Freight data management method and system based on artificial intelligence

Abstract

The invention discloses a freight data management method and system based on artificial intelligence, and relates to the technical field of logistics data management analysis, and the method comprises the steps: collecting multi-source data, carrying out the time correction, calculating the quality score of the multi-source data, constructing the multi-source data into a fixed variable, and splicing the fixed variable into a vector; calculating a fusion reference value of each fixed variable to obtain a data source consistency residual error and a consistency score, performing fusion to obtain an instant score, updating fixed variable credibility, calculating fixed variable fusion based on the updated credibility, and writing the fixed variable fusion into a semantic knowledge graph; calculating a statistical magnitude for each fixed variable and mapping the statistical magnitude into nodes to construct a variable dependency graph; and constructing a window based on a time bucket, processing a window value by using a two-layer full-connection encoder, training the two-layer full-connection encoder based on an SAE loss function, and outputting a hidden variable based on the trained encoder. According to the invention, probabilistic positioning from abnormal phenomena to decision-making root causes is realized.

Description

Technical Field

[0001] This invention relates to the field of logistics data management and analysis technology, and in particular to a freight data management method and system based on artificial intelligence. Background Technology

[0002] In recent years, with the rapid popularization of vehicle-to-everything (V2X), mobile internet, and intelligent logistics platforms, freight companies have formed a large-scale, multi-source, heterogeneous data system in vehicle operation, cargo status, order flow, and terminal operations. This system includes continuous sensor data such as vehicle terminals, OBD / CAN, GPS, and tire pressure / temperature and humidity data, as well as unstructured data such as electronic waybills, loading and unloading barcode scanning, handover and signing records, anomaly handling records, and video / image and text logs. At the same time, the development of cloud-edge collaborative computing and machine learning technologies has enabled the real-time perception and management of in-transit risks to evolve from "post-event statistics" to a closed-loop paradigm of "online monitoring—anomaly alarm—cause tracing and handling—responsibility tracing." To support this paradigm, the industry has introduced time alignment, missing data repair, data cleaning, and quality assessment at the data level, and explored unsupervised anomaly detection, graph neural networks, and knowledge graphs at the model level to improve adaptability and interpretability for complex transportation scenarios (multiple routes, multiple vehicle types, multiple carriers, and multiple environmental disturbances). In particular, for the data governance needs of cross-system collaboration, semantic data organization and relational reasoning capabilities have gradually become important directions for freight data management. However, existing related technologies still have shortcomings. Existing knowledge graph solutions often focus on static entity relationship modeling and lack a high-frequency writing mechanism for in-transit status events and evidence-level confidence expression, making it difficult to connect "numerical symptoms - business semantics - evidence sources" when anomalies occur, resulting in insufficient interpretation and auditing capabilities. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides an artificial intelligence-based freight data management method and system, which solves the problems of existing technologies lacking a high-frequency writing mechanism for in-transit status events and evidence-level confidence expression, as well as insufficient interpretation and auditing capabilities.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a freight data management method based on artificial intelligence, comprising,

[0007] Collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables and concatenate them into a vector;

[0008] Calculate the fusion reference value of each fixed variable to obtain the data source consistency residual and consistency score, and calculate the consistency mean to obtain the real-time score. Update the credibility of fixed variables based on the real-time score, calculate the fixed variable fusion based on the updated fixed variable credibility and write it into the semantic knowledge graph.

[0009] Calculate the statistic for each fixed variable and map it to nodes to construct a variable dependency graph;

[0010] A window is constructed based on a time bucket, and the window value is processed using a two-layer fully connected encoder. The two-layer fully connected encoder is trained based on the SAE loss function, and the latent variables are output based on the trained encoder.

[0011] Node features are constructed based on latent variables. Graph Transformer is used to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes. Anomaly evidence subgraph is constructed based on the anomaly score and the final symptom vector is calculated.

[0012] Define a candidate root cause set, calculate the prior probability and evidence likelihood based on the candidate root cause set, and output the posterior probability to select the most likely root cause.

[0013] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of collecting multi-source data, performing time correction, calculating the quality score of multi-source data, and constructing multi-source data into fixed variables and concatenating them into a vector specifically involves: deploying a freight data gateway in the cloud to collect multi-source data.

[0014] Calculate the clock offset for each data source s ;

[0015] Calculate the correction time for each data entry based on the clock offset. And put it into the 10-second time bucket;

[0016] Calculate the time decay factor for each data point. ;

[0017] The data source s is quality-scored within time bucket b based on the time decay factor. ;

[0018] Constructing multi-source data within the time bucket into fixed variables A vector is formed by concatenating fixed variables from multiple data sources. ;

[0019] For time buckets with missing data and consecutive missing bucket numbers less than or equal to G, use data from the previous time bucket and mark it. If the number of missing data is greater than G, mark the time bucket as empty and record the missing mask. ;

[0020] Quality rating and missing mask Simultaneous splicing to form a mass vector and missing mask vector .

[0021] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the following steps are included: calculating the fusion reference value of each fixed variable to obtain the data source consistency residual and consistency score, fusing them to obtain an instant score to update the credibility of the fixed variables, and calculating the fusion of fixed variables based on the updated credibility and writing it into the semantic knowledge graph: initializing the credibility. Calculate the fusion reference value for each fixed variable. ;

[0022] Calculate source consistency residuals based on fusion reference values. With consistency score ;

[0023] Calculate the consistency mean based on the consistency scores of the fixed variables covered by data source s within the time bucket. The consistency mean was then compared with the quality score of time bucket b. Weighted fusion into instant scoring ;

[0024] Reliability updates based on EWMA ;

[0025] Each fixed variable is weighted and merged based on its update reliability;

[0026] The weighted and fused fixed variables are written into the semantic knowledge graph KG as facts in the form of events.

[0027] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of calculating statistics for each fixed variable and mapping them to nodes to construct a variable dependency graph specifically involves dividing all fixed variables into 5 groups and calculating statistics for each fixed variable i.

[0028] The statistics are concatenated to form the input vector. and input vector Input a Transformer encoder to map and obtain nodes. ;

[0029] According to the node Calculate the cosine similarity and introduce a reliability coefficient;

[0030] Select the top-K nodes with the largest values ​​to build a directed adjacency graph A, which is a variable dependency graph.

[0031] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of constructing a window based on a time bucket and processing the window value using a two-layer fully connected encoder, training the two-layer fully connected encoder based on the SAE loss function, and outputting latent variables based on the trained encoder, specifically includes a time bucket-based vector. Construct a window of length w ;

[0032] From window Extract the window value of the fixed variable i. Using a two-layer fully connected encoder to convert window values Mapped to latent vectors ;

[0033] The two-layer fully connected encoder is trained using the SAE loss function, and the latent variables are output through the trained fully connected encoder.

[0034] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of constructing node features, using Graph Transformer to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes specifically involves: all the latent variables of the fixed vectors Stacking yields a sparse implicit representation matrix h, and the nodes are... The temporal embeddings of variables are stacked to obtain v. The sparse latent representation matrix h is linearly projected and concatenated with the temporal embeddings of variables to obtain the initial node features z. Edge features are defined based on the directed edges. ;

[0035] Graph Transformer is used to compute multi-head attention to update node features;

[0036] Update node features and After element-wise multiplication, the predicted vector is obtained by inputting it into a linear layer. Calculate the predicted vector and the actual vector The predicted loss between;

[0037] The total loss function is obtained by weighting the prediction loss and the SAE loss function. The graph Transformer parameters are trained based on the total loss function, and the prediction vector is output based on the trained graph Transformer.

[0038] Calculate the outlier score of the variable node based on the prediction vector. .

[0039] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of constructing an anomaly evidence subgraph based on anomaly scores and calculating the final symptom vector specifically involves: sorting the anomaly scores in descending order and selecting the top 5 variables. ,right Each node i selects a set of directed neighbors. Form a set of evidence edges and construct an abnormal evidence subgraph ;

[0040] Define the abnormal evidence subgraph Symptom intensity for each variable i ;

[0041] Define the propagation strength for each piece of evidence. ;

[0042] The symptom intensity and the transmission intensity are combined to form the final symptom vector F.

[0043] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, a candidate root cause set is defined, which is manually formulated and fixed as a root cause template library.

[0044] As a preferred embodiment of the artificial intelligence-based freight data management method of the present invention, the step of calculating the prior probability and evidence likelihood based on the candidate root cause set and outputting the posterior probability to select the most likely root cause specifically involves: for each root cause r in the candidate root cause combination, obtaining the prior of the root cause through historical statistics. ;

[0045] Simultaneously, the likelihood of evidence is calculated based on root causes. ;

[0046] Based on prior Likelihood of Evidence Output posterior probability ;

[0047] Based on posterior probability The root cause with the highest probability is selected as the most likely root cause output.

[0048] Secondly, the present invention provides an artificial intelligence-based freight data management system, comprising,

[0049] The data acquisition module is used to collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables, and concatenate them into a vector.

[0050] The credibility analysis module is used to calculate the fusion reference value of each fixed variable to obtain the data source consistency residual and consistency score, and calculate the consistency mean to obtain the real-time score. Based on the real-time score, the credibility of the fixed variables is updated, and the fixed variable fusion is calculated based on the updated credibility of the fixed variables and written into the semantic knowledge graph.

[0051] The dependency graph construction module is used to calculate statistics for each fixed variable and map them to nodes to construct a variable dependency graph.

[0052] The variable mapping module is used to construct windows based on time buckets and process window values ​​using a two-layer fully connected encoder. The two-layer fully connected encoder is trained based on the SAE loss function, and the latent variables are output based on the trained encoder.

[0053] The node anomaly analysis module is used to construct node features based on latent variables, use graph Transformer to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes, construct anomaly evidence subgraph based on the anomaly score and calculate the final symptom vector.

[0054] The root cause analysis module is used to define a set of candidate root causes, calculate prior probabilities and evidence likelihoods based on the set of candidate root causes, and output posterior probabilities to select the most likely root cause.

[0055] The beneficial effects of this invention are as follows: This invention uses multi-source data time correction and quality scoring, maps heterogeneous information to fixed variable vectors, further calculates fusion reference values ​​to obtain consistency residuals / scores, recursively updates the data source credibility and completes adaptive fusion to write into the semantic knowledge graph accordingly, constructs a variable dependency graph based on fixed variable statistics, and explicitly characterizes variable coupling and propagation paths; combines SAE to perform sparse representation and noise reduction compression of window sequences, and then uses graph Transformer multi-head attention for structured collaborative detection, outputs anomaly scores and constructs evidence subgraphs and symptom vectors, enhancing the ability to detect and interpret composite anomalies, and finally calculates prior and evidence likelihoods using the candidate root cause set as boundaries to obtain posterior probabilities, realizing probabilistic localization from abnormal phenomena to decisionable root causes. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of the artificial intelligence-based freight data management method in Example 1.

[0058] Figure 2 This is a structural diagram of the artificial intelligence-based freight data management system in Example 1. Detailed Implementation

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0061] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0062] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a freight data management method based on artificial intelligence, including the following steps:

[0063] S1. Collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables and concatenate them into a vector;

[0064] S1.1 Deploy a freight data gateway in the cloud to collect multi-source data. The freight gateway connects to vehicle terminals, traffic systems and warehouse WMS to collect data. Vehicle terminals collect vehicle location, speed, fuel consumption, tire pressure, number of emergency brakings, number of alarms, etc. The traffic system collects geographical location, transportation route, transportation speed, etc. The warehouse WMS collects cargo data, transportation images, etc.

[0065] Calculate the clock offset for each data source s :

[0066]

[0067] in For server reception time, The time when the source data was generated. The smoothing coefficient is preferably 0.9, because the freight network is subject to sudden fluctuations, and 0.9 is preferred to ensure the stability of the offset. t is time.

[0068] Calculate the correction time for each data entry based on the clock offset. And put it into the 10-second time bucket:

[0069]

[0070] in The time bucket width is 10 seconds. Number the time buckets. The start time of the bucket;

[0071] Calculate the time decay factor for each data point. :

[0072]

[0073] in For the delay in seconds, The attenuation constant is preferably 30s. Real-time freight alarms are required to be effective within a minute time scale, and taking 30s can significantly suppress the impact of delayed data.

[0074] The data source s is quality-scored within time bucket b based on the time decay factor. :

[0075]

[0076] in This serves as a data validity indicator; the value is 1 if the data passes format validation when uploaded to the server, and 0 otherwise. and These are the weighting coefficients. To cut off to ;

[0077] Constructing multi-source data within the time bucket into fixed variables A vector is formed by concatenating fixed variables from multiple data sources. ;

[0078] Specifically, for time-series data, such as continuous variables like vehicle speed, fuel consumption, tire pressure, and location, the average value is taken within the time bucket; for count variables like the number of emergency brakings and alarms, the count value is taken; for visual variables such as transportation images, a lightweight object detection model (YOLO, etc.) is used to detect images and extract cargo visual boxes. The center point of the cargo visual box is extracted, and the distance between the center point and the center point of the visual box in the previous time bucket is calculated as the drift amount. The tilt angle is calculated for the coordinates of the edge endpoints of the cargo visual box, and the probability of cargo integrity is detected by softmax.

[0079] S1.2 For time buckets with missing data and consecutive missing bucket numbers less than or equal to G, use the data from the previous time bucket and mark it. If the number of missing data is greater than G, mark the time bucket as empty and record the missing mask. ;

[0080] Quality rating and missing mask Simultaneous splicing to form a mass vector and missing mask vector .

[0081] S2. Calculate the fusion reference value of each fixed variable to obtain the data source consistency residual and consistency score, fuse them to obtain an instant score to update the credibility of the fixed variables, calculate the fusion of fixed variables based on the updated credibility and write it into the semantic knowledge graph.

[0082] S2.1 Initialize credibility Calculate the fusion reference value for each fixed variable. :

[0083]

[0084] in To provide a data source set for the i-th fixed variable, such as vehicle speed, location, etc. Let i be the fixed variable of the s-th data source;

[0085] Calculate source consistency residuals based on fusion reference values. With consistency score :

[0086]

[0087] in Let be the standard deviation of the i-th fixed variable over the most recent 24 hours. To prevent small quantities from being sold, 10 is preferred. -6 Consistency score ;

[0088] Calculate the consistency mean based on the consistency scores of the fixed variables covered by data source s within the time bucket. The consistency mean was then compared with the quality score of time bucket b. Weighted fusion into instant scoring ;

[0089] Reliability updates based on EWMA :

[0090]

[0091] in The initial value is the quality score, representing the credibility of the previous time bucket. For smoothing coefficients;

[0092] Each fixed variable is weighted and merged based on its update confidence:

[0093]

[0094] in For variable fusion;

[0095] S2.2. Write the weighted and fused fixed variables as facts (such as vehicle speed, location, etc.) into the semantic knowledge graph KG in the form of events;

[0096] Specifically, the semantic knowledge graph (KG) includes entities and relations. Entities include vehicles, drivers, orders, cargo batches, road segments, and service providers. Relationships include transportation, loading, transit, location, handover, maintenance, and reasons for delay. Fixed variables are used as facts to represent events. For example, transporting goods from location A to location B is considered a transportation event and connects the corresponding related entities (such as service providers and orders), with the transportation relations serving as connection edges.

[0097] S3. Calculate the statistic for each fixed variable and map it to nodes to construct a variable dependency graph;

[0098] All fixed variables were divided into 5 groups, including power and fuel consumption, braking and chassis, tire vibration, cold chain in the vehicle compartment, and visual text traffic dispatch. Statistical measures were calculated for each fixed variable i.

[0099]

[0100]

[0101] in The mean, Standard deviation, The missing rate, For the number of time buckets, The autocorrelation coefficient is... Let be the missing mask vector for the i-th fixed variable in time bucket b. This is the concatenated vector of the i-th fixed variable in time bucket b;

[0102] The statistics are concatenated to form the input vector. and input vector Input a Transformer encoder to map and obtain nodes. ;

[0103] It should be noted that the embedding dimension of a single Transformer encoder is 16, the number of attention heads is 2, and the dimension of each head is 8.

[0104] According to the node Calculate cosine similarity and introduce a reliability coefficient for missing data suppression:

[0105]

[0106] in Let be the cosine similarity between nodes i and j. For reliability coefficient, For the suppressed cosine similarity, For node j;

[0107] For each node, select the top-K (the largest K nodes) to construct a directed adjacency graph A, which is a dependency graph.

[0108]

[0109] in For indicator functions, Let be the directed adjacency edge between nodes i and j. Let K be the candidate set of adjacent nodes of node i, and K be the number of nodes. Let be the suppressed cosine similarity between nodes i and j.

[0110] S4. Construct a window based on a time bucket and process the window value using a two-layer fully connected encoder. Train the two-layer fully connected encoder based on the SAE loss function and output the latent variables based on the trained encoder.

[0111] S4.1, Time-bucket-based vectors Construct a window of length w :

[0112]

[0113] From window Extract the window value of the fixed variable i. Using a two-layer fully connected encoder to convert window values Mapped to latent vectors ;

[0114] Specifically, the first fully connected encoder maps the dimension from w to d:

[0115]

[0116] in and To transform variables, and These are the parameters for the first layer fully connected encoder;

[0117] The second fully connected encoder maps the dimension from d to M:

[0118]

[0119] in For mapping variables, As an implicit vector, and These are the parameters for the second-layer fully connected encoder;

[0120] The window length w is preferably 16, and the dimension d of the first fully connected encoder is preferably 32. Too small a dimension will result in underfitting, and too large a dimension will increase latency. 32 is sufficient for an input length of 16. The dimension M of the second fully connected encoder is preferably 128, which can cover the "multivariate anomaly mode" and still allow for real-time inference.

[0121] S4.2. Use the SAE loss function to train the two-layer fully connected encoder and output the latent variables through the trained fully connected encoder.

[0122] It should be noted that the SAE loss function consists of reconstruction error and KL sparsity penalty:

[0123]

[0124]

[0125]

[0126] in For SAE loss function, For reconstruction error, Let i be the value of the i-th fixed variable in dimension u. The value output by the decoder after decoding the hidden variable. For KL sparsity penalty, For target sparsity, , Let be the activation value of the m-th hidden layer neuron in response to input x. This represents the average activation of neurons across the training batch. is the sparsity penalty coefficient, preferably 1, M is the dimension of the second fully connected encoder, and N is the number of fixed variables.

[0127] S5. Construct node features, use graph Transformer to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes, construct an anomaly evidence subgraph based on the anomaly score and calculate the final symptom vector.

[0128] S5.1, all hidden variables of fixed vectors Stacking yields a sparse implicit representation matrix h, and the nodes are... The temporal embeddings of variables are stacked to obtain v. The sparse latent representation matrix h is linearly projected and concatenated with the temporal embeddings of variables to obtain the initial node features z. Edge features are defined based on the directed edges. ;

[0129] Graph Transformer is used to compute multi-head attention to update node features;

[0130] Specifically, for the l-th layer and the c-th attention head, the attention of node i and its neighbor j is calculated. :

[0131]

[0132]

[0133]

[0134]

[0135] in Let be the query matrix of node i at the l-th layer and the c-th attention head. Let be the key matrix of node j in the l-th layer and the c-th attention head. and Let i and j be the node features at layer l. To update edge features, Let i be the set of adjacency matrices of node i. To scale the inner product, , , , , as well as These are trainable parameters;

[0136] Node features are updated based on multi-head attention:

[0137]

[0138]

[0139] in For value matrices, For the node features of the (l+1)th layer, and Here, C represents the number of attention heads;

[0140] Update node features and After element-wise multiplication, the predicted vector is obtained by inputting it into a linear layer. Calculate the predicted vector and the actual vector Predicted loss between:

[0141]

[0142] in For the prediction vector With the actual vector MSE loss between To predict losses;

[0143] The total loss function is obtained by weighting the prediction loss and the SAE loss function. The graph Transformer parameters are trained based on the total loss function, and the prediction vector is output based on the trained graph Transformer.

[0144] Calculate the outlier score of the variable node based on the prediction vector. :

[0145]

[0146] in Let i be the reconstructed value of the i-th variable at the end of the window. Let i be the value of the i-th variable at the end of the window. These are weighting coefficients used to balance prediction and reconstruction.

[0147] S5.2. Sort by outlier score in descending order and select the top 5 variables. ,right Each node i selects a set of directed neighbors. Form a set of evidence edges and construct an abnormal evidence subgraph ;

[0148] Define the abnormal evidence subgraph Symptom intensity for each variable i :

[0149]

[0150] in Missing mask;

[0151] Define the propagation strength for each piece of evidence. :

[0152]

[0153] The symptom intensity and the transmission intensity are combined to form the final symptom vector F.

[0154] S6. Define a candidate root cause set, calculate the prior probability and evidence likelihood based on the candidate root cause set, and output the posterior probability to select the most likely root cause.

[0155] S6.1 Define the candidate root cause set. The candidate root cause set is manually defined and fixed as a root cause template library. A total of 12 categories are set, including slow tire leakage / blowout, brake dragging, suspension abnormality, cold engine failure, loose cargo, damaged seals, dangerous driving by drivers, severe road damage, traffic control delays, abnormality caused by overloading, sensor drift failure, and non-standard loading and unloading operations. Each candidate root cause corresponds to a set of key variables.

[0156] S6.2 Calculate the prior probability and evidence likelihood based on the candidate root cause set, and output the posterior probability. Selecting the most likely root cause refers to obtaining the prior probability of each root cause r in the candidate root cause combination through historical statistics. :

[0157]

[0158] in The number of times root cause r was identified in historical data. For smoothing terms, the value is preferably 1 to ensure smoothness and stability. To identify root causes from historical data The number of times;

[0159] Simultaneously, the likelihood of evidence is calculated based on root causes. ;

[0160] Specifically, the weighted sum of symptom intensity is calculated based on the set of key variables corresponding to the root cause. :

[0161]

[0162] in The set of key variables corresponding to the root cause r. These are the variable weight coefficients;

[0163] Synchronously sum the evidence edge propagation strength of the nodes corresponding to the key variables of the root cause r. :

[0164]

[0165] in Evidence edges for nodes corresponding to key variables of the root cause r;

[0166] Weighted sum of symptom intensity Summing the propagation strength of the evidence The total score is obtained by weighted fusion. Total score Mapping to evidence likelihood :

[0167]

[0168] in The temperature coefficient is preferably 1 to ensure stability in high-risk freight scenarios;

[0169] Based on prior Likelihood of Evidence Output posterior probability :

[0170]

[0171] in The root cause number;

[0172] Based on posterior probability The root cause with the highest probability is selected as the most likely root cause output.

[0173] This embodiment also provides an artificial intelligence-based freight data management system, including:

[0174] The data acquisition module is used to collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables, and concatenate them into a vector.

[0175] The credibility analysis module is used to calculate the fusion reference value of each fixed variable, obtain the data source consistency residual and consistency score, fuse them to obtain an instant score to update the credibility of the fixed variables, calculate the fusion of fixed variables based on the updated credibility, and write them into the semantic knowledge graph.

[0176] The dependency graph construction module is used to calculate statistics for each fixed variable and map them to nodes to construct a variable dependency graph.

[0177] The variable mapping module is used to construct windows based on time buckets and process window values ​​using a two-layer fully connected encoder. The two-layer fully connected encoder is trained based on the SAE loss function, and the latent variables are output based on the trained encoder.

[0178] The node anomaly analysis module is used to construct node features. It uses graph Transformer to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes. Based on the anomaly score, it constructs an anomaly evidence subgraph and calculates the final symptom vector.

[0179] The root cause analysis module is used to define a set of candidate root causes, calculate prior probabilities and evidence likelihoods based on the set of candidate root causes, and output posterior probabilities to select the most likely root cause.

[0180] In summary, this invention uses multi-source data time correction and quality scoring, maps heterogeneous information to fixed variable vectors, further calculates fusion reference values ​​to obtain consistency residuals / scores, recursively updates data source credibility, and uses this to complete adaptive fusion and write into a semantic knowledge graph. Based on fixed variable statistics, it constructs a variable dependency graph to explicitly characterize variable coupling and propagation paths. It combines SAE to perform sparse representation and noise reduction compression of window sequences, and then uses graph Transformer multi-head attention for structured collaborative detection, outputting anomaly scores and constructing evidence subgraphs and symptom vectors to enhance the detection and interpretation capabilities of composite anomalies. Finally, it calculates prior and evidence likelihoods using the candidate root cause set as boundaries to obtain posterior probabilities, realizing probabilistic localization from abnormal phenomena to decisionable root causes.

[0181] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A freight data management method based on artificial intelligence, characterized in that: include, Collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables and concatenate them into a vector; Calculate the fusion reference value for each fixed variable to obtain the data source consistency residual and consistency score, and calculate the consistency mean to obtain the real-time score. Update the credibility of the fixed variables based on the real-time score, calculate the fixed variable fusion based on the updated fixed variable credibility, and write it into the semantic knowledge graph. Calculate the statistic for each fixed variable and map it to nodes to construct a variable dependency graph; A window is constructed based on a time bucket, and the window value is processed using a two-layer fully connected encoder. The two-layer fully connected encoder is trained based on the SAE loss function, and the latent variables are output based on the trained encoder. Node features are constructed based on latent variables. Graph Transformer is used to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes. Anomaly evidence subgraph is constructed based on the anomaly score and the final symptom vector is calculated. Define a candidate root cause set, calculate the prior probability and evidence likelihood based on the candidate root cause set, and output the posterior probability to select the most likely root cause.

2. The freight data management method based on artificial intelligence as described in claim 1, characterized in that: The process of collecting multi-source data, performing time correction, calculating the quality score of the multi-source data, and constructing the multi-source data into fixed variables and concatenating them into a vector specifically involves: deploying a freight data gateway in the cloud to collect multi-source data; Calculate the clock offset for each data source s ; Calculate the correction time for each data entry based on the clock offset. And put it into the 10-second time bucket; Calculate the time decay factor for each data point. ; The data source s is quality-scored within time bucket b based on the time decay factor. ; Constructing multi-source data within the time bucket into fixed variables A vector is formed by concatenating fixed variables from multiple data sources. ; For time buckets with missing data and consecutive missing bucket numbers less than or equal to G, use data from the previous time bucket and mark it. If the number of missing data is greater than G, mark the time bucket as empty and record the missing mask. ; Quality rating and missing mask Simultaneous splicing to form a mass vector and missing mask vector .

3. The freight data management method based on artificial intelligence as described in claim 2, characterized in that: The process of calculating the fusion reference value for each fixed variable to obtain the data source consistency residual and consistency score, and calculating the consistency mean to obtain an immediate score, updating the fixed variable credibility based on the immediate score, and calculating the fixed variable fusion and writing it into the semantic knowledge graph based on the updated fixed variable credibility specifically involves: initializing credibility. Calculate the fusion reference value for each fixed variable. ; Calculate source consistency residuals based on fusion reference values. With consistency score ; Calculate the consistency mean based on the consistency scores of the fixed variables covered by data source s within the time bucket. The consistency mean was then compared with the quality score of time bucket b. Weighted fusion into instant scoring ; Reliability updates based on EWMA ; Each fixed variable is weighted and merged based on its update reliability; The weighted and fused fixed variables are written into the semantic knowledge graph KG as facts in the form of events.

4. The freight data management method based on artificial intelligence as described in claim 3, characterized in that: The specific steps of calculating statistics for each fixed variable and mapping them to nodes to construct a variable dependency graph are as follows: all fixed variables are divided into 5 groups, including power fuel consumption, braking chassis, tire vibration, cold chain in the vehicle compartment and visual text traffic scheduling, and statistics are calculated for each fixed variable i. The statistics are concatenated to form the input vector. and input vector Input a Transformer encoder to map and obtain nodes. ; According to the node Calculate the cosine similarity and introduce a reliability coefficient; Select the top-K nodes with the largest values ​​to build a directed adjacency graph A, which is a variable dependency graph.

5. The freight data management method based on artificial intelligence as described in claim 4, characterized in that: The process involves constructing a window based on time buckets and processing the window values ​​using a two-layer fully connected encoder. The two-layer fully connected encoder is trained using the SAE loss function. The latent variables output by the trained encoder are specifically: a vector based on the time buckets. Construct a window of length w ; From window Extract the window value of the fixed variable i. Using a two-layer fully connected encoder to convert window values Mapped to latent vectors ; The two-layer fully connected encoder is trained using the SAE loss function, and the latent variables are output through the trained fully connected encoder.

6. The freight data management method based on artificial intelligence as described in claim 5, characterized in that: The process of constructing node features based on latent variables, using Graph Transformer to calculate multi-head attention to update node features, and calculating anomaly scores for variable nodes specifically involves: constructing node features based on latent variables of all fixed vectors... Stacking yields a sparse implicit representation matrix h, and the nodes are... The temporal embeddings of variables are stacked to obtain v. The sparse latent representation matrix h is linearly projected and concatenated with the temporal embeddings of variables to obtain the initial node features z. Edge features are defined based on the directed edges. ; Graph Transformer is used to compute multi-head attention to update node features; Update node features and After element-wise multiplication, the input to the linear layer yields the prediction vector. Calculate the predicted vector and the actual vector The predicted loss between; The total loss function is obtained by weighting the prediction loss and the SAE loss function. The graph Transformer parameters are trained based on the total loss function, and the prediction vector is output based on the trained graph Transformer. Calculate the outlier score of the variable node based on the prediction vector. .

7. The freight data management method based on artificial intelligence as described in claim 6, characterized in that: The specific steps for constructing an abnormal evidence subgraph based on abnormal scores and calculating the final symptom vector are as follows: sort the variables in descending order of abnormal scores and select the top 5 variables. ,right Each node i selects a set of directed neighbors. Form a set of evidence edges and construct an abnormal evidence subgraph ; Define the abnormal evidence subgraph Symptom intensity for each variable i ; Define the propagation strength for each piece of evidence. ; The symptom intensity and the transmission intensity are combined to form the final symptom vector F.

8. The freight data management method based on artificial intelligence as described in claim 1, characterized in that: Define a candidate root cause set, which is manually defined and fixed as a root cause template library.

9. The freight data management method based on artificial intelligence as described in claim 8, characterized in that: The process of calculating prior probabilities and evidence likelihoods based on the candidate root cause set and outputting posterior probabilities to select the most likely root cause specifically involves: for each root cause r in the candidate root cause set, obtaining the prior probability of the root cause through historical statistics. ; Simultaneously, the likelihood of evidence is calculated based on root causes. ; Based on prior Likelihood of Evidence Output posterior probability ; Based on posterior probability The root cause with the highest probability is selected as the most likely root cause output.

10. An artificial intelligence-based freight data management system, based on the artificial intelligence-based freight data management method according to any one of claims 1 to 9, characterized in that: include, The data acquisition module is used to collect multi-source data, perform time correction, calculate the quality score of multi-source data, construct multi-source data into fixed variables, and concatenate them into a vector. The credibility analysis module is used to calculate the fusion reference value of each fixed variable to obtain the data source consistency residual and consistency score, and calculate the consistency mean to obtain the real-time score. Based on the real-time score, the credibility of the fixed variables is updated, and the fixed variable fusion is calculated based on the updated credibility of the fixed variables and written into the semantic knowledge graph. The dependency graph construction module is used to calculate statistics for each fixed variable and map them to nodes to construct a variable dependency graph. The variable mapping module is used to construct windows based on time buckets and process window values ​​using a two-layer fully connected encoder. The two-layer fully connected encoder is trained based on the SAE loss function, and the latent variables are output based on the trained encoder. The node anomaly analysis module is used to construct node features based on latent variables, use graph Transformer to calculate multi-head attention to update node features and calculate the anomaly score of variable nodes, construct anomaly evidence subgraph based on the anomaly score and calculate the final symptom vector. The root cause analysis module is used to define a set of candidate root causes, calculate prior probabilities and evidence likelihoods based on the set of candidate root causes, and output posterior probabilities to select the most likely root cause.

Images