent decision unit are as follows:

[0141] (1) Data fusion process of cross-modal data fusion module:

[0142] a) Preprocessing layer:

[0143] Spectral data: Savitzky-Golay filtering (window size 11, polynomial order 2) to remove baseline drift, extract the first 10 principal components (cumulative variance contribution rate ≥ 95%);

[0144] Time series data (temperature and humidity, vibration): calculate 6-dimensional statistical features such as mean, variance, and maximum value through sliding window (window size 10 minutes).

[0145] b) Feature fusion layer:

[0146] Spatiotemporal feature concatenation: combine spectral principal components (10-dimensional), environmental statistical features (6-dimensional), temporal features (hour / week / month, 3-dimensional), and spatial features (latitude and longitude, 2-dimensional) into a 21-dimensional input vector;

[0147] Cross-link correlation: calculate the correlation weight between production link pesticide residue value and storage link humidity through attention mechanism (Attention) (e.g., when pesticide residue exceeds the standard, humidity influence weight increases by 30%).

[0148] (2) Random forest anomaly detection module: random forest model (200 trees, maximum depth 15, Gini index splitting), use feature importance ranking (Mean Decrease Impurity) to select key risk factors (e.g., transportation vibration frequency importance score 0.32).

[0149] (3) Bayesian Network Risk Prediction Module: Bayesian Network constructs a directed acyclic graph (DAG) with 30+ nodes (e.g. “production batch” “warehouse humidity” “transportation duration”), trains conditional probability tables with historical accident data (3217 cases from 2018-2025), and supports cross-link risk transmission path visualization (e.g. “warehouse high temperature → microorganism reproduction → terminal deterioration” path probability 0.68).

[0150] For ease of understanding, the system components and their functions are listed as follows:

[0151]

[0152] Finally, it should be pointed out that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm, characterized in that: The method comprises the following steps: Step S1: perception layer data acquisition: synchronously acquiring heterogeneous data of the physical state, environmental parameters and identity of the food in the circulation link by using a multi-modal sensor array and related information acquisition equipment; Step S2: edge computing data processing: constructing an edge computing layer, and using an NVIDIA Jetson AGX Orin node to perform real-time aggregation on the heterogeneous data of step S1, wherein image data is encrypted by using an AE-256-GCM algorithm, and is subjected to protocol conversion and data standardization by an abnormal initial diagnosis screening module; Step S3: blockchain storage: uploading the standardized data to a blockchain network, and jointly verifying the data by enterprise nodes, supervision nodes and third-party nodes, and generating a time-stamped storage log chain by an intelligent contract engine; Step S4: intelligent decision analysis: performing abnormal detection on the storage data by using a random forest model, and simultaneously performing dynamic prediction on the risk probability by using a Bayesian network, and generating and outputting intelligent decision analysis data including a risk score; Step S5: application layer response: in a consumer client, obtaining desensitized traceability data by scanning a product two-dimensional code, and calculating a risk score in real time by using a Bayesian network; an enterprise management end is synchronously connected to the intelligent decision analysis data, and automatically intercepts a product on the market whose risk score is less than a set threshold, and synchronously generates a recall work order to a manufacturer system; a supervision platform is synchronously connected to the intelligent decision analysis data, and triggers a corresponding disposal process according to the risk score, and leaves a digital mark in the whole process.

2. The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm according to claim 1, characterized in that: In the step S1, the multi-modal sensor array comprises a near-infrared spectrometer, a temperature and humidity sensor, a vibration sensor and a pressure sensor, and the related information acquisition equipment comprises a blockchain electronic lock, a Beidou / GPS dual-mode terminal and a two-dimensional code scanning device. 3.The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm according to claim 1, characterized in that: In the step S3, the specific steps comprise: (1) data on-chain preprocessing: separating and storing operation event data and detection data, the operation event data comprising data collected by the related information acquisition equipment, and the detection data comprising data collected by the multi-modal sensor array, and calculating SHA-256 hash values of the two types of data; (2) block construction: generating a block body containing double hash values, calculating a Merkle root of the block body and constructing a block header; (3) weighted consensus verification: a consensus network is composed of enterprise nodes, supervision nodes and third-party nodes, and an improved Byzantine fault tolerance algorithm is used for verification; (4) storage execution: when the consensus passes, a new block is linked to the storage chain within ≤2 seconds, and the data is tamper-proof. 4.The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm according to claim 1, characterized in that: In the step S4, the specific steps comprise: (1) data preprocessing: dividing the original data into spectrum data and temperature and humidity / vibration data, performing Savitxky-Golay filtering on the spectrum data to eliminate high-frequency noise; and removing data points deviating from the mean value ±3σ from the temperature and humidity / vibration original data by using a 3σ rule for noise removal; (2) Cross-modal feature fusion: The spectral data is compressed from 1024 dimensions to 10 dimensions by PCA dimension reduction to obtain spectral principal component data; the sliding window statistics are performed on the sensor data to calculate the mean, variance and maximum to obtain environmental statistical features; the spectral principal component data, environmental statistical features, time features and spatial features are fused to generate a 21-dimensional input vector: (3) Attention-weighted model collaboration: Based on the 21-dimensional input vector, the feature weights of each link are calculated by the attention mechanism, and the weighted features are input into the random forest model to output the abnormal classification result and confidence; (4) Dynamic risk prediction: When the confidence is greater than 0.8, the Bayesian network is activated to calculate the 72-hour risk probability; (5) Decision output and synchronous push.

5. The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm according to claim 4, characterized in that: The Bayesian network includes a three-level causal model trained based on at least 3000 historical accident data, and the hierarchical structure of the three-level causal model includes: First-level nodes: production links; Second-level nodes: storage links; Third-level nodes: transportation links. 6.The full-chain food safety monitoring method based on multi-modal data fusion and intelligent algorithm according to claim 4, characterized in that: The random forest model uses the SMOTE oversampling technique to handle the sample imbalance problem during training, and the node splitting criterion is that the Gini index variation is greater than 0.

25.

7. The whole-chain food safety monitoring system based on multi-modal data fusion and intelligent algorithm, characterized in that: The system includes a perception layer unit: containing a multi-modal sensor array and related information acquisition equipment, for synchronously acquiring heterogeneous data of the physical state of food in the circulation link, environmental parameters and identity, and each device is connected through an industrial bus; an edge computing unit: an embedded device equipped with an AGX Orin chip, with a data binning module, an AE-256-GCM encryption module and an abnormal initial diagnosis screening module, the data binning module is used to convert continuous numerical features into discrete features for model processing and understanding, the AE-256-GCM encryption module is used for real-time encryption of image data, and the abnormal initial diagnosis screening module is used to realize multi-source data protocol conversion; a blockchain storage unit: including a consensus network composed of enterprise node servers, regulatory node hosts and third-party node clusters, using an improved Byzantine fault tolerance algorithm for verification, the enterprise node server integrates an intelligent contract engine; an intelligent decision unit: the server includes a random forest processing module, a Bayesian network calculation module and a feature fusion module, realizing PCA dimension reduction and attention weight distribution; an application layer interaction unit: including an enterprise management end, a regulatory platform and a consumer client server, which are connected with the server of the intelligent decision unit through an API gateway. 8.The full-chain food safety monitoring system based on multi-modal data fusion and intelligent algorithm according to claim 7, characterized in that: In the blockchain storage unit, the enterprise node server is configured with a double-hash generator; the third-party node cluster is deployed with a Merkle tree verification FPGA.

Images