In-transit transportation abnormity intelligent early warning system for intelligent supply chain

Abstract

The invention relates to the field of image recognition, in particular to an in-transit transportation abnormity intelligent early warning system for an intelligent supply chain. The visual information acquisition module is used for acquiring land visual representation, vehicle motion time sequence signals, operation flow time sequence and intersection point visual representation; the morphological correlation model construction module is used for outputting seaport scene visual representation through semantic segmentation; outputting a sea-land intersection point form correlation model through an image registration algorithm; the event attribution module is used for adding the vehicle motion time sequence signal and the sea transportation environment data to the sea-land intersection point form correlation model to obtain sea-land visual time sequence representation; performing feature analysis to obtain attribution judgment; generating a transportation event attribution heat map through a feature attribution algorithm; and the risk judgment module is used for generating an early warning signal when the cause is judged to be damage and the visual activation degree is lower than a threshold value. According to the method, end-to-end multi-dimensional coupling from geometric modeling to space-time attribution to risk decision is carried out, so that the logicality and credibility during damage distinguishing are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition, specifically to an intelligent early warning system for in-transit transportation anomalies in an intelligent supply chain. Background Technology

[0002] In complex transit environments, real-time and accurate image pattern recognition of cargo visual patterns is crucial for security. Utilizing computer vision technology, especially image pattern recognition, has become a core method. It primarily employs pattern-based semantic segmentation and visual feature extraction techniques to classify and detect specific visual patterns such as damaged containers or broken seals. This is currently one of the mainstream technologies for achieving cargo security monitoring.

[0003] However, existing image pattern recognition schemes suffer from significant deficiencies in pattern classification accuracy when dealing with the specific scenario of complex multimodal transport handover points, making it difficult to meet the needs of proactive security early warning. Their image pattern recognition models used for risk classification suffer from problems such as scene comprehension confusion and limited model generalization ability. Handover points commonly experience visual obstruction leading to difficult or incomplete semantic segmentation; marker wear affects the stability of image feature extraction and the accuracy of pattern classification; and normal traces that are visually similar to cargo anomalies but semantically different directly reduce the accuracy of pattern classification. Existing systems lack sufficient scene context and geometric prior information, limiting the effectiveness of image alignment and correction. Secondly, the system's early warning mechanism relies solely on passive responses to static image patterns. The system only performs snapshot-style burst control when image pattern recognition detects obvious abnormal patterns, lacking the ability to analyze time-series image patterns. How to use cameras or surveillance cameras to capture port environment images, establish a geometric deviation attribution model that can classify transport damage and malicious tampering patterns, integrate land and sea data, identify geometric state pattern mismatches, realize a paradigm shift from passive damage pattern detection to active attribution and security, and improve the ability to automatically identify and control high-risk cargo.

[0004] To address this, a smart early warning system for in-transit anomalies in an intelligent supply chain is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent early warning system for in-transit transportation anomalies in an intelligent supply chain, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent early warning system for in-transit anomalies in an intelligent supply chain, comprising: Visual information acquisition module: At the starting point, a land visual representation is generated using cameras, and the vehicle movement timing signal is acquired by a mobile vision system; at the sea-land junction, the timing of the operation process and the visual representation of the junction point are obtained through video analysis; The morphological association model construction module inputs the visual representation of the junction point into the visual analysis neural network model, performs semantic segmentation and feature extraction, and outputs the visual representation of the harbor scene; based on the time sequence of the operation process, the visual representation of the land and the visual representation of the harbor scene are input into the visual morphological differential coupling model, and the morphological association model of the land-sea junction point is output through the image registration algorithm. Event Attribution Module: Acquires maritime environmental data; Applies vehicle motion time-series signals and maritime environmental data to the land-sea interface morphological correlation model to obtain a land-sea visual time-series representation; Based on the land-sea visual time-series representation, performs spatiotemporal feature analysis through a deep visual model to obtain attribution determination; Generates a transportation event attribution heatmap through a visual feature attribution algorithm. Risk assessment module: When the attribution is determined to be damage and the visual activation of the transportation event attribution heatmap is lower than the preset visual activation threshold, an early warning signal is generated.

[0007] Preferably, the specific process of the visual information acquisition module includes: capturing images of the target object at the origin using a camera to generate multi-angle static images; processing the multi-angle static images using an image feature extraction algorithm to output a land visual representation; acquiring video streams of the target object's movement using a mobile vision system to generate a vehicle motion video stream; processing the vehicle motion video stream based on a motion analysis algorithm to extract motion vectors and output vehicle motion timing signals; capturing the operation process using monitoring cameras deployed at the land-sea junction to generate an operation process video stream; processing the operation process video stream using a video analysis engine to identify key operation events, generate key event timestamps, and construct an operation process timing sequence; extracting key video frames based on the operation process timing sequence to generate a junction point visual representation; measuring transportation environment information using environmental sensors to generate sensor physical readings, and converting the data into a time series through data standardization and timing construction to generate maritime environmental data.

[0008] Preferably, the specific generation process of the visual representation of the harbor scene includes: inputting the visual representation of the junction point into a visual analysis neural network model; extracting features from the visual representation of the junction point through multi-layer convolution operations via the encoder path, and outputting a high-level feature map; receiving the high-level feature map via the decoder path, performing upsampling and multi-scale information fusion, performing pixel-level classification on the fused high-level feature map, and outputting a pixel-level segmentation mask; extracting the target cargo region from the visual representation of the junction point based on the pixel-level segmentation mask; and the pixel data of the target cargo region constituting the visual representation of the harbor scene.

[0009] Preferably, the specific architecture of the visual analysis neural network model includes: an encoder path for receiving the visual representation of the intersection point; the encoder path includes convolutional layers and pooling layers, the convolutional layers performing feature extraction operations on the input data, and the pooling layers downsampling the feature maps output by the convolutional layers; the encoder path outputs high-level feature maps and fine-grained feature maps; a decoder path for receiving the high-level feature maps; the decoder path includes a transposed convolutional layer, the transposed convolutional layer performing upsampling operations on the high-level feature maps to generate upsampled feature maps; skip connections for obtaining the fine-grained feature maps from the encoder path; the decoder path fusing the upsampled feature maps and the fine-grained feature maps to generate fused feature maps; and a classification layer for receiving the fused feature maps, performing pixel-level classification, and outputting the pixel-level segmentation mask.

[0010] Preferably, the specific construction process of the land-sea junction morphological association model includes: receiving the land visual representation and the harbor scene visual representation; extracting feature points from the land visual representation and the harbor scene visual representation using an image registration algorithm, calculating the feature vectors of the feature points and matching them, and generating a transformation matrix based on the matched feature vectors; applying the transformation matrix to the land visual representation and outputting the registered land visual representation; and performing pixel-level comparison between the registered land visual representation and the harbor scene visual representation based on the time-triggered morphological difference calculation of the operation process to generate a morphological difference map, thereby constituting the land-sea junction morphological association model.

[0011] Preferably, the specific generation process of the land-sea visual temporal representation includes: receiving the vehicle motion temporal signal and the maritime environment data; applying causal convolution and dilated convolution through a time series processing model to extract time-dependent features and output sensor temporal feature vectors; performing a spatial copying operation on the sensor temporal feature vectors to match the spatial dimension of the land-sea junction morphological association model to generate a temporal feature map; and concatenating the temporal feature map with the land-sea junction morphological association model in the channel dimension to generate a fused feature map, thus constituting the land-sea visual temporal representation.

[0012] Preferably, the specific generation process of the transportation event attribution heatmap includes: receiving the attribution determination as the target category, obtaining the activation map from the convolutional layer of the deep vision model through the visual feature attribution algorithm, calculating the gradient of the output score of the attribution determination relative to each channel in the activation map, generating channel importance weights, using the channel importance weights to perform weighted summation on the corresponding channels of the activation map to generate a coarse activation map, and enlarging the size of the coarse activation map to form the transportation event attribution heatmap.

[0013] Preferably, the specific architecture of the deep vision model includes: a three-dimensional convolutional layer processing the temporal representation of the land and sea vision and outputting a spatiotemporal feature map; a final convolutional layer receiving the spatiotemporal feature map and outputting a final activation map; a global average pooling layer receiving the final activation map, performing spatial dimension compression, and outputting a high-dimensional feature vector; and a classification layer receiving the high-dimensional feature vector, performing classification calculation, and outputting the attribution determination.

[0014] Preferably, the specific process of the risk determination module includes: receiving the transportation event attribution heatmap; quantifying the transportation event attribution heatmap using the activation calculation model; calculating the average activation value of pixels in the transportation event attribution heatmap; and outputting visual activation; calculating the mean and standard deviation of activation of normal samples based on historical data statistical analysis; and outputting a preset visual activation threshold; determining whether the attribution determination is damage using a logic comparator, and determining whether the visual activation is lower than the preset visual activation threshold, and outputting a Boolean determination result; and generating the warning signal when the Boolean determination result is true.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By inputting the visual representation of the seaport junction into a visual analysis neural network model and performing semantic segmentation, a standardized visual representation of the seaport scene, free from complex background interference, is output. An image registration algorithm is used to process the land and seaport scene visual representations, resolving the geometric distortion problem caused by images taken at different times, angles, and scales. Morphological differences are calculated between the two aligned and standardized visual representations, outputting a morphological correlation model of the sea-land junction. This quantifies geometric deviations, improving the accuracy of subsequent data support for distinguishing between transport damage and tampering.

[0016] 2. The land-sea interface morphology association model is combined with vehicle motion time-series signals representing land data and maritime environmental data representing storm weather data to generate a unified land-sea visual time-series representation. The spatiotemporal features of the fused representation are analyzed by a deep visual model, enabling it to uncover the deep causal relationship between geometric deviations and the transportation process. If a geometric deviation cannot be explained by any time-series data, the system can identify it as a high-risk non-transportation anomaly, improving the system's ability to shift from passive detection to proactive attribution and achieving a paradigm shift.

[0017] 3. When assessing risk, the system relies not only on the attribution results but also on the visual activation of the transportation event attribution heatmap. A warning signal is only generated when the attribution is determined to be damage and the visual activation is below a preset visual activation threshold. Since typical transportation damage is usually obvious and generates high visual activation on the transportation event attribution heatmap, these events generally have lower risk. Conversely, malicious tampering is geometrically subtle and unknown to the model; its visual evidence is spatially local, resulting in a lower average activation value on the transportation event attribution heatmap. By specifically capturing these low-activation, high-risk signals, the generated warning signal becomes a security alert against malicious tampering, rather than a regular damage report. This allows limited manual inspection resources to be concentrated on goods marked by the system, improving the efficiency of automated control of high-risk goods.

[0018] 4. The morphological correlation model of the land-sea junction point output by the morphological correlation model construction module is input into the event attribution module. This geometric deviation is fused with vehicle motion time-series signals and maritime environmental data. The output attribution judgment and transportation event attribution heatmap are then used as input to the risk judgment module. The risk judgment module outputs a warning signal when the attribution is determined to be damage and the visual activation level is lower than the preset visual activation threshold. This end-to-end forced coupling from geometric modeling to spatiotemporal attribution to risk decision-making ensures that the warning signal output by the system is not a simple abnormal alarm, but a security signal that has undergone triple verification of morphological quantification, causal attribution, and risk qualitative analysis. This improves the logical integrity and credibility of the entire system in distinguishing between transportation damage and malicious tampering, thereby realizing the transformation from passive detection to proactive and reliable security warnings and improving the efficiency of automatic deployment for high-risk goods. Attached Figure Description

[0019] Figure 1 This is a structural diagram of an intelligent early warning system for in-transit transportation anomalies in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the construction process of the morphological association model proposed in an embodiment of this invention. Figure 3 This is a flowchart illustrating the risk assessment and early warning process proposed in an embodiment of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-3The present invention provides an intelligent early warning system for in-transit anomalies in a smart supply chain, the specific modules of which are as follows: Visual information acquisition module: At the starting point, a land visual representation is generated using cameras, and the vehicle movement timing signal is acquired by a mobile vision system; at the sea-land junction, the timing of the operation process and the visual representation of the junction point are obtained through video analysis; The morphological association model construction module inputs the visual representation of the junction point into the visual analysis neural network model, performs semantic segmentation and feature extraction, and outputs the visual representation of the harbor scene; based on the time sequence of the operation process, the visual representation of the land and the visual representation of the harbor scene are input into the visual morphological differential coupling model, and the morphological association model of the land-sea junction point is output through the image registration algorithm. Event Attribution Module: Acquires maritime environmental data; Applies vehicle motion time-series signals and maritime environmental data to the land-sea interface morphological correlation model to obtain a land-sea visual time-series representation; Based on the land-sea visual time-series representation, performs spatiotemporal feature analysis through a deep visual model to obtain attribution determination; Generates a transportation event attribution heatmap through a visual feature attribution algorithm. Risk assessment module: When the attribution is determined to be damage and the visual activation of the transportation event attribution heatmap is lower than the preset visual activation threshold, an early warning signal is generated.

[0022] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0023] Example 1

[0024] This application discloses an intelligent early warning system for in-transit transportation anomalies in an intelligent supply chain. (See attached document.) Figure 1 The specific modules proposed in this invention include: a visual information acquisition module: generating a land visual representation through a camera at the starting point, and acquiring vehicle movement timing signals through a mobile vision system; obtaining the operation process timing and the handover point visual representation through video analysis at the land-sea junction; a morphological association model construction module: inputting the handover point visual representation into a visual analysis neural network model, performing semantic segmentation and feature extraction, and outputting a port scene visual representation; based on the operation process timing, inputting the land visual representation and the port scene visual representation into a visual morphological differential coupling model, and outputting a land-sea junction morphological association model through an image registration algorithm; Event Attribution Module: Acquires maritime environmental data; appends vehicle motion time-series signals and maritime environmental data to the land-sea interface morphological association model to obtain a land-sea visual time-series representation; based on the land-sea visual time-series representation, performs spatiotemporal feature analysis through a deep visual model to obtain an attribution determination; generates a transportation event attribution heatmap through a visual feature attribution algorithm; Risk Determination Module: When the attribution determination is damage and the visual activation degree of the transportation event attribution heatmap is lower than a preset visual activation threshold, a warning signal is generated.

[0025] Furthermore, at the starting point, a land visual representation is generated using cameras, and the vehicle movement timing signals are collected by a mobile vision system; at the sea-land junction, the timing of the operation process and the visual representation of the junction are obtained through video analysis; corresponding to the visual information acquisition module; the specific implementation process includes: A morphological baseline for the vehicle at its origin (e.g., the final quality inspection station of an automaker) is established. This is achieved by photographing the target object using a camera at the origin. The camera is configured to capture images at its maximum resolution of 5120 (H) x 5120 (V) pixels, generating multi-angle static images that comprehensively record the initial state of the vehicle body panels, windows, and wheels. These multi-angle static images are then processed using an image feature extraction algorithm, ORB (Oriented Fast and Rotated BRIEF). During instantiation, the algorithm's key parameters are set as follows: a maximum number of feature points of 5000, a pyramid scale factor of 1.2, and a pyramid layer count of 8. The algorithm extracts 5000 key points and their corresponding 32-byte (256-bit) descriptors; this set of descriptors outputs the land visual representation.

[0026] This embodiment employs 3D reconstruction technology, using a structured light scanner to scan the entire vehicle, generating a high-density 3D point cloud. This cloud is then processed by a Poisson reconstruction algorithm to output a 3D mesh model. Combined with a descriptor subset, this forms an enhanced land visual representation. This land visual representation is extended into a composite data structure. During image registration, 2D feature points can be used for matching, and the visual representations of intersection points can be aligned using a camera pose estimation algorithm. This allows for the calculation of morphological differences at the geometric and volumetric levels with sub-millimeter precision, enhancing the detection capability for minute dents or deformations.

[0027] A mobile vision system on a specialized vehicle transporter captures video streams of a moving target object. The device is configured for continuous recording at 1080p resolution and 30fps, generating a video stream of vehicle motion. A motion analysis algorithm, Gunnar Farneback dense optical flow, is used to calculate displacement between consecutive frames of the video stream. The parameters are set as follows: pyramid scale of 0.5, pyramid layers of 3, average window size of 15, number of iterations of 3, polynomial neighborhood size of 5, and polynomial standard deviation of 1.2. The algorithm extracts motion vectors, and the system then calculates the average amplitude (e.g., 4.5 pixels / frame) of all motion vectors at each time step (e.g., every 1 / 30th of a second). This amplitude sequence is output as the vehicle motion time-series signal.

[0028] At the sea-land interface (e.g., a container loading area), the operation process is captured by surveillance cameras deployed at the interface. These cameras operate at a resolution of 5320x3032 pixels, generating a video stream of the operation process. This video stream is processed by a video analytics engine, specifically a 3D convolutional neural network. The engine is trained to identify key operational events (e.g., "lifting begins," "hovering," "placement into the ship's hold / yard," "operation ends") and generates timestamps for these key events upon identification, for example, outputting {event: operation ends, timestamp: 2025-11-08T10:15:30Z}. The collection of these timestamps constitutes the operation sequence. Simultaneously, based on the operation sequence (specifically the "operation ends" event), the system automatically extracts key video frames from the video stream (e.g., a 5320x3032 pixel image), which generates a visual representation of the interface.

[0029] Key environmental variables affecting vehicle condition are collected throughout the transportation process, especially during the maritime transport phase. Environmental sensors, specifically multi-functional sensors fixed in the vehicle chassis or engine compartment, measure transport environment information. These sensors are configured to collect data with specific parameters: a 3-axis accelerometer sampling rate of 100 Hz with a range of ±16 g; a temperature sensor sampling rate of 1 Hz with a range of -20°C to 80°C; and a humidity sensor with a range of 10% to 90% RH. The sensors continuously record physical readings. The collected sensor physical readings are normalized using Z-scores and converted into a time series using time-series construction. The current sensor physical reading (e.g., acceleration value X) is subtracted from the historical normal mean of that sensor channel, and the difference is then divided by the historical standard deviation of that channel. This normalized multivariate time series generates the maritime environmental data.

[0030] Image feature extraction algorithms convert origin images into land visual representations; motion analysis algorithms convert vehicle motion video flow into vehicle motion time-series signals; a video analysis engine parses the operation process video stream into a visual representation of the operation sequence and handover points; and data standardization unifies sensor physical readings of different physical units into dimensionless maritime environmental data. This achieves data preprocessing and alignment at the temporal, spatial, and semantic levels, providing a necessary foundation for subsequent cross-modal fusion analysis.

[0031] Furthermore, the visual representation of the junction point is input into a visual analysis neural network model for semantic segmentation and feature extraction, outputting a visual representation of the harbor scene; corresponding to the morphological association model construction module; the specific implementation process includes: The visual representation of the junction point is input into the visual analysis neural network model. The visual representation of the junction point is a high-resolution RGB image (e.g., a key video frame of 5320x3032 pixels). The visual analysis neural network model is a U-Net architecture, and the encoder path of the U-Net model is configured to use a ResNet-50 backbone network pre-trained on the ImageNet dataset.

[0032] In practice, the visual representation of the junction points is uniformly scaled to 512x512 pixels and normalized before being fed into the encoder path of the visual analysis neural network model. Feature extraction is performed on the visual representation of the junction points through multi-layer convolutional operations via the encoder path. Specifically, the ResNet-50 encoder utilizes its 50 layers of deep convolutions and residual connections to extract hierarchical features from the input car image, ranging from low-level texture to high-level semantics. This process outputs a high-level feature map; for example, a 16x16x2048-dimensional feature tensor is generated at the deepest level of the encoder.

[0033] The high-level feature map is received through the decoder path. The core task of the decoder path is to perform upsampling and multi-scale information fusion. In this embodiment, the upsampling operation is implemented through transposed convolutional layers. Each transposed convolutional layer uses a kernel size of (2,2) and a stride of (2,2), which can magnify the spatial dimensions (height and width) of the feature map by a factor of two. After each upsampling step, the decoder path receives fine-grained feature maps (i.e., features retained before downsampling) from the corresponding layer of the encoder path through skip connections and concatenates the two in the channel dimension. At the end of the decoder path, a 1x1 convolutional layer performs pixel-level classification on the fused high-level feature map. This convolutional layer compresses the multi-channel feature map into a single channel and applies a Sigmoid activation function. The output of the Sigmoid function is a probability map with the same dimensions as the input image (e.g., 512x512x1), where the value of each pixel (ranging from 0.0 to 1.0) represents the probability that the pixel belongs to the car region. By setting a threshold (e.g., 0.5), the probability map is converted into a binary pixel-level segmentation mask (where the pixel value of the car region is 1 and the pixel value of the background region is 0).

[0034] The system extracts the target cargo region from the visual representation of the intersection point based on the pixel-level segmentation mask. This process is achieved by performing a bitmasking operation between the binary mask (upsampled back to the original 5320x3032 resolution) and the original visual representation image of the intersection point, such that all background pixels (mask value 0) are set to black, while all pixels belonging to the car (mask value 1) retain their original RGB values. The pixel data of the target cargo region (i.e., this background-removed car image) constitutes the visual representation of the harbor scene. The U-Net model is trained using the Adam optimizer, with its parameters specifically set as follows: learning rate 0.001, beta1=0.9, beta2=0.999.

[0035] Employing the U-Net architecture, leveraging its feature extraction capabilities in the encoder path and reconstruction capabilities in the decoder path, and fusing multi-scale information through skip connections, semantic segmentation of the vehicle region was achieved. By extracting pixel-level segmentation masks, a clean visual representation of the harbor scene was generated, ensuring that subsequent morphological difference comparisons are vehicle-to-vehicle rather than vehicle-to-background comparisons, thus guaranteeing the robustness and accuracy of damage recognition from the source.

[0036] Furthermore, based on the operational sequence, the visual representations of land and harbor scenes are input into a visual morphological differential coupling model. Through an image registration algorithm, a morphological association model of the land-sea interface is output; corresponding to the morphological association model construction module; see [link / reference]. Figure 2 The specific implementation process includes: The visual morphology differential coupling model (coupling process) receives the land visual representation and the harbor scene visual representation as core inputs. The land visual representation is baseline data generated at the origin (e.g., 5000 ORB feature points and their 32-byte descriptors extracted from a 5120x5120 pixel image). The harbor scene visual representation is a car image generated at the junction point after background interference has been removed (e.g., a 5320x3032 pixel RGB image). Since the shooting angle, distance, and scale of these two car images inevitably differ, an image registration algorithm is first performed.

[0037] In this embodiment, the image registration algorithm is based on the ORB (Oriented Fast and Rotated BRIEF) feature registration process. The system uses the ORB algorithm to extract feature points from the visual representation of the harbor scene. The algorithm parameters are set as follows: maximum number of feature points is 5000, pyramid scale factor is 1.2, and pyramid layer number is 8. Subsequently, the system calculates the feature vectors of the feature points (i.e., 5000 new 32-byte descriptors) and performs matching. The matching process uses Brute-Force (BF) Matcher, and the crossCheck=True parameter is enabled to ensure the symmetry and consistency of the matching, thereby filtering out most weak matches.

[0038] After matching is complete, the system generates a transformation matrix based on the matched feature vectors (e.g., the 800 initial matching points found between the two images). Since the initial matching inevitably contains erroneous matches, the system employs the RANSAC (Random Sample Consensus) algorithm to robustly estimate the transformation. Specifically, the system inputs the coordinates of the matching point pairs into the cv2.estimateAffine2D function, and the RANSAC reprojection threshold is precisely set to 5.0 (pixels), ultimately outputting a high-confidence 2x3 affine transformation matrix that describes the spatial translation, rotation, and scaling of the land visual representation (originating image) relative to the harbor scene visual representation (interchange point image).

[0039] After obtaining the transformation matrix, the system applies it to the original 5120x5120 pixel car image at the origin, outputting a registered land visual representation. This application process is implemented through an affine transformation function, using the transformation matrix as a parameter to geometrically warp the origin image, aligning it spatially with the car shape in the harbor scene visual representation (e.g., 5320x3032 pixels). At this point, the same location in both images (e.g., the center of the left front headlight) has the same pixel coordinates. This embodiment introduces a non-rigid fine registration stage after global registration. The harbor scene visual representation and the registered land visual representation are spatially divided into an NxN grid. Within each corresponding sub-block pair, small-scale feature point matching is performed to obtain matching point pairs. Based on these matching point pairs, a geometric transformation is calculated for each sub-block to obtain a local transformation. The NxN local transformations are fused into a non-rigid deformation field and applied to the registered land visual representation, outputting a finely registered land visual representation. In pixel-level comparison, it can eliminate spurious differences caused purely by different viewpoints and perspectives to the greatest extent, ensuring that the final morphological difference map can more realistically reflect physical damage and reduce the false alarm rate of the system.

[0040] Finally, based on the time-series triggering of the aforementioned workflow, morphological difference calculation is performed. The system performs pixel-level comparisons between the registered land visual representation (aligned origin image) and the harbor scene visual representation (interchange point image). In this embodiment, using L1 absolute difference for each pixel coordinate (x, y) and each color channel (c) in the image, the system calculates the absolute value of the difference between the value of the registered land visual representation at that point and the value of the harbor scene visual representation at that point, generating a morphological difference map. In this map, pixel values ​​are close to 0 (black) in areas where the vehicle's shape has not changed, while areas with physical damage (such as new dents or scratches on the vehicle body) are highlighted (non-zero values). This morphological difference map constitutes the morphological association model of the land-sea interface point.

[0041] A transformation matrix between two images is calculated using an image registration algorithm. This transformation matrix is ​​then applied to the land visual representation to generate a registered land visual representation, aligning it spatially with the harbor scene visual representation. This spatial alignment enables subsequent comparisons, generating a morphological difference map that displays changes in physical form. Simultaneously, calculations triggered by the workflow sequence ensure the temporal correlation between this morphological difference and key operational events.

[0042] Furthermore, maritime environmental data is acquired; vehicle motion time-series signals and maritime environmental data are appended to the land-sea interface morphological association model to obtain a land-sea visual time-series representation; corresponding event attribution module; the specific implementation process includes: The system receives the vehicle motion timing signal (e.g., motion vector amplitude sequence) and the maritime environment data. The maritime environment data is a multivariate time series, for example, a Z-score-normalized multidimensional data stream acquired by a Bosch CISS sensor, including ±16g triaxial acceleration sampled at 100Hz, and temperature and humidity data sampled at 1Hz. The two time series data sources are first time-aligned (e.g., normalized to 100Hz by downsampling or interpolation) and merged into a multi-channel temporal tensor (e.g., a Tx6 tensor, where T is the time step and 6 is the number of channels: 1 motion vector + 3 acceleration axes + 1 temperature + 1 humidity).

[0043] The multi-channel temporal tensor is input into a time series processing model. In this embodiment, the model is a temporal convolutional network that applies causal convolution and dilated convolution to extract temporally dependent features. Causal convolution ensures, through asymmetric padding, that the output at time step t depends only on the input before t. Dilated convolution, when applying the convolution kernel, skips holes between data points determined by the dilation factor d minus 1, exponentially expanding the receptive field without increasing computational cost. The parameters of this temporal convolutional network are set as follows: it consists of 8 stacked residual blocks, each containing two layers of 1D dilated causal convolution with a kernel size of 3 and 64 channels. The dilation factor increases exponentially in consecutive residual blocks. After processing the input temporal tensor, the temporal convolutional network outputs a sensor temporal feature vector at the last time step T.

[0044] The land-sea junction morphological association model (i.e., the morphological difference map, whose original dimension is 5320x3032) is subjected to spatial dimension normalization processing. A bilinear interpolation algorithm is used to resample it to a fixed spatial dimension, such as 256x256, to match the input of the subsequent depth vision model, resulting in a normalized morphological difference map (256x256x1). The sensor temporal feature vector is spatially copied, and this 1x128 vector is tiled or broadcast along the spatial dimensions (height and width) to match the spatial dimension of the land-sea junction morphological association model (i.e., 256x256), generating a temporal feature map with dimensions of 256x256x128, where each spatial location (x, y) contains the same 128-dimensional temporal features. The system concatenates the temporal feature map (256x256x128) with the land-sea interface morphological association model (i.e., a 256x256x1 morphological difference map) along the channel dimension. Two tensors with the same spatial dimension are stacked along the feature channel dimension to generate a fused feature map with a final dimension of 256x256x129. The system repeats this process for the T' key time windows to be analyzed, generating a T'x256x256x129 4D tensor. This 4D tensor, which integrates spatial morphological differences and temporal dynamic features, constitutes the land-sea visual temporal representation.

[0045] A time-series processing model is employed to efficiently extract time-dependent features from vehicle motion time-series signals and maritime environment data through causal convolution and dilated convolution, outputting sensor time-series feature vectors. These time-series features are then appended to each pixel of the land-sea interface morphological association model via spatial copying and channel-dimensional concatenation, generating a land-sea visual time-series representation. This allows subsequent deep vision models to simultaneously analyze "which location" of damage occurred and "what event at which time" caused the damage within the same data structure.

[0046] Furthermore, based on the sea-land visual temporal representation, spatiotemporal feature analysis is performed using a deep visual model to obtain attribution determination; a transportation event attribution heatmap is generated using a visual feature attribution algorithm; corresponding to the event attribution module; the specific implementation process includes: The deep vision model is a (2+1)-dimensional spatiotemporal convolutional network. The land-sea visual temporal representation is input into the model, first through a series of R(2+1)D 3D convolutional layers. Each R(2+1)D block decomposes a k_txkxk 3D convolution into a 1xkxk 2D spatial convolution, followed by a k_tx1x1 1D temporal convolution. The model outputs a spatiotemporal feature map by stacking (e.g., 4) such R(2+1)D modules. The spatiotemporal feature map is fed into a final convolutional layer, which outputs a final activation map (e.g., a T''x8x8x512 4D tensor). A global average pooling layer receives the final activation map and performs spatiotemporal dimension compression, i.e., calculates the average across the time, height, and width dimensions, outputting a fixed-length high-dimensional feature vector (e.g., a 1x512 vector). Finally, the classification layer (in this embodiment, a Dense fully connected layer with a softmax activation function) receives the high-dimensional feature vector, performs classification calculations, and outputs the attribution determination. This attribution determination is a probability distribution, for example: {Category: Damage-Impact, Probability: 0.85}, {Category: Damage-Vibration, Probability: 0.05}, {Category: Normal, Probability: 0.10}. The system uses the category with the highest probability (i.e., "Damage-Impact") as the final attribution determination.

[0047] The training process of the deep vision model is as follows: This embodiment constructs a training dataset containing 36,000 samples. Each sample is a land-sea visual temporal representation and its corresponding attribution label. The dataset is obtained through a combination of simulation and real-world data collection, including normal samples (24,000), which are data collected during normal transportation processes and labeled "normal"; damage-impact samples (8,000), which are generated by applying controlled physical impacts (e.g., using a pendulum to strike a specific area of ​​the cargo) during transportation, while simultaneously recording sensor data and before-and-after images, and labeled "damage-impact"; and damage-vibration samples (4,000), which are generated by fixing the cargo to a vibration table and simulating long-term high-frequency vibration, and labeled "damage-vibration". The classification layer of the deep vision model is configured to output three categories: "normal", "damage-impact", and "damage-vibration". The model is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 16, and 100 training epochs. Malicious alteration (e.g., fine cutting, lock picking) is fundamentally different in physical characteristics from impact or vibration damage in the training data. When such unseen, unknown anomalous data is input into the model, the model incorrectly classifies it as damage. However, because the input features do not match the typical high-confidence features of any known damage category (impact or vibration) learned by the model during training, the model cannot find familiar, strongly correlated evidence, resulting in low confidence in its attribution decision. This low-confidence damage determination, when visualized using the aforementioned visual feature attribution algorithm, manifests as a lack of strong, concentrated highlighted areas on the activation map, leading to a significantly low overall calculated visual activation.

[0048] The system receives the attribution decision (e.g., damage-impact) as the target category. A 3D gradient-weighted class activation map is obtained from the last convolutional layer of the deep vision model. The output score of the attribution decision (logit value before the softmax activation function) is calculated relative to the gradient of each channel in the activation map. These gradients represent the contribution of each channel to the "damage-impact" decision. The gradient of the target category score relative to the gradient of the final activation map is summed over all voxels by global average pooling of the gradients across all spatiotemporal dimensions, and then divided by the total number of voxels. Channel importance weights are generated. After obtaining the weights, the system multiplies each channel importance weight by its corresponding activation map channel, then linearly sums the results for all channels, and finally passes the sum through a ReLU activation function to retain activation features that contribute positively to the target category. Finally, the system enlarges the coarse activation map (e.g., by trilinear interpolation) to restore it to a dimension that matches the input (e.g., T'x256x256), forming the transportation event attribution heatmap. This heatmap is a 3D visualization, with highlighted areas indicating the spatial location and time point in the vehicle morphology difference map where the key event leading to the "damage-impact" attribution occurred.

[0049] A deep vision model is employed, using 3D convolutional layers to perform spatiotemporal feature analysis on the visual temporal representations of land and sea, enabling it not only to identify spatial damage but also to understand temporal dynamics, thus outputting high-level attribution judgments. A visual feature attribution algorithm is used to calculate the gradient of the attribution judgment score relative to the activation map and generate channel importance weights, ultimately producing a transportation event attribution heatmap. Visual evidence is provided, showing which segment of time-series data and which morphological differences jointly led to the model's attribution judgment, enhancing the system's transparency and credibility.

[0050] Furthermore, when the attribution is determined to be damage, and the visual activation level of the transportation event attribution heatmap is lower than a preset visual activation threshold, an early warning signal is generated; this corresponds to the risk assessment module; see [link / reference]. Figure 3 The specific implementation process includes: The system receives the transportation event attribution heatmap. To quantify this heatmap, the system sums the activation intensity values ​​of all voxels in the heatmap (e.g., a T'x256x256 tensor) using an average activation value calculator. This sum is then divided by the total number of voxels to obtain the average activation value, which is output as a scalar value, i.e., visual activation. For example, for a car damage event, the calculated visual activation might be 0.42. This visual activation scalar represents the overall strength of the visual evidence that the deep vision model can find for its attribution determination.

[0051] The system determines a preset visual activation threshold by accessing a historical database containing a large number of normal samples (i.e., confirmed undamaged car transport instances) (N=10,000). The system inputs the complete data from these 10,000 normal samples into the analysis process, generating 10,000 corresponding attribution heatmaps for normal transport events. Since these samples are undamaged, the activation values ​​in their heatmaps should theoretically only represent random noise in the model. The system then calculates the mean and standard deviation of the activation levels for these 10,000 normal samples. For example, the calculated mean is 0.015 and the standard deviation is 0.005. Adding three times the standard deviation of the normal sample activation levels to the mean of the normal sample activation levels yields the preset visual activation threshold.

[0052] Finally, the system uses a logic comparator to determine whether the attribution is an impairment and whether the visual activation level is lower than the preset visual activation threshold. The comparator outputs a Boolean result. When the Boolean result is true, the system generates the warning signal.

[0053] In this embodiment, a spatiotemporal consistency verification mechanism can be introduced into the risk assessment module. The spatial coordinates of the region with the highest activation in the transportation event attribution heatmap are obtained and back-mapped onto the land-sea interface morphological association model. It is then verified whether the spatial coordinates fall on non-zero pixels of the morphological difference map. If the difference in the spatial coordinates is 0, spatial inconsistency is determined. Further, the temporal coordinates are obtained and back-mapped onto maritime environmental data. It is then verified whether sensor readings exist within the temporal coordinates and their neighborhood. If no sensor readings exist, temporal inconsistency is determined. During the logic comparator's determination, when the attribution is determined to be damage, the visual activation is below a threshold, and both spatial and temporal consistency are true, the warning signal is generated. This triple verification ensures that the warning signal is not only based on the model's perception but is also supported by both physical morphological differences and real-time events, filtering out false alarms and enhancing the credibility of the warning signal.

[0054] The average activation value of the transportation event attribution heatmap is quantified by an activation calculation model to obtain visual activation. This visual activation is then compared with a preset visual activation threshold. An early warning signal is generated only when the attribution is determined to be damage and the visual activation is below the preset threshold. This allows the system to intelligently distinguish between known, explainable damage and unknown, novel anomalies that the model cannot understand, ensuring the scarcity and criticality of early warnings.

[0055] This invention provides an intelligent early warning system for in-transit transportation anomalies in a smart supply chain. It collects key node data uniformly through a visual information acquisition module and aligns and differentiates the "baseline morphology" at the origin with the "current morphology" at the junction point spatially through a morphological association model construction module, achieving cross-stage morphological traceability. It combines static morphological differences with dynamic time-series data. The event attribution module analyzes a unified spatiotemporal representation integrating morphological differences, vehicle motion signals, and maritime environment data using a deep visual model. This not only determines whether damage has occurred but also provides a deeper analytical capability for attributing "why the damage occurred." The system quantifies the confidence of the model's attribution judgment by using the visual activation level of the transportation event attribution heatmap. An early warning signal is only generated when the system determines damage but the visual activation level is lower than a preset visual activation threshold obtained based on normal sample statistics. This effectively filters out known, explainable routine damage while specifically capturing novel anomalies that the model cannot understand, have never seen before, and have extremely high potential risks, thus improving the criticality of the early warning signal.

[0056] Example 2

[0057] This second embodiment demonstrates the specific planning process of applying the intelligent early warning system for in-transit transportation anomalies provided by the present invention to the vehicle transfer process; the specific implementation method is as follows: The visual information acquisition module captures images of the vehicle at the origin using cameras, generating multi-angle static images at a resolution of 5120x5120 pixels. A FAST algorithm is used for high-speed detection of key points via image feature extraction, followed by Harris corner measurement to sort and select 4000 optimal points. Finally, a 32-byte rBRIEF descriptor for rotation perception is calculated. During the vehicle's land transport, the mobile vision system is configured to acquire video at 25fps with H.264 encoding, generating a vehicle motion video stream. A motion analysis algorithm, based on polynomial expansion, estimates the displacement field between two frames pixel-by-pixel to obtain the average amplitude of the motion vector, outputting the vehicle motion timing signal. At the land-sea junction, cameras capture the operation process, generating a video stream. The video analysis engine processes this stream, identifies key operational events, and generates key event timestamps; this set constitutes the operational process timing. Based on the operational process timing, the system automatically extracts the key video frames corresponding to the timestamps, generating a visual representation of the junction point. Environmental sensors measure transport environment information. The sampling rate is configured as follows: accelerometer sampling rate is 50Hz, range ±8g; temperature and humidity sensor sampling rate is 0.5Hz. Raw readings are ax=2.5g, ay=0.1g, az=1.0g, temp=18℃. Historical mean μ_accel=0.0g, historical standard deviation σ_accel=0.5g; historical mean μ_temp=22℃, historical standard deviation σ_temp=2℃. The standardized value Z_accel_x is calculated as (2.5-0.0) / 0.5=5.0. Z_temp is calculated as (18-22) / 2=-2.0. This standardized multivariate time series generates maritime environmental data, with the data structure [(timestamp_1, Z_ax, Z_ay, Z_az, Z_temp, Z_hum),...].

[0058] The visual representation of the intersection points is input into a visual analysis neural network model. The input image is downsampled through multiple convolutional and max-pooling operations to extract features and output a high-level feature map. This high-level feature map is then received and upsampled through a transposed convolutional layer. It is then concatenated with a fine-grained feature map from the encoder path via skip connections. A 1x1 convolutional layer performs pixel-level classification on the fused high-level feature map, outputting a pixel-level segmentation mask. Based on this pixel-level segmentation mask, a pixel-by-pixel multiplication with the original visual representation of the intersection points is performed to extract the clean target cargo region. The pixel data of this extracted region constitutes the visual representation of the harbor scene.

[0059] Based on the time-triggered visual morphology differential coupling model of the workflow, the system receives the land visual representation and the harbor scene visual representation. A morphological association model of the land-sea interface is constructed using an image registration algorithm. 4000 ORB features are extracted from the harbor scene visual representation. Using a Brute-Force matcher and NORM_HAMMING distance, 850 matching pairs are obtained between the two sets of feature descriptors. Then, based on the matched feature vectors, the RANSAC algorithm iteratively generates a 2x3 affine transformation matrix, with the ransacReprojThreshold set to 3.0 pixels. This transformation matrix is ​​applied to the land visual representation, outputting the registered land visual representation. A pixel-level comparison is performed between the registered land visual representation and the harbor scene visual representation. At pixel coordinates (100, 150), the registered land visual representation has a pixel value of 50, while the harbor scene visual representation has a pixel value of 200. The morphological difference map at (100, 150) is calculated to have a value of 150. This morphological difference diagram constitutes the morphological correlation model of the land-sea junction point.

[0060] The event attribution module acquires vehicle motion time-series signals and maritime environment data, merges them, and inputs them into a time-series processing model. The sequence is processed through stacked convolutional layers, outputting sensor time-series feature vectors with a data structure of [1x128 vector]. The sensor time-series feature vectors are spatially copied to match the 224x224 spatial dimension of the land-sea interface morphological association model, generating a 224x224x128 time-series feature map. This time-series feature map is then concatenated with the land-sea interface morphological association model along the channel dimension. Tensors are stacked along the channel axis to generate a T'x224x224x129 fused feature map, constituting the land-sea visual time-series representation. A deep vision model is used to perform spatiotemporal feature analysis on the land-sea visual time-series representation. A spatiotemporal feature map is output through a 3D convolutional layer decomposed into 2D spatial convolution and 1D temporal convolution. The spatiotemporal feature map flows through the final convolutional layer, outputting the final activation map. The global average pooling layer receives the final activation map, performs spatial dimension compression, and outputs a high-dimensional feature vector. The classification layer receives this vector and outputs the attribution judgment, with the data structure being {Category: Damage-Vibration, Probability: 0.92}.

[0061] The attribution decision "damage-vibration" is received as the target category. The gradient of the output score of the attribution decision relative to each channel in the activation graph is calculated. Global average pooling is performed on these gradients across all spatiotemporal dimensions to generate channel importance weights. The corresponding channels in the activation graph are weighted and summed using the channel importance weights, and a coarse activation graph is generated using the ReLU activation function. After scaling up the coarse activation graph, the transportation event attribution heatmap is formed.

[0062] The risk assessment module receives the transportation event attribution heatmap and quantifies it using an activation calculation model. The transportation event attribution heatmap is a 10x224x224 tensor with a total of 501760 voxels. The total voxel value is 10035.2. The visual activation is calculated as: 10035.2 / 501760 = 0.020. The output visual activation is 0.020. Based on historical data statistical analysis of N=5000 normal samples, the system sets a preset visual activation threshold, calculates the mean and standard deviation, and obtains a preset visual activation threshold of 0.021. The final judgment is performed through a logic comparator. When the attribution is determined to be damage and the visual activation of the transportation event attribution heatmap (0.020) is lower than the preset visual activation threshold (0.021), the system generates the warning signal.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent early warning system for in-transit anomalies in a smart supply chain, characterized in that, include: Visual information acquisition module: Generates a land visual representation at the origin using a camera, and collects vehicle motion timing signals by a mobile vision system; At the sea-land interface, video analysis is used to obtain the operational sequence and visual representation of the interface. The morphological association model construction module inputs the visual representation of the junction point into the visual analysis neural network model, performs semantic segmentation and feature extraction, and outputs the visual representation of the harbor scene; based on the time sequence of the operation process, the visual representation of the land and the visual representation of the harbor scene are input into the visual morphological differential coupling model, and the morphological association model of the land-sea junction point is output through the image registration algorithm. Event Attribution Module: Acquires maritime environmental data; Applies vehicle motion time-series signals and maritime environmental data to the land-sea interface morphological correlation model to obtain a land-sea visual time-series representation; Based on the land-sea visual time-series representation, performs spatiotemporal feature analysis through a deep visual model to obtain attribution determination; Generates a transportation event attribution heatmap through a visual feature attribution algorithm. Risk assessment module: When the attribution is determined to be damage and the visual activation of the transportation event attribution heatmap is lower than the preset visual activation threshold, an early warning signal is generated.

2. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific process of the visual information acquisition module includes: capturing images of the target object using cameras at the origin to generate multi-angle static images; processing the multi-angle static images using image feature extraction algorithms to output a land visual representation; acquiring video streams of the target object's movement using a mobile vision system to generate a vehicle motion video stream; processing the vehicle motion video stream using motion analysis algorithms to extract motion vectors and output vehicle motion timing signals; capturing the operation process using monitoring cameras deployed at the land-sea junction to generate an operation process video stream; processing the operation process video stream using a video analysis engine to identify key operation events, generate key event timestamps, and construct an operation process timing sequence; extracting key video frames based on the operation process timing sequence to generate a junction point visual representation; measuring transportation environment information using environmental sensors to generate sensor physical readings, converting the data into a time series through data standardization and timing construction to generate maritime environmental data.

3. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific generation process of the visual representation of the harbor scene includes: inputting the visual representation of the junction point into a visual analysis neural network model; extracting features from the visual representation of the junction point through multi-layer convolution operations via the encoder path, and outputting a high-level feature map; receiving the high-level feature map via the decoder path, performing upsampling and multi-scale information fusion, performing pixel-level classification on the fused high-level feature map, and outputting a pixel-level segmentation mask; extracting the target cargo region from the visual representation of the junction point based on the pixel-level segmentation mask; and the pixel data of the target cargo region constituting the visual representation of the harbor scene.

4. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific architecture of the visual analysis neural network model includes: an encoder path for receiving the visual representation of the intersection point; the encoder path includes convolutional layers and pooling layers, the convolutional layers performing feature extraction operations on the input data, and the pooling layers downsampling the feature maps output by the convolutional layers; the encoder path outputs high-level feature maps and fine-grained feature maps; a decoder path for receiving the high-level feature maps; the decoder path includes a transposed convolutional layer, the transposed convolutional layer performing upsampling operations on the high-level feature maps to generate upsampled feature maps; skip connections for obtaining the fine-grained feature maps from the encoder path; the decoder path fuses the upsampled feature maps and the fine-grained feature maps to generate fused feature maps; and a classification layer for receiving the fused feature maps, performing pixel-level classification, and outputting the pixel-level segmentation mask.

5. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific construction process of the land-sea junction morphological association model includes: receiving the land visual representation and the harbor scene visual representation; extracting feature points from the land visual representation and the harbor scene visual representation using an image registration algorithm, calculating the feature vectors of the feature points and matching them, and generating a transformation matrix based on the matched feature vectors; applying the transformation matrix to the land visual representation and outputting the registered land visual representation; and performing pixel-level comparison between the registered land visual representation and the harbor scene visual representation based on the time-triggered morphological difference calculation of the operation process to generate a morphological difference map, thus constituting the land-sea junction morphological association model.

6. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific generation process of the land-sea visual temporal representation includes: receiving the vehicle motion temporal signal and the maritime environment data; applying causal convolution and dilated convolution through a time series processing model to extract time-dependent features and output sensor temporal feature vectors; performing spatial copying operation on the sensor temporal feature vectors to match the spatial dimension of the land-sea junction morphological association model to generate a temporal feature map; and concatenating the temporal feature map with the land-sea junction morphological association model in the channel dimension to generate a fused feature map, thus constituting the land-sea visual temporal representation.

7. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific generation process of the transportation event attribution heatmap includes: receiving the attribution determination as the target category; obtaining the activation map from the convolutional layer of the deep vision model through the visual feature attribution algorithm; calculating the gradient of the output score of the attribution determination relative to each channel in the activation map; generating channel importance weights; using the channel importance weights to perform weighted summation on the corresponding channels of the activation map to generate a coarse activation map; and enlarging the size of the coarse activation map to form the transportation event attribution heatmap.

8. The intelligent early warning system for in-transit anomalies in a smart supply chain according to claim 1, characterized in that, The specific architecture of the deep vision model includes: a three-dimensional convolutional layer processes the temporal representation of the land and sea vision and outputs a spatiotemporal feature map; a final convolutional layer receives the spatiotemporal feature map and outputs a final activation map; a global average pooling layer receives the final activation map, performs spatial dimension compression, and outputs a high-dimensional feature vector; and a classification layer receives the high-dimensional feature vector, performs classification calculation, and outputs the attribution determination.

9. A smart early warning system for in-transit transportation anomalies in an intelligent supply chain according to claim 1, characterized in that, The specific process of the risk determination module includes: receiving the transportation event attribution heatmap; quantifying the transportation event attribution heatmap using the activation calculation model; calculating the average activation value of pixels in the transportation event attribution heatmap; and outputting the visual activation value; calculating the mean and standard deviation of the activation value of normal samples based on historical data statistical analysis; and outputting a preset visual activation threshold; determining whether the attribution determination is damage using a logic comparator, and determining whether the visual activation value is lower than the preset visual activation threshold, and outputting a Boolean determination result; and generating the warning signal when the Boolean determination result is true.

Images