Port cargo hazard source real-time monitoring method and system based on Internet of Things

Abstract

The invention discloses a port cargo hazard source real-time monitoring method and system based on the Internet of Things, and is applied to quality safety monitoring of a port cold chain cargo storage area. The method comprises the following steps: configuring a temperature sensor grid, a humidity sensor grid and a thermal imaging device to acquire a temperature distribution matrix, a humidity distribution matrix and a surface temperature image of a cold chain cargo area; extracting temperature and humidity time sequence change features by using a recurrent neural network, extracting space temperature gradient features of a thermal imaging image by using a deep residual network, and performing feature fusion through an adaptive weighting mechanism; the fused features are constructed into a time window sequence, and a gradient boosting tree model is utilized to evaluate the cargo state abnormity risk level and position a hazard source; according to the invention, potential hazard sources can be timely found and accurately positioned, and the intelligent level and early warning capability of port cargo safety management are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of port cargo monitoring, and particularly relates to a real-time monitoring method and system for port cargo hazard sources based on the Internet of Things. Background Art

[0002] Existing port cargo monitoring systems mainly rely on traditional point sensors for isolated data collection, with problems such as large monitoring blind spots and limited sensing capabilities; especially for temperature-sensitive cold chain goods, this isolated point monitoring method is difficult to detect local temperature and humidity anomalies in a timely manner, resulting in the inability to effectively control the potential safety hazards of the goods' quality. Traditional monitoring systems usually analyze temperature and humidity as independent variables, ignoring the complex interaction between them and their correlation with the characteristics of the goods, greatly reducing the accuracy of monitoring and the timeliness of early warning.

[0003] Existing port cargo monitoring technologies lack the ability of multi-source data fusion and intelligent analysis. Most systems still remain in the stage of simple threshold judgment and single-index monitoring. When cold chain goods are in a complex storage environment, single-index monitoring is difficult to identify potential composite risk factors and accurately locate hazard sources, resulting in managers being unable to take targeted measures when abnormal situations occur. In addition, due to the lack of in-depth mining and learning capabilities for historical data, the existing systems have limited ability to identify abnormal patterns and often have false alarms or missed alarms; although the Internet of Things and artificial intelligence technologies have been widely applied in various fields, their integrated application in the field of port cargo safety monitoring is still in its infancy.

[0004] Existing port cargo monitoring systems have problems such as insufficient real-time performance, single early warning mechanism, and low intelligence level, and are difficult to meet the needs of modern ports for refined and intelligent management of cold chain goods. When multiple batches and types of cold chain goods are stored simultaneously, the lack of accurate spatial positioning and risk grading capabilities greatly increases the management difficulty and potential loss risk. Summary of the Invention

[0005] The present invention constructs an intelligent monitoring system combining a multi-modal sensor network and a deep learning algorithm to achieve the purpose of real-time, accurate monitoring and early warning of temperature and humidity anomalies in the storage area of port cold chain goods.

[0006] The present invention provides a real-time monitoring method for port cargo hazard sources based on the Internet of Things, which is applied to the quality and safety monitoring of the storage area of port cold chain goods. The method includes the following steps: Step S1, configure the sampling intervals of the temperature sensor grid, humidity sensor grid, and thermal imaging device to be the same. Use the temperature sensor grid and humidity sensor grid to collect the temperature distribution matrix and humidity distribution matrix of the cold chain goods area, and use the thermal imaging device to obtain the surface temperature image of the goods; Step S2: Use a recurrent neural network to extract the temporal variation features of temperature and humidity, use a deep residual network to extract the spatial temperature gradient features of the thermal imaging map, and use a variational autoencoder for feature dimensionality reduction; fuse the obtained multi-modal features through a dynamic weight mechanism to construct a comprehensive feature vector; Step S3: Construct the fused comprehensive features into a time window sequence, and based on a pre-trained gradient boosting tree model, evaluate the abnormal risk level of the cargo status in the current monitoring period, and identify and give the location of the hazard source based on the abnormal risk level.

[0007] Further, in step S2, the recurrent neural network adopts a gated recurrent unit (GRU) structure, including: 2 bidirectional GRU layers, each followed by a layer normalization and a Dropout layer, and 1 fully connected output layer; the input is a sequence of temperature and humidity distribution matrices at time points, and the output is a 128-dimensional feature vector; ; ; ; ; where, represents the update gate, represents the reset gate, represents the candidate hidden state, represents the hidden state at the current moment, represents the value of the input sequence at time , , , are the weight matrices of the corresponding gates respectively, , , are the bias terms of the corresponding ones, is the sigmoid activation function, represents the Hadamard product.

[0008] Further, in step S2, the deep residual network adopts a DenseNet-121 structure, including: 1 7x7 convolutional layer, 4 dense blocks, with transition layers between each dense block, 1 global average pooling layer and 1 fully connected layer; the input is the thermal imaging temperature map, and the output is a 256-dimensional feature vector; The mathematical expression of the dense block is: , where represents the composite transformation function, including batch normalization, ReLU activation and convolution operations, represents the output of the th layer, Represents the concatenation of the outputs of all previous layers.

[0009] Furthermore, in step S2, an adaptive weighting mechanism is adopted for feature fusion. The method of feature fusion includes: calculating the uncertainty weights of each modality feature, using an attention gating network to integrate multi-modal features, and generating a context-related fusion representation. The mathematical expression of the adaptive weighting mechanism is:

[0010] The calculation expression of the uncertainty weight is: ; where represents the type of feature, corresponding to the features of different modalities respectively; represents the adaptive weight based on uncertainty; represents the feature variance, reflecting the uncertainty of the feature; represents the attention gating coefficient, represents the global context vector; represents the weight matrix of the gating network, represents the bias vector; represents the finally fused feature representation.

[0011] Furthermore, in step S3, the gradient boosting tree model adopts the LightGBM framework, which contains 100 decision trees, the maximum depth of each tree is 6, the minimum number of samples in a leaf node is 20, and the learning rate is 0.05; the prediction function of the decision tree ensemble can be expressed as: ; ; where represents the predicted value of the sample , represents the th decision tree, represents the input feature vector, represents the loss function, represents the training error term, represents the regularization term, which is used to control the model complexity, represents the sample true label; The abnormal risk probability distribution output by the model is divided into four levels: normal state, slight abnormality, moderate abnormality, and severe abnormality. When it is detected that the abnormal risk levels in three consecutive time windows reach moderate abnormality or above, the system will determine the spatial distribution characteristics of temperature and humidity abnormalities through cluster analysis, and combine with the cargo storage layout diagram to accurately locate the positions of potential hazard sources; Hazard source The positioning adopts the DBSCAN density clustering algorithm, and its mathematical expression is: , where D is the set of sensor data points, is the Euclidean distance function between point and point , and

[0012] is the neighborhood radius.

[0013] Furthermore, in the step S1, the data acquisition frequencies of the temperature sensor grid, the humidity sensor grid, and the thermal imaging device are set to 0.5 Hz; the spatial resolution of the temperature sensor grid is a 12x12 dot matrix, the spatial resolution of the humidity sensor grid is a 10x10 dot matrix, and the temperature image resolution collected by the thermal imaging device is 384x288 pixels.

[0014] Furthermore, in the step S2, the input of the recurrent neural network is the temperature and humidity data of consecutive m time points, where the value range of m is from 15 to 45, and the latent space dimension of the variational autoencoder is 64.

[0015] The classification of the abnormal risk level of the cargo status includes: when the temperature deviation does not exceed ±1.0 °C of the set threshold, the humidity deviation does not exceed ±5%RH, and the temperature difference in the thermal imaging image does not exceed 3.0 °C, it is in a normal state.

[0016] When the temperature deviation is between ±1.0 °C and ±2.0 °C of the set threshold, or the humidity deviation is between ±5%RH and ±10%RH, or the temperature and humidity change rate in a local area is detected to exceed 0.5 °C / hour or 2%RH / hour, and the temperature difference in the thermal imaging image does not exceed 3.0 °C, it is a slight abnormality.

[0017] When the temperature deviation is between ±2.0 °C and ±4.0 °C of the set threshold, or the humidity deviation is between ±10%RH and ±15%RH, or the temperature and humidity change rate in a local area is detected to exceed 1.0 °C / hour or 4%RH / hour, or the temperature difference in the thermal imaging image is between 3.0 °C and 6.0 °C, it is a moderate abnormality.

[0018] When the temperature deviation exceeds ±4.0 °C of the set threshold, or the humidity deviation exceeds ±15%RH, or the temperature and humidity change rate in a local area is detected to exceed 2.0 °C / hour or 8%RH / hour, or the temperature and humidity abnormality indicators show an accelerating deterioration trend within three consecutive time windows, or the temperature difference in the thermal imaging image exceeds 6.0 °C, it is a serious abnormality.

[0019] Based on the same inventive concept, the present invention provides a real-time monitoring system for port cargo hazard sources based on the Internet of Things, which is used to execute the real-time monitoring method for port cargo hazard sources of the present invention. The system includes: a data acquisition module, a feature extraction module, a feature fusion module, and a risk assessment module; The data acquisition module is used to obtain the temperature distribution matrix, humidity distribution matrix, and surface temperature image of the cold chain cargo area. The data acquisition module includes a temperature sensor grid, a humidity sensor grid, and a thermal imaging device.

[0020] The feature extraction module is used to extract time series features and spatial temperature gradient features from the collected data. The feature extraction module includes a recurrent neural network and a deep residual network.

[0021] The feature fusion module is used to perform adaptive weighted fusion on the extracted features to generate a comprehensive feature representation. The feature fusion module adopts an adaptive weighted mechanism based on uncertainty.

[0022] The risk assessment module is used to obtain the abnormal risk level of the cargo status in the current monitoring period based on the fused feature sequence, and identify and give the location of the hazard source based on the abnormal risk level.

[0023] Furthermore, the system further includes an early warning module, which is used to trigger an early warning mechanism when moderate or severe anomalies are detected.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing an Internet of Things architecture in which multi-modal sensor networks work together, and combining advanced deep learning and ensemble learning algorithms, the present invention realizes the comprehensive perception, intelligent analysis, and precise early warning of port cold chain cargo. The system uses a recurrent neural network to extract the time series change features of temperature and humidity, adopts a deep residual network to extract the spatial temperature gradient features of thermal imaging, and performs multi-modal feature fusion through an adaptive weighted mechanism, effectively improving the accuracy and sensitivity of anomaly detection, being able to accurately lock the abnormal area, significantly reducing the manual investigation time and disposal cost; providing comprehensive, real-time, and intelligent technical support for the safety management of port cold chain cargo, effectively reducing the cargo damage rate, and improving the logistics efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is the flowchart of the real-time monitoring method for port cargo hazard sources based on the Internet of Things of the present invention; Figure 2 is the schematic diagram of the composition of the real-time monitoring system for port cargo hazard sources based on the Internet of Things of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be described clearly and completely below. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Embodiment 1 As Figure 1 shown, it is a flowchart of a real-time monitoring method for port cargo hazard sources based on the Internet of Things, which is applied to the quality and safety monitoring of the cold chain cargo storage area in the port. The method includes the following steps: Step S1: Configure the sampling intervals of the temperature sensor grid, humidity sensor grid, and thermal imaging device to be the same. Use the temperature sensor grid and humidity sensor grid to collect the temperature distribution matrix and humidity distribution matrix of the cold chain cargo area, and use the thermal imaging device to obtain the surface temperature image of the cargo.

[0028] The data acquisition frequencies of the temperature sensor grid, humidity sensor grid, and thermal imaging device are set to 0.5 Hz; the spatial resolution of the temperature sensor grid is a 12x12 dot matrix, the spatial resolution of the humidity sensor grid is a 10x10 dot matrix, and the resolution of the temperature image collected by the thermal imaging device is 384x288 pixels.

[0029] In the present invention, the design of the sampling interval consistency of the sensor grid is to ensure the alignment of multi-modal data in the time dimension, which is crucial for subsequent feature fusion and anomaly detection. The sampling frequency of 0.5 Hz is the optimal frequency verified by a large number of experiments. It can capture the subtle trends of temperature and humidity changes without generating excessive data storage and transmission burdens. The dot matrix resolutions of 12x12 and 10x10 enable the system to monitor the small spatial changes in temperature and humidity in the cold chain cargo storage area, especially having a good observation effect on the temperature and humidity stratification phenomenon inside large containers. The thermal imaging resolution of 384x288 pixels can provide sufficiently detailed surface temperature distribution information, which helps to discover internal temperature anomalies that cannot be directly observed on the outer packaging of the cargo.

[0030] Step S2: Use a recurrent neural network to extract the temporal variation features of temperature and humidity, adopt a deep residual network to extract the spatial temperature gradient features of the thermal imaging map, and use a variational autoencoder for feature dimensionality reduction; fuse the obtained multi-modal features through a dynamic weight mechanism to construct a comprehensive feature vector.

[0031] In step S2, the recurrent neural network adopts a gated recurrent unit (GRU) structure, including: 2 bidirectional GRU layers, each followed by a layer normalization and a Dropout layer, and 1 fully connected output layer; the input is The sequence of temperature and humidity distribution matrices at each time point is output as a 128-dimensional feature vector; The mathematical expression of the GRU cell is: ; ; ; ; Among them, represents the update gate, represents the reset gate, represents the candidate hidden state, represents the hidden state at the current time, represents the value of the input sequence at time , , , are the weight matrices of the corresponding gates respectively, , , are the corresponding bias terms, is the sigmoid activation function, represents the Hadamard product.

[0032] The design concept of the GRU cell is to control the information flow through the update gate and the reset gate. Compared with the traditional LSTM structure, the GRU has the advantages of fewer parameters and higher computational efficiency, and is especially suitable for resource-constrained edge computing environments. In the processing of temperature and humidity time series data, the update gate $z_t$ determines the proportion of retaining historical information and accepting new information, while the reset gate $r_t$ controls the influence degree of historical information on the current state. When an abnormal change trend in temperature and humidity is detected, the update gate will automatically adjust its weight, making the model pay more attention to the abnormal signal of the current input. The design of the bidirectional GRU enables the model to consider both past and future information, making it more comprehensive and accurate in judging the abnormal state at the current time.

[0033] The deep residual network adopts the DenseNet-121 structure, which includes: 1 7x7 convolutional layer, 4 dense blocks, with transition layers between each dense block, 1 global average pooling layer, and 1 fully connected layer; the input is a thermal imaging temperature map, and the output is a 256-dimensional feature vector; the DenseNet-121 network is selected for processing thermal imaging maps because its unique dense connection structure has excellent feature reuse and gradient flow characteristics. Compared with traditional convolutional neural networks, DenseNet maximizes the utilization of features by directly connecting each layer to all subsequent layers, greatly reducing the number of parameters and improving computational efficiency at the same time. In the processing of thermal imaging temperature maps, the combination of shallow features (such as edges and textures) and deep features (such as semantic information) is particularly important. Dense connections enable the model to utilize multi-scale features simultaneously and effectively identify temperature anomaly regions of different sizes and shapes. In practical applications, the detection accuracy of this network structure for temperature anomalies on the surface of goods reaches over 95%.

[0034] The mathematical expression of the dense block is: , where represents the composite transformation function, which includes batch normalization, ReLU activation, and convolution operations, represents the output of the th layer, represents the concatenation of the outputs of all previous layers.

[0035] Feature fusion adopts an adaptive weighting mechanism. The methods of feature fusion include: calculating the uncertainty weights of each modality feature, using an attention gating network to integrate multi-modal features, and generating context-related fusion representations; The mathematical expression of the adaptive weighting mechanism is:

[0036] The calculation expression of the uncertainty weight is: ; where represents the th class feature, correspond to features of different modalities respectively; represents the adaptive weight based on uncertainty; represents the feature variance, which reflects the uncertainty of the feature; represents the attention gating coefficient, represents the global context vector; represents the weight matrix of the gating network, represents the bias vector; Represents the finally fused feature representation. In cold chain goods monitoring, data of different modalities may be affected by noise to varying degrees. For example, temperature sensors may be disturbed by local air flow, humidity sensors may be affected by condensed water, and thermal imaging devices may be affected by surface reflections and environmental factors. By calculating the variance of features to estimate their uncertainty, the system can automatically reduce the weights of features with high noise or instability, increase the contribution of more reliable features, thereby improving the overall quality of the fused features and the accuracy of anomaly detection. The introduction of the attention gating network further enhances the context awareness ability of feature fusion, enabling the system to dynamically adjust feature weights according to the currently observed environmental state.

[0037] The input of the recurrent neural network is the temperature and humidity data of consecutive m time points, where the value range of m is from 15 to 45, and the latent space dimension of the variational autoencoder is 64. The recurrent neural network processes the temperature and humidity data of 15 to 45 consecutive time points, which is equivalent to monitoring the state change of goods for 7.5 to 22.5 minutes. The setting of this time window is determined based on the physical characteristics of the temperature and humidity changes of cold chain goods. Experience shows that the temperature and humidity anomalies of most cold chain goods will show obvious change trends within this time range. A too short time window may miss the slowly developing anomaly patterns, while a too long time window will increase the computational burden and reduce the real-time response ability of the system. The latent space dimension of the variational autoencoder is set to 64, which greatly reduces the feature dimension while maintaining sufficient information, reduces the complexity of the subsequent classification model, and improves the generalization ability of the model.

[0038] In step S3, the fused comprehensive features are constructed into a time window sequence. Based on the pre-trained gradient boosting tree model, the anomaly risk level of the goods state in the current monitoring period is evaluated, and the location of the hazard source is identified and given based on the anomaly risk level.

[0039] In cold chain goods monitoring, some anomaly patterns may not be obvious at a single time point, but will show specific development trajectories in continuous observations. As a powerful ensemble learning method, the gradient boosting tree model can handle non-linear feature relationships and imbalanced data sets, and is particularly suitable for scenarios such as cold chain goods anomaly detection where the number of normal samples is much larger than that of abnormal samples. The pre-trained model uses a large amount of historical data, including data of normal operating states and various abnormal situations, ensuring the model's ability to identify various anomaly patterns. The hazard source location function based on the anomaly risk level provides precise intervention guidance for managers, greatly shortening the time for troubleshooting and handling.

[0040] In step S3, the gradient boosting tree model uses the LightGBM framework. The selection of the LightGBM framework is based on its advantages in dealing with high-dimensional features and large-scale data sets. Compared with traditional gradient boosting tree algorithms such as XGBoost, LightGBM adopts a histogram-based decision tree algorithm and a leaf-first growth strategy, which greatly improves the training speed and memory efficiency. The integration setting of 100 decision trees shows the best performance and efficiency balance in experimental verification. The maximum depth of the tree is 6 to prevent overfitting while maintaining sufficient model complexity. The setting of the minimum number of samples per leaf node of 20 ensures the statistical stability of each leaf node and improves the generalization ability of the model. The learning rate of 0.05 enables the model to gradually learn the complex patterns of the data and avoid falling into local optima. It includes 100 decision trees, each with a maximum depth of 6, a minimum number of samples per leaf node of 20, and a learning rate of 0.05; the prediction function of the decision tree ensemble can be expressed as: ; ; where, represents the predicted value of sample , represents the th decision tree, represents the input feature vector, represents the loss function, represents the training error term, represents the regularization term, which is used to control the model complexity, represents sample 's true label; The abnormal risk probability distribution of the model output is divided into four levels: normal state, slight abnormality, moderate abnormality, and severe abnormality. When it is detected that the abnormal risk levels in three consecutive time windows reach moderate abnormality or above, the system will determine the spatial distribution characteristics of temperature and humidity abnormalities through cluster analysis, and combine with the cargo storage layout diagram to accurately locate the positions of potential hazard sources; the division of abnormal risk levels is determined based on the sensitivity of cold chain goods to temperature and humidity changes and the actual damage threshold. For example, for high-value drugs or fresh foods, even a slight temperature deviation may lead to quality degradation or shortened shelf life. The four-level risk classification system provides sufficient fineness, which can not only avoid management fatigue caused by excessive alarms but also ensure timely response to major risks. Among them, the temperature deviation thresholds of ±1.0°C, ±2.0°C, and ±4.0°C are set with reference to international cold chain standards and actual business requirements; the humidity deviation thresholds of ±5%RH, ±10%RH, and ±15%RH take into account the humidity sensitivity differences of different types of goods; the temperature and humidity change rate index is used to capture rapidly deteriorating abnormal conditions, even if the absolute value has not exceeded the safe range.

[0041] Hazard source The positioning adopts the DBSCAN density clustering algorithm. The reason why the DBSCAN density clustering algorithm is applicable to the positioning of hazard sources is that it does not require pre-specifying the number of clusters, can automatically discover high-density regions of arbitrary shapes, and identify noise points. In the actual monitoring of port cold-chain goods, temperature and humidity anomalies often exhibit the characteristics of local concentration, which may be caused by a single heat source or cold source. The DBSCAN algorithm calculates the Euclidean distance between data points and groups the density-connected points into the same cluster, thereby identifying the spatial distribution pattern of abnormal temperature and humidity. The neighborhood radius Eps is a key parameter that determines the maximum distance for points to be grouped into the same cluster. Its mathematical expression is: , where D is the set of sensor data points, is the point and the point 's Euclidean distance function, is the neighborhood radius.

[0042] The length of the time window sequence is K, and the value range of K is from 8 to 20. The overlap rate of the sliding window is 50%; The classification of the risk level of abnormal cargo status includes: When the temperature deviation does not exceed ±1.0 °C of the set threshold, the humidity deviation does not exceed ±5%RH, and the temperature difference in the thermal imaging image does not exceed 3.0 °C, it is in the normal state.

[0043] When the temperature deviation is between ±1.0 °C and ±2.0 °C of the set threshold or the humidity deviation is between ±5%RH and ±10%RH, or the temperature and humidity change rate in a local area is detected to exceed 0.5 °C / hour or 2%RH / hour, and the temperature difference in the thermal imaging image does not exceed 3.0 °C, it is a minor anomaly.

[0044] When the temperature deviation is between ±2.0 °C and ±4.0 °C of the set threshold or the humidity deviation is between ±10%RH and ±15%RH, or the temperature and humidity change rate in a local area is detected to exceed 1.0 °C / hour or 4%RH / hour, or the temperature difference in the thermal imaging image is between 3.0 °C and 6.0 °C, it is a moderate anomaly.

[0045] When the temperature deviation exceeds ±4.0 °C of the set threshold or the humidity deviation exceeds ±15%RH, or the temperature and humidity change rate in a local area is detected to exceed 2.0 °C / hour or 8%RH / hour, or the temperature and humidity anomaly indicators show an accelerating deterioration trend within three consecutive time windows, or the temperature difference in the thermal imaging image exceeds 6.0 °C, it is a severe anomaly.

[0046] The following is an actual application example of the present invention in a large coastal port cold-chain logistics center. The logistics center covers an area of 30,000 square meters and is mainly used for storing temperature-sensitive goods such as imported and exported frozen foods, medicines, and biological products.

[0047] In this cold chain logistics center, a total of 10 sets of monitoring units are deployed in this system. Each set of monitoring units covers a refrigerated area of approximately 500 square meters and is configured as follows: a 12×12 dot matrix temperature sensor grid (144 sensing points in total), a 10×10 dot matrix humidity sensor grid (100 sensing points in total), 2 high-precision thermal imaging devices (384×288 pixel resolution), and 1 edge computing unit (integrating data acquisition, preliminary processing, and transmission functions). All monitoring units are connected to the central control system through a low-power wide area network. The data acquisition frequency is uniformly set to 0.5 Hz. All sensors are calibrated according to ISO standards. The temperature measurement accuracy reaches ±0.1 °C, and the humidity measurement accuracy reaches ±2%RH. The system adopts a hierarchical architecture. The edge computing unit is responsible for data preprocessing and primary feature extraction, and the central server is responsible for high-level feature fusion, model inference, and risk assessment.

[0048] On August 15, 2022, a temperature anomaly occurred in a batch of high-value vaccines stored in the pharmaceutical area. At 09:23:45 in the morning, the temperature sensor grid of the 5th monitoring unit detected that the temperature in a local area (coordinates E6 - F8) began to rise slowly, with an amplitude of approximately 0.3 °C / hour. At this time, it had not exceeded the static threshold preset by the system (±1.0 °C). At the same time, the surface temperature image of this area captured by the thermal imaging device showed a weak temperature gradient change.

[0049] The recurrent neural network (GRU structure) processed the temperature and humidity sequence data of the last 30 time points (about 15 minutes) and extracted a 128-dimensional time series feature vector; the DenseNet-121 network processed the thermal imaging image and extracted a 256-dimensional spatial temperature gradient feature vector; the variational autoencoder reduced these features to a 64-dimensional latent space.

[0050] The system calculates the uncertainty of each modal feature through an adaptive weighting mechanism: the variance of the temperature time series feature is 0.023, the variance of the humidity time series feature is 0.085, and the variance of the heat map feature is 0.042. Based on these uncertainty values, the system calculates that the weight of the temperature time series feature is 0.52, the weight of the humidity time series feature is 0.13, and the weight of the heat map feature is 0.35. After feature fusion, a comprehensive feature vector is generated.

[0051] At 09:38:22 AM, based on sequence analysis with a sliding time window (K = 15, i.e., 7.5 minutes), the abnormal risk probability distribution output by the gradient boosting tree model was: normal state 0.12, slight anomaly 0.68, moderate anomaly 0.19, severe anomaly 0.01. The system determined the risk level as "slight anomaly", generated a low-priority notification, and simultaneously shortened the data processing interval to 15 seconds. At 10:05:47 AM, the system detected that the abnormal probability had been continuously rising for 8 consecutive time windows, and the latest risk probability distribution was: normal state 0.03, slight anomaly 0.24, moderate anomaly 0.63, severe anomaly 0.10. The system raised the risk level to "moderate anomaly" and triggered a second-level warning.

[0052] The system analyzed the spatial distribution of temperature anomalies based on the DBSCAN density clustering algorithm (parameters Eps = 2.5, MinPts = 4), identified an abnormal cluster with the center point located at coordinates E7 - F7, and combined with the goods storage layout diagram to accurately locate that the vaccine refrigerator box with batch number VP22081203 was stored in this area. The system automatically sent SMS, email, and mobile application push notifications to the responsible person, and at the same time marked the abnormal area in the digital twin system of the monitoring center. The management personnel arrived at the scene within 10 minutes after receiving the notice and found that the refrigeration system of this batch of vaccine refrigerator boxes had failed, resulting in a slow rise in the internal temperature.

[0053] The comparative test with the traditional monitoring system demonstrated the obvious advantages of the present invention: the traditional fixed-threshold-based monitoring system did not trigger an alarm until the temperature deviation reached +1.2 °C (at 11:10:35 AM), while this system had already identified the abnormal trend when the temperature deviation was only +0.4 °C (at 09:38:22 AM), discovering potential problems approximately 90 minutes earlier.

[0054] In the 6-month comparative test, the abnormal detection accuracy rate of this system reached 94.7%, and the false alarm rate dropped to 3.2%. For the traditional system, the accuracy rate was 78.5% and the false alarm rate was as high as 15.6%. Especially in detecting slowly developing temperature and humidity anomalies, the advantages of this system were more obvious. This system could locate the abnormal source to a specific goods unit (with an accuracy of about 0.5 meters), while the traditional system could only indicate a general area (with an accuracy of about 5 - 10 meters). In the above case, this system directly pointed out the specific location and batch number of the problematic refrigerator box, greatly shortening the troubleshooting time; according to the statistical data of the logistics center, after using this system, the goods loss caused by temperature and humidity anomalies decreased by 68.4%.

[0055] Example 2 As Figure 2As shown, it is a schematic diagram of the composition of the real-time monitoring system of the port cargo hazard source based on the Internet of Things of the present invention, which is used to execute the real-time monitoring method of the port cargo hazard source of Example 1, and the system includes: a data acquisition module, a feature extraction module, a feature fusion module and a risk assessment module. The system adopts a modular design architecture, and the functional modules exchange data and collaborate functions through standardized interfaces. This design not only improves the scalability and maintainability of the system, but also enables the system to adapt to port cargo storage scenarios of different scales and types. The modular architecture also supports dynamic upgrading and replacement of functions. For example, the algorithm parameters of the feature extraction module can be customized according to the requirements of a specific type of cargo, or the configuration of the data acquisition module can be adjusted according to the infrastructure conditions of different ports. The software part of the system adopts a cloud-edge collaborative architecture, and the data processing tasks are dynamically allocated between the edge device and the cloud server according to the real-time requirements and the availability of computing resources, which not only ensures the real-time response of key warning functions, but also can use the powerful computing resources in the cloud for more complex data analysis and model optimization.

[0056] The data acquisition module is used to obtain the temperature distribution matrix, humidity distribution matrix and surface temperature image of the cold chain cargo area, and the data acquisition module includes a temperature sensor grid, a humidity sensor grid and a thermal imaging device.

[0057] The data acquisition module uses low-power wide area network (LPWAN) technology to realize the networking and data transmission of the sensor network. Compared with traditional wired or short-range wireless solutions, LPWAN technology can cover the entire port area while maintaining low energy consumption, greatly reducing the complexity and cost of infrastructure deployment. The temperature sensor grid uses a high-precision semiconductor temperature sensor array with a temperature measurement accuracy of ±0.1°C, which can detect small temperature changes; the humidity sensor grid uses a capacitive humidity sensor with a measurement range of 0-100%RH and an accuracy of ±2%RH, which is suitable for various humidity environments; the thermal imaging device uses a combination of micro thermal imaging sensors and wide-angle lenses, which can obtain a wide range of surface temperature distribution images without interfering with the normal operation of the cargo. The data acquisition module also integrates self-calibration and fault detection functions, which can automatically identify and compensate for sensor drift to ensure the accuracy and reliability of long-term monitoring data.

[0058] The feature extraction module is used to extract time series features and spatial temperature gradient features from the collected data, and the feature extraction module includes a recursive neural network and a deep residual network.

[0059] The feature extraction module runs on the edge computing unit of the system and adopts a neural network inference framework optimized for low-power devices, enabling efficient feature extraction with limited computing resources. The parameter quantities of the recurrent neural network and the deep residual network are carefully optimized to ensure both the model's expressive ability and control of the computational complexity. In actual deployment, the feature extraction module employs model quantization technology to convert original floating-point operations into integer operations, further improving the computational efficiency and energy consumption ratio. In addition, the feature extraction module also implements the incremental learning function, which can continuously optimize and update the model using newly collected data without interrupting the system operation, enabling the system to adapt to changes in the environment and cargo characteristics and maintain long-term monitoring effects.

[0060] The feature fusion module is used to perform adaptive weighted fusion on the extracted features to generate a comprehensive feature representation, and the feature fusion module adopts an adaptive weighting mechanism based on uncertainty.

[0061] The risk assessment module is used to obtain the risk level of abnormal cargo status in the current monitoring period based on the fused feature sequence using the gradient boosting tree model, and identify and give the location of the hazard source based on the abnormal risk level.

[0062] The system also includes an early warning module for triggering the early warning mechanism when moderate or severe anomalies are detected.

[0063] The early warning module adopts a multi-level and multi-channel early warning strategy, and the system automatically selects an appropriate early warning method according to the detected abnormal risk level. For minor anomalies, the system generates a low-priority notification and prompts the management personnel through the message center of the monitoring platform; for moderate anomalies, the system simultaneously sends SMS, emails, and mobile application push notifications to the relevant responsible persons and initiates more frequent data collection and analysis; for severe anomalies, the system triggers an audible and visual alarm, simultaneously notifies the emergency management department and key decision-makers through a preset emergency communication channel, and automatically generates preliminary emergency disposal suggestions. The early warning module also integrates the early warning confirmation and response tracking functions, which can record the reception, confirmation, and processing process of the early warning information, form a complete early warning response closed-loop, and provide a basis for subsequent system optimization and event backtracking. In addition, the early warning module supports personalized early warning rule configuration, and management personnel can adjust the triggering conditions and notification methods of the early warning according to specific scenarios and requirements, improving the flexibility and practicality of the system.

[0064] The specific embodiments described above further elaborate on the purpose, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A real-time monitoring method for dangerous sources of port cargo based on the Internet of Things is applied to the quality and safety monitoring of the cold chain cargo storage area in the port, which is characterized by: The method comprises the following steps: Step S1, configuring the sampling intervals of the temperature sensor grid, the humidity sensor grid and the thermal imaging device to be consistent, using the temperature sensor grid and the humidity sensor grid to collect the temperature distribution matrix and the humidity distribution matrix of the cold chain cargo area, and using the thermal imaging device to obtain the cargo surface temperature image; Step S2, using a recursive neural network to extract the time series variation characteristics of temperature and humidity, using a deep residual network to extract the spatial temperature gradient characteristics of the thermal image, and using a variational autoencoder to perform feature dimensionality reduction; the acquired multimodal features are fused through a dynamic weight mechanism to construct a comprehensive feature vector; Step S3, construct the fused comprehensive features into a time window sequence, evaluate the abnormal risk level of the cargo status in the current monitoring period based on the pre-trained gradient boosting tree model, and identify and give the location of the danger source based on the abnormal risk level.

2. According to the method of claim 1, the method is characterized in that: In step S2, the recurrent neural network adopts a gated recurrent unit GRU structure, including: 2 bidirectional GRU layers, each followed by a normalization layer and a Dropout layer, and a fully connected output layer; the input is The temperature and humidity distribution matrix sequence at each time point is output as a 128-dimensional feature vector; The mathematical expression of the GRU unit is: ； ； ； ； in, represents the update gate, Represents the reset gate, represents the candidate hidden state, Indicates the hidden state at the current moment, Represents the input sequence at time The value of , , are the weight matrices of the corresponding gates, , , is the corresponding bias term, is the sigmoid activation function, Represents the Hadamard product.

3. The method for real-time monitoring of dangerous sources of port cargo based on the Internet of Things according to claim 2 is characterized in that: In step S2, the deep residual network adopts the DenseNet-121 structure, which includes: 1 7x7 convolutional layer, 4 dense blocks, each dense block has a transition layer, 1 global average pooling layer and 1 fully connected layer; the input is the thermal imaging temperature map, and the output is a 256-dimensional feature vector; The mathematical expression of the dense block is: ,in represents a composite transformation function, including batch normalization, ReLU activation, and convolution operations, Indicates The output of the layer, Represents the concatenation of the outputs of all previous layers.

4. The method for real-time monitoring of dangerous sources of port cargo based on the Internet of Things according to claim 3 is characterized in that: In step S2, feature fusion adopts an adaptive weighting mechanism, and the feature fusion method includes: calculating the uncertainty weight of each modal feature, integrating multimodal features using an attention gating network, and generating a context-related fusion representation; The mathematical expression of the adaptive weighting mechanism is: The calculation expression of the uncertainty weight is: ;in, Indicates Class features, Corresponding to the characteristics of different modes; represents the adaptive weight based on uncertainty; Representation characteristics The variance of reflects the uncertainty of the feature; represents the attention gating coefficient, represents the global context vector; represents the weight matrix of the gating network, represents the bias vector; Represents the final fused feature representation.

5. The method for real-time monitoring of dangerous sources of port cargo based on the Internet of Things according to claim 4 is characterized in that: In step S3, the gradient boosting tree model adopts the LightGBM framework, including 100 decision trees, each tree has a maximum depth of 6, a minimum number of leaf node samples of 20, and a learning rate of 0.05; the prediction function of the decision tree ensemble can be expressed as: ; ;in, Representation sample The predicted value of Indicates A decision tree, represents the input feature vector, represents the loss function, represents the training error term, represents the regularization term, which is used to control the complexity of the model. Representation sample The true label of The model outputs abnormal risk probability distribution, which is divided into four levels: normal state, slight abnormality, moderate abnormality and severe abnormality. When the abnormal risk level of three consecutive time windows reaches moderate abnormality or above, the system will determine the spatial distribution characteristics of temperature and humidity abnormalities through cluster analysis, and accurately locate the potential danger source in combination with the cargo storage layout diagram; Hazard Sources The positioning adopts the density clustering algorithm, and its mathematical expression is: , where D is the set of sensor data points, For point and Point The Euclidean distance function, is the neighborhood radius.

6. A real-time monitoring method for dangerous sources of port cargo based on the Internet of Things according to claim 5, characterized in that: In step S1, the data acquisition frequency of the temperature sensor grid, the humidity sensor grid and the thermal imaging device is set to 0.5 Hz; the spatial resolution of the temperature sensor grid is 12x12 dot matrix, the spatial resolution of the humidity sensor grid is 10x10 dot matrix, and the temperature image resolution collected by the thermal imaging device is 384x288 pixels.

7. A real-time monitoring method for dangerous sources of port cargo based on the Internet of Things according to claim 6, characterized in that: In step S2, the input of the recursive neural network is the temperature and humidity data of m consecutive time points, the value range of m is 15 to 45, and the latent space dimension of the variational autoencoder is 64.

8. The method for real-time monitoring of dangerous sources of port cargo based on the Internet of Things according to claim 1 is characterized in that: In step S3, the length of the time window sequence is K, the value range of K is 8 to 20, and the overlap rate of the sliding window is 50%; The classification of the abnormal risk level of the cargo state includes: when the temperature deviation does not exceed ±1.0°C of the set threshold and the humidity deviation does not exceed ±5%RH, and the temperature difference of the thermal imaging image does not exceed 3.0°C, it is a normal state; When the temperature deviation is between ±1.0℃ and ±2.0℃ of the set threshold, or the humidity deviation is between ±5%RH and ±10%RH, or the detected temperature and humidity change rate in the local area exceeds 0.5℃ / hour or 2%RH / hour, and the temperature difference of the thermal imaging image does not exceed 3.0℃, it is a minor abnormality; When the temperature deviation is between ±2.0℃ and ±4.0℃ of the set threshold, or the humidity deviation is between ±10%RH and ±15%RH, or the detected temperature and humidity change rate in the local area exceeds 1.0℃ / hour or 4%RH / hour, or the temperature difference of the thermal imaging image is between 3.0℃ and 6.0℃, it is a moderate abnormality; When the temperature deviation exceeds the set threshold of ±4.0℃ or the humidity deviation exceeds ±15%RH, or the temperature and humidity change rate in the local area is detected to exceed 2.0℃ / hour or 8%RH / hour, or the temperature and humidity abnormality indicators show an accelerated deterioration trend in three consecutive time windows, or the temperature difference of the thermal imaging image exceeds 6.0℃, it is a serious abnormality.

9. A real-time monitoring system for dangerous sources of port cargo based on the Internet of Things, used to implement the real-time monitoring method for dangerous sources of port cargo according to any one of claims 1 to 8, characterized in that: The system includes: a data acquisition module, a feature extraction module, a feature fusion module and a risk assessment module; The data acquisition module is used to obtain the temperature distribution matrix, humidity distribution matrix and surface temperature image of the cold chain cargo area, and the data acquisition module includes a temperature sensor grid, a humidity sensor grid and a thermal imaging device; The feature extraction module is used to extract time series features and spatial temperature gradient features from the collected data, and the feature extraction module includes a recursive neural network and a deep residual network; The feature fusion module is used to perform adaptive weighted fusion on the extracted features to generate a comprehensive feature representation, and the feature fusion module adopts an adaptive weighting mechanism based on uncertainty; The risk assessment module is used to obtain the abnormal risk level of the cargo status in the current monitoring period based on the fused feature sequence using the gradient boosting tree model, and identify and give the location of the hazard source based on the abnormal risk level.

10. The real-time monitoring system for port cargo hazards based on the Internet of Things according to claim 9 is characterized in that: The system also includes an early warning module for triggering an early warning mechanism when a moderate abnormality or a serious abnormality is detected.

Images