Supply chain risk early warning method and system based on AI

Abstract

The invention relates to the technical field of intelligent logistics, in particular to an AI-based supply chain risk early warning method and system, and the method comprises the following steps: extracting a staggering degree and a jump vector based on a performance time sequence, calculating an abnormal risk through a Transform model in combination with a density mutation node, and generating a performance risk early warning result. According to the method, the plan of the node in the logistics order is compared with the completion time, the performance time sequence is constructed, and change behaviors such as rhythm reversal and execution interval fluctuation are identified, so that rhythm deviation caused by external disturbance in the performance process can be effectively described; through comparison between the performance completion sequence and the node numbers, the displacement degree of performance sequence disturbance is extracted, and the nonlinear sequential relation between the nodes is further captured; the two types of disturbance vectors are jointly coded into a deep expression structure with a time sequence structure, and cross-dimension and cross-channel feature alignment and correlation mapping are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent logistics, and in particular to an AI-based supply chain risk warning method and system. Background Art

[0002] The technical field of intelligent logistics includes intelligent perception, transmission, processing of logistics information, and intelligent decision-making and execution, etc. It is a comprehensive technical system that realizes intelligent management of the entire logistics process based on the Internet of Things, artificial intelligence, big data, etc. The core content of this field includes intelligent scheduling of logistics resources, optimization management of transportation routes, automated coordination of warehousing operations, real-time monitoring of the status of logistics nodes, and construction of a prediction mechanism, etc.

[0003] Among them, the supply chain risk warning method refers to the process of identifying potential abnormal change trends by setting key index monitoring points and comparing and analyzing historical data with real-time operation data for risk events such as interruptions, delays, and inventory imbalances that may occur during the operation of the supply chain. It includes technical matters such as setting of risk identification parameters, extraction of risk characteristic patterns, construction of warning judgment logic rules, and setting of triggering conditions.

[0004] The existing technology mainly relies on the static setting of key indicators and the trend comparison of historical data in risk event monitoring, lacking the ability of dynamic modeling for time disturbances and node sequence changes in performance behaviors, resulting in a lag in response when dealing with complex disturbance patterns; in the process of node anomaly identification, most methods only label the mutations of abnormal values and fail to establish the sequence structure relationship between nodes, making it difficult to identify the systemic risk diffusion caused by node sequence disturbances; density identification is usually based on statistical analysis over a fixed time period, ignoring the volatility of local task loads in time, and easily masking the density variation under high-frequency disturbances; in the process of risk level identification, relying on a single abnormal signal as the judgment basis and failing to conduct collaborative evaluation of multiple features, prone to misjudgment and missed judgment of warning results; for example, in a certain logistics distribution task, although there is no abnormal value jump in the overall node data, there are misalignments in the sequences of multiple performance nodes and task concentration and accumulation. If only relying on traditional index monitoring, the rhythm imbalance and load anomaly existing in the execution chain cannot be identified in time, thus triggering the risk of subsequent logistics interruption and plan deviation, resulting in a reduction in the overall supply chain scheduling efficiency. Summary of the Invention

[0005] The purpose of the present invention is to solve the deficiencies existing in the prior art and propose an AI-based supply chain risk warning method and system.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: An AI-based supply chain risk warning method, comprising the following steps: S1: Obtain the fulfillment time series through the planned time and completion time of the logistics order nodes, and perform staggered trigger point identification processing on the fulfillment status differences to obtain the stagger degree sequence; S2: Construct the fulfillment completion order according to the node completion times in the fulfillment time series, and generate a jump vector by comparing the positions with the node numbers; S3: Input the stagger degree sequence and the jump vector into the embedding layer of the Transformer model to extract non-periodic fulfillment perturbation features, and obtain the non-periodic fulfillment perturbation feature set; S4: Obtain the task distribution according to the node completion times in the fulfillment time series, construct a task pressure time window view for density mutation identification, screen the nodes with density mutations, and form a density mutation node set; S5: Input the density mutation node set and the non-periodic fulfillment perturbation feature set into the decoding layer of the Transformer model, calculate the abnormal risk scores of the density mutation nodes, and generate a supply chain logistics fulfillment risk warning result according to the scores.

[0007] As a further solution of the present invention, the stagger degree sequence includes rhythm inversion positions, sign change nodes, and execution interval fluctuation sections, the jump vector includes sequential offset amounts, reverse execution node numbers, and jump node spacing values, the non-periodic fulfillment perturbation feature set includes multi-channel time alignment vectors, attention focus weights, and fulfillment perturbation pattern encodings, the density mutation node set includes task concentration mutation points, local pressure change sections, and periodic density abnormal windows, and the supply chain logistics fulfillment risk warning result includes risk node numbers, abnormal risk scores, and node risk level identifiers.

[0008] As a further solution of the present invention, the steps for obtaining the stagger degree sequence are specifically as follows: S111: Obtain the planned completion time and actual completion time of each fulfillment node in the supply chain order, calculate the difference between the planned continuation interval and the execution interval between adjacent nodes in the fulfillment time series composed of these two types of time data, and obtain the planned-actual interval difference sequence; S112: Based on the planned-actual interval difference sequence, obtain the positive and negative signs of the differences between each pair of nodes, and perform continuity judgment according to the positive and negative sign change states to obtain the staggered trigger point set; S113: Based on the staggered trigger point set, mark and count the trigger positions corresponding to the fulfillment nodes in the original order to obtain the stagger degree sequence representing the stagger degree value sequence as the rhythm perturbation degree.

[0009] As a further solution of the present invention, the steps for obtaining the jump vector are specifically as follows: S211: Obtain the actual completion time of the nodes in the performance time series, sort each performance node in ascending order according to the completion time, and construct a performance completion order sequence reflecting the actual performance order; S212: Compare the positions of the nodes in the performance completion order sequence with the original number order of the performance nodes, calculate the position difference of the corresponding nodes in the two sequences, and judge the deviation direction and span size according to the position difference to identify the jumping nodes and reverse-order nodes, and generate a jumping vector representing the node order disturbance.

[0010] As a further solution of the present invention, the steps for obtaining the non-periodic performance disturbance feature set are specifically as follows: S311: Based on the jumping vector, the interleaving degree sequence, and the performance time series of the original order nodes, splice the three types of data in the dimension of the performance node position, construct a composite input vector with a unified structure, and obtain a multi-channel performance expression sequence; S312: Input the multi-channel performance expression sequence into the embedding layer of the Transformer model, and call the position encoding mechanism to perform vector-level embedding processing on the node order information to form an encoded input representation with a time sequence structure; S313: Based on the encoded input representation, call the multi-head attention mechanism of the Transformer model to compare the expression content between channels, extract cross-channel cross features, and obtain a non-periodic performance disturbance feature set including node disturbance, rhythm interleaving, and time deviation characteristics.

[0011] As a further solution of the present invention, the steps for obtaining the density mutation node set are specifically as follows: S411: Obtain the completion time of each performance node in the performance time series as the arrival time data of the node task, divide the arrival time data into discontinuous time periods according to the minimum completion time interval between tasks, and construct a set of local time windows covering the entire performance cycle; S412: Based on the set of local time windows, use the kernel density estimation algorithm to calculate the task arrival density within each local time window, and sequentially generate a density vector reflecting the change trend of the task distribution; S413: Based on the density vector, calculate the difference from the average density value of the window at the same position in the cycle record of the corresponding performance node, identify the window positions where the density change amplitude exceeds the amplitude threshold range, extract the nodes located therein, and form a density mutation node set.

[0012] As a further solution of the present invention, the steps for obtaining the supply chain logistics performance risk warning result are specifically as follows: S511: Based on the density mutation node set and the non-periodic performance perturbation feature set, perform feature splicing processing on the two types of data in terms of the performance node dimension to construct a fusion feature sequence including density change information and performance perturbation features; S512: Input the fusion feature sequence into the decoding layer of the Transformer model, call the multi-head attention mechanism to calculate the attention weights of the inter-node fusion features in the fusion feature sequence, extract the focus response values of each node under multi-dimensional perturbations and density backgrounds, and output a risk score representing the degree of node abnormality; S513: Based on the risk score, perform classification judgment according to the set risk judgment threshold, screen the nodes with scores exceeding the threshold and assign corresponding grade identifiers, and generate a supply chain logistics performance risk warning result to reveal risk hazards such as timing imbalance, execution deviation, and node load abnormality existing in the performance process.

[0013] An AI-based supply chain risk warning system, the AI-based supply chain risk warning system is used to execute the above-mentioned AI-based supply chain risk warning method, and the system includes: The performance node analysis module obtains the performance time series through the planned time and completion time of the logistics order nodes, and performs staggered trigger point identification processing on the performance status differences to obtain the stagger degree sequence; The jump vector generation module constructs the performance completion order according to the node completion time in the performance time series, and generates a jump vector by comparing the position with the node number; The perturbation feature extraction module inputs the stagger degree sequence and the jump vector into the embedding layer of the Transformer model to extract non-periodic performance perturbation features, and obtains a non-periodic performance perturbation feature set; The density mutation identification module obtains the task distribution according to the node completion time in the performance time series, constructs a task pressure time window view for density mutation identification, screens the nodes with density mutations, and forms a density mutation node set;

[0014] The risk warning calculation module inputs the density mutation node set and the non-periodic performance perturbation feature set into the decoding layer of the Transformer model, calculates the abnormal risk score of the density mutation nodes, and generates a supply chain logistics performance risk warning result according to the score.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by comparing the planned and completed times of nodes in a logistics order, a fulfillment time series is constructed, and change behaviors such as rhythm reversal and execution interval fluctuation are identified, which can effectively characterize the rhythm deviation caused by external disturbances during the fulfillment process; by comparing the fulfillment completion order with the node number, the displacement degree of the fulfillment order disturbance is extracted to further capture the non-linear time series relationship between nodes; the above two types of disturbance vectors are jointly encoded into a deep expression structure with a time series structure, realizing cross-dimensional and cross-channel feature alignment and correlation mapping, capable of identifying abnormal signals from two dimensions of fulfillment rhythm and order offset, and improving the recognition accuracy of the model for complex disturbance patterns; by means of local time window division and kernel density estimation methods, quantitative analysis of the sudden change in task density is carried out, and density anomaly recognition does not rely on preset thresholds, with higher adaptability and generalization; in the feature fusion stage, a node-level risk assessment system is established through the unified expression structure of fulfillment disturbance features and density mutation features, enabling the abnormal score to have dual background information of fulfillment behavior and load change, thereby significantly improving the accuracy and coverage breadth of risk recognition; in the process of early warning generation, a risk level identifier is dynamically constructed according to the results of multi-dimensional feature extraction, making the early warning result interpretable and having a hierarchical guidance function, and being able to provide a more valuable decision-making basis for supply chain scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the working process of the present invention; Figure 2 is a flowchart of step S1 of the present invention; Figure 3 is a flowchart of step S2 of the present invention; Figure 4 is a flowchart of step S3 of the present invention; Figure 5 is a flowchart of step S4 of the present invention; Figure 6 is a flowchart of step S5 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0018] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0019] Please refer to Figure 1 , the present invention provides a technical solution: an AI-based supply chain risk warning method, including the following steps: S1: Obtain the performance time series through the planned time and completion time of the logistics order nodes, and perform staggered trigger point identification processing on the performance status differences to obtain the stagger degree sequence; S2: Construct the performance completion order according to the node completion time in the performance time series, and generate a jump vector by comparing the positions with the node numbers; S3: Input the stagger degree sequence and the jump vector into the embedding layer of the Transformer model for non-periodic performance perturbation feature extraction to obtain a non-periodic performance perturbation feature set; S4: Obtain the task distribution according to the node completion time in the performance time series, construct a task pressure time window view for density mutation identification, screen the nodes with density mutations, and form a density mutation node set; S5: Input the density mutation node set and the non-periodic performance perturbation feature set into the decoding layer of the Transformer model, calculate the abnormal risk score of the density mutation nodes, and generate a supply chain logistics performance risk warning result according to the score; The stagger degree sequence includes the rhythm reversal position, the sign change node, and the execution interval fluctuation section. The jump vector includes the sequence offset, the reverse execution node number, and the jump node spacing value. The non-periodic performance perturbation feature set includes the multi-channel time alignment vector, the attention focus weight, and the performance perturbation pattern encoding. The density mutation node set includes the task concentration mutation point, the local pressure change section, and the periodic density abnormal window. The supply chain logistics performance risk warning result includes the risk node number, the abnormal risk score, and the node risk level identifier.

[0020] Please refer to Figure 2 , the specific steps for obtaining the stagger degree sequence are as follows: S111: Obtain the planned completion time and the actual completion time of each fulfillment node in the supply chain order. Calculate the difference between the planned continuation interval and the execution interval between adjacent nodes in the fulfillment time series composed of these two types of time data to obtain the planned-actual interval difference sequence. Obtain the planned completion time and the actual completion time of each fulfillment node in the supply chain order. First, retrieve the order numbers involved and their associated fulfillment node information from the order management system. The data of each node should include the planned completion time (such as 14:00 on March 12, 2025) and the actual completion time of the node (such as 10:00 on March 13, 2025). Then, arrange all the fulfillment nodes in the same order and traverse each adjacent pair of nodes in turn. Between node i and node i+1, calculate the time difference of the planned completion time (for example, if the plan for node 1 is 14:00 on March 12 and for node 2 is 14:00 on March 13, the planned continuation interval is 24 hours). Similarly, calculate the time difference between the actual completion times of the two nodes (for example, if the actual for node 1 is 10:00 on March 13 and for node 2 is 13:00 on March 14, the execution interval is 27 hours). Then, calculate the difference between the two intervals, that is, the execution interval minus the planned continuation interval. This difference is used as the time deviation value in this section of fulfillment. Repeat the above process to obtain the planned-actual interval difference sequence between all nodes in turn. Uniformly use hours as the unit of time to maintain consistency in subsequent calculation processes. For example, if there are 5 fulfillment nodes in order number A123, a total of 4 adjacent node pairs will be generated, and finally 4 difference data such as [3, -2, 1, -4] will be obtained, where 3 means the actual continuation is extended by 3 hours, -2 means the continuation is completed 2 hours ahead of schedule compared to the plan. Obtain the complete difference sequence in this way.

[0021] S112: Based on the planned-actual interval difference sequence, obtain the positive and negative signs of the differences between each pair of nodes. Make a continuity judgment according to the change state of the positive and negative signs to obtain the set of interleaved trigger points. Based on the planned-actual interval difference sequence, such as the sequence [3, -2, 1, -4], first obtain the sign of each difference to form a sign sequence [+, -, +, -]. Then, judge the sign sequence item by item, and analyze the change between the signs of two adjacent items. If there is an alternation between positive and negative (such as + to - or - to +), it is regarded as an interleaving trigger point. For example, the first item to the second item changes from + to -, which is recorded as interleaving trigger point 1. Then the second item to the third item changes from - to +, which is recorded as interleaving trigger point 2. And so on. The entire sequence is traversed to determine all possible interleaving positions, thereby forming an interleaving trigger point set. The elements are recorded according to the index in the original sequence. For example, the trigger point index in this example is [1, 2, 3]. This set can be used to further analyze the rhythm change characteristics between the fulfillment nodes, assign a number or label to each staggered record, and form annotation results such as "staggered point @ node 2", "staggered point @ node 3", etc. In the data record table, the staggered annotation status is recorded for each node, and whether it is an staggered node pair is recorded in an array or Boolean value, such as [True, True, True, False], indicating that the first three pairs of nodes are all staggered change points. Finally, all recognition results are integrated to form a complete set of staggered trigger points.

[0022] S113: Based on the staggered trigger point set, the trigger positions corresponding to the fulfillment nodes in the original sequence are marked and counted, and an staggered degree sequence representing the staggered degree numerical sequence as the rhythm disturbance degree is obtained; Based on the set of staggered trigger points, the trigger positions corresponding to the fulfillment nodes in the original sequence are marked and counted. First, in the original fulfillment node list, the identified staggered points are marked with serial numbers. For example, if the original fulfillment nodes are [N1, N2, N3, N4, N5] and the staggered trigger points are [1, 2, 3], the marking results are that nodes N2, N3, and N4 are marked as "staggered node 1", "staggered node 2", and "staggered node 3" respectively. Then the number of staggered trigger points in the overall sequence is counted, such as 3 in this example. On this basis, a sequence of staggered degrees is formed with nodes as indexes, and its value represents the number of staggered trigger points in the node. The interleaving frequency of the point. If a node appears in multiple interleaving trigger points at the same time, the interleaving degree of the node increases by 1. For example, node N3 is the common node of the node pairs (N2, N3) and (N3, N4), so the interleaving degree is 2. After statistics of all nodes, the interleaving degree sequence is obtained, such as [0, 1, 2, 1, 0]. The interleaving degree sequence can be further normalized. Assuming the highest interleaving degree is 2, the normalized sequence is [0, 0.5, 1.0, 0.5, 0]. This sequence is the interleaving degree sequence of the rhythm disturbance degree. Each value clearly reflects the degree of rhythm instability reflected by the corresponding node in the entire fulfillment chain.

[0023] See also Figure 3 , the steps to obtain the jump vector are as follows: S211: Obtain the actual completion time of nodes in the performance fulfillment time series, sort each performance fulfillment node in ascending order according to the completion time, and construct a performance fulfillment order sequence reflecting the actual order of performance fulfillment; To obtain the actual completion time of nodes in the performance fulfillment time series, it is necessary to record the completion time information of each performance fulfillment node. This information usually comes from the timestamp data generated in the on-site execution system. For example, the actual completion time of node N1 is 15:00 on March 12, 2025, node N2 is 09:00 on March 11, 2025, node N3 is 18:00 on March 13, 2025, node N4 is 12:00 on March 11, 2025, and node N5 is 08:00 on March 14, 2025. Arrange all nodes in ascending order according to the timestamps from earliest to latest, and the actual order of performance fulfillment is N2 (09:00) → N4 (12:00) → N1 (15:00) → N3 (18:00) → N5 (08:00@14th). Construct a new sequence [N2, N4, N1, N3, N5] to reflect the true order of performance fulfillment of each node. And this sorting result can be used as the basis for subsequent jump identification. It is necessary to retain the original numbering order for subsequent comparison. For example, the original order is [N1, N2, N3, N4, N5]. Through sequence mapping operations, record the position of each node in the actual order. For example, N1 was originally in the 1st place and is in the 3rd place in the actual order, N2 was originally in the 2nd place and is now in the 1st place. In this way, the complete sorting mapping correspondence relationship is obtained.

[0024] S212: Compare the positions of nodes in the performance fulfillment order sequence and the original numbering order of performance fulfillment nodes, calculate the difference in positions of corresponding nodes in the two sequences, and judge the deviation direction and span size according to the position difference to identify jump nodes and reverse-order nodes, and generate a jump vector representing the disturbance of node order; Compare the order of performance completion sequence with the original number order of performance nodes to correspond the node positions. Compare each node's position in the original sequence and the actual performance sequence one by one, record the position difference and calculate its positive and negative directions as well as the absolute value of the difference. For example, for node N1, its original position is 1, the actual performance position is 3, and the position difference is +2, indicating that it is completed two positions later. For N2, the original is 2 and the actual is 1, with a difference of -1, indicating that it is completed 1 position earlier. Process all nodes in this way to obtain a complete position difference vector such as [+2, -1, 0, -2, +1]. Then analyze each value in the difference vector. When the absolute value of the difference is greater than or equal to 2, record it as a jump node. For example, N1(+2) and N4(-2) are jump nodes. If the difference is negative and the absolute value exceeds 1, record it as a reverse order node. For example, N4(-2) meets the condition. Mark and classify the jump types of nodes in this way, construct a jump vector structure for all nodes, and each item in the vector is the marked value of the corresponding node. If encoded with {0: no disturbance, 1: jump, 2: reverse order}, the vector result is [1, 0, 0, 2, 0], indicating that node N1 is a jump, N4 is a reverse order, and the remaining nodes are normal order nodes. At the same time, construct a complete disturbance vector and output the corresponding disturbance type and degree value for each node.

[0025] Please refer to Figure 4 , the steps for obtaining the non-periodic performance disturbance feature set are specifically as follows: S311: Based on the jump vector, the interleaving degree sequence, and the performance time series of the original order nodes, perform splicing processing on the three types of data in the dimension of the performance node position to construct a composite input vector with a unified structure, and obtain a multi-channel performance expression sequence; Based on the jump vector, interleaving degree sequence, and fulfillment time sequence, first align them according to the fulfillment node positions to ensure that the three data dimensions correspond in the node sequence. The jump vector is used to represent the perturbation state of each node in the fulfillment order. For example, the jump state from node 1 to node 5 is [1, 0, 0, 2, 0]; the interleaving degree sequence represents the intensity of the rhythm perturbation during the fulfillment process, such as [0.0, 0.5, 1.0, 0.5, 0.0]; the fulfillment time sequence records the actual completion time of the nodes. Taking a fixed reference time as the starting point (e.g., 00:00 on March 11, 2025), convert the completion time of each node into a time difference (in hours). For example, if node 1 is completed at 15:00 on March 12, the difference is 39 hours, and if node 2 is completed at 09:00 on March 11, the difference is 9 hours. Thus, we get [39, 9, 66, 12, 80]. Concatenate the three types of data in sequence according to the node order. Each node forms a three-dimensional vector including three dimensions. The result is: node 1 is [1, 0.0, 39], node 2 is [0, 0.5, 9], node 3 is [0, 1.0, 66], node 4 is [2, 0.5, 12], node 5 is [0, 0.0, 80]. Finally, construct a 5-row and 3-column composite input matrix.

[0026] S312: Input the multi-channel fulfillment expression sequence into the embedding layer of the Transformer model, and call the position encoding mechanism to perform vector-level embedding processing on the node order information to form an encoded input representation with a temporal structure; When inputting the multi-channel fulfillment expression sequence into the Transformer model, first perform mapping processing on the three-dimensional vector of each node. Usually, use a fully connected layer to map each three-dimensional input to a high-dimensional vector with a fixed dimension, such as 64 dimensions. The processing process is as follows: The original three-dimensional vector of node 1 is [1, 0.0, 39], which is linearly mapped to a 64-dimensional vector A1 through a linear mapping; the input [0, 0.5, 9] of node 2 is mapped to vector A2, and so on, obtaining a total of 5 high-dimensional representations A1 to A5. To retain the order information of the nodes, introduce a position encoding mechanism. Generate a group of encoding vectors P1 to P5 with the same dimension as the mapping vector according to the index position of the nodes in the sequence (i.e., 1 to 5). Then add the embedding vector of each node to its corresponding position encoding vector to form embedding vectors E1 to E5 that integrate the order information. For example, if the logistics process includes five nodes: node 1 is the warehousing station, node 2 is the sorting center, node 3 is the transportation center, node 4 is the transfer hub, and node 5 is the terminal distribution point, then these nodes exhibit different perturbation behaviors due to rhythm, order, and time differences during the fulfillment process. The above five embedded vectors E1 to E5 are used as the model inputs with time and order semantics.

[0027] S313: Based on the encoded input representation, call the multi-head attention mechanism of the Transformer model to compare the expression content between channels, extract cross-channel cross features, and obtain a set of non-periodic performance perturbation features including node perturbation, rhythm interleaving, and time deviation characteristics; After receiving the node embedding vector sequence with position encoding, the Transformer model first analyzes the dependency relationships between nodes through the multi-head attention mechanism. Each attention head calculates the correlation between the embedding vectors for each pair of nodes. By constructing query vectors, key vectors, and value vectors, it calculates the attention degree of one node to all other nodes. For example, node 3 has the largest rhythm perturbation during transportation, with a jump value of 0, an interleaving degree of 1.0, and a time difference of 66 hours. Such a node may generate higher attention scores for node 1 (warehousing node) and node 5 (terminal node) in the attention mechanism, indicating that its abnormal position may affect the performance stability of the front and rear nodes. Multiple attention heads can capture cross features between different channels in parallel, such as the relationship between jumps and time, and the synchronization between the interleaving degree and the performance time. The features extracted by all attention heads are fused and fed into a feed-forward neural network to further refine high-order semantics, resulting in a set that includes the non-periodic perturbation features shown by each node in the performance process. The finally output feature set can reflect the dynamic changes in multiple dimensions such as rhythm chaos, order changes, and time delays between nodes. For example, node 4 is a transfer hub. Due to sudden traffic problems, the actual performance order is advanced, but the interleaving degree is not high. Its perturbation features are shown as feature points dominated by order perturbation after being processed by the model. However, for node 3, since time delay and rhythm jumps occur simultaneously, its feature vector shows as a typical multi-perturbation node.

[0028] Please refer to Figure 5 , and the steps for obtaining the density mutation node set are specifically as follows: S411: Obtain the completion time of each performance node in the performance time series as the arrival time data of the node task. Divide the arrival time data into discontinuous time periods according to the minimum completion time interval between tasks, and construct a set of local time windows covering the entire performance cycle;

[0029] Obtain the actual completion time of each node in the performance fulfillment time series as the task arrival time data. First, retrieve the completion timestamps of each node from the system. For example, the completion times of nodes 1 to 5 are 09:00 on March 11, 2025, 12:00 on March 11, 2025, 15:00 on March 12, 2025, 18:00 on March 13, 2025, and 08:00 on March 14, 2025. Calculate the time differences between the completion times of adjacent nodes, which are 3 hours, 27 hours, 27 hours, and 14 hours respectively. Take the minimum interval value of 3 hours as the division unit of the local time window, and divide the entire performance fulfillment cycle discontinuously according to this interval. The performance fulfillment cycle starts from the earliest 09:00 on March 11, 2025, and ends at 08:00 on March 14, 2025, with a total duration of 71 hours. Therefore, it is divided into 24 non-overlapping 3-hour windows: [The 1st window is from 09:00 to 12:00 on March 11, the 2nd window is from 12:00 to 15:00... the 24th window is from 05:00 to 08:00 on March 13]. Classify all task completion times into the corresponding windows to form a set of local time windows.

[0030] S412: Based on the set of local time windows, use the kernel density estimation algorithm to calculate the task arrival density within each local time window, and sequentially generate a density vector reflecting the changing trend of task distribution; On the basis of the constructed set of local time windows, use the kernel density estimation algorithm to estimate the arrival density of tasks in each window. Considering the non-uniform distribution characteristics of node tasks in the performance fulfillment scenario, the following kernel density estimation formula is introduced: ; The parameter explanations are as follows: : Represents the task density estimation value at time point , which is a continuous expression of the task distribution in the current window; : Is the number of task nodes within the current local time window, and this value may be different in each window; : The basic bandwidth parameter, which controls the diffusion degree of the kernel function. The larger the value, the higher the smoothing degree. In this article, it is set to 1 hour according to the density of the performance fulfillment time series; : Represents the actual completion time of the th task node (in hours, and the time difference is calculated based on the performance fulfillment start time as the benchmark); : Is the node adaptive bandwidth adjustment term, which is used to correct the bandwidth in combination with the performance fulfillment perturbation characteristics. If the jump vector value corresponding to the node is 2 or the interleaving degree value is greater than 0.8, then set hours, otherwise , which is used to reflect the expansion of the local influence range of the perturbed node on density estimation; : is the standard normal kernel function, defined as: , which is used to calculate the density contribution value of a certain task arrival point to the target time point.

[0031] Taking the first time window from 09:00 to 12:00 on March 11, 2025 as an example, this window includes Node 1 (arrival time is 09:00, jump value is 0, interleaving degree is 0.0) and Node 2 (arrival time is 12:00, jump value is 0, interleaving degree is 0.5). Since neither of the two nodes meets the perturbation condition, so . Set the basic bandwidth hours, and calculate the density estimate at the midpoint 10:30 of the window: , and look up from the standard normal distribution table , to get: . Calculate the center points of each time window in this way, and successively obtain the complete density vector, such as: [0.1295, 0.1102, 0.0300, 0,...]. This density vector reflects the dynamic distribution trend of tasks during the entire performance cycle and serves as the basis for subsequent density mutation analysis.

[0032] S413: Based on the density vector, calculate the difference with the average density value of the corresponding window at the same position in the cycle record of the performance node, identify the window positions where the density change amplitude exceeds the amplitude threshold range, extract the nodes involved, and form a density mutation node set; Based on the above density vector, calculate the difference between the density value of each time window and the historical average density at the same window position in the cycle record. The historical data comes from the average value of multiple task arrival records in the historical performance cycle. For example, the current density of Window 1 is 0.1295 and the historical average density is 0.08, then the difference is +0.0495. Set the density change amplitude threshold to 0.04. If the absolute value of the current difference exceeds this threshold, it is regarded as a density mutation window. In this example, the difference of the first window is +0.0495 > 0.04, which is marked as a mutation window. Extract the nodes covered by this window, namely Node 1 and Node 2, to form a density mutation node set. After traversing all windows in turn, obtain the complete mutation node index set, such as [1, 2, 4], which means that the task arrival density of Node 1, 2, and 4 has changed significantly during their respective time periods.

[0033] Please refer to Figure 6 , and the specific steps for obtaining the supply chain logistics performance risk warning result are as follows: S511: Based on the density mutation node set and the non-periodic performance perturbation feature set, perform feature splicing processing on the two types of data in the dimension of the performance node, and construct a fusion feature sequence including density change information and performance perturbation features; Based on the density mutation node set and the non-periodic performance perturbation feature set, it is necessary to perform feature splicing processing on the two in the dimension of performance nodes. The density mutation node set reveals the abnormality of task centrality through the change of window density, while the perturbation feature set is derived from the extraction of multi-dimensional perturbation features such as node jump, stagger, and time offset by the Transformer model. Each node can be represented as a high-dimensional vector. For example, the perturbation feature is a 64-dimensional dense representation, and the density change is represented by whether it is a mutation node (boolean value) and its original difference (real number) as 2 dimensions. Combine these two types of information to construct a 66-dimensional fusion feature vector for each node, that is, each node includes, for example: [perturbation feature 1,..., perturbation feature 64, density mutation boolean flag, density change value], and organize them in sequence along the performance nodes to form a fusion feature sequence. For example, node 1 is [0.23,..., 0.75, 1, 0.049], node 2 is [0.11,..., 0.60, 1, 0.036], node 3 is [0.04,..., 0.52, 0, 0.015], and finally form a matrix input of the form N×66 for decoder processing.

[0034] S512: Input the fusion feature sequence into the decoding layer of the Transformer model, call the multi-head attention mechanism to calculate the attention weights of the fusion features between nodes in the fusion feature sequence, extract the focus response values of each node under the multi-dimensional perturbation and density background, and output the risk score representing the degree of node abnormality; After inputting the fusion feature sequence into the decoding layer of the Transformer model, the model uses the multi-head attention mechanism to extract the focus response degree of each node under the background of the fusion perturbation feature and density information. The response value of each node represents the attention intensity it receives in the global sequence, and the focus response value is calculated by the following formula: ; The definitions of each letter are as follows: : The focus response value of the th performance node. The higher the value, the more likely this node is to be affected by the response aggregation of other nodes in the perturbation and density dimensions; : The fusion feature vector of the th performance node, including non-periodic perturbation features (such as jump, stagger, time offset) and density mutation features (boolean value, difference) extracted from multi-channel performance expressions; : The fusion feature vector of the th performance node, which is the feature of other target nodes focused by node j; : Query matrix, which maps to the query vector , used to initiate an attention operation on external information, with a dimension of ; : The key matrix maps to a key vector , which is used to respond to query operations and has the same dimension ; : The density correction vector of node , which represents the enhancement of the node's attention due to density mutation, is defined as: , : Whether node is a density mutation node. If it is a mutation node, it is 1; otherwise, it is 0; : The density mutation amplitude of node (the difference between the current window density and the historical mean, in density units per hour); : The attention weight of node to node , and the calculation method is: , , : The unnormalized attention score, which measures the feature correlation between node and ; : The dimensionality scaling factor of the query / key vector to ensure numerical stability of the model; : The total number of nodes, which is equal to the number of fulfillment nodes in the fulfillment sequence; : The traversal index, which is used to calculate attention normalization for all nodes; : Represents the number of the first node in the current sequence input data, used to indicate the minimum position index.

[0035] Assume that the fulfillment sequence includes 3 nodes (N = 3), and each fused feature vector is reduced to 3 dimensions for calculation demonstration: Node 1: , where 0.049 is the density difference and the mutation flag is 1, , , so . Node 2: , . Node 3: , a non-mutation node, , the identity matrix, .

[0036] Calculate the attention response value of node 1 : Calculate the unnormalized attention score: ; ; ; Calculate attention weights : ; 、 ; Calculate the focused response value: , finally, the focused response value of node 1 is 0.0917, reflecting the risk intensity of the node.

[0037] S513: Based on the risk score, perform classification judgment according to the set risk judgment threshold, screen the nodes with scores exceeding the threshold and assign corresponding level labels, generate the early warning result of the supply chain logistics fulfillment risk, and use it to reveal the risk hidden dangers such as timing imbalance, execution deviation and node load abnormality in the fulfillment process; Based on the focused response value (i.e., risk score ) of each fulfillment node output by the Transformer decoding layer, perform the classification judgment process to identify the key risk nodes in the fulfillment process. The system pre-sets a set of risk judgment thresholds for grading the risk score results. Let the main threshold be , this value is the critical point obtained through the statistical distribution of abnormal node scores in historical fulfillment data, and is used to distinguish normal and abnormal nodes; at the same time, a hierarchical threshold system is introduced to further divide the score interval into three levels: low-risk interval: , the node score is within the normal fluctuation range; medium-risk interval: , there are medium-level disturbances in the node, which need to be focused on; high-risk interval: , the node may be in a state of execution bottleneck, rhythm chaos or overloaded pressure, triggering a strong early warning mechanism. Judge each node score, screen out the nodes exceeding the main threshold , and assign a risk level label according to the score interval it belongs to. Combining the score of node 1 in the above example as , which is in the medium-risk interval, so it is labeled as "medium risk". If the score of node 2 is 0.125 and the score of node 3 is 0.060, the corresponding labels are "high risk" and "low risk" respectively. Finally, bind the risk level information of all nodes with the node identifier and output it as a set of risk early warning result structures. Example: Node 1 (transfer warehouse): medium risk, Node 2 (city distribution hub): high risk, Node 3 (terminal distribution point): low risk. This structure can be written into the fulfillment process tracking record by the system log recording module, or directly trigger downstream scheduling optimization actions by the risk early warning engine to realize the early identification and dynamic adjustment of hidden dangers such as timing imbalance, fulfillment deviation and abnormal node pressure. The entire classification and output process ensures the closed-loop application of the previous disturbance modeling, feature extraction and response calculation links.

[0038] An AI-based supply chain risk warning system, which is used to execute the above-mentioned AI-based supply chain risk warning method. The system includes: The performance node analysis module obtains the performance time series through the planned time and completion time of the logistics order nodes, and performs staggered trigger point identification processing on the performance status differences to obtain the stagger degree sequence; The jump vector generation module constructs the performance completion order according to the node completion time in the performance time series, and generates a jump vector by comparing the positions with the node numbers; The perturbation feature extraction module inputs the stagger degree sequence and the jump vector into the embedding layer of the Transformer model for extracting non-periodic performance perturbation features, and obtains a non-periodic performance perturbation feature set; The density mutation identification module obtains the task distribution according to the node completion time in the performance time series, constructs a task pressure time window view for density mutation identification, screens the nodes with density mutation, and forms a density mutation node set; The risk warning calculation module inputs the density mutation node set and the non-periodic performance perturbation feature set into the decoding layer of the Transformer model, calculates the abnormal risk score of the density mutation nodes, and generates a supply chain logistics performance risk warning result according to the score.

[0039] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical content of the technical solution of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. An AI-based supply chain risk early warning method, characterized in that, Including the following steps: S1: Obtain the fulfillment time series through the planned time and completion time of the logistics order nodes, and perform staggered trigger point identification processing on the fulfillment status differences to obtain the staggering degree sequence; S2: Construct the fulfillment completion order according to the node completion time in the fulfillment time series, and generate a jump vector by comparing the positions with the node numbers; S3: Input the staggering degree sequence and the jump vector into the embedding layer of the Transformer model for non-periodic fulfillment perturbation feature extraction to obtain the non-periodic fulfillment perturbation feature set; S4: Obtain the task distribution according to the node completion time in the fulfillment time series, construct a task pressure time window view for density mutation identification, screen the nodes with density mutation, and form a density mutation node set; S5: Input the density mutation node set and the non-periodic fulfillment perturbation feature set into the decoding layer of the Transformer model, calculate the abnormal risk score of the density mutation nodes, and generate a supply chain logistics fulfillment risk warning result according to the score.

2. The AI-based supply chain risk warning method according to claim 1, wherein The staggering degree sequence includes the rhythm inversion position, the sign change node, and the execution interval fluctuation section. The jump vector includes the sequential offset, the reverse execution node number, and the jump node spacing value. The non-periodic fulfillment perturbation feature set includes the multi-channel time alignment vector, the attention focus weight, and the fulfillment perturbation pattern encoding. The density mutation node set includes the task concentration mutation point, the local pressure change section, and the periodic density abnormal window. The supply chain logistics fulfillment risk warning result includes the risk node number, the abnormal risk score, and the node risk level identifier.

3. The AI-based supply chain risk warning method according to claim 1, wherein The specific steps for obtaining the staggering degree sequence are as follows: S111: Obtain the planned completion time and the actual completion time of each fulfillment node in the supply chain order, calculate the difference between the planned continuation interval and the execution interval between adjacent nodes in the fulfillment time series composed of the two types of time data to obtain the planned-actual interval difference sequence; S112: Based on the planned-actual interval difference sequence, obtain the positive and negative signs of the differences between each node, and make a continuity judgment according to the positive and negative sign change status to obtain the staggered trigger point set; S113: Based on the staggered trigger point set, mark and count the trigger positions corresponding to the fulfillment nodes in the original order to obtain the staggered degree sequence representing the staggered degree value sequence as the rhythm perturbation degree.

4. The AI-based supply chain risk warning method according to claim 3, wherein The specific steps for obtaining the jump vector are as follows: S211: Obtain the actual completion time of the nodes in the fulfillment time series, sort each fulfillment node in ascending order according to the completion time, and construct a fulfillment completion order sequence reflecting the actual fulfillment order; S212: Compare the positions of the nodes in the fulfillment completion order sequence and the original number order of the fulfillment nodes, calculate the difference in the positions of the corresponding nodes in the two sequences, and judge the offset direction and span size according to the difference in positions to identify the jump nodes and reverse nodes, and generate a jump vector representing the node order perturbation.

5. The AI-based supply chain risk warning method according to claim 4, wherein, The specific steps for obtaining the non-periodic fulfillment perturbation feature set are as follows: S311: Based on the jump vector, the interleaving degree sequence, and the fulfillment time series of the original order nodes, perform splicing processing on the three types of data in the dimension of the fulfillment node position to construct a composite input vector with a unified structure, and obtain a multi-channel fulfillment expression sequence; S312: Input the multi-channel fulfillment expression sequence into the embedding layer of the Transformer model, and call the position encoding mechanism to perform vector-level embedding processing on the node order information to form an encoded input representation with a time series structure; S313: Based on the encoded input representation, call the multi-head attention mechanism of the Transformer model to compare the expression content between channels, extract cross-channel cross features, and obtain an aperiodic fulfillment perturbation feature set including node perturbation, rhythm interleaving, and time deviation characteristics.

6. The AI-based supply chain risk warning method according to claim 5, wherein The specific steps for obtaining the density mutation node set are as follows: S411: Obtain the completion time of each fulfillment node in the fulfillment time series as the arrival time data of the node task, and divide the arrival time data into discontinuous time periods according to the minimum completion time interval between tasks to construct a set of local time windows covering the entire fulfillment cycle; S412: Based on the set of local time windows, use the kernel density estimation algorithm to calculate the task arrival density within each local time window, and sequentially generate density vectors reflecting the change trend of the task distribution; S413: Based on the density vectors, calculate the difference with the average density value of the corresponding position window in the cycle record of the fulfillment node, identify the window positions where the density change amplitude exceeds the amplitude threshold range, extract the corresponding nodes, and form a density mutation node set.

7. The AI-based supply chain risk warning method according to claim 6, wherein The specific steps for obtaining the supply chain logistics fulfillment risk warning result are as follows: S511: Based on the density mutation node set and the aperiodic fulfillment perturbation feature set, perform feature splicing processing on the two types of data in the dimension of the fulfillment node to construct a fusion feature sequence including density change information and fulfillment perturbation features; S512: Input the fusion feature sequence into the decoding layer of the Transformer model, call the multi-head attention mechanism to calculate the attention weights of the fusion features between nodes in the fusion feature sequence, extract the focus response values of each node under the background of multi-dimensional perturbation and density, and output a risk score representing the abnormal degree of the node; S513: Based on the risk score, perform classification judgment according to the set risk judgment threshold, screen the nodes with scores exceeding the threshold and assign corresponding level identifiers to generate a supply chain logistics fulfillment risk warning result, which is used to reveal the risk hidden dangers of time series imbalance, execution deviation, and node load abnormality existing in the fulfillment process.

8. An AI-based supply chain risk early warning system, characterized in that, According to the AI-based supply chain risk warning method described in any one of claims 1-7, the system includes: The fulfillment node parsing module obtains the fulfillment time series through the planned time and completion time of the logistics order nodes, and performs interleaving trigger point identification processing on the fulfillment status differences to obtain the interleaving degree sequence; The jump vector generation module constructs the fulfillment completion order according to the node completion time in the fulfillment time series, and generates a jump vector by comparing the positions with the node numbers; The perturbation feature extraction module inputs the interleaving degree sequence and the jump vector into the embedding layer of the Transformer model for non-periodic performance perturbation feature extraction, obtaining a non-periodic performance perturbation feature set; The density mutation identification module obtains the task distribution according to the node completion time in the performance time sequence, constructs a task pressure time window view for density mutation identification, screens the nodes with density mutation, and forms a density mutation node set; The risk warning calculation module inputs the density mutation node set and the non-periodic performance perturbation feature set into the decoding layer of the Transformer model, calculates the abnormal risk score of the density mutation nodes, and generates a supply chain logistics performance risk warning result according to the score.

Images