A timing diagram-based industry chain risk identification method and system

Abstract

The application discloses a kind of industry chain risk identification method and system based on timing diagram, it is related to industry chain risk identification field.Acquire industry chain data and construct risk factor graph with time series information, obtain risk factor timing diagram;Acquire industry chain risk data;Acquire industry chain state sequence and industry chain state transition diagram, and according to industry chain state sequence and industry chain state transition diagram, risk factor timing diagram is extracted, and risk factor timing sequence data is obtained;Industry chain risk data and risk factor timing sequence data are respectively carried out static embedding and dynamic embedding, and static embedding vector and dynamic embedding vector are obtained respectively;Dynamic embedding vector and risk factor timing diagram are carried out joint information embedding, and the result vector set of joint information embedding is obtained;Static embedding vector and the result vector set of joint information embedding are carried out information fusion based on attention mechanism, and the fusion result is obtained;Industry chain risk is finally identified and predicted.

Description

Technical Field

[0001] This invention relates to the field of supply chain risk identification, and in particular to a supply chain risk identification method and system based on time sequence diagrams. Background Technology

[0002] Supply chain risks are influenced by numerous disruptive factors, and their content, modes of impact, and consequences can change in complex ways as the supply chain is structured, operated, upgraded, and reshaped. Therefore, accurately perceiving and identifying these diverse and complexly evolving supply chain risks is an important scientific challenge.

[0003] Existing research on supply chain risk assessment methods is relatively simple, with two popular techniques. The risk logic tree method is a qualitative analysis approach. It starts with the outcome, breaks down the parent risk concept into different sub-risk concepts, and then works backward from the causes of the risk outcome to construct a tree-like framework for risk reasoning. The other commonly used risk identification method is the Analytic Hierarchy Process (AHP). AHP is a comprehensive evaluation method that combines qualitative and quantitative characteristics and is widely used in risk assessment in specific fields. However, these two existing methods are mainly based on the economic field and rely on mathematical tools for analysis. This results in an evaluation system that only considers economics, exhibiting a certain degree of bias and failing to fully utilize the rich information in supply chain risk data. Furthermore, due to the primary use of mathematical methods, existing techniques cannot effectively model and analyze supply chain risks under large amounts of complex and interference-laden data, hindering accurate risk identification. In today's big data era, machine learning and deep learning methods are increasingly widely used. However, in the field of supply chain risk identification, modern artificial intelligence methods and techniques are still in their infancy, lacking comprehensive and mature technical methods for supply chain risk identification. There is an urgent need to propose new artificial intelligence methods and theories for supply chain risk identification. Summary of the Invention

[0004] The problem this invention aims to solve is to address the shortcomings of the prior art by providing a supply chain risk identification method and system based on time sequence diagrams, thereby resolving the issue that existing methods cannot fully utilize the rich information in supply chain risk data, resulting in an inability to accurately identify supply chain risks.

[0005] To achieve the above-mentioned objectives of the present invention, a first aspect of the present invention provides a supply chain risk identification method based on time sequence diagrams, the method comprising the following steps:

[0006] Step 1: Obtain industry chain data, and then construct a risk factor graph with time series information based on the industry chain data, thereby obtaining a risk factor time series graph; the risk factor graph is a directed acyclic graph, where nodes represent risk factors existing in the industry chain, and the edge relationships between nodes represent the relationships between risk factors; the risk factor time series graph is composed of risk factor graphs under different timestamps;

[0007] Step 2: Obtain supply chain risk data as the data source for supply chain risk identification; the supply chain risk data refers to data containing risk factors that exist in the actual supply chain.

[0008] Step 3: Obtain the industrial chain state sequence and industrial chain state transition diagram, and extract the risk factor time series diagram based on the industrial chain state sequence and industrial chain state transition diagram to obtain risk factor time series data; the industrial chain state transition diagram is a state transition diagram based on the state changes in the upstream and downstream when the industrial chain is operating normally; the industrial chain state sequence is a record of the state transitions of a segment of industrial chain operation;

[0009] Step 4: Statically embed the supply chain risk data to obtain a static embedding vector; dynamically embed the time series data of risk factors to obtain a dynamic embedding vector; the static embedding is embedding without considering the time factor; the dynamic embedding is embedding that considers the time factor.

[0010] Step 5: Perform joint information embedding on the dynamic embedding vector and the risk factor time series diagram to obtain a set of result vectors for joint information embedding; the joint information embedding is the embedding of different information after fusion.

[0011] Step 6: Perform attention-based information fusion on the static embedding vector and the result vector set of joint information embedding to obtain the fusion result;

[0012] Step 7: Based on the integration results, conduct the final identification and prediction of supply chain risks.

[0013] Furthermore, according to the aforementioned supply chain risk identification method based on time sequence diagrams, step 1 includes the following specific steps:

[0014] Step 1.1: Acquire supply chain data and perform preliminary processing on the supply chain data to obtain preliminary processed supply chain data including time series information and risk factors;

[0015] Step 1.2: Perform unified processing on the time series information in the pre-processed industry chain data to obtain unified time series information;

[0016] Step 1.3: Based on the unified time series information and risk factors in the industrial chain data, establish a risk factor diagram, and then obtain a risk factor time series diagram.

[0017] Furthermore, according to the aforementioned method for identifying supply chain risks based on time-series diagrams, step 5 includes the following specific steps:

[0018] Step 5.1: Based on the dynamic embedding vector, perform a random walk on the risk factor time series graph to obtain a subgraph of the risk factor time series graph, which is simply referred to as the risk factor time series subgraph;

[0019] Step 5.2: Concatenate the time series subgraph of risk factors and the dynamic embedding vector to obtain the concatenation result, and then obtain the concatenation result set;

[0020] Step 5.3: Pre-train the BERT language model to obtain the trained BERT language model and use the model to embed the concatenated result set to obtain the result vector of joint information embedding, and then obtain the result vector set of joint information embedding.

[0021] Furthermore, according to the aforementioned supply chain risk identification method based on time series diagrams, the preliminary processing method is as follows: The acquired supply chain data is cleaned to remove noise generated by irrelevant information, retaining only time series information and risk factors; then, by sampling ten data points near the missing data, the average value of these ten data points is used to fill in the missing data, thus completing the missing data. Missing data with excessive missing information that is difficult to fill in is deleted as abnormal data; finally, the supply chain data is standardized and normalized to complete the preliminary processing of the supply chain data, resulting in the pre-processed supply chain data.

[0022] Furthermore, according to the aforementioned method for identifying supply chain risks based on time series diagrams, the method for unifying the time series information is as follows: The pre-processed supply chain data is statistically analyzed according to time granularity. Specifically, the time series information is first divided into four time granularities: month, day, hour, and minute. Then, the time series information is classified and stored according to its different time granularities. Finally, the time series information of the four time granularities is statistically analyzed separately. From the statistical results, time series information with a quantity reaching a set value is selected, and the smallest time granularity is found as the unified granularity of the time series information. Time series information with a time granularity smaller than the unified granularity is integrated into the time series information with a unified granularity. Time series information with a time granularity larger than the unified granularity is not used.

[0023] A second aspect of the present invention provides a supply chain risk identification system based on time sequence diagrams, comprising:

[0024] The risk factor time series diagram construction module is used to acquire industry chain data and construct a risk factor time series diagram based on the industry chain data; and send the risk factor time series diagram to the risk factor time series sequence data extraction module and the joint embedding module.

[0025] The risk factor time series data extraction module is used to receive the risk factor time series diagram sent by the risk factor time series diagram construction module and extract the risk factor time series diagram to obtain risk factor time series data; and send the risk factor time series data to the dynamic embedding module.

[0026] The static embedding module is used to acquire supply chain risk data and embed the data to obtain static embedding vectors; the static embedding vectors are then sent to the attention-based fusion module.

[0027] The dynamic embedding module is used to receive the risk factor time series data sent by the risk factor time series data extraction module, perform embedding processing on the risk factor time series data to obtain a dynamic embedding vector, and send the dynamic embedding vector to the joint embedding module.

[0028] The joint embedding module receives the dynamic embedding vector sent by the dynamic embedding module and the risk factor time series graph sent by the risk factor time series graph construction module; it performs joint information embedding on the dynamic embedding vector and the risk factor time series graph to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the attention-based fusion module.

[0029] The attention-based fusion module receives the joint information embedding result vector sent by the joint embedding module and the static embedding vector sent by the static embedding module; performs attention-based information fusion on the joint information embedding result vector and the static embedding vector to obtain the fusion result and sends the fusion result to the risk factor prediction module.

[0030] The risk factor prediction module receives the fusion results sent by the attention-based fusion module, and performs the final identification and prediction of supply chain risks based on the fusion results to obtain the supply chain risk prediction results.

[0031] Furthermore, according to the aforementioned supply chain risk identification system based on time-series diagrams, the risk factor time-series diagram construction module further includes:

[0032] The data processing module is used to acquire industry chain data and perform data cleaning, missing data completion, time series information processing, and risk factor relationship calculation to obtain a preliminary risk factor map; the preliminary risk factor map is then sent to the knowledge mapping module.

[0033] The knowledge mapping module receives the preliminary risk factor map sent by the data processing module, and performs knowledge mapping between the preliminary risk factor map and the professional industrial chain risk knowledge graph to construct a risk factor time series map; it then sends the risk factor time series map to the risk factor time series sequence data extraction module and the subgraph extraction module.

[0034] Furthermore, according to the aforementioned supply chain risk identification system based on time sequence diagrams, the joint embedding module further includes:

[0035] The subgraph extraction module is used to receive the risk factor time series diagram sent by the knowledge mapping module and extract the risk factor time series subgraph from the risk factor time series diagram; and send the risk factor time series subgraph to the embedding module.

[0036] The embedding module receives the risk factor time series subgraph sent by the subgraph extraction module and the dynamic embedding vector sent by the dynamic embedding module; it concatenates the risk factor time series subgraph and the dynamic embedding vector and performs joint information embedding through the BERT language model to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the GRU network module.

[0037] Furthermore, according to the aforementioned supply chain risk identification system based on time-series graphs, the attention-based fusion module further includes:

[0038] The GRU network module receives the joint information embedding result vector sent by the embedding module, performs time series information perception on the joint information embedding result vector, obtains the output result of the GRU network, and sends the output result of the GRU network to the attention mechanism module.

[0039] The attention mechanism module receives the output of the GRU network sent by the GRU network module and the static embedding vector sent by the static embedding module. It then uses the attention mechanism to fuse the output of the GRU network and the static embedding vector to obtain the fusion result. Finally, it sends the information fusion result to the risk factor prediction module.

[0040] Compared with the prior art, the technical solution adopted in this invention has the following technical effects:

[0041] The method and system of this invention consider the characteristics of supply chain data changing over time and introduce a risk factor time series graph. It fully learns the time series information and risk factors on the risk factor time series graph. It performs preliminary embedding of supply chain risk data and risk factor time series data through static embedding and dynamic embedding, respectively. Through joint information embedding, it fully integrates the risk factor time series subgraph and the risk factor time series data. It can accurately perceive and identify diverse supply chain risk data with complex evolution mechanisms. Furthermore, through an attention mechanism, it adjusts the attention weight of the supply chain risk data and the result vector after joint information embedding, thereby making full use of existing supply chain risk data and achieving accurate prediction of supply chain risks. Attached Figure Description

[0042] Figure 1 This is a structural diagram of a supply chain risk identification system based on a time sequence diagram, as described in an embodiment of the present invention.

[0043] Figure 2 This is a flowchart of a supply chain risk identification method based on time sequence diagrams in an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram illustrating the construction of a risk factor time series diagram in an embodiment of the present invention;

[0045] Figure 4 (a) is a schematic diagram of static embedding in an embodiment of the present invention; (b) is a schematic diagram of dynamic embedding in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the embedding of joint information in an embodiment of the present invention;

[0047] Figure 6 This is a schematic diagram of information fusion based on an attention mechanism in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram illustrating risk factor prediction in an embodiment of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0050] Figure 1 The system shown is a supply chain risk identification system based on time sequence diagrams. The system includes:

[0051] The risk factor time series diagram construction module is used to acquire industry chain data and construct a risk factor time series diagram based on the industry chain data; and send the risk factor time series diagram to the risk factor time series sequence data extraction module and the joint embedding module.

[0052] The risk factor time series data extraction module is used to receive the risk factor time series diagram sent by the risk factor time series diagram construction module and extract the risk factor time series diagram to obtain risk factor time series data; and send the risk factor time series data to the dynamic embedding module.

[0053] The static embedding module is used to acquire supply chain risk data and embed the data to obtain static embedding vectors; the static embedding vectors are then sent to the attention-based fusion module.

[0054] The dynamic embedding module is used to receive the risk factor time series data sent by the risk factor time series data extraction module, perform embedding processing on the risk factor time series data to obtain a dynamic embedding vector, and send the dynamic embedding vector to the joint embedding module.

[0055] The joint embedding module receives the dynamic embedding vector sent by the dynamic embedding module and the risk factor time series graph sent by the risk factor time series graph construction module; it performs joint information embedding on the dynamic embedding vector and the risk factor time series graph to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the attention-based fusion module.

[0056] The attention-based fusion module receives the joint information embedding result vector sent by the joint embedding module and the static embedding vector sent by the static embedding module; performs attention-based information fusion on the joint information embedding result vector and the static embedding vector to obtain the fusion result and sends the fusion result to the risk factor prediction module.

[0057] The risk factor prediction module is used to receive the fusion results sent by the attention-based fusion module, and to perform the final identification and prediction of the industrial chain risks based on the fusion results, so as to obtain the industrial chain risk prediction results.

[0058] Furthermore, the risk factor time series diagram construction module also includes:

[0059] The data processing module is used to acquire industry chain data and perform data cleaning, missing data completion, time series information processing, and risk factor relationship calculation to obtain a preliminary risk factor map; the preliminary risk factor map is then sent to the knowledge mapping module.

[0060] The knowledge mapping module receives the preliminary risk factor diagram sent by the data processing module, and performs knowledge mapping between the preliminary risk factor diagram and the professional industrial chain risk knowledge graph to construct a risk factor time series diagram; it then sends the risk factor time series diagram to the risk factor time series sequence data extraction module and the subgraph extraction module.

[0061] Furthermore, the joint embedding module also includes:

[0062] The subgraph extraction module is used to receive the risk factor time series diagram sent by the knowledge mapping module and extract the risk factor time series subgraph from the risk factor time series diagram; and send the risk factor time series subgraph to the embedding module.

[0063] The embedding module receives the risk factor time series subgraph sent by the subgraph extraction module and the dynamic embedding vector sent by the dynamic embedding module; it concatenates the risk factor time series subgraph and the dynamic embedding vector and performs joint information embedding through the BERT language model to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the GRU network module.

[0064] Furthermore, the attention-based fusion module also includes:

[0065] The GRU network module receives the joint information embedding result vector sent by the embedding module, performs time series information perception on the joint information embedding result vector, obtains the output result of the GRU network, and sends the output result of the GRU network to the attention mechanism module.

[0066] The attention mechanism module receives the output of the GRU network sent by the GRU network module and the static embedding vector sent by the static embedding module. It then uses the attention mechanism to fuse the output of the GRU network and the static embedding vector to obtain the fusion result. Finally, it sends the information fusion result to the risk factor prediction module.

[0067] This embodiment provides a supply chain risk identification method based on time sequence diagrams, such as... Figure 2 As shown, the method includes the following steps:

[0068] Step 1: Obtain industry chain data, and then construct a risk factor graph with time series information based on the industry chain data to obtain a risk factor time series graph; the risk factor graph is a directed acyclic graph, where nodes represent risk factors existing in the industry chain, and the edge relationships between nodes represent the relationships between risk factors; the risk factor time series graph is composed of risk factor graphs under different timestamps.

[0069] The method for constructing a risk factor time series diagram in this embodiment is as follows: Figure 3 As shown, it includes:

[0070] Step 1.1: Acquire supply chain data and perform preliminary processing on the supply chain data to obtain preliminary processed supply chain data including time series information and risk factors;

[0071] The preliminary processing method is as follows: First, the acquired industry chain data is cleaned to remove noise caused by irrelevant information, retaining only time series information and risk factors. Then, the missing data is filled in by sampling the ten nearest data points and averaging these ten nearest data points. Missing data with excessive missing information that is difficult to fill in is deleted as outlier data. Finally, the industry chain data is standardized and normalized to complete the preliminary processing, resulting in pre-processed industry chain data.

[0072] Step 1.2: Perform unified processing on the time series information in the pre-processed industry chain data to obtain unified time series information;

[0073] The method for unifying time series information is as follows: After preliminary processing, the industry chain data is statistically analyzed according to time granularity. Specifically, the time series information is first divided into four time granularities: month, day, hour, and minute. Then, the time series information is categorized and stored according to its different time granularities. Finally, the time series information at each of the four time granularities is statistically analyzed. From the statistical results, time series information with a quantity reaching a set value is selected, and the smallest time granularity is identified as the unified granularity. Time series information with a time granularity smaller than the unified granularity is integrated into the time series information with the unified granularity. Time series information with a time granularity larger than the unified granularity is not used. In this embodiment, the unified time granularity is day. Time series information with a time granularity of hour and minute is integrated into the time series information with a time granularity of day, resulting in all time series information with a time granularity of day. Time series information with a time granularity of month is not used.

[0074] Step 1.3: Based on the unified time series information and risk factors in the industrial chain data, establish a risk factor diagram, and then obtain a risk factor time series diagram;

[0075] The method for establishing a risk factor graph and then obtaining a risk factor time series graph is as follows: Based on the risk factors in the unified time series information and industrial chain data, the risk factors corresponding to each time series information are taken as nodes in the risk factor graph of that time stamp. The relationship between the nodes in the risk factor graph of that time stamp is calculated to obtain the edges connecting these nodes, forming a preliminary risk factor graph. Then, the preliminary risk factor graph is mapped with a professional industrial chain risk knowledge graph to add actual industrial chain risks to the preliminary risk factor graph, resulting in a final risk factor graph. The risk factor graphs of different time stamps are arranged in chronological order to obtain the risk factor time series graph G.

[0076] In this embodiment, based on the unified time series information and risk factors in the supply chain data, the risk factors corresponding to different dates are used as nodes in the risk factor graph for that date. The Point Mutual Information (PMI) algorithm is used to calculate the correlation and interdependence between two nodes, obtaining the edges connecting these nodes to form a preliminary risk factor graph. Then, the preliminary risk factor graph is mapped to a professional supply chain risk knowledge graph. Specifically, the preliminary risk factor graph and the supply chain risk knowledge base are mapped to a vector space using an embedding method. The similarity between the preliminary risk factor graph and the information in the supply chain risk knowledge base is calculated, and those with a similarity greater than 0.6 are used for knowledge mapping. The preliminary risk factor graph is mapped to the final risk factor graph through the supply chain risk knowledge graph. The risk factor graphs of different dates are arranged in chronological order to obtain the risk factor time series graph G.

[0077] Step 2: Obtain supply chain risk data as the data source for supply chain risk identification; the supply chain risk data refers to data containing risk factors that exist in the actual supply chain.

[0078] Step 3: Obtain the industrial chain state sequence and industrial chain state transition diagram, and extract the risk factor time series diagram based on the industrial chain state sequence and industrial chain state transition diagram to obtain risk factor time series data; the industrial chain state transition diagram is a state transition diagram based on the state changes in the upstream and downstream when the industrial chain is operating normally; the industrial chain state sequence is a record of the state transitions of a segment of industrial chain operation;

[0079] The method for extracting risk factor time series data from the risk factor time series diagram is as follows: the industrial chain state sequence is compared with the industrial chain state transition diagram, that is, the actual state of industrial chain operation is compared with the standard state of industrial chain operation. If there are subsequences in the industrial chain state sequence that do not match the subsequences in the industrial chain state transition diagram, then these subsequences in the industrial chain state sequence have generated state transition deviations. The subsequences that appear multiple times in these subsequences that have generated state transition deviations are taken as the extraction range. Data that is the same as the subsequences in the extraction range is extracted from the risk factor time series diagram, thereby obtaining the risk factor time series data.

[0080] Step 4: Statically embed the supply chain risk data to obtain a static embedding vector; dynamically embed the time series data of risk factors to obtain a dynamic embedding vector; the static embedding is embedding without considering the time factor; the dynamic embedding is embedding that considers the time factor.

[0081] like Figure 4As shown, supply chain risk data is the risk relationship data between supply chains. It is static data in the time series information changes. Therefore, static embedding without time attributes is used to embed the supply chain risk data. In this embodiment, the static embedding method is as follows: use the Word2Vec algorithm to perform word embedding on the supply chain risk data, and map the supply chain risk data to a vector space. Words and phrases with similar semantics and grammar in the supply chain risk data will be closer in the space. The static embedding vector is obtained and denoted as S0. The subscript 0 indicates that the static embedding method only performs embedding once.

[0082] Risk factor time series data contains a large amount of time series information. Therefore, dynamic embedding considering time factors is performed on the risk factor time series data. In this embodiment, word2vec word embedding is performed on the risk factor time series data at each timestamp, mapping the risk factor time series data to a vector space. Words and phrases with similar semantics and grammar will be closer in the space. Several dynamic embeddings are performed according to the above method, and the dynamic embedding vector obtained from each embedding is denoted as D. i , where i∈(1,n), i represents a dynamic embedding in the nth dynamic embedding process, and n is the number of dynamic embeddings.

[0083] Step 5: Perform joint information embedding on the dynamic embedding vector and the risk factor time series diagram to obtain a set of result vectors for joint information embedding; the joint information embedding is the embedding after fusing different information.

[0084] like Figure 5 As shown, the dynamically embedded vector D i The joint information embedding is performed with the risk factor time series diagram G. The dynamic embedding is performed n times in step 4. Therefore, the joint information embedding in this embodiment will also be performed n times.

[0085] Specifically, the steps include the following:

[0086] Step 5.1: Based on the dynamic embedding vector, perform a random walk on the risk factor time series graph to obtain a subgraph of the risk factor time series graph, which is simply referred to as the risk factor time series subgraph;

[0087] In this embodiment, the embedding vector of each node in the risk factor time series graph G is first calculated and compared with the dynamic embedding vector D. i By comparison, we can find the embedding vector and the dynamic embedding vector D. i The node in the risk factor time series graph G with the highest similarity is used to perform a random walk. In this implementation, the number of hops is 3. The subgraph of the risk factor time series graph is obtained through the random walk, which is simply referred to as the risk factor time series subgraph. The risk factor time series subgraph corresponding to a certain time series information is denoted as G. iPerform n random walks in the manner described above.

[0088] Step 5.2: Concatenate the time series subgraph of risk factors and the dynamic embedding vector to obtain the concatenation result, and then obtain the concatenation result set;

[0089] In this embodiment, the risk factor time series subgraph G is used. i and dynamic embedding vector D i The splicing method is shown in formula (1):

[0090]

[0091] in Let Ci be the concatenation symbol, and let Ci be the concatenation result. A total of n concatenation operations are performed, and the result is denoted as the concatenation result set C, as shown in formula (2):

[0092] C = (C1,...,C) i ,...,Cn),i∈(1,n) (2)

[0093] Step 5.3: Pre-train the BERT language model to obtain the trained BERT language model and use the model to embed the concatenated result set to obtain the result vector of joint information embedding, and then obtain the result vector set of joint information embedding.

[0094] In this embodiment, the publicly available pre-training parameters are first obtained, followed by the acquisition of a larger corpus of supply chain risks. The BERT language model is then fine-tuned using this corpus to enhance its generalization and cognitive abilities regarding supply chain risks. This results in a trained BERT language model, which is then used to embed the splicing result set C, fully capturing the contextual information of the time series. The set of vectors resulting from the joint information embedding is denoted as H, as shown in formulas (3) and (4).

[0095] H = BERT(C) (3)

[0096] H = (H1,...,H) i ,...,Hn),i∈(1,n) (4)

[0097] Among them, H i The result vector is the result of embedding the joint information corresponding to the input concatenation result Ci.

[0098] By capturing the changes in supply chain risk factors over time and identifying the risk relationships between time-series subplots of risk factors within the risk factor time-series graph, more information can be provided for subsequent supply chain risk prediction. It is worth noting that during the embedding of the concatenated result set C using the BERT language model, the embedding effect is significantly influenced by whether the BERT language model is pre-trained on a relevant corpus.

[0099] Step 6: Perform attention-based information fusion on the static embedding vector and the result vector set of joint information embedding to obtain the fusion result;

[0100] In this embodiment, as follows Figure 6 As shown, the result vector set H of the joint information embedding is first input into the recurrent neural network in the order of time series information. The recurrent neural network used in this embodiment is the GRU network. The GRU network uses update gate and reset gate to reduce training parameters and improve the calculation speed to a certain extent. Since the result vector set H of the joint information embedding has clear unidirectional time series information, the update gate and reset gate of the GRU network can preserve the time series information for a long time and will not be cleared over time or removed because it is irrelevant to the prediction. Therefore, the GRU network is used to further process the long-term time series information in the industrial chain risk time series information, accurately perceive the changes in time series information, and obtain the output result of the GRU network, as shown in formula (5):

[0101] V1,...,V i ,…,V n =GRU(H1,...,H) i ,…,H n (5)

[0102] Where V i H is the result vector set H embedded by the joint information. i The output of the GRU network is obtained by using it as input to the GRU recurrent neural network.

[0103] Then, the attention mechanism is used to process the output V of the GRU network. i The fusion result O is obtained by fusing the static embedding vector S0 with the static embedding vector S0. V This allows for the full capture of dynamic and static information regarding risks in the industrial chain, enabling precise perception of these risks and enhancing the ability to perceive and identify them.

[0104] The fusion method is shown in formula (6). The attention mechanism automatically learns the importance of different parts of the data in the industrial chain, captures all effective information from the industrial chain risk data and the time series data of risk factors, and performs attention weight allocation.

[0105] O v =attention(S0,V1,...,V) i ,…,V n (6)

[0106] Here, α is the attention weight, which is assigned based on the importance of the supply chain risk data and the supply chain risk time series data sequence information. The information fusion of supply chain risk data and risk factor time series is performed based on the attention weight.

[0107] The attention weights are in the form of the following formula (7):

[0108] α=(α0,α1,...,α i ,...,α n (7)

[0109] Where α0 is the attention weight of S0, α i For V i The attention weights are i∈(1,n).

[0110] Step 7: Based on the fusion results, conduct the final identification and prediction of supply chain risks;

[0111] like Figure 7 As shown, the fusion result O v The input is fed into the risk factor prediction layer, which includes a multilayer perceptron and a softmax activation function. The output results are then distributed according to probability, transforming the problem of identifying risk factors in the industrial chain into a multi-classification problem. The probability of occurrence of different industrial chain risks is obtained, and the five industrial chain risks with the highest probability of occurrence are selected and output in order of probability. The probability of occurrence of industrial chain risks is used for prediction.

Claims

1. A supply chain risk identification method based on time sequence diagrams, characterized in that, Includes the following steps: Step 1: Obtain industry chain data, and then construct a risk factor graph with time series information based on the industry chain data to obtain a risk factor time series graph; the risk factor graph is a directed acyclic graph, where nodes represent risk factors existing in the industry chain, and the edge relationships between nodes represent the relationships between risk factors; The risk factor time series diagram is composed of risk factor diagrams at different timestamps; Step 2: Obtain supply chain risk data as the data source for supply chain risk identification; the supply chain risk data refers to data containing risk factors that exist in the actual supply chain. Step 3: Obtain the industrial chain state sequence and industrial chain state transition diagram, and extract the risk factor time series diagram based on the industrial chain state sequence and industrial chain state transition diagram to obtain risk factor time series data; the industrial chain state transition diagram is a state transition diagram based on the state changes in the upstream and downstream when the industrial chain is operating normally; the industrial chain state sequence is a record of the state transitions of a segment of industrial chain operation; Step 4: Statically embed the supply chain risk data to obtain a static embedding vector; dynamically embed the time series data of risk factors to obtain a dynamic embedding vector; the static embedding is embedding without considering the time factor; the dynamic embedding is embedding that considers the time factor. Step 5: Perform joint information embedding on the dynamic embedding vector and the risk factor time series diagram to obtain a set of result vectors for joint information embedding; the joint information embedding is the embedding of different information after fusion. Step 5 includes the following specific steps: Step 5.1: Based on the dynamic embedding vector, perform a random walk on the risk factor time series graph to obtain a subgraph of the risk factor time series graph, which is simply referred to as the risk factor time series subgraph; Step 5.2: Concatenate the time series subgraph of risk factors and the dynamic embedding vector to obtain the concatenation result, and then obtain the concatenation result set; Step 5.3: Pre-train the BERT language model to obtain the trained BERT language model and use the model to embed the concatenated result set to obtain the result vector of joint information embedding, and then obtain the result vector set of joint information embedding. Step 6: Perform attention-based information fusion on the static embedding vector and the result vector set of joint information embedding to obtain the fusion result; Step 7: Based on the integration results, conduct the final identification and prediction of supply chain risks.

2. The supply chain risk identification method based on time sequence diagrams according to claim 1, characterized in that, Step 1 includes the following specific steps: Step 1.1: Acquire supply chain data and perform preliminary processing on the supply chain data to obtain preliminary processed supply chain data including time series information and risk factors; Step 1.2: Perform unified processing on the time series information in the pre-processed industry chain data to obtain unified time series information; Step 1.3: Based on the unified time series information and risk factors in the industrial chain data, establish a risk factor diagram, and then obtain a risk factor time series diagram.

3. The supply chain risk identification method based on time sequence diagrams according to claim 2, characterized in that, The preliminary processing method is as follows: First, the acquired industry chain data is cleaned to remove noise caused by irrelevant information, retaining only time series information and risk factors. Then, by sampling the ten data points closest to the missing data, the average value of these ten data points is used to fill in the missing data. Missing data with excessive missing information that is difficult to fill in is deleted as outlier data. Finally, the industry chain data is standardized and normalized to complete the preliminary processing, resulting in pre-processed industry chain data.

4. The supply chain risk identification method based on time sequence diagrams according to claim 2, characterized in that, The method for unifying time series information is as follows: After preliminary processing, the industry chain data is statistically analyzed according to time granularity. Specifically, the time series information is first divided into four time granularities: month, day, hour, and minute. Then, the time series information is categorized and stored according to its different time granularities. Finally, the time series information at each of the four time granularities is statistically analyzed. From the statistical results, time series information with a quantity reaching a set value is selected, and the smallest time granularity is identified as the unified granularity. Time series information with a time granularity smaller than the unified granularity is integrated into the time series information with the unified granularity. Time series information with a time granularity larger than the unified granularity is not used.

5. A supply chain risk identification system based on time sequence diagrams, characterized in that, include: The risk factor time series diagram construction module is used to acquire industry chain data and construct a risk factor time series diagram based on the industry chain data. Send the risk factor time series diagram to the risk factor time series sequence data extraction module and the joint embedding module; The risk factor time series data extraction module is used to receive the risk factor time series diagram sent by the risk factor time series diagram construction module, and extract the risk factor time series diagram according to the industrial chain state sequence and industrial chain state transition diagram to obtain risk factor time series data. Send the time series data of risk factors to the dynamic embedding module; The static embedding module is used to acquire supply chain risk data and embed the supply chain risk data to obtain static embedding vectors. Send the static embedding vector to the attention-based fusion module; The dynamic embedding module is used to receive the risk factor time series data sent by the risk factor time series data extraction module, perform embedding processing on the risk factor time series data to obtain a dynamic embedding vector, and send the dynamic embedding vector to the joint embedding module. The joint embedding module receives the dynamic embedding vector sent by the dynamic embedding module and the risk factor time series graph sent by the risk factor time series graph construction module; it performs joint information embedding on the dynamic embedding vector and the risk factor time series graph to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the attention-based fusion module. The joint embedding module further includes: The subgraph extraction module is used to receive the risk factor time series diagram and extract the risk factor time series subgraph from the risk factor time series diagram; and send the risk factor time series subgraph to the embedding module. The embedding module receives the risk factor time series subgraph sent by the subgraph extraction module and the dynamic embedding vector sent by the dynamic embedding module; it concatenates the risk factor time series subgraph and the dynamic embedding vector and performs joint information embedding through the BERT language model to obtain the joint information embedding result vector; and it sends the joint information embedding result vector to the GRU network module. The attention-based fusion module receives the joint information embedding result vector sent by the joint embedding module and the static embedding vector sent by the static embedding module; performs attention-based information fusion on the joint information embedding result vector and the static embedding vector to obtain the fusion result and sends the fusion result to the risk factor prediction module. The risk factor prediction module receives the fusion results sent by the attention-based fusion module, and performs the final identification and prediction of supply chain risks based on the fusion results to obtain the supply chain risk prediction results.

6. The supply chain risk identification system based on time sequence diagrams according to claim 5, characterized in that, The risk factor time series diagram construction module further includes: The data processing module is used to acquire industry chain data and perform data cleaning, missing data completion, time series information processing, and risk factor relationship calculation to obtain a preliminary risk factor map; the preliminary risk factor map is then sent to the knowledge mapping module. The knowledge mapping module receives the preliminary risk factor map sent by the data processing module, and performs knowledge mapping between the preliminary risk factor map and the professional industrial chain risk knowledge graph to construct a risk factor time series map; it then sends the risk factor time series map to the risk factor time series sequence data extraction module and the subgraph extraction module.

7. A supply chain risk identification system based on time sequence diagrams according to claim 5, characterized in that, The attention-based fusion module further includes: The GRU network module receives the joint information embedding result vector sent by the embedding module, performs time series information perception on the joint information embedding result vector, obtains the output result of the GRU network, and sends the output result of the GRU network to the attention mechanism module. The attention mechanism module receives the output of the GRU network sent by the GRU network module and the static embedding vector sent by the static embedding module. It then uses the attention mechanism to fuse the output of the GRU network and the static embedding vector to obtain the fusion result. Finally, it sends the information fusion result to the risk factor prediction module.

Images