Fruit and vegetable quality dynamic monitoring system and method based on KAN network and cloud side end cooperation

Abstract

The invention discloses a fruit and vegetable quality dynamic monitoring system and method based on KAN network and cloud side end cooperation. The system comprises an equipment end, an edge end and a cloud end, a sensing layer collects fruit and vegetable storage and transportation environment data, and the edge end utilizes a KAN enhancement model to infer fruit and vegetable rot indexes in real time; and the cloud completes model training, parameter aggregation and hash abstract verification. The method comprises the following steps that: a KAN enhancement model adopts three types of time sequence architectures, namely LSTM, BiLSTM and TCN, and a KAN layer is fused at the rear ends of the three types of time sequence architectures; carrying out weight amplitude importance-based sorting on the trained KAN enhancement model, and pruning a target layer according to a set proportion; and collecting real-time data in a fruit and vegetable storage and transportation environment, after preprocessing and interval optimization, inputting the data into the trained and pruned KAN enhancement model, and outputting a fruit and vegetable rot index. Intelligent and credible management of the whole process of fruit and vegetable quality monitoring from data acquisition, modeling to early warning is realized.

Description

Technical Field

[0001] This invention belongs to the field of smart agriculture and Internet of Things sensing technology. Specifically, it relates to a dynamic monitoring system and method for fruit and vegetable quality that combines multi-gas sensing, a deep time-series model based on Kolmogorov-Arnold Network (KAN) enhancement, and a cloud-edge-device collaborative deployment strategy. It is applicable to online freshness monitoring and early warning of perishable agricultural products such as fruits and vegetables during storage, warehousing, and transportation. Background Technology

[0002] Fruits and vegetables are among the most consumed agricultural products globally, and their post-harvest storage and distribution processes directly impact the economic efficiency of the supply chain and food safety levels. After harvesting, fruits and vegetables retain their living physiological characteristics, continuously releasing carbon dioxide through respiration and metabolic processes. ), ethylene ( Fruits and vegetables are highly susceptible to spoilage in cold storage or transportation environments due to factors such as temperature and humidity, gaseous accumulation, microbial contamination, and mechanical stress. This can lead to quality deterioration, browning, off-odors, and microbial decay. Especially in long-term storage and large-scale distribution scenarios, even minor fluctuations in environmental parameters can trigger a surge in respiration intensity, metabolic abnormalities, or pathogen spread, making it difficult to accurately identify the spoilage process in its early stages. This makes spoilage detection and prediction a key scientific issue and a major pain point in fruit supply chain management.

[0003] Currently, fruit and vegetable spoilage monitoring relies heavily on univariate methods such as temperature and humidity, neglecting the multidimensional coupling characteristics of the spoilage process. Studies have shown that... An increase in concentration often indicates vigorous respiration or the onset of anaerobic metabolism. A decrease may indicate increased oxygen consumption or abnormal gas exchange within tissues; As key hormones inducing maturation and aging, VOCs significantly accelerate softening and decay; changes in VOC types and concentrations reflect cell rupture, enzymatic oxidation, and fungal metabolic processes. Therefore, the combined dynamic changes of multiple gas components are more effective than single variables in reflecting the potential risk of putrefaction. However, these gas signals exhibit significant nonlinearity, strong coupling, and weak prior knowledge; their changes are influenced by temperature and humidity, substrate type, pathogen type, and mechanical damage, presenting complex temporal patterns. This makes it difficult for traditional linear models to achieve high-precision predictions.

[0004] On the other hand, existing corruption prediction models generally rely on centralized computing, transmitting all sensor data to the cloud for prediction. As storage facilities, processing plants, and logistics vehicles continue to expand, the number of sensors is growing exponentially, leading to a simultaneous increase in cloud transmission load and computational pressure. This results in increasingly prominent issues related to real-time performance, bandwidth costs, and data security risks. While some deep learning models perform well in experimental environments, their complex structures and large number of parameters make them difficult to deploy directly on resource-constrained edge devices. High cloud inference latency and insufficient edge inference computing power make it difficult for corruption prediction to operate stably in real-world scenarios. Furthermore, difficulties in model updates and inconsistent data distribution across different nodes in a multi-point distributed environment also limit the scalability and engineering feasibility of existing systems.

[0005] From a modeling perspective, traditional deep learning models (such as LSTM, GRU, and TCN) still generally employ fixed activation functions and linear weight mapping when processing multi-gas signals. The spoilage process of fruits and vegetables involves multiple physiological stages (from a stable phase to a respiratory climacteric phase and then to a rapid spoilage phase), with each stage exhibiting different nonlinear characteristics. The fixed function space of traditional neural networks often struggles to flexibly adapt to these stages. Furthermore, the models suffer from severe parameter redundancy, significant training oscillations, slow convergence, insufficient interpretability, high sensitivity to noise, and unstable generalization ability, all of which limit their practical deployment.

[0006] In engineering applications, traditional monitoring equipment is generally large in size, consumes a lot of energy, and has a single data interface. It requires professional maintenance and is not suitable for large-scale, distributed, and long-term deployment. Systems lacking cloud-edge-device collaboration capabilities cannot achieve a collaborative closed loop of real-time edge inference, cloud training and updates, and lightweight terminal operation. In addition, the long-term reliance on cloud transmission and storage for large amounts of environmental data not only puts pressure on bandwidth but also makes it difficult to meet data security and privacy requirements in some scenarios.

[0007] In summary, the main challenges of monitoring and predicting fruit and vegetable spoilage at present are as follows: (1) The spoilage process is significantly coupled with multiple factors, and there are complex nonlinear relationships between gas composition, temperature and humidity and physiological processes, which are difficult to effectively model using traditional methods; (2) A single environmental parameter cannot support high-precision spoilage prediction, and deep fusion based on multi-gas, multi-modal and multi-time series information is required; (3) The deep model structure is huge, difficult to adapt to edge devices, lacks real-time performance, and has high deployment costs; (4) The centralized computing method in the cloud results in strong network dependence, high latency and weak data security, which is not conducive to large-scale implementation; (5) The interpretability and robustness of existing spoilage prediction models are insufficient, and their adaptability to different storage and transportation environments is limited. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a dynamic monitoring system and method for fruit and vegetable quality based on KAN network and cloud-edge-device collaboration. This system balances real-time performance, prediction accuracy, model interpretability, and engineering deployability, aiming to improve the early spoilage identification capability during storage and transportation, reduce losses, and ensure food safety.

[0009] By combining multi-gas sensing, multimodal temporal modeling, KAN functionalized edge mapping, structured pruning, and cloud-edge collaborative management based on Kubernetes / KubeEdge, this invention accurately captures the nonlinear coupling relationships of the spoilage process while meeting the engineering requirements of real-time inference from edge devices and long-term adaptive updates of the system. The system is suitable for online quality monitoring and early warning of perishable fruits and vegetables in warehousing, transportation, and cold chain scenarios.

[0010] Note that the description of these objectives does not preclude the existence of other objectives. One aspect of the invention does not require achieving all of the above objectives. Objectives other than those described above can be extracted from the description, drawings, and claims.

[0011] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0012] A dynamic monitoring system for fruit and vegetable quality based on KAN network and cloud-edge-device collaboration includes a perception layer, an edge computing layer, and a cloud optimization layer.

[0013] The sensing layer includes temperature and humidity sensors and gas concentration sensors, which are used to collect multi-dimensional data in the fruit and vegetable storage and transportation environment.

[0014] The edge computing layer uses an i.MX6UL processor and integrates a 4G communication module. The data collected by the perception layer is preprocessed and optimized in intervals before being used as input to the KAN augmentation model. The KAN augmentation model performs real-time inference and outputs the fruit and vegetable spoilage index.

[0015] The cloud optimization layer performs model training, parameter aggregation, and hash digest verification, and then sends the optimized KAN enhanced model back to the edge nodes.

[0016] A method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration:

[0017] A KAN augmentation model is constructed, which adopts three types of time-series architectures: LSTM, BiLSTM and TCN. The KAN network layer is used to replace the fully connected layer in the back end of the three types of time-series architectures.

[0018] Train the KAN augmentation model in the cloud, sort the trained KAN augmentation model based on the importance of weight magnitude, and prune the target layer according to a set ratio.

[0019] Real-time data from the storage and transportation environment of fruits and vegetables is collected, preprocessed, and optimized for intervals. The data is then input into a trained and pruned KAN-enhanced model, which outputs a fruit and vegetable spoilage index.

[0020] Furthermore, the KAN enhancement model includes the LSTM-KAN model, the BiLSTM-KAN model, and the TCN-KAN model; the LSTM-KAN model consists of a two-level long short-term memory network, used to extract dynamic features across time; the BiLSTM-KAN model consists of a forward long short-term memory network and a backward long short-term memory network, used to extract dynamic features of the time series on the forward and backward time axes, respectively; and the TCN-KAN model is used to extract local temporal convolution features.

[0021] Furthermore, the functionalized mapping layer of the KAN network uses cubic B-spline basis functions to construct a functionalized edge connection structure, and maps the output. ,in Let k be the basis function of the B-spline. Here, M represents the number of splines and the learnable weights.

[0022] Furthermore, after pruning, the root hash digest of the model before pruning is calculated. Calculate the Merkle root summary of the model after pruning. After pruning, the model divides the weights into n blocks, and the local hash of each block is: , ;in, This is a summary of the model roots before pruning. This represents the complete set of weights consisting of all network weight parameters saved after training, before the model performs structured pruning. This represents the i-th subset of weight parameters obtained by dividing the entire set of network weight parameters retained after pruning according to a predetermined block rule after the model has completed structured pruning; || represents the hash join operation.

[0023] Execute the inheritance verification function Verify( , If the function returns True, it confirms that the source of the pruned model is reliable and the inheritance chain is valid.

[0024] Furthermore, edge nodes rely on local data Calculate the gradient of the loss function Generate local increments And upload it to the cloud during the synchronization period; the cloud uploads it based on the node weight. Complete weighted aggregation to form a global model. Cloud-based root hash calculation and upload hash with the node Compare; if This will automatically trigger the model re-aggregation process; among which, The learning rate is adapted for this edge node. This represents the parameter matrix of the KAN augmentation model in the cloud after the t-th round of training, where N is the number of edge nodes. For node weight coefficients, This is the hash tolerance threshold for performance summary comparison.

[0025] Furthermore, the input feature vector of the trained and pruned KAN augmentation model Output fruit and vegetable spoilage risk prediction values The prediction results generate a hash digest. The data is uploaded to the cloud and incremental parameter training is performed; among them, The set of variable intervals with the lowest RMSE. This is the complete set of weights consisting of all network weight parameters of the current KAN augmentation model.

[0026] Furthermore, the preprocessing includes temporal alignment and synchronization correction, denoising and outlier removal, normalization and missing value completion. The interval optimization involves dividing the preprocessed data into a preset number of intervals, with each interval corresponding to a set of statistical features, thereby generating several candidate sets of interval combinations. For each candidate set of combinations, partial least squares regression is used in the cloud to model the validation set root mean square error curve through cross-validation. The root mean square error corresponding to different numbers of latent variables is recorded. By extracting local minima and sorting them, several high-performance interval combinations are selected. The set of variable intervals that minimizes the cross-validation root mean square error is selected as the input to the trained and pruned KAN augmentation model.

[0027] Furthermore, a smoothing regularization term is introduced after each round of training. Suppress oscillations between adjacent weights; the joint loss function for training is defined as... Minimize the updated parameters through backpropagation After the update, a hash digest is generated. Used for cloud storage; cloud training employs a gradient descent algorithm with momentum. And generate a parameter summary after each round of training. The summary is written to a trusted log chain along with a timestamp, enabling traceability of the training process and rollback in case of anomalies; among which, The regularization coefficient is . For learnable weights, This represents the mean squared error loss between the model's predicted output and the measured values ​​of fruit and vegetable quality. This refers to the complete set of weights consisting of all network weight parameters of the current KAN augmentation model. For learning rate, This is the momentum coefficient.

[0028] Furthermore, a set of cloud computing performance metrics If twice in a row Then, retraining and canary deployment will be performed, where, Let be the coefficient of determination in the t-th training iteration. Let be the root mean square error in the t-th training iteration. The threshold value is used.

[0029] Through the above steps, this method realizes an integrated design of the entire process from data acquisition, feature modeling, cloud-edge collaborative training, energy consumption control, performance evaluation to trusted verification. Its core lies in using the functional mapping of the KAN network and the blockchain hash verification mechanism to achieve high-precision prediction of fruit and vegetable quality changes and trusted tracking of model evolution, which significantly improves the intelligence, stability and feasibility of the system.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. This invention comprises three parts: an edge data acquisition device, an edge computing node, and a cloud management platform, constructing a cloud-edge-device collaborative dynamic monitoring system for fruit and vegetable quality. The edge device collects temperature and humidity data in real time through multi-source sensors. , , and Environmental parameters are analyzed, and a local data preprocessing module is used to perform time-series alignment and synchronization correction, noise reduction and outlier removal, normalization and missing value completion. Edge nodes perform real-time quality status inference based on the KAN augmented model, and upload the inference results and key monitoring data to the cloud periodically. The cloud platform aggregates data from multiple nodes, removes outliers, assesses feature importance, and retrains the model. The system automatically triggers multi-level risk warnings based on predicted values ​​and threshold ranges, and distributes the optimized model to each edge node, realizing cloud-edge collaborative monitoring and warning. The system ensures the credibility of the entire data and model transmission process through blockchain hash digest and Merkle Tree verification mechanisms, realizing closed-loop management from environmental monitoring and status assessment to risk warning, and possessing high real-time performance, high stability, and high prediction accuracy.

[0032] 2. This invention deeply integrates multi-source environmental sensing technology with edge intelligent inference. Addressing the problems of existing monitoring methods relying on single parameters, high response latency, and poor adaptability to fluctuating environments, it designs a multi-parameter sensor array to achieve temperature, humidity, and... , , and Multi-dimensional, high-frequency acquisition of key factors. The KAN network enhances the model's nonlinear expressive power by utilizing functional mapping and spline basis function structures, maintaining stable prediction output even when dealing with sensor noise and data heterogeneity. After each round of model training, a smoothing regularization term and confidence constraint mechanism are introduced into the loss function to limit excessive fluctuations in the prediction output, thereby improving the model's prediction accuracy and robustness in fruit and vegetable quality prediction. Compared to traditional schemes based on fixed regression or shallow neural networks, this invention can maintain high-precision quality change prediction capabilities over a long period in dynamic cold chain environments.

[0033] 3. This invention combines a cloud-edge collaborative architecture with an edge computing model. Through structured pruning and model compression strategies, it optimizes the size and computational load of the trained deep learning model, enabling the model to be successfully deployed and run on low-cost, low-computing-power edge devices. The cloud handles model training, structured pruning, and model version iteration management. The pruning results are verified using SHA256 and Merkle Tree to form a trusted version, which is then distributed to edge nodes, enabling model updates, lightweighting, and security management. After receiving the lightweight model, the edge nodes execute on-site inference, achieving localized real-time prediction and effectively reducing dependence on cloud computing power and network bandwidth. Through a collaborative approach of centralized cloud management and local execution at the edge, this invention achieves consistent deployment and efficient operation of multi-node edge devices, improving inference speed and resource utilization efficiency, and is suitable for quality dynamic monitoring tasks in large-scale edge scenarios.

[0034] 4. This invention leverages the Kubernetes and Kube Edge framework to build a unified cloud-based collaborative management system, supporting integrated control of cloud training, model pruning, version release, and batch distribution. The cloud platform can centrally schedule and monitor the status of edge nodes in different cold chain, warehousing, or transportation scenarios, enabling automatic model updates and health checks. Edge nodes perform online inference based on deployed lightweight models and periodically transmit results for continuous training and model self-iteration. The system utilizes hash signatures and blockchain digest comparisons during task distribution and model migration to ensure operational security and traceability. Compared to traditional distributed monitoring systems, this invention possesses strong scalability, high maintainability, and adaptive learning capabilities, enabling efficient collaborative operation in multi-site, multi-node environments, significantly enhancing the engineering application value and long-term availability of fruit and vegetable quality monitoring systems.

[0035] Note that the description of these effects does not preclude the existence of other effects. An embodiment of the invention does not necessarily have all the aforementioned effects. Effects other than those described above can be readily observed and extracted from the description, drawings, claims, etc. Attached Figure Description

[0036] To more intuitively illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings required for the embodiments or related technologies are now briefly described. Obviously, the drawings described below are only a part of the embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without creative effort.

[0037] Figure 1 This is a diagram illustrating the overall architecture of a dynamic monitoring system for fruit and vegetable quality based on KAN network and cloud-edge-device collaboration as described in this invention.

[0038] Figure 2 This is a hardware architecture diagram of the edge computing gateway for dynamic monitoring of fruit and vegetable quality as described in this invention.

[0039] Figure 3 This is a software architecture diagram of the edge computing gateway for dynamic monitoring of fruit and vegetable quality as described in this invention.

[0040] Figure 4 This is an architecture diagram of the intelligent monitoring system for fruit and vegetable storage environment based on a cloud-edge-device collaborative system as described in this invention.

[0041] Figure 5 This is a structural diagram of the fruit and vegetable quality detection model based on KAN optimization described in this invention. Detailed Implementation

[0042] To more clearly illustrate the objectives, technical solutions, and advantages of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings and using fruits and vegetables as the detection objects. The examples described are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0043] Example 1

[0044] An edge computing gateway system for monitoring and predicting the quality of fruit and vegetable storage environments is provided. This system is used for real-time acquisition of multi-source environmental data, edge intelligent processing, cloud-based collaborative management, and deployment of predictive models in fruit and vegetable storage and cold chain logistics scenarios. Figure 1 As shown, the system integrates an edge computing gateway hardware platform, an edge-side intelligent analysis software framework, and a cloud-based Kubernetes cluster to build a collaborative system from device-side perception and edge computing to cloud-based decision optimization, thereby realizing a dynamic management mechanism for real-time perception, local decision-making, cloud-based optimization, and continuous upgrades.

[0045] like Figure 2As shown, the fruit and vegetable storage environment monitoring gateway system includes at least one set of sensing layer interface modules, network communication interface modules, edge computing processing modules, and edge-cloud communication modules. The sensing layer interface modules are used to access environmental sensor devices through various industrial interface protocols, including temperature and humidity sensors, gas concentration sensors, etc. , , The system collects key parameters from the storage environment, including VOCs; the network communication interface module is used to realize communication between the gateway and the cloud platform, including a 4G module, a WiFi module and an Ethernet interface; the edge computing processing module realizes lightweight model inference on the edge side based on the .NET cross-platform running framework; the edge-cloud communication module is used to realize data interoperability between the fruit and vegetable storage environment monitoring gateway system and the IoT cloud platform, preferably using the MQTT protocol for efficient and low-power data transmission.

[0046] Among them, see Figure 2 The hardware system of the fruit and vegetable storage environment monitoring gateway includes a shell, a main control processing unit (i.MX6UL processor), an RS232 interface, an RS485 interface, an I2C interface, a ZigBee communication module, a 4G communication module, a WiFi module, an Ethernet interface, a power management module, and a debugging serial port. The RS232 interface, RS485 interface, I2C interface, and ZigBee communication module are used for wired / wireless data interaction with the sensing layer sensors; the 4G communication module, WiFi module, and Ethernet interface are used for cloud communication and data upload; the debugging serial port is used for system maintenance and firmware upgrades; and the power management module provides regulated power to the entire system.

[0047] Through the integrated design of the above hardware components, the fruit and vegetable storage environment monitoring gateway system can operate stably in the storage environment, realizing real-time acquisition, uploading and local analysis of multi-source parameters such as temperature and humidity, oxygen concentration, carbon dioxide concentration, ethylene concentration and VOC concentration.

[0048] The outer shell is made of high-strength ABS engineering plastic and reinforced with a metal shielding layer to enhance the device's shock resistance, dustproofing, and electromagnetic interference shielding performance, ensuring the long-term stable operation of the fruit and vegetable storage environment monitoring gateway system in complex environments such as cold chain storage. The shell structure adopts a modular design, resulting in a more rational interface layout and better heat dissipation performance.

[0049] The main control processing unit is based on the NXP i.MX6UL high-efficiency processor, which adopts the ARM Cortex-A7 architecture and features low power consumption, high stability, and strong real-time performance. This processor has a rich set of built-in peripheral interfaces, facilitating flexible integration with various industrial communication protocols. It is suitable for data acquisition and lightweight inference tasks in fruit and vegetable storage environment monitoring gateway systems.

[0050] The sensing layer interface module includes an RS232 interface, an RS485 interface, an I2C interface, and a ZigBee communication module. The RS232 interface is suitable for point-to-point real-time data acquisition from temperature and humidity sensors; the RS485 interface uses the ModBus-RTU protocol to acquire data on oxygen and carbon dioxide; the I2C interface enables networked data acquisition from ethylene and VOC gas sensors; and the ZigBee communication module connects wireless sensor nodes, enabling flexible deployment in large-scale warehousing environments.

[0051] The network communication interface module includes a 4G communication module, a WiFi communication module, and an Ethernet interface. The 4G communication module is used to report data to the cloud in the absence of a fixed network; the WiFi communication module is suitable for the local area network environment of the warehouse area to achieve high-speed data transmission; and the Ethernet interface ensures stable communication in high-reliability scenarios.

[0052] The fruit and vegetable storage environment monitoring gateway also includes an independent power management module, which is used to realize overvoltage protection, temperature control protection and automatic power supply switching mechanism, so as to ensure that the system can still maintain basic acquisition tasks and data caching in the event of power grid fluctuations or sudden power outages.

[0053] like Figure 3 As shown, the software architecture of the fruit and vegetable storage environment monitoring gateway system is built on the cross-platform .NET framework and consists of a data layer, a controller layer, an interface driver layer, and an edge computing layer. The software architecture is highly portable and can run stably on Linux and embedded operating systems, without being limited by hardware platforms.

[0054] The data layer is used to implement data access, data storage and data processing functions. It achieves efficient reading and writing of sensor data and storage of historical information through a built-in caching mechanism and a local database.

[0055] The controller layer includes a device controller, a variable controller, and a system configuration controller. The device controller is used to uniformly manage the status of sensor devices and data acquisition tasks. The variable controller is used to manage the thresholds, statuses, and alarm configurations of various sensor variables. The system configuration controller is used to maintain the operating parameters of the gateway system.

[0056] The interface driver layer includes device address and status interfaces, method interfaces, and Mod Bus driver modules, which can realize unified management of the underlying sensors. By encapsulating protocol details through the driver layer, upper-layer services do not need to care about the underlying communication protocol, thus improving system scalability.

[0057] The edge computing layer further integrates the Kube Edge edge computing framework, including Edge Core and CloudCore components. Edge Core is deployed at the edge gateway, supporting device management, data collection, edge inference, and service scheduling; Cloud Core is deployed in a Kubernetes cluster in the cloud, enabling device lifecycle management, model synchronization, and application deployment.

[0058] like Figure 4 As shown, the edge computing layer comprises three parts: the cloud layer, the edge layer, and the device layer.

[0059] The equipment layer collects real-time warehouse environment data through multi-source sensors, including temperature and humidity. concentration, The concentrations of ethylene, VOCs, and other organic compounds are recorded and uploaded to the edge gateway via RS485, RS232, I2C, or ZigBee interfaces.

[0060] The edge layer includes an edge communication module, a metadata manager, an event bus, and a container manager. The edge communication module enables bidirectional communication with the device layer for real-time data access; the metadata manager maintains device information, sensor attributes, and status; the event bus handles data flow and triggers business logic; and the container manager deploys lightweight machine learning models (i.e., KAN-enhanced models) to achieve local intelligent prediction, including functions such as corruption risk prediction and environmental anomaly detection.

[0061] The cloud layer includes a Kubernetes controller, a scheduler, a cloud device management service, and a model training service. The Kubernetes cluster is used for unified management of cloud resources, enabling containerized application deployment. The scheduler allocates model training tasks, data processing tasks, and management services based on cluster resource usage status, service priorities, and load conditions to achieve efficient utilization of cloud resources and stable system operation. The cloud device management service handles device registration, status monitoring, and policy distribution. The model training service trains LSTM, BiLSTM, TCN, and KAN-integrated prediction models based on extensive historical environmental data and distributes the updated models to the edge via the Kube Edge framework.

[0062] The cloud layer and edge layer use the MQTT protocol to achieve lightweight data uploading and command issuance. The Mapper driver adaptation layer transmits sensor data to the cloud in a standardized format, improving the system's compatibility and scalability.

[0063] Through the aforementioned cloud-edge-device collaborative design, the system achieves real-time perception, local intelligent prediction, centralized cloud optimization, and continuous model upgrades. The edge device can independently complete prediction tasks even in unstable network environments, improving system reliability; while the cloud utilizes massive computing power to optimize models and perform global scheduling, thus constructing a highly robust and real-time intelligent warehouse environment monitoring system.

[0064] Example 2

[0065] This embodiment uses apples as the subject, and is based on the fruit and vegetable storage environment monitoring gateway system described in Embodiment 1, combined with the temperature and humidity data collected in Embodiment 1. , , This invention utilizes multi-source time-series environmental data such as VOCs to construct and deploy a time-series deep learning prediction and lightweight deployment scheme for cold chain / warehousing environments. The method focuses on (1) preprocessing and feature construction of multi-source data; (2) time-series feature extraction based on LSTM, BiLSTM and TCN; (3) replacing the traditional fully connected layer with Kolmogorov-Arnold Network (KAN) as the output mapping module to improve nonlinear expression capability; (4) using structured pruning to optimize the model for edge optimization and distributing it to resource-constrained edge nodes for real-time inference and early warning; and (5) realizing model training, version management and distribution updates in a cloud-edge collaborative manner. This embodiment forms a closed-loop workflow of "collection-correction-feature extraction-KAN mapping-prediction-early warning" between the remote monitoring and early warning platform and the edge inference node, ensuring both prediction accuracy and deployment feasibility. This invention will describe the purpose and implementation principle of each part in detail, and cite the experimental results obtained by this invention at key points to verify the effectiveness and quantifiable benefits of this embodiment.

[0066] Multi-source data preprocessing aims to transform the raw high-frequency time-series signals uploaded from the fruit and vegetable storage environment monitoring gateway terminal (for apple storage environment monitoring and quality prediction) into stable inputs usable for fruit and vegetable quality detection models (i.e., KAN-enhanced models). The raw samples recorded temperature, relative humidity, and other parameters at a frequency of 1 Hz. concentration, concentration, The system includes channels for concentration and VOC response. First, the data undergoes time-series alignment and synchronization correction, followed by denoising and outlier removal. This involves smoothing high-frequency noise using a moving window-based median filter and a low-pass filter, and for sudden anomalies, employing local interpolation and threshold-based removal strategies to eliminate outliers caused by sensor drift or communication jitter. Multi-source data preprocessing and feature construction simultaneously normalize the data (based on the mean and standard deviation of the training set) and perform missing data completion (linear or piecewise interpolation) to ensure consistency and numerical stability between training and online inference. Edge-side preprocessing reduces uplink bandwidth and improves data quality for cloud training, facilitating subsequent feature selection and modeling.

[0067] Interval optimization: This is a high-order variable selection operation performed in the cloud or at the edge to remove redundant features and noisy variables and generate a high-quality input set. Based on the experimental process, this embodiment adopts an interval-based feature partitioning strategy: the preprocessed data is divided into preset intervals (e.g., short-term 5s, medium-term 30s, long-term 300s), with each interval corresponding to a set of statistical features, thereby generating several candidate sets of interval combinations. For each candidate combination, partial least squares (PLS) regression is used for modeling in the cloud, and the root mean square error (RMSE) curve of the validation set is calculated through cross-validation. The RMSE corresponding to different numbers of latent variables is recorded, and several high-performance interval combinations are selected by extracting local minima and sorting them. Finally, the variable interval set that minimizes the cross-validation RMSE (RMSECV) is selected as the input for the fruit and vegetable quality detection model to ensure that the input covers temperature and humidity, etc. , , The process optimizes key dimensions such as VOC (Volume, Orientation, and Consistency) while removing non-informative or noisy channels to the greatest extent possible. This interval optimization process has been proven to reduce input dimensionality and training complexity while maintaining information integrity, thus ensuring the training convergence and generalization ability of downstream models. Furthermore, it provides input data for fruit and vegetable quality detection models. Generate timestamps and calculate SHA256 hash digests. This is used for subsequent cloud data consistency verification and log chain recording, achieving trusted identification before data upload. In addition, each interval selection and verification process generates corresponding optimization log records, including parameter selection, verification error, and optimal interval set identifier, which are written to the trusted log database in the form of hash digests. The cloud compares and verifies these digests to ensure the traceability and verification consistency of feature selection results.

[0068] like Figure 5As shown. For temporal feature extraction, experiments were conducted using three temporal architectures: LSTM, BiLSTM, and TCN, which were then used as candidate frameworks. LSTM and BiLSTM focus on capturing long-term dependencies and gated memory mechanisms to express nonlinear temporal sequences, making them suitable for processing temperature, humidity, and gas accumulation signals with significant long-range correlations. TCN efficiently extracts local and cross-scale temporal dependencies through one-dimensional causal convolution and dilated convolution structures, demonstrating excellent parallel training capabilities and gradient propagation stability. To balance expressive power and deployment efficiency, this embodiment employs a two-layer LSTM stacked structure on top of LSTM and BiLSTM, and constructs several temporal blocks on top of TCN (each convolutional block contains a one-dimensional convolutional layer Conv1D, a ReLU activation function, and a Dropout regularization unit), where Dropout (0.3) is used for regularization. Network design details include the number of hidden units (adjustable), number of layers, kernel width, and dilation rate configuration. During training, tanh activation (inside LSTM), the Adam optimizer, a batch size of 64, and the MSE loss function are used. The number of training iterations is set to 500 in the experiments to ensure convergence. Baseline training on three types of basic models provides a performance benchmark. Experimental results show that TCN outperforms conventional LSTM and BiLSTM even without enhancements, exhibiting faster convergence and lower RMSEP.

[0069] To enhance the model's ability to fit complex nonlinear relationships and improve training convergence characteristics, a novel hybrid network architecture is designed. In this embodiment, a Kolmogorov-Arnold Network (KAN) module is fused to the backend of three types of temporal backbones to replace the traditional fully connected layers. The structure of KAN is as follows: Figure 5 As shown in the KAN Layer section, the functionalized mapping layer of the KAN network uses cubic B-spline basis functions to construct a functionalized edge connection structure, mapping the output. ,in Let k be the basis function of the B-spline. M represents the number of splines and the weights are the learnable weights. The core idea of ​​KAN is to replace the fixed linear weights along the network edges with parameterized univariate functions (such as trainable splines based on B-splines), so that each edge carries a trainable nonlinear mapping, thus "edge-ifying" the nonlinear transformation rather than node-ifying it. Structurally, the KAN layer receives temporal feature vectors from LSTM / BiLSTM / TCN, and after interacting with the KAN edge matrix, the output activation is obtained by summing. The function of each edge is composed of a linear combination of residual basis functions and several B-spline basis functions, and its coefficients are updated based on gradient descent. The technical effects brought by the introduction of the KAN module include: improving function approximation ability, accelerating the reduction of training loss, improving robustness in the presence of noise, and significantly reducing system bias in certain intervals. Preferably, after each KAN module training, the system calculates the hash digest of the model weight matrix. This summary is stored in the cloud to prevent structural drift due to data or code deviations.

[0070] The LSTM-KAN model consists of two-stage Long Short-Term Memory networks (LSTM1 and LSTM2) used to extract dynamic features across time. Input feature sequence After processing by the LSTM layer, a high-dimensional time state vector is generated. The output is then fed into the KAN layer; the KAN layer uses a set of differentiable B-spline basis functions to perform a nonlinear mapping on the LSTM output, defined as:

[0071]

[0072] in, Let k be the cubic B-spline basis function of the LSTM-KAN model. For the learnable weights of the LSTM-KAN model, This represents the number of splines in the LSTM-KAN model.

[0073] The BiLSTM-KAN model consists of a forward long short-term memory network and a backward long short-term memory network, used to extract dynamic features of time series along the forward and backward time axes, respectively. Its input feature sequence is defined as... The input sequence is fed into forward and backward LSTM networks for parallel processing. The forward LSTM models the input sequence sequentially, while the backward LSTM models the same sequence in reverse chronological order. The forward and backward LSTMs output corresponding hidden state feature vectors. Subsequently, the hidden state features output by the forward and backward LSTMs are concatenated along the feature dimension to form a unified bidirectional temporal feature representation. The concatenated bidirectional feature vector... = The input is fed into the KAN layer for nonlinear mapping, and the output of the BiLSTM-KAN model is defined as:

[0074]

[0075] in, Let k be the cubic B-spline basis function of the BiLSTM-KAN model. For the learnable weights of the BiLSTM-KAN model, This represents the number of splines in the BiLSTM-KAN model.

[0076] The TCN-KAN model is used to extract local temporal convolutional features. Input feature matrix. The convolution output passes through multiple temporal blocks in sequence. After being concatenated, the data is input to the KAN layer. The output of TCN-KAN is defined as follows:

[0077]

[0078] in, Let k be the cubic B-spline basis function of the TCN-KAN model. For TCN output features, Let k be the basis function of the spline. For the learnable weights of the TCN-KAN model, This represents the number of splines in the LSTM-KAN model.

[0079] This embodiment compares LSTM-KAN, BiLSTM-KAN, and TCN-KAN (KAN augmentation model) under the same training settings. The results show that the KAN augmentation model not only significantly improves prediction relevance (Rp) and goodness of fit (Rc), but also exhibits a faster convergence rate and a lower final loss value in the training curve. In particular, TCN-KAN performs best (the prediction set correlation coefficient reaches 0.9679, and the RMSEP is the lowest). These quantitative indicators are based on the dataset used in this embodiment and can serve as direct evidence of the effectiveness of model augmentation.

[0080] The model training and cross-validation design employs a strategy combining hierarchical random partitioning and K-fold cross-validation to ensure model generalization. The training and test sets are randomly partitioned at 70% and 30% respectively. During training, K-fold cross-validation is further applied to the training set for hyperparameter tuning (number of hidden units, learning rate, dropout rate, KAN spline parameters, etc.) and early stopping detection. Performance evaluation metrics primarily include the coefficient of determination (COP) between the training and test sets. , ) and RMSEC and RMSEP are used to quantitatively measure the fit and prediction error for unknown samples. Furthermore, by plotting scatter plots of predicted and measured values, the distribution of prediction residuals, and training loss curves, the bias performance and robustness of different models in different CDI intervals can be analyzed in depth. Experimental results show that, compared with the baseline model, the KAN-enhanced model significantly improves on most metrics, with TCN-KAN achieving the best results ( = 0.9667, = 0.9679, RMSEP = 0.0436), while LSTM-KAN and BiLSTM-KAN also showed significant improvements in accuracy and convergence. This performance verification provides data support for subsequent deployment of model selection and pruning schemes.

[0081] To adapt the model to resource-constrained edge devices and meet real-time inference requirements, this embodiment employs structured pruning as a lightweighting strategy. Unlike unstructured pruning, structured pruning removes parameters at the channel, filter, and neuron levels as a whole, while maintaining a dense matrix output, facilitating hardware acceleration. Specifically, the trained KAN augmentation model is sorted in the cloud based on the importance of weight magnitudes (global L1 norm), and the input weights of the target layers (including convolutional, fully connected, and recurrent layers) are pruned according to a set ratio (e.g., removing 30% of the lowest magnitude weights). For the KAN module, its spline components are retained to maintain its core nonlinear expressive power, and only its optional linear projections or redundant parts are conservatively pruned. After pruning, the KAN augmentation model is retrained to restore or improve its performance before pruning. Furthermore, after pruning, the system calculates the Merkle root hash digest before and after pruning; the model generates a root hash digest before pruning. After pruning, the model divides the weights into n blocks and calculates the local hash of each block: ( ), and calculate the Merkle root summary of the pruned model. ,in, This represents the complete set of weights consisting of all network weight parameters saved after training, before the model performs structured pruning. This represents the i-th subset of weight parameters obtained by dividing the entire set of network weight parameters retained after structured pruning according to a predetermined block division rule. This is a summary of the model roots before pruning. For the local hash of the i-th block, This is the Merkle root summary of the pruned model, where || represents the hash join operation. The inherited verification function Verify( is executed.) , If the response returns True, it confirms that the source of the pruned model is trustworthy and the inheritance chain is legitimate. This verification record, along with the digital signature, is written to the blockchain log to ensure the complete traceability of the model evolution process. Experimental results show that pruning is significantly effective in improving inference latency: in the example, the inference time of TCN-KAN after pruning decreased from 2.071s to 1.187s, an improvement of approximately 42.68%. Simultaneously, the accuracy reduction was effectively controlled under KAN enhancement; the pruned TCN-KAN still maintained a high Rp and low RMSEP, proving that the KAN module helps offset the information loss caused by pruning, thus achieving a better balance between inference efficiency and prediction accuracy. After the above pruning and retraining process is executed in the cloud, the lightweight model version is pushed to edge nodes for online inference via a cloud-edge distribution mechanism.

[0082] To systematically evaluate the comprehensive performance of the prediction model constructed in this invention in the task of modeling apple quality deterioration, comparative experiments were conducted on traditional time-series deep learning models, optimized models incorporating the KAN structure, and lightweight models further optimized through structured pruning, based on the same dataset partitioning method and a unified training and validation process. Evaluation metrics such as the coefficient of determination (R²) and root mean square error (RMSE) were calculated on the training and test sets, respectively, to quantitatively compare the fitting ability, generalization performance, and prediction stability of each model, thus comprehensively reflecting the impact of different model structures and optimization strategies on the accuracy of apple quality prediction. The resulting comparison of model prediction performance is shown in Table 1. These results intuitively demonstrate the improvement effect of the KAN structure on the model's nonlinear expression ability and prediction accuracy, as well as the technical advantages of maintaining or even further optimizing model performance under appropriate pruning strategies.

[0083] Table 1 Modeling results for each model

[0084]

[0085] Based on this, to verify the actual effect of the structured pruning method proposed in this invention on improving model deployment efficiency, further tests and comparative analyses were conducted on the inference runtime of various models before and after pruning. During the tests, under the same hardware environment and the same inference conditions, the time required for a single forward inference of the model was statistically analyzed, and the reduction ratio and speedup ratio of the model inference time before and after pruning were calculated, thereby quantifying the degree of improvement of model running efficiency by structured pruning. By comparing the inference acceleration effect of different models while maintaining prediction performance, the effective balance achieved by this invention between "prediction accuracy and running efficiency" was verified. The statistical results of the inference time and acceleration effect of each model before and after pruning are shown in Table 2. The results show that the models after structured pruning all achieved varying degrees of improvement in running speed, and are particularly suitable for edge device deployment scenarios with limited computing power.

[0086] Table 2. Inference time and speedup effect before and after pruning for each model.

[0087]

[0088] The edge inference and local early warning module is deployed in a container manager on edge devices (e.g., based on i.MX6UL or a similar embedded platform). It is responsible for receiving preprocessed time-series data from Apple's warehouse environment monitoring gateway and performing real-time predictions. The core responsibilities of the edge device include: low-latency model forward inference, threshold-based level mapping and early warning triggering logic, and uploading batch or event-level data to the cloud when network availability is available. Predictive output is generated in the form of a continuous decay index (e.g., Composite Decay Index, CDI). The system maps the CDI to four decay levels (fresh, slightly, moderate, and severe) and uses quantile thresholds or empirical thresholds (generally dividing the CDI range evenly) to trigger alarms. When a predicted value exceeds a certain level threshold, the edge device can immediately generate a local early warning, triggering control commands (e.g., temporarily lowering the temperature, increasing ventilation, or activating cargo securing indicators), and simultaneously sending event logs and original channel snapshots back to the cloud for subsequent analysis and model updates. This ability to respond instantly locally significantly reduces the risk of network latency or connection interruptions, while reducing cloud load and saving network bandwidth by offloading most of the computing load to the edge.

[0089] Cloud-edge collaborative management and model lifecycle control are crucial for ensuring the system's long-term adaptability. The cloud utilizes a Kubernetes (K8s) cluster to expand model training resources, aggregate data, and manage versions. Edge nodes maintain a two-way control channel with the cloud via KubeEdge or similar Cloud Core / Edge Core frameworks: the cloud periodically aggregates data from multiple edge nodes, performs unified incremental sample labeling and retraining, and evaluates the performance of new model versions on validation sets and cross-node data. When a new model outperforms the current deployment version (based on a comprehensive metric of Rp, RMSEP, and latency tradeoffs), the cloud triggers model packaging, structured pruning, and distributes the lightweight binary model to the target edge nodes via a secure distribution channel. Edge nodes can perform inference validation locally upon receiving a new model and automatically roll back. If the new model performs worse locally than the current version, the system will revert to the previous stable version and record the rollback event in a trusted log chain to prevent performance degradation from impacting online monitoring, thus achieving a closed-loop management strategy of "cloud optimization - edge validation - secure switchover." Furthermore, the cloud platform is responsible for aggregating cross-node performance statistics, anomaly analysis, and long-term trend modeling to support decision-makers in developing better thresholds and control strategies. This collaborative mechanism enables continuous model optimization, rapid deployment, and on-site safety verification, ensuring the system's adaptability under different seasons, production locations, and transportation conditions. Further, during the cloud-edge collaborative training phase, the cloud platform utilizes local incremental data uploaded by edge nodes... Execution parameter aggregation , where node weight Calculated based on a combination of node sample capacity, network latency, and communication quality. and This is the complete weight set consisting of all network weight parameters saved after training in rounds t and t+1. The cloud calculates the root hash after each aggregation. and node summary If the comparison matches, the version number is synchronized; otherwise, a rollback mechanism is triggered to ensure consistency of global model parameters. In the cloud-edge model management section, the system generates a trusted log chain for each round of global training. The logs include parameter snapshots, performance metrics, and timestamps, and are written to a trusted cloud database using SHA256 digests, forming a replayable and verifiable record of model evolution. The cloud further uses digital signatures to authenticate the model version, and edge nodes verify the signature and digest matching upon receiving the data, thereby achieving secure inheritance and trusted deployment of model versions.

[0090] The remote monitoring and early warning platform displays key environmental parameters and prediction results in the form of dashboards, time-series curves, heatmaps, and event logs. The interface supports backtracking queries by device, vehicle, or batch, and provides early warning level color codes (blue / yellow / orange / red) and suggested control lists. The system supports multi-channel notifications (SMS, email, and mobile application push), and records manual confirmation actions to complete the "prediction-response-verification" closed loop. It also provides a model performance monitoring view, allowing technicians to monitor model performance in real time and decide whether to trigger cloud retraining or on-site calibration.

[0091] Although this embodiment uses apples as an example, the method is universal and can be extended to other cold chain applications for fruits and vegetables. The model architecture, feature extraction pipeline, and pruning strategy can all be adjusted according to the physiological characteristics of the target fruits and vegetables and the sensor layout; the cloud-edge collaboration framework and model distribution mechanism can also be extended to canary release, rollback, and policy-based hierarchical distribution according to actual operation and maintenance requirements. Through the above-mentioned integrated "multi-sensor-temporal extraction-KAN enhancement-pruning lightweight-cloud-edge collaboration" process, this embodiment demonstrates quantifiable advantages in both theory and engineering implementation: it improves prediction accuracy, accelerates training convergence, reduces edge inference latency, and maintains high stability after pruning, thus providing an operable and scalable technical solution for real-time quality monitoring and early warning in real-world cold chain / warehousing environments.

[0092] Example 3

[0093] This embodiment focuses on fruits and vegetables, employing the fruit and vegetable storage environment monitoring gateway system described in Embodiment 1 and the method described in Embodiment 2 to achieve real-time monitoring and early warning management of the entire cold chain logistics process for fruits and vegetables. Figure 1 As shown, the specific steps are as follows:

[0094] As shown in Figure 1, this embodiment proposes a dynamic monitoring method for fruit and vegetable quality based on KAN network and cloud-edge-device collaboration, used to monitor temperature, humidity, and other parameters within the storage space. , , This system continuously collects, analyzes in real-time, and dynamically controls multiple environmental variables, including VOCs. This embodiment utilizes a collaborative approach of data acquisition at the device level, intelligent inference at the edge, and centralized training and optimization in the cloud to construct an intelligent warehouse environment management system with real-time performance, distributed computing capabilities, and continuous model upgrades. This system can predict fruit and vegetable spoilage trends, generate environmental control strategies, and issue early warning information in actual warehouse scenarios, thereby significantly improving the stability and safety of stored fruit and vegetable quality.

[0095] The device is used for real-time acquisition of multi-source sensor data in the warehouse environment. The device includes temperature sensors, humidity sensors, and... sensor, sensor, Sensors, including VOC sensors, are installed at multiple points within the warehouse space to capture dynamic changes in environmental parameters at different locations. The collected raw data undergoes initial filtering and calibration at the device level and is then uploaded to the edge gateway node via MQTT or serial communication. Further, each set of collected data... Before uploading, a timestamp and batch identifier are added, and a data digest is calculated. This summary is used for subsequent cloud storage to ensure the integrity and tamper-proof nature of uploaded data during transmission and storage.

[0096] The edge computing unit, serving as the system's local intelligent core, preprocesses the raw data collected from the devices and executes real-time inference of the KAN-enhanced model locally. Utilizing a low-power computing platform based on the i.MX6UL, the edge unit runs LSTM-KAN-pruned, BiLSTM-KAN-pruned, and TCN-KAN-pruned models through containerized deployment to extract the temporal dynamic features of environmental data and predict fruit and vegetable spoilage trends. By performing inference tasks at the edge, the system's response time is significantly reduced under unstable network or high latency conditions, ensuring that environmental control commands are output within millisecond timescales, thereby maintaining the dynamic stability of the storage environment. Furthermore, after receiving data uploaded from the devices, the edge unit first preprocesses the multi-source sensor data, and then the edge node invokes the locally deployed KAN-enhanced model for real-time inference, outputting the fruit and vegetable spoilage index (SI) and its corresponding warning level. This embodiment's model incorporates three basic structures: LSTM, BiLSTM, and TCN, respectively suited for long sequence dependency inference, bidirectional dependency inference, and local feature convolution extraction. Furthermore, this embodiment introduces a KAN (Kolmogorov–Arnold Network) module at the end of the model to replace the traditional fully connected layer, giving the model stronger nonlinear expressive capabilities. The KAN module achieves combined mapping of complex features through a multivariate decomposition function, enabling the system to maintain high prediction accuracy even when facing multi-gas coupled change processes. KAN enhances the model's input feature vector. , Given the set of variable intervals with the lowest RMSE, output the predicted value of fruit and vegetable spoilage risk. ,in The complete weight set consists of all network weight parameters of the current KAN augmentation model; a hash digest is generated from the prediction results. The data is uploaded to the cloud and incremental parameter training is performed.

[0097] The KAN-enhanced model at the edge runs via Docker containers, enabling hot updates of the model—meaning a new model version can be installed without stopping the edge gateway's main program. The edge can locally output adjustment commands based on predicted decay trends, such as starting fans, activating dehumidification modules, and increasing oxygen supply, to achieve real-time environmental control. This edge-based intelligent control improves the system's response speed to abnormal environmental fluctuations, avoiding control lag caused by cloud latency. Furthermore, digital signature verification and Merkle Tree root digest verification are performed during model container updates to ensure the model files have not been tampered with and the version source is trustworthy; only after successful verification can the model be activated and deployed; otherwise, the system automatically rolls back to the previous stable version.

[0098] The cloud platform serves as the core of the system's global optimization, storing long-term, multi-scenario environmental data and performing model training, evaluation, and version management. It utilizes the Kubernetes (k8s) platform to build a distributed model training cluster, executing large-scale deep learning model training tasks through multi-GPU nodes. The cloud integrates long-term data from different storage locations, using a unified data format and standardized processing strategies to build a stable data foundation. During training, the cloud rigorously evaluates the model using performance metrics such as RMSE, MAE, and R², and optimizes parameters such as hidden layer dimensions, number of convolutional kernels, and KAN function structure through hyperparameter search strategies to achieve optimal prediction results. A smoothing regularization term is introduced after each training round. (in (Regularization coefficient) suppresses oscillations between adjacent weights, achieving stable output and high generalization under nonlinear temporal mappings. This regularization term is also enabled during local retraining at edge nodes to ensure smooth parameter convergence. The joint loss function for training is defined as... ( (This represents the mean squared error loss between the model's predicted output and the measured values ​​of fruit and vegetable quality), which is minimized through backpropagation to update the parameters. After the update, a hash digest is generated. Used for cloud storage. Furthermore, cloud training employs a gradient descent algorithm with momentum. (in For learning rate, (momentum coefficients), and a parameter summary is generated after each training round. The summary is written to a trusted log chain along with a timestamp, enabling traceability of the training process and rollback in case of anomalies.

[0099] After training, the cloud distributes the latest version of the KAN augmented model to the edge devices via the kube Edge framework. Two-way communication is maintained between the cloud and the edge, allowing the edge devices to report model inference results, device status, and anomalies in real time. This enables the cloud to evaluate model performance and continuously optimize it based on actual usage. The cloud provides unified management of multiple model versions and has canary release capabilities, meaning a new model can be tested on a subset of edge nodes first, and then globally pushed after performance stabilizes. Simultaneously, the cloud uses an asynchronous message queue mechanism to prevent model distribution interruptions due to network jitter. If digest inconsistencies or communication failures are detected, the system automatically executes breakpoint resumption and rollback mechanisms to ensure the integrity and stability of model synchronization. (Cloud computing performance metrics set) If twice in a row ( Let be the coefficient of determination in the t-th training iteration; Let be the root mean square error in the t-th training iteration; If a threshold value is set to 0.5 in this embodiment, retraining and canary deployment are performed to ensure stable model performance. A Merkle Tree is constructed in the cloud to store the summary set. This enables end-to-end consistency verification of data and models; when an anomaly is detected or the delay exceeds the threshold, the system automatically performs task migration and updates the cloud status table, achieving dynamic, reliable, and traceable control of the entire process of fruit and vegetable quality monitoring.

[0100] This embodiment employs a cloud-edge-device collaborative architecture, enabling the reasonable distribution of data acquisition, real-time inference, and model training workloads across different levels. The device handles data acquisition, the edge handles inference, and the cloud handles training. This system design avoids the high latency issues inherent in traditional cloud architectures under unstable network conditions, and also avoids the insufficient computing power problems that arise from placing all computational tasks on the device. This collaborative system can run in parallel across multiple warehouse scenarios, achieving cross-regional data sharing and unified model upgrades through unified cloud scheduling, thus improving the overall maintainability of the system.

[0101] The cloud-edge-device collaborative method in this embodiment forms a stable closed-loop control process: the device collects multi-source data → the edge performs fusion and inference → the cloud performs training and optimization → the model is updated and distributed → the edge performs local intelligent control → the cloud monitors the system status in real time. This closed-loop structure realizes a continuous learning mechanism of "real-time perception - local decision-making - cloud optimization - collaborative upgrade", enabling the warehousing system to dynamically evolve with environmental changes.

[0102] The cloud provides a visual monitoring interface that displays environmental trends, evolution of the corruption index, equipment operating status, and early warning information, and supports unified management of different warehouse nodes. Early warning events at the edge are synchronously uploaded to the cloud, which can push alarm information to managers or automatically trigger emergency strategies when necessary, achieving multi-level linkage response.

[0103] This embodiment enables fruit and vegetable storage environment monitoring to have higher real-time performance, stronger predictive capabilities, and more stable system reliability. During long-term operation, the cloud continuously updates the model based on the collected large-scale historical data, giving the system self-learning and adaptive capabilities, thereby ensuring that the predictive model can be applied to environmental changes in different seasons, different batches of fruits and vegetables, and different storage conditions.

[0104] Although embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention, and all such modifications and variations fall within the protection scope of the claims of the present invention.

Claims

1. A dynamic monitoring system for fruit and vegetable quality based on KAN network and cloud-edge-device collaboration, characterized in that, The system comprises a perception layer, an edge computing layer, and a cloud optimization layer; The sensing layer includes temperature and humidity sensors and gas concentration sensors, which are used to collect multi-dimensional data in the fruit and vegetable storage and transportation environment. The edge computing layer uses an i.MX6UL processor and integrates a 4G communication module. The data collected by the perception layer is preprocessed and optimized in intervals before being used as input to the KAN augmentation model. The KAN augmentation model performs real-time inference and outputs the fruit and vegetable spoilage index. The cloud optimization layer performs model training, parameter aggregation, and hash digest verification, and then sends the optimized KAN enhanced model back to the edge nodes.

2. A method for a dynamic monitoring system for fruit and vegetable quality based on a KAN network and cloud-edge-device collaboration as described in claim 1, characterized in that: A KAN augmentation model is constructed, which adopts three types of time-series architectures: LSTM, BiLSTM and TCN. The KAN network layer is used to replace the fully connected layer in the back end of the three types of time-series architectures. Train the KAN augmentation model in the cloud, sort the trained KAN augmentation model based on the importance of weight magnitude, and prune the target layer according to a set ratio. Real-time data from the storage and transportation environment of fruits and vegetables is collected, preprocessed and optimized, and then input into a trained and pruned KAN-enhanced model to output a fruit and vegetable spoilage index.

3. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, The KAN enhancement models include LSTM-KAN, BiLSTM-KAN, and TCN-KAN models; the LSTM-KAN model consists of a two-level long short-term memory network, used to extract dynamic features across time. The BiLSTM-KAN model consists of a forward long short-term memory network and a backward long short-term memory network, which are used to extract the dynamic features of the time series on the forward and backward time axes, respectively; the TCN-KAN model is used to extract local temporal convolution features.

4. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, The functionalized mapping layer of the KAN network uses cubic B-spline basis functions to construct a functionalized edge connection structure, and maps the output. ,in Let k be the basis function of the B-spline. Here, M represents the number of splines and the learnable weights.

5. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, After pruning, calculate the root hash digest of the model before pruning. Calculate the Merkle root summary of the model after pruning. After pruning, the model divides the weights into n blocks, and the local hash of each block is: , ;in, This is a summary of the model roots before pruning. This represents the complete set of weights consisting of all network weight parameters saved after training, before the model performs structured pruning. This represents the i-th subset of weight parameters obtained by dividing the entire set of network weight parameters retained after pruning according to a predetermined block rule after the model has completed structured pruning; || represents the hash join operation. Execute the inheritance verification function Verify( , If the function returns True, it confirms that the source of the pruned model is reliable and the inheritance chain is valid.

6. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, Edge nodes based on local data Calculate the gradient of the loss function Generate local increments And upload it to the cloud during the synchronization period; the cloud uploads it based on the node weight. Complete weighted aggregation to form a global model. Cloud-based root hash calculation and upload hash with the node Compare; if This will automatically trigger the model re-aggregation process; among which, The learning rate is adapted for this edge node. This represents the parameter matrix of the KAN augmentation model in the cloud after the t-th round of training, where N is the number of edge nodes. For node weight coefficients, This is the hash tolerance threshold for performance summary comparison.

7. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, The input feature vector of the trained and pruned KAN augmentation model Output fruit and vegetable spoilage risk prediction values The prediction results generate a hash digest. The data is uploaded to the cloud and incremental parameter training is performed; among them, The set of variable intervals with the lowest RMSE. This is the complete set of weights consisting of all network weight parameters of the current KAN augmentation model.

8. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 7, characterized in that, The preprocessing includes temporal alignment and synchronization correction, denoising and outlier removal, normalization and missing value completion. The interval optimization involves dividing the preprocessed data into a preset number of intervals, with each interval corresponding to a set of statistical features, thereby generating several candidate sets of interval combinations. For each candidate set of combinations, partial least squares regression is used in the cloud to model the validation set root mean square error curve through cross-validation. The root mean square error corresponding to different numbers of latent variables is recorded. By extracting local minima and sorting them, several high-performance interval combinations are selected. The set of variable intervals that minimizes the cross-validation root mean square error is selected as the input to the trained and pruned KAN augmentation model.

9. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, A smoothing regularization term is introduced after each training round. Suppress oscillations between adjacent weights; the joint loss function for training is defined as... Minimize the updated parameters through backpropagation After the update, a hash digest is generated. Used for cloud storage; cloud training employs a gradient descent algorithm with momentum. And generate a parameter summary after each round of training. The summary is written to a trusted log chain along with a timestamp, enabling traceability of the training process and rollback in case of anomalies; among which, The regularization coefficient is . For learnable weights, This represents the mean squared error loss between the model's predicted output and the measured values ​​of fruit and vegetable quality. This refers to the complete set of weights consisting of all network weight parameters of the current KAN augmentation model. For learning rate, This is the momentum coefficient.

10. The method for dynamic monitoring of fruit and vegetable quality based on KAN network and cloud-edge-device collaboration according to claim 2, characterized in that, Cloud computing performance metrics set If twice in a row Then, retraining and canary deployment will be performed, where, Let be the coefficient of determination in the t-th training iteration. Let be the root mean square error in the t-th training iteration. The threshold value is used.

Images