A multimodal freight whole life cycle state monitoring and early warning platform

Abstract

The present application relates to a kind of multimodal freight full life cycle state monitoring and early warning platform, belong to freight transport monitoring field.To solve the problem of heterogeneous perception data space-time dislocation and abnormal lack of cause and effect traceability, the present platform realizes the space-time alignment of cross-modal discrete and continuous data through the smoothing dynamic time warping and time transfer penalty mechanism in space-time alignment engine;Adaptive update baseline using dynamic baseline extraction unit perception transition event;By cause and effect dependence graph construction unit, based on relative entropy and delay cross-covariance, construct abnormal cause and effect directed acyclic graph;Finally, the mathematical anomaly is converted into on-site intervention instruction by context evaluation decision engine.The present application eliminates multi-source data island, realizes high robustness and has business executable state monitoring closed loop.

Description

Technical Field

[0001] This invention relates to the field of cargo transportation status monitoring technology, and in particular to a multimodal transport cargo full life cycle status monitoring and early warning platform. Background Technology

[0002] Multimodal transport, as an important organizational method for the long-distance cross-border transportation of modern bulk commodities and high-value goods, integrates various transportation resources such as roads, railways, inland waterways, ocean shipping, and air transport to break through the limitations of single modes of transport in terms of geographical coverage and cost structure. However, multimodal transport networks are highly dispersed in space and exhibit strong dynamic changes in time. Throughout their transportation lifecycle, goods often experience drastic environmental changes, frequent physical transshipments, and a relay of responsibility between different carriers, making real-time tracking of cargo status extremely difficult.

[0003] Existing cargo status monitoring solutions are mostly limited to a single mode of transportation, relying primarily on isolated sensor networks, such as using independent satellite positioning modules for track tracking or temperature and humidity tags to record environmental data. When containers are unloaded from ocean-going freighters to port yards and then transferred to dedicated rail freight lines, the change in transport vehicles often disrupts the monitoring chain. Furthermore, the differences in data structures, communication protocols, and sampling frequencies among the sensing devices used by different carriers result in a fragmented and isolated information landscape throughout the supply chain. To compensate for these monitoring blind spots, stakeholders often need to repeatedly replace monitoring terminals or perform manual data alignment at different stages of transportation. This not only increases operating costs but also makes it difficult to meet the timely delivery requirements for bulk and sensitive goods due to the lack of a globally unified status awareness mechanism.

[0004] In multimodal transport scenarios, the system simultaneously receives a large amount of continuous time-series data, such as acceleration and continuous ambient temperature recorded by inertial measurement units, as well as discrete asynchronous event data, such as sudden changes in engine audio, abnormal deformation images captured by cameras, and RFID scanning events at customs checkpoints. Since these data are defined on different spatial coordinate systems and time axes, comparing and fusing them requires quantifying both spatial and temporal differences, which is quite challenging. Traditional time alignment methods often rely on simple linear interpolation or absolute timestamp matching. When goods are delayed in deep-water ports or railway yards with signal obstructions, or when timestamps are disordered due to transmission delays from different sensors, conventional dynamic time warping algorithms are prone to feature misalignment because they ignore temporal offsets. This leads to incorrect alignment of later results with earlier causes, causing subsequent multimodal fusion and anomaly detection to fail due to contaminated input features.

[0005] Meanwhile, the deterioration of cargo condition in multimodal transport environments is often the cascading result of multi-physical field coupling. Existing technologies generally rely on isolated static thresholds for early warning. Normal vibrations generated by cargo traveling on rough roads trigger meaningless alarms once they exceed the rigid threshold. Furthermore, when cargo experiences slow tilting and deformation due to improper stacking, the initial physical manifestations are weak, and a single displacement sensor often fails to reach the alarm threshold. More critically, existing solutions lack calculations of the joint probability distribution between primary and secondary sensor data. When anomalies actually occur, they cannot construct a dependency graph reflecting the logical relationships between anomalies, leaving managers only able to see isolated alarm points, making it difficult to trace the propagation path and root cause of the anomalies, and thus unable to provide clear instructions to on-site personnel.

[0006] The aforementioned difficulties are coupled at the software logic level: the misalignment of spatiotemporal data directly disrupts the temporal sequence between abnormal events, which is the basis for causal inference; the lack of an isolated early warning mechanism for causal dependency graphs prevents the system from correcting spatiotemporal alignment deviations through logical reverse reasoning. Therefore, how to simultaneously solve the asynchronous spatiotemporal alignment of cross-modal heterogeneous data and the construction and contextualized early warning of causal graphs for cross-modal anomalies at the algorithmic level, thereby obtaining robust, low-false-positive, and business-executable full lifecycle monitoring capabilities, has become a pressing technical problem in this field. Summary of the Invention

[0007] To overcome the shortcomings of existing multimodal transport cargo status monitoring solutions, such as spatiotemporal misalignment in heterogeneous data fusion and lack of causal logic in anomaly tracing, the present invention aims to provide a multimodal transport cargo full lifecycle status monitoring and early warning platform, which aims to achieve accurate spatiotemporal alignment of cross-modal sensing data and contextualized closed-loop early warning based on causal correlation graphs.

[0008] Therefore, the present invention adopts the following technical solution: A multimodal transport cargo lifecycle status monitoring and early warning platform, comprising: An edge aggregation gateway is used to access the main sensor data stream that generates continuous values ​​and the secondary sensor event stream that generates discrete features, and to perform time-series normalization processing on them respectively, outputting a first normalized matrix and a second normalized matrix with the same length order of magnitude. The spatiotemporal alignment engine is used to map the first and second normalized matrices to a unified hidden feature space. When calculating the similarity cost of each time step of the two modal sequences, a longitudinal and transverse deviation penalty term is applied. The smoothed time warp function is obtained through smoothed dynamic time warping to align the time index, and the multimodal spatiotemporal alignment tensor is fused and output to eliminate the spatiotemporal misalignment in the process of joint operation field and multi-source heterogeneous data fusion. The dynamic baseline extraction unit is used to monitor the dynamic characteristics of the multimodal spatiotemporal alignment tensor to sense physical transition events, reset the historical benchmark based on the transition events and relearn the baseline model of the cargo's normal steady state, calculate the degree of deviation between the alignment feature vector of the current time slice and the baseline model, and output a local anomaly set when the limit is exceeded, thereby avoiding normal environmental changes from being misjudged as anomalies. The causal dependency graph construction unit is used to instantiate candidate anomalies in the local anomaly set into directed acyclic graph nodes. Based on the relative entropy of the empirical probability distribution of the first and second nodes and the delayed cross-covariance of the deviation intensity sequence, the causal transmission relationship is judged. When the edge construction condition is met, a directed edge from the first node to the second node is created, generating a dynamic anomaly causal association network graph matrix to distinguish between normal bumps and real physical risks caused by improper cargo stacking. The context evaluation decision engine is used to extract the main causal path from the dynamic abnormal causal relationship network graph matrix, match its feature vector with a pre-set expert experience pattern library to determine the physical failure mode, and generate a context warning message containing on-site intervention instructions, so that the monitoring is transformed from a simple status record into a closed-loop early warning that can guide on-site handling.

[0009] As a preferred embodiment of the present invention, the edge aggregation gateway includes a timing standardization module; For the main sensor data stream, the time-series normalization module uses a sliding window averaging algorithm with a forgetting factor to filter out high-frequency noise, and uses cubic spline interpolation to fill in the missing time points of low-frequency signals to obtain the first normalization matrix. For the sub-sensor event stream, the timing normalization module sets a sliding time window of fixed width to convert asynchronous events into a structured frequency density tensor by accumulating distribution frequency statistics, so as to obtain the second normalization matrix; The timing standardization module is also used to continuously monitor the relative drift of the timestamps of the main and secondary data streams. When the relative drift exceeds the preset asynchronous tolerance threshold, or the sampling frequency ratio of the main and secondary data streams exceeds the preset frequency ratio threshold, the dimensionality reduction and smooth alignment coroutine is triggered, and the control is transferred to the spatiotemporal alignment engine, thereby providing a unified length basis for subsequent network processing.

[0010] As a preferred embodiment of the present invention, the spatiotemporal alignment engine includes a spatiotemporal modal adapter and a distance cost matrix unit; The spatiotemporal modality adapter includes two parallel processing branches, each equipped with a backbone feature extraction network with frozen pre-training parameters and a self-attention projection layer connected in series thereafter, used to compress the first normalized matrix and the second normalized matrix and map them into hidden layer feature vectors with corresponding sequence lengths. The distance cost matrix unit is used to calculate the similarity cost of the two sets of hidden layer feature vectors at each time step using Euclidean distance. For each coordinate point on the dynamic programming path, its deviation from the ideal synchronization diagonal is first calculated. Then, a time transition penalty term consisting of the square of the deviation and the penalty intensity weight coefficient is added to the similarity cost as the corrected transition cost. When the deviation exceeds the tolerance threshold corresponding to the maximum physical tolerance delay, hard pruning logic is triggered, and the transition path cost of the coordinate point is given an infinite penalty value to overcome excessive temporal distortion and strictly constrain feature matching within a reasonable physical delay window.

[0011] As a preferred embodiment of the present invention, the spatiotemporal alignment engine further includes a smooth dynamic time warping module and a progressive modal complementarity adapter; The smoothing dynamic time warping module receives the corrected transition cost and replaces the hard minimum operation in the dynamic programming process with a soft minimum operation based on the expectation approximation of the exponential distribution to perform state transition recursion, so as to derive the smoothing time warp function in reverse and then perform stretching or squeezing operations on the time index of the modal sequence. The progressive modal complementarity adapter utilizes a shallow adaptation sublayer with shared weights to calculate the dot product attention weight matrix of the two sets of feature sequences after time index alignment. The dot product attention weight matrix is ​​then multiplied by the value projection matrix of the submodal features to obtain preliminary fused features, achieving shallow cross-modal information penetration. The preliminary fused features are then fed into a deep adaptation sublayer, where an intramodal attention mechanism is used to dynamically adjust the weight distribution of the information flow, outputting the multimodal spatiotemporal alignment tensor. This improves the model's robustness to high-frequency noise and automatically captures the intrinsic rhythms of heterogeneous features in a high-dimensional manifold space.

[0012] As a preferred embodiment of the present invention, the dynamic baseline extraction unit includes a transition sensing module and a steady-state baseline learning module; The transition sensing module extracts the velocity feature branch and vibration envelope feature branch of the vehicle from the multimodal spatiotemporal alignment tensor. When it is determined that the velocity feature branch continuously falls below the zero threshold and the vibration power spectral density jumps when it restarts, the module determines that a multimodal transport physical transition has occurred based on the distributed consensus state switching logic and clears the historical benchmark cache. After the physical transition determination, the steady-state baseline learning module initiates a baseline relearning silent period of a preset duration. During this silent period, the mean vector and covariance matrix are continuously statistically analyzed using the alignment feature vector on the main sensor side to construct a non-parametric Gaussian baseline model as the baseline model, ensuring that the anomaly detection mechanism can adapt to the physical dynamic differences of different carriers.

[0013] As a preferred embodiment of the present invention, the dynamic baseline extraction unit further includes a local anomaly detection module; The local anomaly detection module calculates the Mahalanobis distance between the latest time slice aligned feature vector and the mean and covariance matrix of the non-parametric Gaussian line model in real time, and uses this distance as the overall feature offset metric. When the Mahalanobis distance exceeds the dimensionless detection threshold adjusted by an adaptive multiple, the active time window where the anomaly is located is tracked starting from the trigger time. The alignment feature observation sequence and Mahalanobis deviation intensity sequence of each slice within the active time window, together with the global absolute timestamp and the normalized deviation confidence, are encapsulated and the local anomaly set is output. This effectively takes into account the correlation between multidimensional variables to accurately reflect the true deviation of the cargo status.

[0014] As a preferred technical solution of the present invention, the causal dependency graph construction unit includes an abnormal node instantiation module and a divergence edge construction module; The abnormal node instantiation module encapsulates each candidate abnormality in the local abnormality set into a node object, and extracts the normalized deviation intensity of each node within the active time window, after discretization according to a unified interval division, the empirical probability distribution defined in the common event space, and the one-dimensional deviation intensity sequence characterizing the abnormality intensity. The divergence edge building module traverses any first-mover node and subsequent-mover node falling within the graph time window, calculates the relative entropy between the empirical probability distributions of the two nodes, and the delay cross-covariance between the deviation intensity sequences after time-shift alignment with the time interval between the occurrences of the two nodes as the delay amount; when the delay cross-covariance is positive and its ratio to the relative entropy is greater than the minimum connection energy threshold for graph construction, it is determined that there is a causal propagation path and a directed edge is created, thereby realizing the mining of real physical causal chains through statistical divergence.

[0015] As a preferred technical solution of the present invention, the edge weight of the directed edge created by the divergence edge building module is jointly determined by the product of the deviation confidence of the first node, the deviation confidence of the second node, and the time delay attenuation coefficient. The causal dependency graph construction unit also includes a propagation delay correction module. The propagation delay correction module obtains macroscopic coupling parameters representing the congestion intensity of external logistics nodes, constructs a divisor factor using the congestion intensity and coupling enhancement coefficient, and performs attenuation correction on the time delay attenuation coefficient. This allows for dynamic amplification of the tolerance span of abnormal edge construction on the time axis when continuous congestion occurs, thus avoiding the breakage of logical associations due to transit delays.

[0016] As a preferred technical solution of the present invention, the context evaluation decision engine includes a graph traversal matching module, a stacking and packaging evaluation module, a payload center evaluation module, and a semantic message generation module; The graph traversal matching module uses a graph traversal mechanism to extract the main causal path with the highest weight integral from the dynamic abnormal causal association network graph matrix, calculates its composite confidence sum, and performs similarity calculation between the path feature vector and the fault matrix in the expert experience pattern library. The stacked packaging evaluation module is used to determine whether the goods stack collapses or the packaging fails when the physical failure evolution mode of vibration and morphological variation is matched; the load center evaluation module is used to collect the compression of the vehicle's multi-point independent suspension and combine it with the vehicle's effective load constant when the image offset and vehicle attitude angle are abnormal, and then calculate the current total load weight and the distribution coordinates of the center of gravity in the horizontal plane of the platform through joint force calculation. The semantic message generation module is used to combine the physical judgments obtained by each evaluation module with the preset safety scheduling knowledge rules to generate a contextual warning message containing suggestions for subsequent logistics node operations and center of gravity adjustment instructions, thus transforming the underlying mathematical anomalies into action strategies with business guidance.

[0017] As a preferred embodiment of the present invention, the platform further includes a closed-loop feedback optimization unit; The closed-loop feedback optimization unit receives the actual damage report and manual handling results at the transportation endpoint as the truth label; it maps the sum of composite confidence of the main causal path extracted by the context evaluation decision engine to a predicted anomaly edge probability with a value in the range of zero to one through a normalized mapping function with learnable parameters, and constructs a cross-entropy loss function between the predicted anomaly edge probability and the truth label. With minimizing the cross-entropy loss function as the optimization objective, backpropagation is performed to update the learnable parameters of the normalized mapping function. The weight parameters in the spatiotemporal alignment engine are updated using gradient descent based on the self-supervised alignment objective provided by smooth dynamic time warping. For the minimum connection energy threshold of the causal dependency graph construction unit and the adaptive multiple of the dynamic baseline extraction unit, which are located after the non-differentiable graph traversal and pattern matching stages and cannot have their gradients directly back propagated by the cross-entropy loss, the adjustment is made using a gradient-free optimization method with the cross-entropy loss and false positive statistics as feedback. At the same time, the false positive rate of high-frequency warnings that have not caused cargo damage is statistically analyzed according to the waybill cycle. When the false positive rate is higher than the tolerance limit, the connection energy tolerance range is automatically reduced and the abnormal trigger sensitivity is narrowed, giving the platform continuous evolution and false alarm suppression capabilities.

[0018] Compared with the prior art, the present invention has the following beneficial effects: To overcome the shortcomings of existing multimodal transport cargo status monitoring solutions in heterogeneous data fusion and anomaly tracing, this invention provides a multimodal transport cargo full lifecycle status monitoring and early warning platform. This platform organizes the processing of sensing data access, cross-modal spatiotemporal alignment, dynamic baseline extraction, causal network construction, and contextual evaluation feedback using a data bus and microservice architecture. At the underlying level, it achieves spatiotemporal alignment of cross-modal data through a smooth dynamic time warping algorithm with a quadratic time offset penalty and a progressive adapter. Furthermore, it enables causal tracing and contextualized early warning of cross-modal anomalies through a directed acyclic graph constructed using statistical divergence.

[0019] Leveraging the aforementioned spatiotemporal alignment mechanism, the platform, without replacing the underlying sensing hardware, can project low-frequency discrete event streams and high-frequency continuous physical quantities into a unified high-dimensional spatiotemporal feature manifold with only approximately 0.93 trillion trainable parameters. This eliminates spatiotemporal misalignment during the fusion of joint transport sites and multi-source heterogeneous data. Even if some sensor data is lost or delayed for a considerable period due to tunnel obstruction or port signal shielding, the platform can still maintain the continuity of freight trajectories and state matrices on the time axis. When the mode of transport changes, the platform can automatically detect the change and adjust the alignment strategy accordingly, ensuring that subsequent anomaly detection is based on time-uncontaminated features, thereby improving the accuracy and real-time performance of state perception.

[0020] Leveraging the aforementioned causal graph mechanism, the platform characterizes the logical relationships between different modal anomalies using the joint probability distribution of primary and secondary sensor data and a directed acyclic graph. By calculating relative entropy and delay propagation coefficients, it determines whether a physical causal relationship exists between anomalies, thereby distinguishing between vibrations caused by normal road surface bumps and structural resonances caused by improper cargo stacking, filtering out a significant proportion of invalid alarms. When a real physical risk is detected, the platform combines the vehicle's effective load capacity constant with image evaluation results to output executable contextual instructions, such as suggesting re-stacking at the next hub or suggesting load redistribution. This transforms monitoring from simple status recording into a closed-loop early warning system that can guide on-site handling. Simultaneously, based on on-site handling and end-point damage assessment results, the platform continuously corrects network parameters and edge thresholds through backpropagation, improving the true positive recognition rate while suppressing false alarms, helping to reduce cargo damage rates and the costs associated with ineffective scheduling. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of a preferred embodiment of the multimodal transport cargo lifecycle status monitoring and early warning platform of the present invention.

[0022] Figure reference numerals: 100, Edge aggregation gateway; 110, Main sensor access module; 120, Secondary sensor access module; 130, Temporal normalization module; 200, Spatiotemporal alignment engine; 210, Spatiotemporal modal adapter; 220, Distance cost matrix unit; 230, Smooth dynamic time warping module; 240, Progressive modal complementarity adapter; 300, Dynamic baseline extraction unit; 310, Transition perception module; 320, Steady-state baseline learning module; 330, Local anomaly detection module; 400, Causal dependency graph construction unit; 410, Anomaly node instantiation module; 420, Divergence edge construction module; 430, Propagation delay correction module; 500, Context evaluation decision engine; 510, Graph traversal matching module; 520, Stacking and packaging evaluation module; 530, Load center evaluation module; 540, Semantic message generation module; 600, Closed-loop feedback optimization unit. Detailed Implementation

[0023] To enable those skilled in the art to clearly understand and implement the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments are only used to explain the present invention and not to limit its scope of protection. The software computing core of the multimodal transport cargo life cycle status monitoring and early warning platform involved in this embodiment is preferably deployed on a graphics processor cluster with distributed tensor computing capabilities, and the elastic scaling and fault transfer of each algorithm microservice are realized by means of a container orchestration and scheduling framework. Its front end is connected to the data reporting gateways deployed on each multimodal transport node and transport vehicle. It should be noted that the multimodal transport cargo life cycle status monitoring and early warning platform does not logically rely on any fixed hardware wiring under a single transport mode. Instead, it uses a data bus and microservice architecture as the skeleton, and organizes the processing processes such as the access of raw sensing data, cross-modal spatiotemporal alignment, dynamic baseline extraction, causal network construction and even context evaluation feedback into a unidirectional cascaded processing flow. At the same time, it is supplemented by the stepwise propagation of the loss function along the reverse path, so that the platform has the deep learning capability to continuously evolve with the accumulation of freight data.

[0024] like Figure 1 As shown, the multimodal transport cargo lifecycle status monitoring and early warning platform, from data entry to early warning exit, includes an edge aggregation gateway 100, a spatiotemporal alignment engine 200, a dynamic baseline extraction unit 300, a causal dependency graph construction unit 400, a context evaluation and decision engine 500, and a closed-loop feedback optimization unit 600. These components are interconnected using a standardized data state flow protocol. The output of one processing stage becomes the input source for the next. The truth labels output by the context evaluation and decision engine 500 and the closed-loop feedback optimization unit 600 flow back along the reverse gradient path to the aforementioned network layers, forming a self-correcting closed loop throughout the entire processing chain.

[0025] The edge aggregation gateway 100 is responsible for accessing and time-series standardizing raw IoT data. Internally, it includes a main sensor access module 110, a secondary sensor access module 120, and a time-series standardization module 130. The main sensor access module 110 monitors data streams reported by devices generating continuous values, including a six-axis inertial measurement unit (IMU) for characterizing the three-dimensional spatial attitude of the carrier, a temperature and humidity probe, and a satellite positioning module that samples on a fixed schedule. The secondary sensor access module 120 receives event streams reported by devices generating irregular discrete features, including a color camera recording the deformation state of cargo packaging, an acoustic detector capturing instantaneous collision stress changes, and RFID barcode scanning events recording changes in transportation nodes. The communication of the edge aggregation gateway 100 is driven concurrently by several asynchronous coroutines to avoid high-frequency reporting from one type of device blocking other links. Simultaneously, it categorizes the accessed data sources into two modal groups—continuous values ​​and discrete events—based on the aforementioned distinctions, laying the foundation for subsequent differentiated resampling.

[0026] The timing normalization module 130 constructs a virtual global standard time axis in memory and applies different normalization strategies to the two sets of data accordingly. For continuous data from the main sensor access module 110, considering the significant differences in sampling frequencies between different devices—for example, the sampling rate of a vehicle-mounted six-axis inertial measurement unit can reach 100 times per second, while the sampling rate of the maritime satellite positioning module may be as low as once per 100 seconds due to power-saving strategies—the timing normalization module 130 uses a sliding window averaging algorithm with a forgetting factor to filter out high-frequency noise. Its update form can be expressed as follows: In the formula and These are the original observation value and the filtered output value at the current moment, respectively. The forgetting factor, whose value is closer to a certain threshold, indicates a stronger dependence on historical observations and more effective suppression of high-frequency noise. For missing time points in low-frequency signals, cubic spline interpolation is used to fill in the gaps, ensuring the continuous sequence remains smooth and continuous on the time axis. For discrete events from the secondary sensor access module 120, the time-series normalization module 130 sets a fixed-width sliding time window to convert asynchronous pulse or text records into structured frequency density tensors using cumulative distribution frequency statistics. This ensures that the two types of data, originally of vastly different sequence lengths, have similar lengths before entering the neural network. Simultaneously, the time-series normalization module 130 continuously monitors the relative drift of the timestamps of the two data streams. In the formula and These represent the first and second data streams, respectively. The timestamps of the data units to be aligned on the global standard timeline. The sampling sequence number increments along this time axis; once Exceeding the preset asynchronous tolerance threshold Or the sampling frequency ratio of the primary and secondary data streams If the frequency ratio exceeds the preset threshold, it is determined that the two can no longer directly enter the subsequent neural network. Therefore, the dimensionality reduction and smoothing alignment coroutine is forcibly triggered, and the control is transferred to the spatiotemporal alignment engine 200.

[0027] The two sets of feature sequences processed by the temporal normalization module 130 are fed into the spatiotemporal alignment engine 200 to solve the spatiotemporal misalignment problem in the process of multi-modal joint operation and multi-source heterogeneous data fusion. The spatiotemporal alignment engine 200 has two parallel spatiotemporal modality adapters 210, which respectively receive the normalization matrix from the main sensor side and the normalization matrix from the secondary sensor side. Each spatiotemporal modality adapter 210 consists of a backbone feature extraction network with frozen pre-trained parameters and a lightweight self-attention projection layer connected in series. This projection layer uses matrix multiplication to uniformly compress and map heterogeneous signals with vastly different physical dimensions and semantics to, for example, a 256-dimensional hidden feature space. This initially bridges the dimensional gap without destroying the inherent characteristics of each modality, and outputs the hidden layer feature vector. and ,in , These are the sequence lengths of the continuous and discrete modes after mapping, respectively. Thanks to the freezing of the backbone network parameters and the lightweight design of the projection layer, the spatiotemporal modal adapter 210 and the subsequent progressive modal complementarity adapter 240 require only about 0.93 megabytes of trainable parameters to complete the unified projection of low-frequency discrete event streams and high-frequency continuous physical quantities in the cloud or at the edge node.

[0028] After obtaining the two sets of feature vectors, the distance cost matrix unit 220 allocates a two-dimensional floating-point matrix space in memory and calculates the similarity cost of the two sets of feature vectors at each time step using Euclidean distance. To overcome the excessive temporal distortion caused by the blind pursuit of the shortest path in traditional dynamic time warping, a diagonal offset constraint is introduced in the distance cost matrix element 220. For each coordinate point on the dynamic programming path... First, calculate its longitudinal and transverse deviations relative to the ideal synchronous diagonal. Then, a time transition penalty term that increases quadratically with the degree of deviation is applied, so that the modified transition cost is expressed as In the formula The weighting coefficients are used to control the intensity of the penalty. That is, the aforementioned longitudinal and transverse deviation. The square of; when Exceeding the coefficient corresponding to the maximum physical tolerance delay At that time, the distance cost matrix unit 220 triggers hard pruning logic, assigning an infinite penalty value to the cost of the transition path, thereby forcing feature matching to fall only within a reasonable physical time delay window.

[0029] The smoothing dynamic time warping module 230 receives the aforementioned penalized distance cost matrix and replaces the hard minimum operation, which is non-differentiable and prone to local misalignment in traditional dynamic programming, with a smooth soft minimum operation. Its state transition recursion can be written as follows:

[0030] in This indicates alignment to coordinate points during dynamic programming. The accumulated soft regularization cost over time, the soft minimum operator is defined as follows: By calculating the expectation of all possible alignment paths according to an exponential distribution, the smoothing dynamic time warping module 230 inversely derives a smooth and globally optimal time warp function, and uses this function to stretch or compress the time index of the modal sequence, so that the time axis, which was originally disordered due to tunnel obstruction or deep-water port signal shielding, can be accurately aligned. The smoothing coefficient... The degree to which the soft minimum approximates the hard minimum can be flexibly tuned according to the noise level and alignment accuracy requirements of different transport segments. The smaller the value, the closer the alignment result is to the deterministic shortest path, while the larger the value, the smoother the gradient and the more robust it is to noise.

[0031] After the timeline alignment is completed, the progressive modal complementarity adapter 240 initiates progressive complementarity fusion from shallow to deep. This module first calculates the dot product attention weight matrix of the two sets of feature sequences using a shallow adaptation sublayer with shared weights.

[0032] In the formula and These are the query projection matrix and the key projection matrix, which are applied to the two modal features, respectively. Both are learnable parameters. The dimension of the key vector after projection, divided by This is used to scale the dot product result and suppress gradient instability caused by excessively large values; subsequently, the attention weight matrix and the submodal feature are projected onto the matrix. Multiplying the value matrices obtained from the mapping yields preliminary fusion features. In the formula The value projection matrix acting on the submodal features is a learnable parameter; this allows for the mutual penetration of shallow information between the two modalities, and then the preliminary fused features are... The data is fed into a deep adaptation sublayer with layer normalization and nonlinear activation. The weight distribution of the information flow is dynamically adjusted using an intramodal attention mechanism. Finally, a multimodal spatiotemporal alignment tensor with a rigorous temporal structure and highly aggregated spatial features is output. Each column shares a homogeneous relative time axis index. Because the above alignment process automatically captures the inherent rhythm of continuous tensors and discrete event streams in the high-dimensional manifold space of the hidden layer, the spatiotemporal alignment engine 200 can complete high-precision data fusion and missing data filling in an environment without manual alignment annotation; especially when the transportation mode changes, such as seamlessly switching from road to rail or sea, the spatiotemporal alignment engine 200 can automatically sense the switch and adaptively adjust the alignment strategy, so that even if some sensor data is lost or delayed for a long time, the continuity of the freight trajectory and state matrix throughout the entire life cycle can still be maintained.

[0033] Multimodal spatiotemporal alignment tensor After entering the dynamic baseline extraction unit 300, the platform no longer relies on absolute physical thresholds, but instead dynamically restarts statistical baseline learning based on the switching events of transportation modes to avoid misjudging normal environmental changes as anomalies. The transition perception module 310 utilizes the multimodal spatiotemporal alignment tensor... The dynamic characteristic branches of the vehicle are extracted, such as the satellite positioning velocity and engine vibration envelope, and the change events of the carrier are authenticated using a distributed consensus state switching logic similar to the Raft protocol. When a certain velocity characteristic is continuous If the vibration power spectral density drops below the near-zero threshold within seconds and then undergoes a significant jump upon restarting (e.g., a sudden change from low-frequency, long-wave ocean swaying to high-frequency rail joint impact), the transfer sensing module 310 determines that a multimodal transport physical transfer has occurred. It then clears the historical reference cache and suspends the anomaly detection thread, putting the system into a period of... The baseline relearning quiescent period is defined. During this quiescent period, the steady-state baseline learning module 320 continuously calculates the mean vector and covariance matrix of the high-dimensional aligned tensor from the main sensor side, such as the result of fusing heavy vehicle suspension system pressure and inertial measurement data. This constructs a non-parametric Gaussian baseline model representing the current normal steady state of the cargo, with its mean value denoted as . Covariance is .

[0034] As new data continues to flow in, the local anomaly detection module 330 calculates the latest time slice feature vector in real time. Mahalanobis distance between the dynamic baseline and the baseline

[0035] In the formula This is the aligned feature vector of the current time slice. This metric naturally incorporates the correlation between multidimensional feature variables, and therefore reflects the overall shift in cargo status better than a rigid threshold for dimension-by-dimensional comparison. When Exceeding the dimensionless detection threshold That is, satisfying At that time, among them The dimensionless detection threshold is determined by taking the square root of the upper quantile of the chi-square distribution with degrees of freedom equal to the feature dimension, and then adjusting it using a dimensionless adaptive sensitivity coefficient. Scaling and tuning yields the following: Under the non-parametric Gausky line assumption, the square of the Mahalanobis distance approximately follows a chi-square distribution. Therefore, the chi-square quantile is used to characterize the threshold, which is both dimensionless and has a clear statistical significance. Since the Mahalanobis distance, after being normalized by the inverse covariance matrix, no longer contains the physical dimensions of the original observations, it is compared with the threshold, which is also dimensionless. Direct comparison avoids the inconsistency of dimensions and threshold failure caused by multiplying dimensionless quantities with dimensional scales. The local anomaly detection module 330 continuously tracks the active time window of the anomaly starting from the trigger time, and records the alignment feature observation sequence of each slice within the window, the corresponding Mahalanobis deviation intensity sequence, along with the global absolute timestamp, the dimension identifier of the anomaly feature involved, and the deviation confidence. They are all encapsulated as structured records for output and aggregated into a local anomaly set. Among them, the deviation from confidence level It is obtained by normalizing the deviation intensity relative to the trigger threshold.

[0036] Local anomaly set The process proceeds to the causal dependency graph construction unit 400 to mitigate massive false alarms and achieve expert-level source tracing and diagnosis. The anomaly node instantiation module 410 instantiates local anomaly sets. Each candidate anomaly, along with its characteristic observation sequence and deviation intensity sequence within its active time window, is encapsulated as a node object in the graph data structure. Based on this, two types of statistics are pre-derived for each node: one is an empirical probability distribution defined in the common event space, obtained by discretizing the normalized deviation intensity within the window according to a unified interval division; the other is a one-dimensional deviation intensity sequence characterizing the change of the node's anomaly intensity over time. Meanwhile, the anomaly node instantiation module 410 maintains a directed acyclic graph model with sliding time window characteristics in memory, automatically performing garbage collection on expired nodes that exceed the time window to constrain the graph size and ensure the timeliness of inference.

[0037] The divergence edge building module 420 uses a background traversal coroutine to process any two nodes that fall sequentially into the graph on the time axis. and An investigation was conducted, among which those meeting the requirements were... Considering that isolated single-time-slice feature vectors are insufficient to characterize the correlation at the distribution level, the divergence edge building module 420 instead calculates based on the statistics derived from the node activity window: on the one hand, it utilizes relative entropy, i.e., KL divergence.

[0038] Evaluate the empirical probability distribution of two nodes , The information difference between them, in the formula , That is, the aforementioned empirical probability distribution derived from the normalized deviation strength of the two-node windows based on the same interval division and defined in the same discrete event space, summing the index. Traverse each discrete interval under the unified interval division; because the two empirical distributions are based on the same support set, the relative entropy calculation is mathematically well-defined and can be directly implemented in programming, without dimension mismatch; on the other hand, after time-shifting and aligning the deviation intensity sequence with the occurrence time interval of the two nodes as the delay, calculate the delay cross-covariance of the two. This is used to measure whether the intensity fluctuations of the initiating node are propagated to the subsequent nodes under the corresponding time delay. The edge construction logic function is set accordingly. In the formula To prevent the denominator from approaching zero with a minimal positive value, max(0,·) represents a non-negative truncation, ensuring that the edge-building logic function takes a positive value only when there is a positive causal propagation from the initiating node to the subsequent node, while negative correlations are truncated to zero and no edge is built; when the deviation intensity sequences of the initiating and subsequent nodes are significantly positively correlated at the corresponding time delay, i.e. When the value is positive and relatively large, and the KL divergence of the empirical distributions of both is small, Take the larger value. When Constructing a minimum connectivity energy threshold across the graph At that time, the divergence edge building module 420 determined that there was a causal propagation path between the two nodes in physical and logical terms, and therefore created a path from the edge building module 420. point to The weight of a directed edge is determined by the product of the deviation confidences of the two nodes and the time delay decay coefficient, denoted as . and The deviation confidence levels are respectively and That is, each of the two The weight of the directed edge can then be expressed as: In the formula This is the time delay attenuation coefficient; thus, a dynamic anomaly causal relationship network graph matrix is ​​gradually generated. For example, when the damaged and deformed packaging inside the container identified by the visual model on the secondary sensor side, and the lateral acceleration impact of the railway track change recorded by the primary sensor side, show extremely small KL divergence in the empirical distribution of the above statistics and significantly positive delayed cross-covariance deviating from the intensity sequence, the divergence edge building module 420 determines that the two are adjacent links on the same physical causal chain, rather than isolated occasional alarms.

[0039] Considering that fluctuations in logistics efficiency at multimodal transport transshipment nodes directly alter the propagation rate of abnormal states, the propagation delay correction module 430 maps externally acquired macroscopic logistics coupling parameters into the directed acyclic graph model, such as the congestion level or waiting time for transshipment at port container yards, and corrects the time decay strength in the edge weights, making it represent as follows: In the formula Characterizing the current congestion intensity, This is the coupling enhancement coefficient. Under prolonged congestion and stagnation, the propagation delay correction module 430 dynamically amplifies the tolerance span of anomalies on the time axis based on this coefficient, thereby ensuring that even minor anomalies separated by several days can be accurately captured by the graph and associated, avoiding the breakage of the causal chain due to transit delays.

[0040] Contextual evaluation decision engine 500 uses dynamic anomaly causal relationship network graph matrix The system takes a pre-built expert knowledge mapping dictionary as input and translates the underlying mathematical anomalies into actionable strategies with business guidance. This mapping dictionary records the tolerance ranges for thresholds such as weight, deformation, and temperature. The graph traversal matching module 510 utilizes graph traversal mechanisms such as depth-first search to match the dynamic anomaly causal relationship network graph matrix. The system extracts the anomalous evolutionary chain with the highest weighted integral, i.e., the main causal path, and calculates the sum of composite confidence for this longest connected path. Then, it performs similarity calculations between the feature vector of the extracted chain and the fault matrix in the expert experience pattern library to determine its corresponding physical failure mode. When the sum of the composite confidence exceeds the system's preset global defense line, the context evaluation decision engine 500 executes a multi-path routing mechanism, assigning different evolutionary chains to the corresponding evaluation subsystems for processing.

[0041] Specifically, when the main causal path closely matches the evolution pattern of continuous low-frequency vibration, slight increase in static tilt angle to visual deformation, or manifests as abnormal diffusion of loading and unloading stress to continuous micro-displacement and image deformation, the stacking and packaging evaluation module 520 determines that the bottom cargo has physically collapsed or the packaging has failed. Therefore, it no longer outputs a simple tilt angle exceeding the limit code, but instead generates a contextual instruction with a clear action direction through the semantic message generation module 540. For example, it informs the operator that cargo packaging deformation and displacement caused by improper stacking have been detected, and suggests that the cargo be unpacked and re-stacked or the damaged unit be repackaged at the next logistics transfer node to prevent the cargo from being tipped over and damaged during subsequent sea transport. When the main causal path contains camera reference image offset and abnormal suspension compression, or indicates that a specific vehicle spring is over-compressed and causes the vehicle's horizontal attitude deflection angle to continuously increase, the load center evaluation module 530 collects the compression of each independent suspension support point of the vehicle (e.g., the four corner suspensions of the vehicle). Combining the elastic coefficient and the upper limit constant of the effective load capacity set by the vehicle at the factory, it converts the compression of each support point into the bearing capacity of the corresponding support point, sums the bearing capacity of each support point to inversely deduce the total weight currently loaded, and inversely deduces the distribution coordinates of the center of gravity in the horizontal plane of the platform by taking the torque of the bearing capacity of each support point relative to the geometric center of the platform and solving the torque balance. Once it is determined that the weight exceeds the limit or the center of gravity deviates excessively from the geometric center in the horizontal plane of the platform, the load center evaluation module 530 triggers an emergency level dynamic alarm, and the semantic message generation module 540 presents corresponding instructions on the dashboard of the driver's or dispatcher's interactive terminal. For example, if the loading capacity exceeds the limit, the vehicle or cargo compartment is at risk of bottoming out and loss of control, and it is recommended to immediately remove some smaller goods to reduce the limit below the limit and redistribute the center of gravity to the geometric center. The semantic message generation module 540 integrates the diagnostic conclusions and intervention actions of the above-mentioned evaluation subsystems into a semantically rich contextual warning message, which is then pushed to the terminal application of the carrier or cargo owner, so that on-site personnel can obtain clear handling basis instead of a screen full of isolated alarm codes.

[0042] The closed-loop feedback optimization unit 600 constitutes the reverse output of the entire processing chain. After the aforementioned warning work order containing contextual association is issued and executed, the results of manual handling, such as confirmation of damage or confirmation of false alarm, and the cargo actual damage verification report issued by the transportation destination, will be sent back to the closed-loop feedback optimization unit 600 as truth labels. The closed-loop feedback optimization unit 600 uses the truth label as a supervision signal. First, it normalizes the sum of composite confidence scores of the main causal paths extracted by the context evaluation decision engine 500 to a predicted anomaly edge probability ranging from zero to one using a Sigmoid function with learnable parameters (or a Softmax function when multiple failure modes coexist). Then, it constructs a cross-entropy loss between the predicted anomaly edge probability and the actual occurrence reflected by the truth label, with minimizing this loss as the optimization objective. Since the cross-entropy loss occurs after non-differentiable steps such as graph traversal and expert mode matching, its gradient cannot be backpropagated to the front-end deep network through the aforementioned discrete logic. Therefore, it is only used to directly update the learnable parameters of the Sigmoid normalization mapping itself. The weight parameters of the fully connected layers in the spatiotemporal modality adapter 210 and the progressive modality complementarity adapter 240 are trained independently by gradient descent, relying on the self-supervised alignment objective provided by smooth dynamic time warping. As for the connection energy threshold used by the divergence edge building module 420 when building edges, Sensitivity coefficient of local anomaly detection module 330 Instead of gradient backpropagation, the cross-entropy loss and false positive rate statistics are used as feedback, and a gradient-free feedback optimization method is employed to gradually adjust the parameters. It should be noted that the alignment learning on the spatiotemporal alignment engine 200 side relies on the self-supervised objective provided by smooth dynamic time warping, which can be completed without manual alignment labeling and does not depend on gradient backpropagation of the final decision label. In contrast, the closed-loop supervised correction here, after obtaining the true null label, only applies gradient fine-tuning to the differentiable normalized mapping parameters and adjusts the aforementioned discrete thresholds using a feedback optimization method. These two processes perform their respective functions without conflict. Considering the trade-off between true positive identification and false positive suppression, the above loss function assigns adjustable weights to the missed and false positive terms respectively, allowing parameter updates to achieve a balance between improving the true positive identification rate and maintaining the overall confidence level. In addition, the closed-loop feedback optimization unit 600 sets a settlement judgment mechanism according to the waybill cycle, and counts the number of warnings issued due to isolated high-frequency vibrations within the current cycle and the number of times that did not actually cause cargo damage. Once the ratio of the two, i.e. the false positive rate, continues to be higher than the tolerance limit, the connection energy fault tolerance range when building edges in the directed acyclic graph is automatically reduced and the sensitivity of local anomaly triggering is reduced. This allows the multimodal transport cargo life cycle status monitoring and early warning platform to gradually learn to ignore normal environmental noise that does not cause physical damage, and achieve the evolution goal of becoming more and more accurate with use.

[0043] It should be noted that the specific implementations in the above embodiments are merely examples, and those skilled in the art can make various equivalent substitutions or reasonable modifications without departing from the concept of the present invention. The color camera connected to the secondary sensor access module 120 can be replaced with an event camera to more sensitively capture instantaneous deformations and reduce data bandwidth usage; the Mahalanobis distance used by the local anomaly detection module 330 can also be replaced with a distance metric or reconstruction error metric based on a kernel function, as long as it can take into account the correlation between multidimensional features; the smoothing coefficient of the smoothing dynamic time warping module 230 and the secondary penalty weight in the distance cost matrix unit 220 in the spatiotemporal alignment engine 200 can be adaptively tuned according to the noise characteristics and delay tolerance of different transportation segments; the relative entropy and delay cross-covariance used by the divergence edge building module 420 when evaluating node association can also be replaced with equivalent metrics that can reflect distribution similarity and temporal transmission relationships, such as maximum information coefficient, mutual information, or time delay correlation coefficient; the depth-first search used by the graph traversal matching module 510 can be replaced with weighted longest path search or cluster search to balance efficiency and recall in large-scale graphs; the consensus logic based on the transition perception module 310 is not limited to Raft-like approaches. The protocol can be implemented using any mechanism that can achieve consensus authentication for carrier switching events in a distributed environment. In terms of deployment, the computing core of the multimodal transport cargo lifecycle status monitoring and early warning platform can be centrally deployed on a cloud-based graphics processing unit cluster, or the spatiotemporal alignment engine 200 and dynamic baseline extraction unit 300 can be deployed to edge computing nodes next to the transport vehicle to shorten loopback latency and adapt to weak network environments. The monitored objects are not limited to bulk commodities; they are also applicable to temperature- and vibration-sensitive medical supplies, cold chain goods, etc. All such modifications and substitutions are natural extensions and refinements of the technical solution of this invention and still fall within the scope of protection claimed by this invention.

Claims

1. A multimodal transport cargo full lifecycle status monitoring and early warning platform, characterized in that, include: An edge aggregation gateway is used to access the main sensor data stream that generates continuous values ​​and the secondary sensor event stream that generates discrete features, and to perform time-series normalization processing on them respectively, outputting a first normalized matrix and a second normalized matrix with the same length order of magnitude. The spatiotemporal alignment engine maps the first and second normalized matrices to a unified hidden feature space. When calculating the similarity cost of each time step of the two modal sequences, it applies a longitudinal and transverse deviation penalty term. After smoothing dynamic time warping, it obtains a smooth time warp function to align the time index and fuses the output multimodal spatiotemporal alignment tensor. The dynamic baseline extraction unit is used to monitor the dynamic characteristics of the multimodal spatiotemporal alignment tensor to sense physical transition events, reset the historical benchmark based on the transition events and relearn the baseline model of the cargo's normal steady state, calculate the degree of deviation between the alignment feature vector of the current time slice and the baseline model, and output a local anomaly set when the limit is exceeded. The causal dependency graph construction unit is used to instantiate candidate anomalies in the local anomaly set into directed acyclic graph nodes. It judges the causal transmission relationship based on the relative entropy of the empirical probability distribution of the first node and the second node, and the delayed cross-covariance of the deviation intensity sequence. When the edge construction condition is met, it creates a directed edge from the first node to the second node, and generates a dynamic anomaly causal association network graph matrix. The context evaluation decision engine is used to extract the main causal path from the dynamic abnormal causal relationship network graph matrix, match its feature vector with a pre-set expert experience pattern library to determine the physical failure mode, and generate a context warning message containing on-site intervention instructions.

2. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 1, characterized in that, The edge aggregation gateway includes a timing standardization module; For the main sensor data stream, the time-series normalization module uses a sliding window averaging algorithm with a forgetting factor to filter out high-frequency noise, and uses cubic spline interpolation to fill in the missing time points of low-frequency signals to obtain the first normalization matrix. For the sub-sensor event stream, the timing normalization module sets a sliding time window of fixed width to convert asynchronous events into a structured frequency density tensor by accumulating distribution frequency statistics, so as to obtain the second normalization matrix; The timing standardization module is also used to continuously monitor the relative drift of the timestamps of the main and secondary data streams. When the relative drift exceeds the preset asynchronous tolerance threshold, or the sampling frequency ratio of the main and secondary data streams exceeds the preset frequency ratio threshold, the dimensionality reduction and smooth alignment coroutine is triggered, and the control is transferred to the spatiotemporal alignment engine.

3. The multimodal transport cargo full life cycle status monitoring and early warning platform as described in claim 1, characterized in that, The spatiotemporal alignment engine includes a spatiotemporal modality adapter and a distance cost matrix unit; The spatiotemporal modality adapter includes two parallel processing branches, each equipped with a backbone feature extraction network with frozen pre-training parameters and a self-attention projection layer connected in series thereafter, used to compress the first normalized matrix and the second normalized matrix and map them into hidden layer feature vectors with corresponding sequence lengths. The distance cost matrix unit is used to calculate the similarity cost of the two sets of hidden layer feature vectors at each time step using Euclidean distance. For each coordinate point on the dynamic programming path, its longitudinal and transverse deviations relative to the ideal synchronization diagonal are first calculated. Then, a time transition penalty term consisting of the square of the longitudinal and transverse deviations and the penalty intensity weight coefficient is added to the similarity cost as the corrected transition cost. When the deviation in length and width exceeds the tolerance threshold corresponding to the maximum physical tolerance delay, hard pruning logic is triggered, and the transfer path cost of that coordinate point is given an infinite penalty value.

4. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 3, characterized in that, The spatiotemporal alignment engine also includes a smooth dynamic time warping module and a progressive modal complementarity adapter. The smoothing dynamic time warping module receives the corrected transition cost and replaces the hard minimum operation in the dynamic programming process with a soft minimum operation based on the expectation approximation of the exponential distribution to perform state transition recursion, so as to derive the smoothing time warp function in reverse and then perform stretching or squeezing operations on the time index of the modal sequence. The progressive modal complementarity adapter uses a shallow adaptation sublayer with shared weights to calculate the dot product attention weight matrix of the two sets of feature sequences after time index alignment, and multiplies the dot product attention weight matrix with the value projection matrix of the submodal features to obtain preliminary fusion features, realizing shallow information cross-modal penetration; and sends the preliminary fusion features into a deep adaptation sublayer, uses an intramodal attention mechanism to dynamically adjust the weight distribution of the information flow, and outputs the multimodal spatiotemporal alignment tensor.

5. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 1, characterized in that, The dynamic baseline extraction unit includes a transition sensing module and a steady-state baseline learning module; The transition sensing module extracts the velocity feature branch and vibration envelope feature branch of the vehicle from the multimodal spatiotemporal alignment tensor. When it is determined that the velocity feature branch continuously falls below the zero threshold and the vibration power spectral density jumps when it restarts, the module determines that a multimodal transport physical transition has occurred based on the distributed consensus state switching logic and clears the historical benchmark cache. After the physical transition determination, the steady-state baseline learning module starts a baseline relearning silent period of a preset duration. During this silent period, the mean vector and covariance matrix are continuously statistically calculated using the alignment feature vector on the main sensor side to construct a non-parametric Gaussian baseline model as the baseline model.

6. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 5, characterized in that, The dynamic baseline extraction unit also includes a local anomaly detection module; The local anomaly detection module calculates the Mahalanobis distance between the latest time slice aligned feature vector and the mean and covariance matrix of the non-parametric Gaussian line model in real time, and uses this distance as the overall feature offset metric. When the Mahalanobis distance exceeds the dimensionless detection threshold adjusted by an adaptive multiple, the active time window where the anomaly is located is tracked starting from the trigger time. The alignment feature observation sequence and Mahalanobis deviation intensity sequence of each slice within the active time window, together with the global absolute timestamp and the normalized deviation confidence, are encapsulated and the local anomaly set is output.

7. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 6, characterized in that, The causal dependency graph construction unit includes an abnormal node instantiation module and a divergence edge construction module. The abnormal node instantiation module encapsulates each candidate abnormality in the local abnormality set into a node object, and extracts the normalized deviation intensity of each node within the active time window, after discretization according to a unified interval division, the empirical probability distribution defined in the common event space, and the one-dimensional deviation intensity sequence characterizing the abnormality intensity. The divergence edge building module traverses any first-mover node and subsequent-mover node falling within the graph time window, calculates the relative entropy between the empirical probability distributions of the two nodes, and the delay cross-covariance between the deviation intensity sequences after time-shift alignment with the time interval between the occurrences of the two nodes as the delay amount; when the delay cross-covariance is positive and its ratio to the relative entropy is greater than the minimum connection energy threshold for graph construction, it is determined that there is a causal propagation path and a directed edge is created.

8. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 7, characterized in that, The weight of the directed edge created by the divergence edge building module is determined by the product of the deviation confidence of the first node, the deviation confidence of the second node, and the time delay decay coefficient. The causal dependency graph construction unit also includes a propagation delay correction module. The propagation delay correction module obtains macroscopic coupling parameters representing the congestion intensity of external logistics nodes, constructs a divisor factor using the congestion intensity and coupling enhancement coefficient, and performs attenuation correction on the time delay attenuation coefficient to dynamically amplify the tolerance span of abnormal edge construction on the time axis when continuous congestion occurs.

9. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in claim 1, characterized in that, The context evaluation decision engine includes a graph traversal matching module, a stacking and packaging evaluation module, a payload center evaluation module, and a semantic message generation module. The graph traversal matching module uses a graph traversal mechanism to extract the main causal path with the highest weight integral from the dynamic abnormal causal association network graph matrix, calculates its composite confidence sum, and performs similarity calculation between the path feature vector and the fault matrix in the expert experience pattern library. The stacked packaging evaluation module is used to determine whether the goods stack collapses or the packaging fails when the physical failure evolution mode of vibration and morphological variation is matched; the load center evaluation module is used to collect the compression of the vehicle's multi-point independent suspension and combine it with the vehicle's effective load constant when the image offset and vehicle attitude angle are abnormal, and then calculate the current total load weight and the distribution coordinates of the center of gravity in the horizontal plane of the platform through joint force calculation. The semantic message generation module is used to combine the physical judgments obtained by each evaluation module with the preset safety scheduling knowledge rules to generate a contextual warning message containing suggestions for subsequent logistics node operations and center of gravity adjustment instructions.

10. The multimodal transport cargo lifecycle status monitoring and early warning platform as described in any one of claims 1 to 9, characterized in that, The platform also includes a closed-loop feedback optimization unit; The closed-loop feedback optimization unit receives the actual damage report and manual handling results at the transportation endpoint as the truth label; it maps the sum of composite confidence of the main causal path extracted by the context evaluation decision engine to a predicted anomaly edge probability with a value in the range of zero to one through a normalized mapping function with learnable parameters, and constructs a cross-entropy loss function between the predicted anomaly edge probability and the truth label. With minimizing the cross-entropy loss function as the optimization objective, backpropagation is performed to update the learnable parameters of the normalized mapping function. The weight parameters in the spatiotemporal alignment engine are updated using gradient descent based on the self-supervised alignment objective provided by smooth dynamic time warping. For the minimum connection energy threshold of the causal dependency graph construction unit, which is located after the non-differentiable graph traversal and pattern matching stages and cannot have its gradient directly back propagated by the cross-entropy loss, and the adaptive multiple of the dynamic baseline extraction unit, the adjustment is made using a gradient-free optimization method with the cross-entropy loss and false positive statistics as feedback. At the same time, the high-frequency warning false positive rate that has not caused cargo damage is statistically analyzed according to the waybill cycle. When the false positive rate is higher than the tolerance limit, the connection energy fault tolerance range is automatically reduced and the abnormal trigger sensitivity is narrowed.

Images