Express package informatization monitoring method and system based on Internet of Things

Abstract

According to the express package informatization monitoring method and system based on the Internet of Things, edge calculation and deep learning technologies are combined and utilized, and real-time, whole-course and intelligent monitoring and early warning of express packages are achieved. The method specifically comprises the steps of installing a sensor and a micro camera on an express package, and collecting package information and environment data; and the edge computing device is responsible for performing real-time data collection and primary processing nearby so as to optimize the use of computing resources and improve the data processing speed. In addition, a zero sample learning model is introduced, and new and non-appearing conditions are predicted through a rule learned by a training sample. In addition, by training a deep learning model and monitoring the state of the express parcel in real time, early warning is immediately given to potential problems. According to the method, the parcel monitoring efficiency, precision and real-time performance are greatly improved, and the accuracy and efficiency of express delivery operation are guaranteed.

Description

Technical Field

[0001] The present invention belongs to the field of express logistics, and more specifically relates to a method and system for information-based monitoring of express packaging based on the Internet of Things. Background Art

[0002] In the past, the monitoring of express parcels mainly relied on manual inspection, with couriers regularly checking the status of parcels. This method is inefficient and error-prone, and can only be checked at fixed time points, unable to achieve full-process monitoring. With the rapid development of society and the prosperity of the express industry, this manpower-based parcel management method can no longer meet the high-efficiency and precision requirements of the express industry.

[0003] The trend of modern technology development is informatization, intelligence, and networking. With the development of Internet of Things (IoT) technology, the popularization of various sensors, and the gradual maturity of edge computing and cloud computing technologies, these technologies have brought new opportunities to the express industry.

[0004] Through the deployment of various sensors and micro cameras, IoT technology can track and collect the status information and environmental data of parcels in real time. Edge computing technology can quickly and real-time process a large amount of collected data nearby, reducing the computing tasks on the cloud and improving the overall data processing efficiency.

[0005] Moreover, with the rapid development of deep learning technology in recent years, especially the proposal of zero-shot learning methods, it can learn the rules from training samples and generalize to seemingly unrelated data sets, enabling the model to predict new situations that did not appear in the original training samples. This provides a new solution for the dynamic, complex, and ever-changing real-time monitoring tasks of express parcels.

[0006] Therefore, how to effectively combine cutting-edge technologies such as IoT, edge computing, and deep learning to build an information-based monitoring system for express parcels with high efficiency, good accuracy, and strong real-time performance to meet the growing monitoring needs, improve the accuracy and efficiency of express operations, and ensure the safety of customers' parcels has become an important research direction in the modern express industry. Summary of the Invention

[0007] The present invention will solve the problems existing in the traditional express parcel monitoring method in terms of efficiency, accuracy, and real-time performance. By using IoT technology, edge computing, and deep learning technology, it realizes real-time, full-process, and intelligent monitoring of express parcels, accurately warns of possible risks, and improves the efficiency and security of express services.

[0008] To achieve the above object, the present invention is implemented by the following technical solutions: The method includes:

[0009] Install different sensors and micro cameras on the express delivery package to collect package information and environmental data;

[0010] The edge computing device performs real-time data collection and preliminary processing nearby, reducing the computing load on the cloud and improving the data processing speed;

[0011] Build a deep learning model: Build a zero-shot learning model that learns the patterns from the training samples and generalizes to seemingly unrelated data sets;

[0012] Train the deep learning model: Use the zero-shot learning algorithm to train the collected data so that the model can predict new situations that did not appear in the original training samples;

[0013] Status monitoring and early warning: Use the trained model to monitor the status of the express delivery package in real time. If the model predicts a problem, it will immediately give an early warning, such as warning that the package has been shaken or the temperature is too high.

[0014] In one solution, the environmental data includes location, temperature, humidity, vibration, and light.

[0015] In one solution, the data collection and preliminary processing include: data filtering and noise reduction, location data analysis, preliminary image analysis, and vibration data analysis.

[0016] In one solution, the zero-shot learning model includes:

[0017] The model architecture includes the following components:

[0018] Semantic embedding space: Embed the semantic information of the category or status into a high-dimensional vector space using a word vector model;

[0019] a c =Embedding(Semantic_Info c )

[0020] where a c is the semantic embedding vector of category c.

[0021] Visual embedding space: Project the fused feature vector z into the semantic embedding space for cross-space matching:

[0022] v=W v z+b v

[0023] where W v and b v are the weight matrix and bias vector of the linear transformation respectively, and v is the visual embedding vector;

[0024] Compatibility function: Define a compatibility function F(v, a c ), which is used to measure the similarity between visual embeddings and semantic embeddings; the compatibility function is as follows:

[0025] F(v, a c ) = v · Wa c

[0026] where W is the learned weight matrix used to capture the relationship between the two.

[0027] Loss function design: To train the model to maximize the compatibility score of the correct class, a contrastive loss is adopted:

[0028]

[0029] where γ is the margin hyperparameter and c′ is the negative sample class, ensuring that the score of the correct class c is at least γ higher than the score of the negative sample c′.

[0030] In one scheme, the training of the deep learning model includes: adopting a margin contrastive loss to ensure that the matching score of the correct class with the sample is at least higher than the score of the wrong class by a margin γ;

[0031] Margin contrastive loss function:

[0032]

[0033] where γ is the margin hyperparameter and c′ is the negative sample class, ensuring that the score of the correct class c is at least γ higher than the score of the negative sample c′; F(v x , a c ) is the compatibility score between the visual embedding and the semantic embedding;

[0034] Model training process

[0035] Forward propagation: Input data: Input the fused feature vector z of each sample in the training set into the visual embedding network to generate a visual embedding vector v:

[0036] v = W v z + b v

[0037] Semantic embedding: Use the predefined semantic embedding vector a c to represent each class;

[0038] Compatibility scoring: Calculate the compatibility score F(v, a c ) for each pair of visual embeddings and semantic embeddings;

[0039] Loss calculation: Based on the margin contrast loss function, calculate the loss of each sample and accumulate the total loss of the entire training set:

[0040]

[0041] Backpropagation and parameter update:

[0042] Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and update the parameter θ through an optimization algorithm (Adam):

[0043]

[0044] where η is the learning rate, is the gradient of the loss function with respect to the parameters;

[0045] Regularization and prevention of overfitting: Introduce the L2 regularization term to prevent the model from overfitting the training data: where λ is the regularization strength coefficient.

[0046] In one solution, the state monitoring and early warning introduce a multi-level risk assessment model to classify different types and degrees of anomalies and accurately locate potential risks:

[0047]

[0048] where K is the number of risk categories, m k is the number of risk indicators for the k-th category, w k,i and S k,i (t) are the weight and score respectively.

[0049] On the other hand, an information-based monitoring system for express delivery packages based on the Internet of Things, the system is applicable to the method, and the system includes sensors and micro cameras installed on the express delivery packages for collecting package information and environmental data; a data collection device with edge computing function for real-time data collection and preliminary processing; a zero-shot learning model for learning and generalizing seemingly unrelated data sets according to the rules learned from the training samples; a deep learning model training device for training the collected data using the zero-shot learning algorithm; and an early warning device for using the trained model to monitor the status of express delivery packages in real time. When the model predicts a problem, an early warning is given in real time through the early warning device.

[0050] Advantages of the present invention:

[0051] By utilizing Internet of Things technology, edge computing and deep learning technology, the present invention can greatly improve the efficiency of express delivery package monitoring. Real-time monitoring, automated processing and intelligent early warning can avoid excessive manual intervention and improve the running speed of the work process.

[0052] Through the training and application of deep learning models, the prediction accuracy of the status of express packages can be improved. The application of micro cameras and sensors helps to obtain more objective and detailed real-time data, improving the accuracy of analysis results.

[0053] The real-time monitoring and intelligent early warning functions of the present invention can help detect potential risks earlier, improve the processing speed of abnormal situations, and effectively ensure the safety of express packages.

[0054] By performing data processing on edge devices, the computing load on the cloud can be effectively reduced, efficient allocation of computing tasks can be achieved, and the overall performance of the system can be improved. Description of the Drawings

[0055] Figure 1 is a flowchart of the method of the present invention;

[0056] Figure 2 is a flowchart of establishing a deep learning model of the present invention;

[0057] Figure 3 is a flowchart of training a deep learning model of the present invention. Detailed Embodiments

[0058] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Typical embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0059] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Typical embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0060] As Figure 1 shown, an information-based monitoring method for express packaging based on the Internet of Things includes:

[0061] S1. Sensor configuration: Install different sensors (such as position, temperature, humidity, vibration, light sense, etc.) and micro cameras on the express packaging, and these devices will be responsible for collecting package information and environmental data.

[0062] Determine the environmental parameters to be monitored, such as location, temperature, humidity, vibration, and light, etc., and whether an image capture function is required. Based on these requirements, it is crucial to select the appropriate sensor type and model. For perishable goods, it is preferred to select high-precision temperature and humidity sensors; for valuable or fragile items, high-sensitivity vibration sensors and positioning modules should be configured with emphasis.

[0063] S2. Edge computing devices: These devices collect and preprocess real-time data locally, reducing the computing load on the cloud and improving the data processing speed. For example, perform primary analysis on the images captured by cameras.

[0064] Edge computing devices play a key role in collecting and preprocessing real-time data locally, thus reducing the computing load on the cloud and increasing the data processing speed. The following is the detailed process method implemented by edge computing devices.

[0065] S201. Device selection and deployment

[0066] First, suitable edge computing devices need to be selected. This includes high-performance microprocessors (such as processors with ARM architecture), sufficient memory and storage space, and modules supporting wireless communication (such as Wi-Fi, Bluetooth, LoRa, or NB-IoT). The devices should have the characteristics of low power consumption and high efficiency to meet the requirements of long-term operation and different transportation environments.

[0067] S202. Data collection and preprocessing

[0068] Edge computing devices collect multi-dimensional data of packages in real time by connecting various sensors and cameras. This data may include temperature, humidity, vibration, light intensity, and images, etc. To improve the data processing efficiency, the collected data needs to be preprocessed first. The specific steps include data filtering, noise reduction, and compression.

[0069] (1) Data filtering and noise reduction:

[0070] For continuous sensor data, filters can be used to remove noise. A low-pass filter (such as a moving average filter) can be used to smooth the temperature and humidity data:

[0071]

[0072] where y[n] is the filtered signal, x[n] is the original signal, and M is the filter window size.

[0073] (2) Preliminary data processing and analysis

[0074] Edge computing devices need to perform preliminary analysis on the collected data to extract useful information and reduce the amount of data transmitted to the cloud.

[0075] Location data analysis: For the location data collected by the GPS module, the Kalman Filter can be used to estimate and predict the location to improve the positioning accuracy.

[0076]

[0077] Among them, is the state vector, P is the error covariance matrix, K is the Kalman gain, and z k is the measurement value, and A, B, H, Q, and R are the system matrices and noise covariance matrices.

[0078] Initial image analysis: For the images captured by the camera, the edge device can perform preliminary image processing and analysis, such as object detection and classification. Apply a pre-trained CNN model (such as MobileNet) for feature extraction.

[0079] Vibration data analysis: For the vibration sensor data, the time-domain signal can be converted into a frequency-domain signal through the Fourier Transform (FFT) to analyze the vibration frequency characteristics:

[0080]

[0081] Through frequency-domain analysis, abnormal vibration frequencies can be detected, thereby determining whether the package has been subjected to abnormal impacts.

[0082] Based on the preliminary analysis results, the edge computing device makes local decisions. If the temperature exceeds the set threshold or abnormal vibrations are detected, the system can immediately issue a warning signal, reducing the dependence on cloud resources. The warning rules can be implemented through logical judgments:

[0083] If T>T threshold or V>V threshold , trigger alert

[0084] Among them, T is the temperature, V is the vibration intensity, and T threshold and V threshold are the preset thresholds.

[0085] As Figure 2 shown, S3. Establish a deep learning model: Establish a deep learning model architecture, namely a zero-shot learning model, which can learn and generalize from seemingly unrelated data sets based on the patterns learned from the training samples.

[0086] Aiming to develop a Zero-Shot Learning (ZSL) model. By leveraging the patterns learned from the training samples, this model can learn and generalize from seemingly unrelated datasets, thereby achieving accurate prediction and classification of new or unseen package states. The following are the detailed implementation steps of this process, as well as the mathematical formulas and theoretical bases involved.

[0087] S301. Dataset Preparation and Integration

[0088] Combining the previous two steps (S1 and S2), the system has collected a large amount of multi-dimensional data from sensors and edge computing devices. This data includes environmental parameters (such as temperature, humidity, vibration, light intensity), location data, and image information. To build a deep learning model, this data is sorted and labeled.

[0089] Data Preprocessing:

[0090] Data Cleaning: Remove missing values, outliers, and noisy data to ensure data quality. For example, use statistical methods to detect and remove abnormal data points that exceed 3 standard deviations.

[0091] Data Standardization: Normalize data with different dimensions to make them have the same scale, including min-max normalization and Z-score normalization.

[0092] Feature Extraction and Fusion: Sensor Data Feature Extraction: Extract statistical features (such as mean, variance) and temporal features (such as autocorrelation) from sensor data such as temperature and humidity.

[0093] Image Data Feature Extraction: Use pre-trained convolutional neural networks (such as ResNet, VGG) to extract high-dimensional image feature vectors:

[0094] Data Fusion: Fuse features from different sources to form a unified feature representation. It includes feature concatenation and attention mechanism:

[0095] z = Concat(z sensor , z image )

[0096] where z sensor and z image are the feature vectors of sensor and image data respectively, and z is the fused feature vector.

[0097] S302. Zero-Shot Learning Model Architecture Design

[0098] Zero-shot learning aims to enable the model to recognize categories or states that do not appear in the training set. The core of the S3 stage is to design an effective ZSL model architecture, which usually includes the following components:

[0099] Semantic embedding space: Embed the semantic information of categories or states (such as text descriptions, attribute vectors) into a high-dimensional vector space. Use a word vector model (such as Word2Vec).

[0100] a c = Embedding(Semantic_Info c )

[0101] where a c is the semantic embedding vector of category c.

[0102] Visual embedding space: Project the fused feature vector z into the semantic embedding space for cross-space matching:

[0103] v = W v z + b v

[0104] where W v and b v are the weight matrix and bias vector of the linear transformation respectively, and v is the visual embedding vector.

[0105] Compatibility function: Define a compatibility function F(v, a c ) to measure the similarity between the visual embedding and the semantic embedding. The compatibility function is as follows:

[0106] F(v, a c ) = v · W a c

[0107] where W is the learned weight matrix used to capture the relationship between the two.

[0108] Loss function design: To train the model to maximize the compatibility score of the correct category, use the contrastive loss:

[0109]

[0110] where γ is the margin hyperparameter and c' is the negative sample category, ensuring that the score of the correct category c is at least γ higher than the score of the negative sample c'.

[0111] S303, Zero-shot Inference and Application

[0112] In the deployment stage, the model needs to make inferences on newly emerging unseen categories or states. Based on the compatibility function learned in the training stage, the model calculates the compatibility scores between the new samples and the semantic embeddings of all possible categories and selects the category with the highest score as the prediction result:

[0113]

[0114] Among them, C unseen is the set of unseen categories, is the predicted category.

[0115] Deploy the trained deep learning model to the cloud or edge server to ensure that it can process the data stream from edge devices in real time. During the deployment process, the computational complexity and response time of the model need to be considered, and model compression and quantization are carried out if necessary:

[0116] W quant = round(W × 2 q )

[0117] Among them, q is the quantization bit number, reducing the model storage and computational requirements.

[0118] The process of establishing the zero-shot learning deep learning model (S3) is closely combined with the sensor configuration (S1) and the real-time data processing of the edge computing device. Through the system's data preparation, model architecture design, training optimization, and deployment implementation, the ZSL model can still maintain efficient and accurate prediction capabilities when facing new or unseen package states. Through continuous optimization and iteration, the model can adapt to the dynamically changing logistics environment, significantly improving the intelligence and reliability of the entire express packaging informatization monitoring system.

[0119] Such as Figure 3 shown, S4, Training the deep learning model: Use the zero-shot learning algorithm to train the collected data so that the model can predict new situations that did not appear in the original training samples.

[0120] This step is based on the zero-shot learning (ZSL) model established in the S3 stage, and uses a specific training algorithm to train the collected multi-dimensional data to ensure that the model can not only accurately identify the categories or states that appear in the training set, but also effectively predict and classify new situations that did not appear in the original training samples.

[0121] S401, Data preparation and division. In the S3 stage, the data preprocessing, feature extraction, and fusion have been completed. In the S5 stage, these prepared data need to be divided first for effective training, validation, and testing.

[0122] Data division:

[0123] Training Set: Contains a large number of labeled samples for parameter optimization of the model. Usually accounts for 70%-80% of the total data.

[0124] Validation Set: Used to adjust model hyperparameters and prevent overfitting, usually accounting for 10%-15% of the total data.

[0125] Test Set: Used to evaluate the performance of the model on unseen data, usually accounting for 10%-15% of the total data.

[0126] S402, Zero-Shot Learning Algorithm Selection and Configuration

[0127] Zero-shot learning aims to enable the model to handle categories or states not present in the training set. In this system, an embedding method based on a compatibility function is adopted. The core idea is to map visual features and semantic features into the same embedding space and evaluate the matching degree between them through the compatibility function.

[0128] Compatibility function:

[0129] F(v,a c )=v · Wa c

[0130] where v is the visual embedding vector, a c is the semantic embedding vector of category c, and W is the learned weight matrix.

[0131] S403, Loss Function Design and Optimization Objectives

[0132] To train the ZSL model to make the compatibility score of the correct category higher than that of the wrong category, contrastive loss is usually adopted. In this system, margin contrast loss is adopted to ensure that the matching score of the correct category with the sample is at least higher than that of the wrong category by a margin γ.

[0133] Margin contrast loss function:

[0134]

[0135] where γ is the margin hyperparameter, c′ is the negative sample category, ensuring that the score of the correct category c is higher than that of the negative sample c′ by at least γ.

[0136] F(v x ,a c ) is the compatibility score between the visual embedding and the semantic embedding.

[0137] S404, Model Training Process

[0138] Forward propagation:

[0139] Input data: The fused feature vector z of each sample in the training set is input into the visual embedding network to generate the visual embedding vector v:

[0140] v = W v z + b v

[0141] Semantic Embedding: Represent each category using a predefined semantic embedding vector a c

[0142] Compatibility Scoring: Calculate the compatibility score F(v, a c ) for each pair of visual and semantic embeddings.

[0143] Loss Calculation: Based on the margin contrast loss function, calculate the loss for each sample and accumulate the total loss over the entire training set:

[0144]

[0145] Backpropagation and Parameter Update:

[0146] Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters and update the parameters θ using an optimization algorithm (Adam):

[0147]

[0148] where η is the learning rate, is the gradient of the loss function with respect to the parameters.

[0149] Regularization and Overfitting Prevention: Introduce an L2 regularization term to prevent the model from overfitting the training data:

[0150]

[0151] where λ is the regularization strength coefficient.

[0152] S405. Model Validation and Hyperparameter Tuning: During training, use the validation set to evaluate the model's performance and adjust hyperparameters (such as learning rate, margin, regularization coefficient) to optimize the model's performance. Performance evaluation metrics: Accuracy: The proportion of correctly classified samples. Recall: The ability of the model to identify the correct class. F1 Score: The harmonic mean of accuracy and recall.

[0153] Through methods such as cross - validation, ensure the generalization ability of the model and avoid overfitting or underfitting on the validation set. Through this series of steps, the deep - learning model can not only accurately identify existing categories but also effectively generalize to new package states that do not appear in the training set, greatly improving the intelligence level and adaptability of the entire express - package informatization monitoring system.

[0154] ​S5, Status Monitoring and Early Warning: Use the trained model to monitor the status of express packages in real time. If the model predicts possible problems, it will immediately give an early warning, such as warning that the package may be shaken, the temperature is too high, etc.

[0155] Once potential anomalies are detected, the system needs to quickly take the following steps to ensure the safety of the package: Early Warning Generation: According to the anomaly detection results, generate corresponding early warning information, including key data such as anomaly type, time, location, etc. For example:

[0156] Vibration Early Warning: The package is shaken beyond the normal range during transportation, which may cause damage to the internal items.

[0157] Temperature Early Warning: The temperature of the environment where the package is located exceeds the safe range, which may affect the quality of temperature-sensitive items.

[0158] Early Warning Notification: Send the early warning information to relevant personnel in real time through multiple channels (such as text messages, emails, mobile app notifications, etc.) to ensure timely response. For example:

[0159] Automated Response Measures: In some scenarios, the system can automatically trigger response measures, such as adjusting the transportation route, adjusting the warehouse environment parameters, or starting backup equipment, to minimize the impact of anomalies on the package status.

[0160] Visualization and User Interface: To facilitate user monitoring and management, the system provides an intuitive visualization interface to display real-time monitoring data, prediction results, and early warning information.

[0161] User Interaction: Users can view detailed package status information through the interface, adjust the early warning threshold, configure notification methods, and provide feedback on early warning events to continuously optimize the system performance.

[0162] Edge Computing Deployment: Allocate some computing tasks to edge devices to reduce data transmission latency and improve real-time response capabilities:

[0163] Adopt a distributed computing framework (such as Apache Kafka, Apache Spark) for data stream processing and model inference to ensure that the system still operates efficiently under large-scale data streams. Model Update and Iteration:

[0164] Multi-level Risk Assessment: Introduce a multi-level risk assessment model to classify different types and degrees of anomalies and accurately locate potential risks:

[0165]

[0166] Among them, K is the number of risk categories, m k is the number of risk indicators for the kth category, w k,i and Sk,i (t) are the weight and score respectively.

[0167] By utilizing a trained zero-shot learning deep learning model, a real-time and efficient express package status monitoring and early warning system is constructed. This process covers the full process implementation from real-time data collection, preprocessing, model inference to anomaly detection and early warning. Key mathematical models and algorithms, such as compatibility functions, risk score calculation, and dynamic threshold adjustment, ensure the stability and accuracy of the system in various complex environments. Through the timely response of the early warning mechanism, the system can effectively guarantee the safety of express packages and improve the intelligent level of logistics management. Combining the visualization interface and continuous optimization strategy, the comprehensive monitoring and intelligent early warning of the express package status are realized, providing solid technical support for the entire information monitoring system.

[0168] Embodiment 2:

[0169] An information monitoring system for express packaging based on the Internet of Things. The system is applicable to the method described above. The system includes sensors and micro cameras installed on the express packaging for collecting package information and environmental data; a data collection device with edge computing function for real-time data collection and preliminary processing; a zero-shot learning model for learning and generalizing seemingly unrelated data sets according to the rules learned from training samples; a deep learning model training device for training the collected data using the zero-shot learning algorithm; and an early warning device for using the trained model to monitor the status of express packages in real time. When the model predicts a problem, an early warning is given in real time through the early warning device.

[0170] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above embodiments of the methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0171] It should be understood that the detailed description of the technical solutions of the present invention with the help of the preferred embodiments above is illustrative rather than restrictive. Those of ordinary skill in the art can modify the technical solutions recorded in each embodiment or perform equivalent substitution on some of the technical features based on reading the specification of the present invention; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A method for monitoring express packaging information based on the Internet of Things, characterized in that: The method includes: Install different sensors and micro cameras on express packages to collect package information and environmental data; Edge computing devices collect and process data in real time nearby, improving data processing speed; Building a deep learning model: Building a zero-shot learning model; Training deep learning models: Using zero-shot learning algorithms to train the collected data, the model can predict new situations that have not appeared in the original training samples; Status monitoring and early warning: Use the trained model to monitor the status of express packages in real time. If the model predicts a problem, it will immediately give an early warning.

2. According to the express packaging information monitoring method based on the Internet of Things according to claim 1, it is characterized in that: The environmental data includes location, temperature, humidity, vibration and light.

3. According to the express packaging information monitoring method based on the Internet of Things according to claim 1, it is characterized in that: The data collection and preliminary processing include: data filtering and noise reduction, position data analysis, preliminary image analysis, and vibration data analysis.

4. According to the express packaging information monitoring method based on the Internet of Things according to claim 1, it is characterized in that: The zero-shot learning model includes the following components: Semantic embedding space: embed the semantic information of categories or states into a high-dimensional vector space, using a word embedding model; Visual embedding space: Project the fused feature vector z into the semantic embedding space for cross-space matching: v=W v z+b v Among them, W v and b v are the weight matrix and bias vector of linear transformation respectively, and v is the visual embedding vector; Compatibility function: Define a compatibility function F(v,a c ), which is used to measure the similarity between visual embedding and semantic embedding; the compatibility function is as follows: F(v, a c )=v · Time c Among them, W is the learned weight matrix, which is used to capture the relationship between the two; Loss function design: using contrast loss: Among them, γ is the marginal hyperparameter, c′ is the negative sample category, ensuring that the score of the correct category c is higher than the score of the negative sample c′ by at least γ.

5. According to the express packaging information monitoring method based on the Internet of Things according to claim 1, it is characterized in that: The training of the deep learning model includes the following steps: Forward propagation: Input the fused feature vector z of each sample in the training set into the visual embedding network to generate a visual embedding vector v: Semantic embedding: Using predefined semantic embedding vector a c For each category, indicate; Compatibility score: Calculate the compatibility score F(v,a) for each pair of visual embedding and semantic embedding c ); Loss calculation: Based on the marginal contrast loss function, the loss of each sample is calculated and the total loss of the entire training set is accumulated: Backpropagation and parameter update: Use the backpropagation algorithm to calculate the gradient of the loss function with respect to the model parameters, and update the parameters θ through the optimization algorithm: Where η is the learning rate, is the gradient of the loss function with respect to the parameters; Regularization and prevention of overfitting: The L2 regularization term is introduced to prevent the model from overfitting the training data.

6. The express packaging information monitoring method based on the Internet of Things according to claim 1 is characterized in that: The aforementioned status monitoring and early warning introduces a multi-level risk assessment model to grade abnormalities of different types and degrees and accurately locate potential risks: Where K is the number of risk categories, m k is the number of risk indicators of the kth category, w k,i and S k,i (t) are weight and score respectively.

7. An express packaging information monitoring system based on the Internet of Things, the system is applicable to the method according to any one of claims 1 to 6, characterized in that: The system includes sensors and micro cameras installed on express packages to collect package information and environmental data; a data collection device with edge computing capabilities to perform real-time data collection and preliminary processing; A zero-shot learning model is used to learn and generalize to seemingly unrelated datasets based on the patterns learned from the training samples; A deep learning model training device that uses a zero-shot learning algorithm to train the collected data; The early warning device is used to monitor the status of express packages in real time using the trained model. When the model predicts that there is a problem, a warning is given in real time through the early warning device.

Images