A multi-system data unified logistics asset internet of things monitoring processing system

Abstract

The present application relates to the technical field of logistics asset internet of things monitoring and data unification, in particular to a multi-system data unified logistics asset internet of things monitoring processing system, comprising a graph construction unit, used for real-time acquisition of multi-source heterogeneous data streams, and for aggregated generation of dynamic time series graph snapshots; a heuristic quantification unit, used for taking scores as pseudo labels; a time series graph neural network unit, used for generation of high-dimensional state vectors for independent nodes; a model training unit, used for making the similarity of high-dimensional state vectors in vector space reproduce heuristic contextual similarity; a unified asset aggregation unit, used for solving high-dimensional state vectors and aggregating them into a unified asset view; and a hidden event inference unit, used for generation of hidden event alarms, and for determination of normal state when an abnormal inference score does not exceed an alarm threshold. The present application solves the technical problem that business system identification and internet of things identification need to be manually annotated to be aligned, and realizes automatic association of heterogeneous data and construction of a unified asset view.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things monitoring and data unification of logistics assets, in particular to a logistics asset Internet of Things monitoring processing system for multi-system data unification. BACKGROUND

[0002] With the development of modern logistics systems, asset monitoring is highly dependent on multi-system collaboration, such as warehouse management system (WMS), transportation management system (TMS), and Internet of Things (IoT) sensors. This diversified data source leads to a serious data island problem.

[0003] Currently, a key technical problem is that the system cannot automatically and accurately associate different sources of heterogeneous identity identifiers (IDs) to the same physical asset. For example, the SKU code in the business system and the sensor ID in the Internet of Things system may refer to the same physical goods, but they are independent at the data level, lack effective automatic association means, and often rely on manual annotation. The lack of such association hinders the formation of a unified asset view, making it difficult to achieve cross-system data fusion and advanced event inference. Therefore, how to automatically align heterogeneous IDs and break down data silos to achieve accurate asset monitoring and hidden event inference has become a key technical problem in the field that needs to be solved. SUMMARY

[0004] To solve the above technical problems, the present application provides a logistics asset Internet of Things monitoring processing system for multi-system data unification. Specifically, the technical solution of the present application includes:

[0005] A graph construction unit is used to collect multi-source heterogeneous data streams in real time, parse entities and contexts in the data streams into independent nodes and edges, and aggregate to generate dynamic time series graph snapshots.

[0006] A heuristic quantification unit is used to analyze the original context information in the dynamic time series graph snapshots, calculate the heuristic context similarity scores between independent nodes, and use the scores as pseudo labels.

[0007] A time series graph neural network unit is used to process the dynamic time series graph snapshots and generate high-dimensional state vectors for the independent nodes.

[0008] A model training unit is used to dynamically construct training samples based on the heuristic context similarity scores and train the time series graph neural network unit using a pre-set cross-modal context contrast loss function, so that the similarity of the high-dimensional state vectors in the vector space reproduces the heuristic context similarity.

[0009] A unified asset aggregation unit is used to use the trained time series graph neural network unit to solve the high-dimensional state vectors and align the independent nodes of different sources based on the similarity between the high-dimensional state vectors to aggregate into a unified asset view.

[0010] The hidden event inference unit is configured to establish a behavior baseline of the unified asset view, calculate a real-time anomaly inference score in combination with a deviation degree of the behavior baseline and a closeness degree of a preset risk factor, and generate a hidden event alarm when the anomaly inference score exceeds an alarm threshold, or determine a normal state when the anomaly inference score does not exceed the alarm threshold.

[0011] Preferably, the system further adopts a preset vector similarity function for calculating the similarity between high-dimensional state vectors.

[0012] Preferably, the heuristic quantification unit calculates the heuristic context similarity score in the following analysis process:

[0013] The time similarity, the space similarity and the graph topology similarity between nodes are calculated respectively.

[0014] The time similarity is used to represent the closeness of the time stamps of the node events.

[0015] The space similarity is used to represent the closeness of the nodes in the geographical space.

[0016] The graph topology similarity is used to represent whether the nodes share similar neighbors.

[0017] The time similarity, the space similarity and the graph topology similarity are weighted and summed according to a preset domain knowledge weight to obtain the heuristic context similarity score.

[0018] Preferably, the model training unit dynamically constructs training samples in the following process:

[0019] For anchor nodes, nodes with a heuristic context similarity score greater than a preset high heuristic similarity threshold are selected as positive samples.

[0020] Nodes with a heuristic context similarity score less than a preset low heuristic similarity threshold are selected as a negative sample set.

[0021] Preferably, the cross-modal context contrast loss function is used for:

[0022] High-dimensional state vectors corresponding to the anchor nodes, the positive samples and the negative sample set are obtained respectively.

[0023] A preset vector similarity function is used to calculate the vector similarity between the anchor nodes and the positive samples, and the vector similarity between the anchor nodes and each node in the negative sample set.

[0024] By minimizing the loss function, the vector similarity between the anchor nodes and the positive samples is increased, and the vector similarity between the anchor nodes and the negative samples is decreased.

[0025] Preferably, the process of aligning independent nodes by the unified asset aggregation unit is as follows:

[0026] The latest high-dimensional state vectors of all nodes are calculated by using the trained time series graph neural network unit.

[0027] The vector neighbor search is performed by using the preset vector similarity function.

[0028] The node pair with a vector similarity greater than the preset identity alignment threshold is identified as an equivalent relationship pointing to the same physical entity.

[0029] The node pair with a vector similarity not greater than the preset identity alignment threshold is identified as a non-equivalent relationship.

[0030] Preferably, the process of establishing a behavior baseline by the hidden event inference unit is as follows:

[0031] The high-dimensional state vectors corresponding to the unified asset view under the historical normal working condition are obtained.

[0032] The behavior baseline is obtained by performing time series smoothing or statistical modeling on the high-dimensional state vectors under the historical normal working condition.

[0033] Preferably, the calculation process of the anomaly inference score is as follows:

[0034] The current high-dimensional state vector of the asset is obtained.

[0035] The trajectory deviation metric between the current high-dimensional state vector and the behavior baseline is calculated.

[0036] The maximum vector similarity between the current high-dimensional state vector and each node in the predefined risk factor node set is calculated by using the preset vector similarity function.

[0037] The trajectory deviation metric and the maximum vector similarity are weighted and summed according to the preset deviation weight and proximity weight to obtain the anomaly inference score.

[0038] Preferably, the trajectory deviation metric is defined as 1 minus the similarity value between the current high-dimensional state vector and the behavior baseline calculated by using the preset vector similarity function.

[0039] Preferably, the graph construction unit collects multi-source heterogeneous data streams from warehouse management systems, transportation management systems, and Internet of Things sensors.

[0040] Among them, the business ID from the warehouse management system or the transportation management system and the sensor ID from the Internet of Things sensor are created as independent nodes in the graph during initial analysis.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] 1. The system solves the technical problem that business system identification and Internet of Things identification need to be manually annotated to align, realizes automatic association of heterogeneous data and construction of unified asset view by learning cross-modal context semantic representation;

[0043] 2. The system adopts a self-supervised training method, calculates heuristic pseudo-labels by analyzing time, space and topology and other original context information to guide model training, and avoids dependence on expensive and difficult to obtain manually annotated alignment data sets;

[0044] 3. The system realizes high-level hidden event inference; by establishing a dynamic behavior baseline for unified assets, and combining the deviation of assets from the baseline and the proximity to risk factors for double criteria, it can accurately find complex events that are difficult to monitor by traditional rules;

[0045] 4. The system improves the convergence and robustness of model training; by setting high and low heuristic similarity thresholds to dynamically construct positive and negative samples, discarding difficult samples that are also difficult for rules to distinguish, and ensuring the quality of training samples. BRIEF DESCRIPTION OF DRAWINGS

[0046] The application will be further explained below in conjunction with the accompanying drawings and embodiments:

[0047] Figure 1 is a structural diagram of the system of the application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the application clearer and more apparent, the application will be further described in detail below in conjunction with specific embodiments.

[0049] Embodiment 1:

[0050] Please refer to Figure 1 A multi-system data unified logistics asset Internet of Things monitoring processing system, comprising:

[0051] A graph construction unit for real-time acquisition of multi-source heterogeneous data streams, parsing entities and contexts in the data streams into independent nodes and edges, and aggregating to generate dynamic time series graph snapshots;

[0052] A heuristic quantization unit for analyzing original context information in the dynamic time series graph snapshots, calculating heuristic context similarity scores between independent nodes, and taking the scores as pseudo-labels;

[0053] A time series graph neural network unit for processing dynamic time series graph snapshots to generate high-dimensional state vectors for independent nodes;

[0054] The model training unit is configured to dynamically construct training samples according to the heuristic context similarity score, and train the time series graph neural network unit by using a preset cross-modal context contrast loss function, so that the similarity of the high-dimensional state vectors in the vector space reproduces the heuristic context similarity.

[0055] The unified asset aggregation unit is configured to calculate the high-dimensional state vectors by using the trained time series graph neural network unit, align the independent nodes from different sources according to the similarity between the high-dimensional state vectors, and aggregate the independent nodes into a unified asset view.

[0056] The hidden event inference unit is configured to establish a behavior baseline of the unified asset view, calculate a real-time anomaly inference score by combining the deviation of the behavior baseline and the proximity of the preset risk factors, and generate a hidden event alarm when the anomaly inference score exceeds an alarm threshold, or determine that the system is in a normal state when the anomaly inference score does not exceed the alarm threshold.

[0057] The embodiment provides a multi-system data unified logistics asset Internet of Things monitoring processing system; the system aims to solve the data island problem caused by the diversified logistics data sources such as warehouse management systems WMS, transportation management systems TMS and Internet of Things IoT in the prior art, and in particular, the technical problem that different systems cannot automatically and accurately associate heterogeneous IDs such as SKU codes of business systems and sensor IDs of Internet of Things systems to the same physical asset; the application constructs a dynamic time series graph, and uses an innovative self-supervised training method based on heuristic pseudo-labels, so that a time series graph neural network T-GNN learns a cross-modal context semantic representation, thereby realizing accurate alignment of heterogeneous IDs and subsequent hidden event inference.

[0058] In the embodiment, the system specifically includes six core units that work together to achieve the above-mentioned purposes.

[0059] The graph construction unit aims to capture data streams from multiple logistics subsystems with different structures in real time, and represent the original data streams as a dynamic time series graph structure without pre-identity alignment; the unit collects multi-source heterogeneous data streams in real time, and parses entities and contexts in the data streams into independent nodes and edges in the graph; the key point is that the system does not know whether the nodes from different sources, such as SKU-123 of WMS and Sensor-789 of IoT, point to the same physical entity at this time, and they are created as independent nodes; the unit generates a dynamic time series graph snapshot at each timestamp Aggregate all nodes and edges to generate a dynamic time series graph snapshot ;

[0060] The heuristic quantification unit aims to not rely on neural networks, but to analyze the dynamic time series graph snapshot generated by the graph construction unit The original context information is used to calculate a quantized score for the probability that any two independent nodes, such as SKU-123 and Sensor-789, point to the same physical entity; this unit analysis The original context information is used to calculate independent nodes. and Heuristic context similarity score between This score The pseudo-labels used as training units in subsequent models serve to guide the neural network in learning what contextual similarity is.

[0061] The Temporal Graph Neural Network (T-GNN) unit is the core semantic representation module of this system; its purpose is to process dynamic temporal graph snapshots generated by the graph construction unit. The discrete ID, time-varying features (such as sensor readings), and complex spatiotemporal topological relationships of each node in the graph are encoded and embedded into a low-dimensional, dense, high-dimensional state vector that can capture deep semantics. This unit employs a time-series graph neural network model. For example, GCRN or TGAT model structures, for processing dynamic time series snapshots. ;for Each independent node in This unit generates a high-dimensional state vector for it. This generation process can be represented by the following standard state update paradigm:

[0062]

[0063] in, : Refers to a node The state vector of the previous time step is derived from the state vector of this unit at the previous time step. The calculation is used to realize the time dependency;

[0064] : Refers to a node In the current snapshot The set of neighboring nodes in the graph originates from the graph topology resolved by the graph building unit. Provided to provide graph context;

[0065] : Refers to a node exist The original characteristics of the moment, such as real-time readings from IoT sensors, are collected and provided by the graph construction unit;

[0066] : Refers to the state update function of T-GNN;

[0067] This paradigm leverages the powerful nonlinear fitting capabilities of neural networks to capture nodes. In graph structure and time series The complex spatiotemporal dependencies are uniformly encoded into semantic vectors. ;

[0068] The model training unit aims to utilize the pseudo-labels generated by the heuristic quantization unit. To supervise and optimize the temporal graph neural network units The parameters; this unit is based on heuristic context similarity scores. Training samples are dynamically constructed; a pre-defined cross-modal contextual contrastive loss function is used. For time-series graph neural network units Training is performed; the ultimate goal of training is to force the T-GNN to... Learn a representation method that generates high-dimensional state vectors for nodes. Similarity in vector space, such as cosine similarity It can reproduce the heuristic context similarity calculated by the heuristic quantization unit. ;

[0069] The unified asset aggregation unit is the identity alignment execution module of this invention; its purpose is to solve the initial heterogeneous ID alignment problem by utilizing a fully trained temporal graph neural network unit; this unit employs a trained temporal graph neural network unit. Solve for the latest high-dimensional state vector of all nodes in the system, such as SKU-123 and Sensor-789. and This unit is based on the similarity between these high-dimensional state vectors. Align independent nodes from different sources; if two vectors are sufficiently close in semantic space, their similarity is considered. If they exceed a certain threshold, the system determines that they point to the same physical entity, and thus aggregates them into a unified asset view;

[0070] The Hidden Event Inference Unit is a higher-level application module of this system, used to achieve intelligent monitoring beyond simple threshold alarms. Its purpose is to establish a normal behavior model of assets using the aligned unified asset view and its high-dimensional state vector generated by the unified asset aggregation unit. By monitoring behavioral deviations and risk proximity, it infers hidden events, such as cold chain disruptions, that are difficult to detect using traditional rules. This unit utilizes historical normal operating condition data to generate the unified asset view, i.e., the aligned asset vector. Establish a behavioral baseline This unit combines the current status of the asset. Relative to behavioral baseline The deviation from the target value and the asset's proximity to pre-defined risk factors, such as ambient temperature unloading areas, are used to calculate a real-time anomaly inference score. When the abnormal inference score Exceeding a preset alarm threshold At that time, the alarm threshold These are preset hyperparameters, which are pre-defined based on the business's risk tolerance or obtained through analysis of historical alarm verification sets. The ROC curve is used for calibration. For example, if a cold chain pallet not only deviates from the normal cold storage trajectory but also approaches the room temperature zone abnormally, the system will generate a hidden event alarm. If the score does not exceed the alarm threshold, it is judged as a normal state.

[0071] The solution in this embodiment does not rely on any manually labeled ID-aligned datasets. Instead, it uses a self-supervised approach to allow the system to automatically learn equivalent definitions from the context of the data itself, including time, space, and topology. This enables end-to-end multi-system data fusion and advanced event monitoring.

[0072] Example 2:

[0073] The system also employs a preset vector similarity function to calculate the similarity between high-dimensional state vectors.

[0074] This embodiment further defines Embodiment 1; in the scheme of Embodiment 1, both the unified asset aggregation unit and the hidden event inference unit need to calculate the similarity between high-dimensional state vectors when performing alignment and calculating scores; in this embodiment, the system also uses a preset vector similarity function. This function is used to solve for any two high-dimensional state vectors, such as and Similarity between them;

[0075] The vector similarity function It is a mathematical metric used to quantify the relationship between two vectors. In this invention, the degree of proximity in dimensional space serves as a standard for calculating semantic similarity; in this embodiment, this function... This can be specifically implemented as cosine similarity, which is calculated as follows:

[0076]

[0077] in It is the dot product of vectors. It is a vector The Euclidean norm;

[0078] By defining and uniformly using a preset vector similarity function, such as cosine similarity, the application ensures that the measurement standard of similarity is completely consistent in the three key steps of model training, asset aggregation and event inference; such consistency is the guarantee of training convergence and inference accuracy, avoiding the decline of model performance or logical confusion caused by inconsistent measurement standards.

[0079] Embodiment 3:

[0080] The analysis process of the heuristic quantization unit to calculate the heuristic context similarity score is as follows:

[0081] The time similarity, spatial similarity and graph topology similarity between nodes are calculated respectively;

[0082] Among them, the time similarity is used to represent the closeness of the event timestamps of the nodes;

[0083] The spatial similarity is used to represent the closeness of the nodes in geographical space;

[0084] The graph topology similarity is used to represent whether the nodes share similar neighbors;

[0085] According to the preset domain knowledge weight, the time similarity, spatial similarity and graph topology similarity are weighted and summed to obtain the heuristic context similarity score.

[0086] This embodiment is a specific implementation of the heuristic quantization unit in Embodiment 1; the core task of this unit is to calculate the heuristic context similarity score ;

[0087] In this embodiment, the heuristic quantization unit calculates the analysis process of The unit extracts the original context information of nodes and from the dynamic time sequence diagram snapshot , and calculates three independent similarities: time similarity , spatial similarity and graph topology similarity ;

[0088] The time similarity is used to represent the closeness of the event timestamps of nodes and on the time axis; for example, it can be defined as , where is the difference between the event timestamps of the two nodes, and is the time decay constant;

[0089] The spatial similarity is used to represent the closeness of the nodes and spatial proximity; e.g., if node has a recorded location from WMS as Shanghai Warehouse A Gate, and node has GPS coordinates from IoT also near Gate A, then proximity 1; otherwise, proximity 0;

[0090] graph topology similarity , which aims to characterize whether two nodes and share similar neighbors in the graph structure; e.g., if SKU-123 and Sensor-789 are both connected to the same node Vehicle-GTE-556, their topology similarity will be high; this similarity can be computed using graph theory metrics such as Jaccard similarity or Adamic-Adar index;

[0091] To ensure the robustness of the heuristic quantification unit, the system handles missing context information: when any pair of nodes is missing context information in a certain dimension, e.g., node lacks spatial coordinates leading to being unable to be computed, or node lacks neighbors leading to being unable to be computed, the similarity score in that dimension is assigned a pre-defined neutral value, e.g., 0 or 0.5; this ensures that can always be computed, avoiding the failure of the subsequent training process due to partial data missing;

[0092] The unit weights the above three similarities by pre-defined domain knowledge weights , and obtains the final heuristic context similarity score ;

[0093]

[0094] where ;

[0095] These weights are pre-defined hyperparameters according to specific business scenarios, e.g., in the cold-chain logistics scenario, the geographical location of assets and whether they are on the same refrigerated truck are highly related in topology, thus and may be set to have higher weights than ; these weights can be optimized and determined through cross-validation on the annotated validation set;

[0096] The embodiment disassembles the ambiguous context similarity into three quantifiable dimensions of time, space and topology, and introduces the field knowledge weight, so that the generated pseudo label can more accurately reflect the real relevance of the physical world, providing higher quality supervision signals for the subsequent training of T-GNN.

[0097] Embodiment 4:

[0098] The process of dynamically constructing training samples by the model training unit is as follows:

[0099] For the anchor node, a node with a heuristic context similarity score greater than a preset high heuristic similarity threshold is selected as a positive sample;

[0100] A node with a heuristic context similarity score less than a preset low heuristic similarity threshold is selected as a negative sample set.

[0101] The embodiment is a specific implementation of the model training unit in embodiment 1, which details the process of dynamically constructing training samples;

[0102] In the training process, the model training unit selects a node as an anchor node from a training batch ; based on the heuristic context similarity score calculated in embodiment 3 , the unit dynamically matches positive and negative samples for the anchor node

[0103] For the anchor node , the unit selects a node in , which satisfies ; i.e. the preset high heuristic similarity threshold is a value close to 1, for example 0.9; the selected is regarded as a positive sample;

[0104] The unit selects a set of nodes in , each node in the set satisfies ; i.e. the preset low heuristic similarity threshold is a value close to 0, for example 0.1; the set is regarded as a negative sample set;

[0105] The high heuristic similarity threshold and the low heuristic similarity threshold ​are preset hyperparameters, whose values are determined by cross-validation tuning on the validation dataset; their role is to establish a buffer zone to discard difficult samples that are also difficult for the heuristic rules to distinguish, and only keep high-confidence positive samples with highly similar contexts and high-confidence negative samples with completely irrelevant contexts for training;

[0106] By setting and two different thresholds to dynamically construct positive and negative samples, the present application significantly improves the quality of training samples, prevents the T-GNN model from learning the noise brought by the heuristic rules, and enhances the convergence and robustness of model training.

[0107] Embodiment 5:

[0108] The cross-modal context contrast loss function is used to:

[0109] Respectively obtain the high-dimensional state vectors corresponding to the anchor node, the positive sample set and the negative sample set;

[0110] Adopt a preset vector similarity function to calculate the vector similarity between the anchor node and the positive sample, and the vector similarity between the anchor node and each node in the negative sample set;

[0111] By minimizing the loss function, the vector similarity between the anchor node and the positive sample is increased, and the vector similarity between the anchor node and the negative sample is reduced.

[0112] This embodiment details the specific mechanism and role of the cross-modal context contrast loss function ;

[0113] The loss function is an improved contrast learning loss such as InfoNCE; its calculation process is:

[0114] Obtain the high-dimensional state vectors corresponding to the anchor node , the positive sample from Embodiment 4 and the negative sample set from Embodiment 4; these vectors are calculated by the time series graph neural network unit , that is, ;

[0115] Adopt a preset vector similarity function such as the cosine similarity as described in Embodiment 2 to respectively calculate the vector similarity between the anchor node and the positive sample , and the vector similarity between the anchor node and each node in the negative sample set ;

[0116] The loss function The objective is to increase the anchor node vector similarity with positive samples while decreasing the anchor node vector similarity with all negative samples The loss function can be specifically expressed as:

[0117]

[0118] wherein, are output by the temporal graph neural network unit ;

[0119] is the vector similarity function defined in embodiment 2;

[0120] denotes the temperature hyperparameter; it is a positive scalar, which is set by tuning experiments on the validation set; its role is to control the sharpness of the loss function, i.e., the model’s ability to distinguish difficult negative samples;

[0121] By minimizing , the output of the T-GNN network is forced to fit the heuristic rule ; after training, the T-GNN unit has a semantic decoding capability: even if the original IDs and data modalities of two nodes, e.g., SKU-123 and Sensor-789, are completely different, as long as they share similar context in the physical world, i.e. the value is high, the T-GNN unit and can generate highly close vectors in the vector space for them, thereby laying the foundation for identity alignment.

[0122] Embodiment 6:

[0123] The process of aligning independent nodes by the unified asset aggregation unit is as follows:

[0124] Using the trained temporal graph neural network unit, calculate the latest high-dimensional state vectors of all nodes;

[0125] Using the preset vector similarity function, perform vector neighbor search;

[0126] Identify the node pairs with a vector similarity greater than the preset identity alignment threshold as equivalent relationships pointing to the same physical entity;

[0127] ​​The node pair with a vector similarity less than the preset identity alignment threshold is identified as a non-equivalent relationship.

[0128] This embodiment is a further application of embodiment 2 and a specific implementation of the process of aligning independent nodes by the unified asset aggregation unit in embodiment 1.

[0129] The alignment process is performed after the T-GNN model training is completed, and the specific steps are as follows:

[0130] The trained time series graph neural network unit is used to calculate the latest high-dimensional state vector of all nodes .

[0131] A preset vector similarity function such as the cosine similarity described in embodiment 2 is used to perform efficient vector neighbor search in an N-dimensional vector space, for example, using index technology such as FAISS.

[0132] For any node pair found , if their vector similarity is greater than a preset identity alignment threshold , the system identifies the node pair as an equivalent relationship. is a high-confidence threshold, for example, 0.95, which is obtained by testing the alignment accuracy and recall rate of the model on the validation set;

[0133] If the vector similarity is not greater than , the system identifies them as a non-equivalent relationship.

[0134] This embodiment realizes automatic and high-confidence alignment of heterogeneous IDs by setting a strict identity alignment threshold in a high-dimensional vector space; it successfully associates heterogeneous IDs such as SKU-123 and Sensor-789 to the same physical entity, completely breaking down the data silos between WMS, TMS, and IoT systems, and forming a unified asset view.

[0135] Embodiment 7:

[0136] The process of establishing a behavior baseline by the hidden event inference unit is as follows:

[0137] Obtain the high-dimensional state vector corresponding to the unified asset view under historical normal working conditions.

[0138] Perform time series smoothing or statistical modeling on the high-dimensional state vector under the historical normal working conditions to obtain the behavior baseline. This embodiment is a specific implementation of the process of establishing a behavior baseline by the hidden event inference unit in embodiment 1.

[0139] In order to determine whether the current behavior of the asset is abnormal, it is necessary to define what is normal; in this embodiment, a behavior baseline is established The process is as follows:

[0140] Obtain the unified asset view of the transportation period under the historical normal working condition, for example, confirmed according to the historical work order, without any alarm and accident, that is, the corresponding high-dimensional state vector thereof ;

[0141] Time series smoothing or statistical modeling is performed on the high-dimensional state vector sequence under the historical normal working condition ; for example, a sliding average or exponential smoothing method can be used to generate a smoothed baseline trajectory , which represents the average normal state of the asset at ; or a statistical model such as Gaussian Mixture Model (GMM) can be used to model the distribution of the historical , at which time the behavior baseline may be represented as a distribution center or a confidence interval;

[0142] By smoothing or modeling the historical normal data, the present application establishes a dynamic baseline that can represent the regular mode of the asset; compared with the traditional method using a fixed threshold, the baseline established based on the high-dimensional state vector takes into account the multi-dimensional attributes of the asset, and the definition of normal is more accurate and robust.

[0143] Embodiment 8:

[0144] The calculation process of the anomaly inference score is as follows:

[0145] Obtain the current high-dimensional state vector of the asset;

[0146] Calculate the trajectory deviation metric between the current high-dimensional state vector and the behavior baseline;

[0147] Using a preset vector similarity function, calculate the maximum vector similarity between the current high-dimensional state vector and each node in the pre-defined risk factor node set;

[0148] According to the preset deviation weight and proximity weight, the trajectory deviation metric and the maximum vector similarity are weighted and summed to obtain the anomaly inference score;

[0149] The trajectory deviation metric is defined as 1 minus the similarity value between the current high-dimensional state vector and the behavior baseline calculated using the preset vector similarity function.

[0150] This embodiment is a further deepening of embodiment 7, which details the calculation process of the anomaly inference score ;

[0151] exist At any given time, the hidden event inference unit calculates the real-time anomaly inference score according to the following steps. :

[0152] Obtain the current high-dimensional state vector of the asset. This vector is calculated in real time by the T-GNN unit;

[0153] Solve the current high-dimensional state vector With behavioral baseline Trajectory deviation measurement from Example 7 ;

[0154] Trajectory Deviation Measurement Defined as Subtract using a preset vector similarity function The similarity value calculated from Example 2 is defined by the following formula: ;

[0155] This metric, for example, when When the similarity is cosine That is, the cosine distance is dimensionless; if the current state Compared with baseline Highly similar, then Approaching 0 means no deviation; conversely, if the deviation is greater than 0, then... Close to 1;

[0156] Using a preset vector similarity function Solve for the current high-dimensional state vector With the predefined set of risk factor nodes Each node in the middle The maximum vector similarity;

[0157] Risk factor node set Its source is a predefined set of node vectors representing known risk sources based on domain knowledge; for example, in a cold chain scenario, It includes feature vectors of all nodes such as the ambient temperature unloading area and the sun-exposed parking lot. ;

[0158] The formula for risk proximity is defined as follows: This value is used to quantify assets. How semantically close it is to the most dangerous source of risk;

[0159] Based on the preset deviation weight and close to weight The two metrics are then weighted and summed to obtain the final anomaly inference score. ;

[0160] Its formula is defined as:

[0161] and It is a dimensionless hyperparameter, which is derived from preset based on the business's sensitivity to risk or determined through experimental optimization.

[0162] To ensure dimensional consistency, in this formula As a dimensionless abnormal fraction, its constituent terms and Both are dimensionless similarity or distance metrics, with weights. and It is also a dimensionless scalar;

[0163] Defined here It is a linear weighted sum of the trajectory deviation metric and the maximum vector similarity, which is a simplified implementation; in other embodiments, It can also be defined as a nonlinear combinatorial function, for example... ,in Modeling can be done using, for example, exponential functions or small neural networks, to capture more complex dose-response relationships, such as physical scenarios where risk increases exponentially with proximity, thereby improving physical fidelity.

[0164] This embodiment proposes a dual-criteria anomaly scoring mechanism based on both deviation from the baseline and proximity risk; only when an asset, such as a cold chain pallet, deviates from its normal cold storage trajectory will an anomaly be identified. The vector increases, and its current state begins to approach that of the ambient temperature unloading zone. When increased, the abnormal inference score This will significantly increase and trigger an alarm; this mechanism can accurately infer complex hidden events such as cold chain disruptions that cannot be determined by data from a single sensor.

[0165] Example 9:

[0166] The graph construction unit collects multi-source heterogeneous data streams from warehouse management systems, transportation management systems, and IoT sensors;

[0167] In this process, the business ID from the warehouse management system or transportation management system, and the sensor ID from the IoT sensor, are created as independent nodes in the graph during the initial parsing.

[0168] This embodiment further defines the data source and node creation method of the map construction unit in Embodiment 1;

[0169] In this embodiment, the map construction unit explicitly collects multi-source heterogeneous data streams from the following data sources:

[0170] Warehouse Management System (WMS): provides business data such as SKU (Stock Keeping Unit), bin location, inbound order number, outbound order number, etc.

[0171] Transportation Management System (TMS): provides business data such as waybill number, vehicle ID, driver information, route planning, estimated time of arrival (ETA), etc.

[0172] Internet of Things (IoT) sensors: provide real-time physical data such as GPS coordinates, temperature readings, humidity readings, vibration sensor ID, device status, etc.

[0173] At the initial resolution, a key technical process is that the business ID from WMS or TMS, such as SKU-123 of WMS, and the sensor ID from IoT, such as Sensor-789, are created as independent nodes in the graph at the initial resolution.

[0174] The embodiment explicitly defines the data heterogeneous business system ID vs. IoT physical ID processed by the present application; by deliberately retaining this heterogeneity at the beginning of graph construction, treating them as independent nodes, the present application converts the ID alignment problem into a subsequent graph node semantic similarity calculation problem; this technical route of first accepting heterogeneity and then achieving unity through T-GNN and contrastive learning is the core prerequisite for the present application to automatically connect multiple system data.

[0175] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A multi-system data unified logistics asset Internet of Things monitoring processing system, characterized in that, Comprise: A graph construction unit for real-time collection of multi-source heterogeneous data streams, parsing entities and contexts in the data streams into independent nodes and edges, and generating dynamic time sequence graph snapshots by aggregation; A heuristic quantification unit for analysis of original context information in the dynamic time sequence graph snapshots, calculation of heuristic context similarity scores between independent nodes, and use of the scores as pseudo-labels; A time sequence graph neural network unit for processing of the dynamic time sequence graph snapshots and generation of high-dimensional state vectors for the independent nodes; A model training unit for dynamic construction of training samples based on the heuristic context similarity scores, and training of the time sequence graph neural network unit using a preset cross-modal context comparison loss function, so that similarity of the high-dimensional state vectors in the vector space reproduces the heuristic context similarity; A unified asset aggregation unit for use of the trained time sequence graph neural network unit to solve the high-dimensional state vectors, alignment of independent nodes of different sources based on similarity between the high-dimensional state vectors, and aggregation into a unified asset view; A hidden event inference unit for establishment of a behavior baseline of the unified asset view, calculation of real-time anomaly inference scores in combination with deviation of the behavior baseline and proximity of preset risk factors, and generation of a hidden event alarm when the anomaly inference scores exceed an alarm threshold, and determination of a normal state when the anomaly inference scores do not exceed the alarm threshold.

2. The multi-system data unified logistics asset Internet of Things monitoring processing system according to claim 1, characterized in that, The system also uses a preset vector similarity function for solving similarity between the high-dimensional state vectors.

3. The multi-system data unified logistics asset Internet of Things monitoring processing system of claim 1, wherein, The analysis process of the heuristic quantification unit for calculation of the heuristic context similarity scores is as follows: Time similarity, spatial similarity and graph topology similarity between nodes are calculated respectively; The time similarity is used to represent proximity of node event timestamps; The spatial similarity is used to represent proximity of nodes in geographical space; The graph topology similarity is used to represent whether the nodes share similar neighbors; The time similarity, spatial similarity and graph topology similarity are weighted and summed according to a preset domain knowledge weight to obtain the heuristic context similarity scores.

4. The multi-system data unified logistics asset Internet of Things monitoring processing system according to claim 3, characterized in that, The process of dynamic construction of training samples by the model training unit is as follows: For anchor nodes, nodes with heuristic context similarity scores greater than a preset high heuristic similarity threshold are selected as positive samples; Nodes with heuristic context similarity scores less than a preset low heuristic similarity threshold are selected as a negative sample set.

5. The multi-system data unified logistics asset Internet of Things monitoring processing system according to claim 4, characterized in that, The cross-modal context comparison loss function is used to: High-dimensional state vectors corresponding to the anchor nodes, positive samples and the negative sample set are obtained respectively; A preset vector similarity function is used to calculate vector similarity between the anchor nodes and the positive samples, and vector similarity between the anchor nodes and each node in the negative sample set; By minimizing the loss function, the vector similarity between the anchor nodes and the positive samples is increased, and the vector similarity between the anchor nodes and the negative samples is decreased.

6. The multi-system data unified logistics asset Internet of Things monitoring processing system of claim 2, wherein, The process of alignment of independent nodes by the unified asset aggregation unit is as follows: The trained time sequence graph neural network unit is used to calculate the latest high-dimensional state vectors of all nodes; A preset vector similarity function is used for vector neighbor search; Node pairs with vector similarity greater than a preset identity alignment threshold are identified as equivalent relationships pointing to the same physical entity; The node pair with a vector similarity less than the preset identity alignment threshold is identified as a non-equivalent relationship.

7. The multi-system data unified logistics asset Internet of Things monitoring processing system of claim 1, wherein, The process of the hidden event inference unit establishing the behavior baseline is as follows: Obtain the high-dimensional state vector corresponding to the unified asset view under the historical normal working condition; Perform time series smoothing or statistical modeling on the high-dimensional state vector under the historical normal working condition to obtain the behavior baseline.

8. The multi-system data unified logistics asset Internet of Things monitoring processing system according to claim 7, characterized in that, The calculation process of the anomaly inference score is as follows: Obtain the current high-dimensional state vector of the asset; Solve the trajectory deviation metric between the current high-dimensional state vector and the behavior baseline; Using a preset vector similarity function, solve the maximum vector similarity between the current high-dimensional state vector and each node in the predefined risk factor node set; According to the preset deviation weight and proximity weight, the trajectory deviation metric and the maximum vector similarity are weighted and summed to obtain the anomaly inference score.

9. The multi-system data unified logistics asset Internet of Things monitoring processing system according to claim 8, characterized in that, The trajectory deviation metric is defined as 1 minus the similarity value between the current high-dimensional state vector and the behavior baseline calculated using the preset vector similarity function.

10. The multi-system data unified logistics asset Internet of Things monitoring processing system of claim 1, wherein, The graph construction unit collects multi-source heterogeneous data streams from warehouse management systems, transportation management systems, and Internet of Things sensors; Among them, the business ID from the warehouse management system or the transportation management system, and the sensor ID from the Internet of Things sensor, are created as independent nodes in the graph during initial analysis.

Images