Artificial Intelligence-Based Multidimensional Fluctuation Prediction and Inventory Optimization Method for Supply Chain

Abstract

The present invention provides a multi-dimensional fluctuation prediction and inventory optimization method for supply chain based on artificial intelligence, which relates to the technical field of supply chain management. The method includes obtaining multi-dimensional data and preprocessing it to generate a standardized time series dataset and a knowledge graph, using wavelet transform for multi-scale decomposition to extract fluctuation features, combining graph attention network and multi-head self-attention mechanism to construct a state matrix, predicting the risk conduction path based on historical fluctuation pattern matching and graph convolutional network, and generating a risk prevention and control plan. The present invention can accurately identify the fluctuation features and risk conduction path in the supply chain system, and improve the accuracy of early warning and the timeliness of response.

Description

Technical Field

[0001] The present invention relates to supply chain management technology, and particularly to a multi-dimensional fluctuation prediction and inventory optimization method for supply chain based on artificial intelligence. Background Art

[0002] With the increasing complexity of the global supply chain network and the explosion of multi-dimensional data, the fluctuation risks faced by the supply chain system are constantly intensifying. Existing supply chain prediction methods mainly adopt time series analysis and statistical modeling, which cannot effectively handle data missing and abnormal problems, and have limitations in multi-dimensional data fusion, resulting in insufficient extraction of fluctuation characteristics and difficult to reflect the actual fluctuation trend of the supply chain system in the prediction results.

[0003] Existing supply chain fluctuation prediction methods often only focus on the data analysis of a single dimension, ignoring the mutual influence and correlation relationship among multiple elements in the supply chain system, resulting in insufficient accuracy and reliability of the prediction results. This single-dimensional analysis method is difficult to comprehensively reflect the overall operation state of the supply chain system and cannot effectively identify potential risk factors. Most current prediction models adopt static data processing methods, failing to fully consider the dynamic evolution characteristics of the supply chain system, and having weak early warning capabilities for emergencies and abnormal fluctuations. This static analysis method is difficult to capture the changes in the market environment in a timely manner, resulting in the problem of lag in early warning signals. Traditional inventory optimization methods lack in-depth analysis of the supply chain network structure and fail to effectively utilize advanced technologies such as knowledge graphs to mine the implicit knowledge and experience rules in the supply chain. This makes the formulation of risk prevention and control plans lack systematicness and foresight, and it is difficult to achieve precise inventory control and risk management.

[0004] Therefore, there is an urgent need for a supply chain fluctuation prediction and inventory optimization method that can integrate multi-dimensional data, make full use of artificial intelligence technology and domain knowledge, realize the accurate prediction of the fluctuation trend of the supply chain system and the accurate identification of the risk conduction path, so as to improve the risk prevention and control ability of the supply chain system. Summary of the Invention

[0005] An embodiment of the present invention provides a multi-dimensional fluctuation prediction and inventory optimization method for supply chain based on artificial intelligence, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention,

[0007] A multi-dimensional fluctuation prediction and inventory optimization method for supply chain based on artificial intelligence is provided, including:

[0008] Obtain multi-dimensional data in the supply chain system and perform preprocessing, use the piecewise linear interpolation method to supplement missing data, generate a standardized time series data set, and construct a supply chain knowledge graph based on the standardized time series data set;

[0009] Based on the standardized time-series dataset, the wavelet transform method is used to decompose the data at multiple scales, extract the trend component, periodic component, and random component in the time-series data, and calculate the node fluctuation characteristic index value;

[0010] Taking the node fluctuation characteristic index value as the input, through the knowledge graph-enhanced graph attention network for feature aggregation to obtain the fluctuation feature vector, calculating the association strength between different nodes based on the multi-head self-attention mechanism, and constructing a state matrix reflecting the overall wave dynamic trend of the supply chain based on the fluctuation feature vector and the association strength;

[0011] Performing time-series matching between the historical fluctuation pattern and the current state matrix, and outputting the risk probability distribution;

[0012] Inputting the risk probability distribution into the graph convolutional network, performing feature propagation based on the supply chain knowledge graph, extracting the local structure feature matrix in the supply chain network through multi-layer spatial convolutional operations, combining the expert rules in the supply chain knowledge graph, identifying the key path of risk conduction, and generating a risk warning signal including the conduction probability;

[0013] According to the risk warning signal, call the preset emergency response strategy library to generate the corresponding risk prevention and control plan.

[0014] In an alternative embodiment,

[0015] Obtain multi-dimensional data in the supply chain system and perform preprocessing, use the piecewise linear interpolation method to supplement missing data, generate a standardized time-series dataset, and construct a supply chain knowledge graph based on the standardized time-series dataset, including:

[0016] Obtain multi-dimensional data in the supply chain system, where the multi-dimensional data includes supplier production data, logistics transportation data, warehousing inventory data, and market demand data;

[0017] According to the multi-dimensional data, establish a feature distribution matrix, use the local density calculation method to obtain the density feature vector, calculate the local deviation value of the data point based on the density feature vector, and compare it with the adaptive threshold to identify the first-round abnormal data, and record the location information of the first-round abnormal data in the abnormal index list;

[0018] Extract the time-series data segments around the first-round abnormal data according to the abnormal index list, calculate the change trend coefficient between the data points in the time-series data segments, construct a data change feature matrix, and use the data change feature matrix to perform secondary verification on the first-round abnormal data to generate the final abnormal dataset;

[0019] Based on the location information of the final abnormal data set, divide the data range to be supplemented in the multi-dimensional data, calculate the numerical difference and time span of the data points at both ends of the data range to be supplemented, adaptively generate the supplementary data weight according to the numerical difference and time span, and use the supplementary data weight to perform piecewise linear interpolation calculation to obtain the supplementary data set;

[0020] Fill the supplementary data set into the multi-dimensional data, and perform normalization processing on the filled data to generate a standardized data matrix;

[0021] Extract supply chain entity features based on the standardized data matrix, calculate the multi-dimensional correlation degree between entities to obtain a correlation strength matrix, and construct a supply chain knowledge graph according to the correlation strength matrix, where the nodes represent supply chain entities, and the weight of the connecting edges is determined by the correlation strength matrix.

[0022] In an alternative embodiment,

[0023] Based on the standardized time series data set, perform multi-scale decomposition on the data using the wavelet transform method, extract the trend component, periodic component and random component in the time series data, and calculate the node fluctuation characteristic index values including:

[0024] Calculate the energy concentration degree and information entropy of different wavelet basis functions based on the standardized supply chain time series data set, construct a multi-objective optimization function according to the energy concentration degree and information entropy, and obtain the optimal wavelet basis function by solving the multi-objective optimization function;

[0025] Perform adaptive multi-level decomposition on the standardized supply chain time series data set, calculate the scale coefficient and wavelet coefficient at each decomposition level, reconstruct the trend component according to the scale coefficient, reconstruct the periodic component and random component according to the wavelet coefficient, and use the energy distribution characteristics of the trend component, periodic component and random component at different scales as the fluctuation characteristic index values.

[0026] In an alternative embodiment,

[0027] Calculating the energy concentration degree and information entropy of different wavelet basis functions based on the standardized supply chain time series data set, constructing a multi-objective optimization function according to the energy concentration degree and information entropy, and obtaining the optimal wavelet basis function by solving the multi-objective optimization function includes:

[0028] Construct a wavelet basis function candidate set, and perform wavelet transform on the supply chain time series data in the standardized supply chain time series data set using each wavelet basis function in the wavelet basis function candidate set to obtain the corresponding wavelet coefficients;

[0029] Calculate the energy value of each wavelet basis function based on the wavelet coefficients and perform normalization processing to obtain the energy distribution probability, and calculate the energy concentration degree of each wavelet basis function according to the energy distribution probability;

[0030] Calculate the information entropy of each wavelet basis function based on the energy distribution probability, and obtain the relative entropy feature of each wavelet basis function by using the ratio of the information entropy to the theoretical maximum entropy;

[0031] Use the wavelet coefficients and the corresponding wavelet basis functions to perform signal reconstruction to obtain a reconstructed signal, and calculate the reconstruction error between the reconstructed signal and the supply chain time series data;

[0032] Weight and combine the energy concentration degree, relative entropy feature and reconstruction error to construct a multi-objective optimization function, and use the gradient descent method to iteratively optimize the multi-objective optimization function. When the adjacent iteration difference of the multi-objective optimization function is less than the preset convergence threshold and the function variance meets the stability requirement, select the optimal wavelet basis function from the wavelet basis function candidate set.

[0033] In an alternative embodiment,

[0034] Take the node fluctuation feature index value as the input, perform feature aggregation through a knowledge graph-enhanced graph attention network to obtain a fluctuation feature vector, and calculate the association strength between different nodes based on the multi-head self-attention mechanism. The state matrix reflecting the overall wave dynamic trend of the supply chain constructed based on the fluctuation feature vector and the association strength includes:

[0035] Obtain the fluctuation feature index value of the supply chain node, and perform feature embedding on the fluctuation feature index value and the position encoding information to obtain a node representation vector;

[0036] Extract node association information from the supply chain knowledge graph, construct a knowledge-enhanced graph attention network based on the node association information. The knowledge-enhanced graph attention network fuses the node association information and the node representation vector to calculate the attention weight between nodes, and performs feature aggregation on the node representation vector according to the attention weight to generate a node-level fluctuation feature vector;

[0037] Input the fluctuation feature vector into a multi-head self-attention network. The multi-head self-attention network performs feature transformation on the fluctuation feature vector through multiple groups of feature mapping matrices to obtain a query vector, a key vector and a value vector, calculates the association strength between different nodes based on the query vector and the key vector, and performs weighted fusion on the value vector by using the association strength to generate an association feature vector of the node;

[0038] Construct a node state vector based on the fluctuation feature vector and the correlation feature vector, combine the node state vectors to form a state matrix, and optimize the state matrix by imposing multi-scale consistency constraints and temporal consistency constraints to obtain a state matrix reflecting the overall wave trend of the supply chain.

[0039] In an alternative embodiment,

[0040] Perform temporal matching between the historical fluctuation pattern and the current state matrix, and the output risk probability distribution includes:

[0041] Decompose the historical fluctuation data into a trend component, a periodic component, and a random component through data decomposition, calculate the features of the trend component, periodic component, and random component respectively to obtain trend features, periodic features, and random features, and construct a historical fluctuation pattern feature set;

[0042] Determine the feature weights by calculating the information entropy of each feature in the historical fluctuation pattern feature set, and use the feature weights to perform weighted fusion on the historical fluctuation pattern feature set to obtain a historical fluctuation pattern matrix;

[0043] Construct a multi-layer progressive similarity calculation network, calculate the similarity between the current state matrix and the historical fluctuation pattern matrix in each layer of the network, and use the similarity result of each layer as the calculation weight of the next layer to obtain an inter-layer progressive temporal matching weight;

[0044] Adaptively adjust the time window size according to the distribution characteristics of the temporal matching weights, resample the current state matrix based on the time window size, and perform multi-layer progressive similarity calculations at different scales. Stop the calculation when the difference between the temporal matching weights obtained from two adjacent calculations is less than a preset weight threshold;

[0045] Non-linearly combine the temporal matching weights at different scales to obtain a comprehensive matching score, use the comprehensive matching score to perform probability mapping on the risk indicators in the historical fluctuation pattern matrix, and output a risk probability distribution.

[0046] In an alternative embodiment,

[0047] Input the risk probability distribution into a graph convolutional network, perform feature propagation based on the supply chain knowledge graph, extract the local structure feature matrix in the supply chain network through multi-layer spatial convolutional operations, and combine the expert rules in the supply chain knowledge graph to identify the key path of risk conduction and generate a risk warning signal including the conduction probability:

[0048] Input the risk probability distribution into the graph convolutional network, construct an adjacency matrix based on the entity relationships in the supply chain knowledge graph. The adjacency matrix dynamically adjusts the node relationships using two-way adaptive weights, and performs multi-layer spatial convolutional operations on the node features through the adjacency matrix. Each layer of convolution introduces a residual structure and a skip connection with a gating mechanism, and obtains a local structure feature matrix through processing by a weight matrix and a non-linear activation function;

[0049] Calculate the structural similarity between nodes using the local structure feature matrix, attenuate the importance of historical conduction rules based on a time-series decay function, and obtain the rule confidence by dynamically combining the structural similarity and the decayed expert rule scores;

[0050] Establish multi-hop conduction links between nodes based on the local structure feature matrix, construct a set of conduction paths from the risk source node to the target node, and calculate the risk diffusion coefficient of each conduction path. The risk diffusion coefficient is calculated based on the risk probability magnitudes of the nodes on the path, the structural similarity between nodes, and the rule confidence. Select the conduction path with the largest risk diffusion coefficient as the key risk conduction path;

[0051] For adjacent nodes on the key risk conduction path, calculate the conduction intensity by combining the time-series conduction mode and the spatial conduction mode between nodes. Take the product of the conduction intensity and the risk probability distribution of the nodes as the path conduction probability, and generate a risk warning signal including the key risk conduction path and the path conduction probability.

[0052] In an alternative embodiment,

[0053] Calculating the conduction intensity by combining the time-series conduction mode and the spatial conduction mode between adjacent nodes on the key risk conduction path includes:

[0054] For adjacent nodes on the key risk conduction path, extract the historical fluctuation patterns of the adjacent nodes based on a sliding time window. Perform multi-dimensional time-series matching of the historical fluctuation patterns with the current state matrix through fluctuation frequency matching, fluctuation amplitude matching, and fluctuation trend matching, and introduce fluctuation mutation detection to obtain a fluctuation mutation coefficient. Weightedly combine the multi-dimensional time-series matching results and the fluctuation mutation coefficient to obtain a time-series conduction feature;

[0055] Extract the business association intensity, capital flow intensity, and information transfer intensity of adjacent nodes based on the current state matrix, and calculate a conduction impedance coefficient in combination with the time-series conduction feature. The conduction impedance coefficient characterizes the risk resistance ability of the nodes. Modify the business association intensity, capital flow intensity, and information transfer intensity through the conduction impedance coefficient to obtain a spatial conduction feature;

[0056] The timing conduction feature and the spatial conduction feature are adaptively weighted and combined to obtain a basic conduction intensity. Based on the basic conduction intensity, a conduction critical value is calculated. When the basic conduction intensity exceeds the conduction critical value, a cascade influence coefficient is calculated;

[0057] The basic conduction intensity and the cascade influence coefficient are combined to obtain the conduction intensity between adjacent nodes.

[0058] In the second aspect of the embodiments of the present invention,

[0059] A kind of electronic device is provided, including:

[0060] A processor;

[0061] A memory for storing instructions executable by the processor;

[0062] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0063] In the third aspect of the embodiments of the present invention,

[0064] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0065] In this embodiment, by constructing a supply chain knowledge graph and combining a graph attention network and a multi-head self-attention mechanism, the multi-dimensional fluctuation characteristics and the complex correlation relationships between nodes in the supply chain system can be effectively captured, and the accuracy and interpretability of fluctuation prediction are improved. Wavelet transform is used for multi-scale decomposition, and local structure features are extracted through a graph convolutional network. Combining expert rules to identify risk conduction paths can timely discover potential risks and trace their propagation paths, effectively enhancing the risk prevention and control ability of the supply chain system. A preset emergency response strategy library is established, and corresponding prevention and control plans can be quickly generated according to risk warning signals, improving the emergency handling efficiency of the supply chain system for emergencies, reducing inventory management costs, and enhancing the overall resilience of the supply chain. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a schematic flowchart of the method for multi-dimensional fluctuation prediction and inventory optimization of the supply chain of artificial intelligence in the embodiments of the present invention;

[0067] Figure 2 It is a schematic diagram of the construction of the supply chain knowledge graph in the embodiments of the present invention;

[0068] Figure 3 It is a distribution diagram of the prediction accuracy of the wave dynamic trend in the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0070] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0071] Figure 1 It is a schematic flowchart of the method for multi-dimensional fluctuation prediction and inventory optimization of the supply chain of artificial intelligence in the embodiments of the present invention. As Figure 1 shown, the method includes:

[0072] S101. Obtain multi-dimensional data in the supply chain system and perform preprocessing, use the piecewise linear interpolation method to supplement missing data, generate a standardized time series data set, and construct a supply chain knowledge graph based on the standardized time series data set;

[0073] Based on the standardized time series data set, use the wavelet transform method to perform multi-scale decomposition on the data, extract the trend component, periodic component, and random component in the time series data, and calculate the node fluctuation characteristic index value;

[0074] Take the node fluctuation characteristic index value as the input, perform feature aggregation through a knowledge graph-enhanced graph attention network to obtain a fluctuation feature vector, calculate the association strength between different nodes based on the multi-head self-attention mechanism, and construct a state matrix reflecting the overall wave dynamic trend of the supply chain based on the fluctuation feature vector and the association strength;

[0075] Perform time series matching between the historical fluctuation pattern and the current state matrix, and output the risk probability distribution;

[0076] Input the risk probability distribution into the graph convolutional network, perform feature propagation based on the supply chain knowledge graph, extract the local structure feature matrix in the supply chain network through multi-layer spatial convolutional operations, combine the expert rules in the supply chain knowledge graph, identify the key path of risk conduction, and generate a risk warning signal including the conduction probability;

[0077] Call the preset emergency response strategy library according to the risk warning signal to generate a corresponding risk prevention and control plan.

[0078] Among them, the emergency response strategy library is a preset set of risk response plans, which contains response measures for different supply chain risks, such as inventory adjustment, supplier replacement, logistics optimization, etc. When the system detects a risk, it can automatically call the most appropriate strategy from the emergency response strategy library and provide it to the decision maker for reference, so as to take corresponding risk control measures. For example, in the case of supplier delivery delay, the system can select "purchase from alternative suppliers" or "adjust the production plan" from the strategy library as the response strategy.

[0079] In an alternative implementation, multi-dimensional data in the supply chain system is obtained and preprocessed. The piecewise linear interpolation method is used to supplement missing data, and a standardized time series data set is generated. Building a supply chain knowledge graph based on the standardized time series data set includes:

[0080] Obtain multi-dimensional data in the supply chain system, and the multi-dimensional data includes supplier production data, logistics transportation data, warehouse inventory data, and market demand data;

[0081] According to the multi-dimensional data, a feature distribution matrix is established. The local density calculation method is used to obtain the density feature vector. Based on the density feature vector, the local deviation value of the data point is calculated and compared with the adaptive threshold to identify the first-round abnormal data. The location information of the first-round abnormal data is recorded in the abnormal index list;

[0082] Extract the time series data segments around the first-round abnormal data according to the abnormal index list, calculate the change trend coefficient between the data points in the time series data segments, construct a data change feature matrix, and use the data change feature matrix to perform secondary verification on the first-round abnormal data to generate the final abnormal data set;

[0083] Based on the location information of the final abnormal data set, a data interval to be supplemented is divided in the multi-dimensional data. Calculate the numerical difference and time span between the data points at both ends of the data interval to be supplemented. The supplementary data weight is adaptively generated according to the numerical difference and time span. Use the supplementary data weight for piecewise linear interpolation calculation to obtain the supplementary data set;

[0084] Fill the supplementary data set into the multi-dimensional data, and perform normalization processing on the filled data to generate a standardized data matrix;

[0085] Extract supply chain entity features based on the standardized data matrix, calculate the multi-dimensional correlation degree between entities to obtain the correlation strength matrix, and construct a supply chain knowledge graph according to the correlation strength matrix, where the nodes represent supply chain entities, and the weight of the connection edge is determined by the correlation strength matrix.

[0086] Exemplarily, to obtain multi-dimensional data in a supply chain system, it is necessary to collect supply chain-related information from different data sources. The multi-dimensional data includes supplier production data, logistics transportation data, warehousing inventory data, and market demand data. Supplier production data involves information such as production progress, production capacity utilization rate, and raw material inventory, which is used to evaluate the production capacity and supply stability of suppliers. Logistics transportation data includes information such as transportation time, transportation cost, route planning, and vehicle scheduling, reflecting the efficiency and cost of material flow in the supply chain. Warehousing inventory data covers content such as inventory level, inventory turnover rate, and safety stock, which helps to judge the rationality of supply chain inventory management. Market demand data involves information such as sales forecast, customer order volume, and market trend, providing a market orientation for supply chain decision-making.

[0087] Based on the multi-dimensional data, a feature distribution matrix is established. The feature distribution matrix is used to describe the statistical characteristics of the data, including eigenvalue features such as mean, variance, skewness, and kurtosis, for subsequent analysis. A local density calculation method is used to obtain a density feature vector. This method measures the relative density of data points in the overall data by calculating the distribution density of data points in a local area. Based on the density feature vector, the local deviation value of the data point is calculated. The local deviation value reflects the degree of deviation of the data point from adjacent data and is used to identify abnormal data. When the local deviation value exceeds the set adaptive threshold, the data point is determined to be abnormal data. The adaptive threshold is dynamically adjusted according to the overall distribution characteristics of the data to ensure the accuracy of anomaly detection. The data points determined to be abnormal are recorded in the anomaly index list.

[0088] According to the anomaly index list, time series data segments around the first-round abnormal data are extracted. The time series data segment refers to a set of data points at multiple time points before and after the abnormal data, aiming to further verify the rationality of the abnormal data by observing the change trend of the data. The change trend coefficient between the data points in the time series data segment is calculated. This coefficient is used to measure the change amplitude and direction of adjacent data points, so as to identify whether the abnormal data conforms to the overall trend. A data change feature matrix is constructed. This matrix is used to comprehensively describe the change pattern of the data, including features such as data change rate, change amplitude, and trend direction. The first-round abnormal data is verified twice using the data change feature matrix. By comparing the historical data pattern and the current data trend, the real abnormal data is further screened out, and the final abnormal data set is generated.

[0089] Based on the location information of the final abnormal data set, a data interval to be supplemented is divided in the multi-dimensional data. The data interval to be supplemented refers to a continuous time period containing abnormal data, and its two ends are bounded by normal data points. Calculate the numerical difference and time span between the data points at both ends of the data interval to be supplemented. The numerical difference represents the change amplitude of the data in the time dimension, and the time span represents the interval of the abnormal data on the time axis. Generate a supplementary data weight adaptively according to the numerical difference and time span. The supplementary data weight is used to adjust the calculation result of the supplementary data to make it conform to the overall change trend of the data. Use the supplementary data weight to perform piecewise linear interpolation calculation. Piecewise linear interpolation is a data filling method that estimates the interpolation by connecting adjacent normal data points and according to the weight to obtain a supplementary data set.

[0090] Fill the supplementary data set into the multi-dimensional data, and perform normalization processing on the filled data to generate a standardized data matrix. The purpose of normalization processing is to eliminate the numerical differences between different data dimensions and enable the data to be calculated and compared on the same scale. Common normalization methods include min-max normalization, mean normalization, etc. The standardized data matrix is a structured data format after normalization processing, which is convenient for subsequent feature extraction and analysis.

[0091] Extract supply chain entity features based on the standardized data matrix. Supply chain entity features include information such as supplier supply capacity, logistics distribution efficiency, inventory management level, market demand fluctuations, etc. Calculate the multi-dimensional correlation degree between entities to obtain a correlation strength matrix. The correlation strength matrix is used to quantify the degree of mutual influence between different supply chain entities. When calculating the correlation strength, multiple factors such as geographical location, transaction frequency, and supply chain cooperation relationship can be considered. Construct a supply chain knowledge graph according to the correlation strength matrix. Among them, the nodes represent supply chain entities, such as suppliers, warehouses, logistics companies, retailers, etc., and the weight of the connecting edge is determined by the correlation strength matrix, indicating the degree of dependence between entities. The supply chain knowledge graph can intuitively describe the supply chain structure, which is convenient for subsequent risk assessment and optimization decision-making.

[0092] As Figure 2As shown in the figure, the figure shows the construction process of a supply chain knowledge graph based on multi-dimensional data. Among them, Supplier A is used as the core node, which is connected to Logistics Provider B through a transportation relationship and to Warehouse C through a storage relationship, forming the basic business network of the supply chain. Nodes D and E are connected to Logistics Provider B through a supply relationship and a collaboration relationship, and nodes F and G are connected to Warehouse C through a scheduling relationship and a distribution relationship, constituting the complete supply chain network structure. The weights of the connection edges between each node are determined by the association strength matrix, which calculates the multi-dimensional association degree between entities based on the standardized data matrix. In this way, the knowledge graph not only describes the topological structure between supply chain entities, but also quantifies the business association degree between entities, providing a basis for subsequent risk conduction analysis.

[0093] In this embodiment, through the anomaly detection and supplementation of multi-dimensional data, the data quality and integrity are improved, providing a reliable data basis for subsequent analysis. By adopting a dual verification method of local density and change trend, the accuracy of anomaly detection is improved. The supplementation method combining piecewise linear interpolation and adaptive weights ensures the continuity and rationality of the data. The standardization process eliminates the dimensional differences of data in different dimensions, making the data comparable. The calculation of multi-dimensional association degree comprehensively depicts the complex relationships between supply chain entities, and the constructed knowledge graph intuitively shows the supply chain network structure. The design of adaptive thresholds and weights improves the robustness of the method, enabling it to adapt to the dynamic changes of data distribution. The construction of the knowledge graph provides strong support for supply chain optimization and decision-making. Compared with traditional simple mean filling or interpolation methods, it can better restore the time series characteristics of the data. By constructing a supply chain knowledge graph, the association relationships of each link in the supply chain are quantified, enabling supply chain risk assessment and optimization to be based on more comprehensive data support, and improving the scientificity and feasibility of supply chain management decisions.

[0094] In an alternative implementation manner, based on the standardized time series data set, the wavelet transform method is used to perform multi-scale decomposition on the data, extract the trend component, periodic component, and random component in the time series data, and calculate the node fluctuation characteristic index values including:

[0095] Calculate the energy concentration and information entropy of different wavelet basis functions based on the standardized supply chain time series data set, construct a multi-objective optimization function according to the energy concentration and information entropy, and obtain the optimal wavelet basis function by solving the multi-objective optimization function;

[0096] Perform adaptive multi-level decomposition on the standardized supply chain time series data set, calculate the scale coefficient and wavelet coefficient at each decomposition level, reconstruct the trend component according to the scale coefficient, reconstruct the periodic component and random component according to the wavelet coefficient, and use the energy distribution characteristics of the trend component, periodic component, and random component at different scales as the fluctuation characteristic index values.

[0097] Exemplarily, the time - series dataset of the supply chain is standardized. By subtracting the mean and dividing by the standard deviation, the data distribution becomes more regular, facilitating subsequent processing. For example, if the original data is [100, 120, 95, 130], after standardization, standardized values such as [-0.5, 0.2, -0.8, 1.1] can be obtained.

[0098] Next, different types of wavelet basis functions are selected, including Haar, Daubechies, Symlets, etc., to decompose the standardized data. Calculate the energy concentration degree of each wavelet basis function, which is determined by calculating the ratio of the sum of the squares of the coefficients at different scales to the total energy. At the same time, calculate the information entropy, which reflects the uniformity of the coefficient distribution. For example, if the energy distribution of a certain wavelet basis function at three scales is [0.6, 0.3, 0.1], the corresponding information entropy can be calculated to be 1.45.

[0099] Construct a multi - objective optimization function, with maximizing the energy concentration degree and minimizing the information entropy as the optimization objectives. Solve for the optimal wavelet basis function through optimization methods such as genetic algorithms. For example, the optimization results for 10 different wavelet basis functions show that the db4 wavelet has the best comprehensive performance.

[0100] Use the optimal wavelet basis function to perform multi - level decomposition on the data. At each decomposition level, scale coefficients and wavelet coefficients are obtained. The scale coefficients reflect the low - frequency trend characteristics of the data, and the wavelet coefficients reflect the high - frequency detail characteristics. Taking three - layer decomposition as an example, three groups of scale coefficients [a3, a2, a1] and wavelet coefficients [d3, d2, d1] are obtained.

[0101] Reconstruct the trend component based on the scale coefficients, reflecting the overall change trend of the data. Reconstruct the periodic component and the random component based on the wavelet coefficients at different levels, reflecting the characteristics of periodic fluctuations and random perturbations. For example, the trend component of a certain time - series data after decomposition shows an upward trend, the periodic component shows a period of about 30 days, and the random component fluctuates within the range of ±0.2.

[0102] Calculate the energy distribution characteristics of the trend component, periodic component, and random component at different scales as fluctuation characteristic indicators. These include the energy proportion and energy concentration degree of each component. For example, the energy proportion of the trend component is 0.7, the periodic component is 0.2, and the random component is 0.1, indicating that the data is mainly affected by the trend.

[0103] In this embodiment, multi-scale decomposition of time-series data is achieved through wavelet transform to accurately extract the trend, periodicity, and random features in the data, providing a reliable basis for supply chain fluctuation analysis. An energy concentration degree and information entropy are used to construct a multi-objective optimization function to select the optimal wavelet basis function, improving the decomposition effect and making each component obtained by decomposition more physically meaningful. A fluctuation index system is constructed based on the energy distribution characteristics of components at different scales to comprehensively characterize the supply chain fluctuation characteristics and provide a quantitative basis for supply chain risk prevention and control.

[0104] In an alternative embodiment, the energy concentration degree and information entropy of different wavelet basis functions are calculated based on a standardized supply chain time-series data set, a multi-objective optimization function is constructed according to the energy concentration degree and information entropy, and obtaining the optimal wavelet basis function by solving the multi-objective optimization function includes:

[0105] Construct a candidate set of wavelet basis functions, and perform wavelet transform on the supply chain time-series data in the standardized supply chain time-series data set using each wavelet basis function in the candidate set of wavelet basis functions to obtain corresponding wavelet coefficients;

[0106] Calculate the energy value of each wavelet basis function based on the wavelet coefficients and perform normalization processing to obtain an energy distribution probability, and calculate the energy concentration degree of each wavelet basis function according to the energy distribution probability;

[0107] Calculate the information entropy of each wavelet basis function based on the energy distribution probability, and obtain the relative entropy feature of each wavelet basis function using the ratio of the information entropy to the theoretical maximum entropy;

[0108] Use the wavelet coefficients and the corresponding wavelet basis functions to perform signal reconstruction to obtain a reconstructed signal, and calculate the reconstruction error between the reconstructed signal and the supply chain time-series data;

[0109] Weight and combine the energy concentration degree, relative entropy feature, and reconstruction error to construct a multi-objective optimization function, and use the gradient descent method to iteratively optimize the multi-objective optimization function. When the adjacent iteration difference of the multi-objective optimization function is less than a preset convergence threshold and the function variance meets the stability requirement, select the optimal wavelet basis function from the candidate set of wavelet basis functions.

[0110] Exemplarily, construct a candidate set of wavelet basis functions, where a wavelet basis function is a mathematical function used for signal processing that can decompose an original signal into different frequency components for further analysis and processing. Select a set of candidate wavelet basis functions from multiple known wavelet basis functions, and perform a transformation on the standardized supply chain time-series data using the candidate wavelet basis functions to obtain corresponding wavelet coefficients.

[0111] Calculate the energy values of the wavelet coefficients corresponding to each wavelet basis function and normalize the energy values. The purpose of normalization is to ensure that wavelet energies at different scales can be compared within the same reference range. Based on the normalized energy values, calculate the distribution probability of the energy over different wavelet basis functions, which reflects the degree of adaptation of different wavelet basis functions to the time series data.

[0112] Calculate the energy concentration degree of each wavelet basis function based on the energy distribution probability. The energy concentration degree is used to measure the concentration performance ability of a certain wavelet basis function in signal analysis. The higher the concentration degree, the stronger the decomposition ability of the wavelet basis function for the signal. At the same time, calculate the information entropy corresponding to each wavelet basis function based on the energy distribution probability. The information entropy is used to measure the complexity of the signal. The information entropies of different wavelet basis functions can reflect their ability to extract the characteristics of the supply chain time series data. Based on the ratio of the calculated information entropy value to the theoretical maximum entropy, determine the relative entropy feature of the wavelet basis function, which is used to measure the degree of preservation of the time series data characteristics by the wavelet basis function.

[0113] Use the calculated wavelet coefficients and the corresponding wavelet basis functions to perform signal reconstruction on the supply chain time series data to obtain the reconstructed signal. Calculate the reconstruction error between the reconstructed signal and the original supply chain time series data. The reconstruction error reflects the information loss situation during the process of wavelet basis function transforming and reconstructing the signal. The smaller the error, the higher the fitting degree of the wavelet basis function to the signal.

[0114] Perform weighted combination of the energy concentration degree, relative entropy feature, and reconstruction error to construct a multi-objective optimization function. This optimization function is used to comprehensively evaluate the advantages and disadvantages of different wavelet basis functions in the supply chain time series data analysis. Use the gradient descent method to iteratively optimize the multi-objective optimization function. The gradient descent method is a commonly used optimization algorithm. By calculating the function gradient and adjusting the parameters along the gradient direction, the optimization objective gradually converges. When the difference between adjacent iterations of the multi-objective optimization function is less than the preset convergence threshold and the variance of the optimization function meets the stability requirements, it is considered that the optimization process reaches a stable state. Select the optimal wavelet basis function from the candidate wavelet basis function set. The optimal wavelet basis function refers to the wavelet basis function with the best performance under the evaluation of the multi-objective optimization function, which can be used for further analysis and prediction of the subsequent supply chain time series data.

[0115] The prior art usually uses fixed wavelet basis functions for the decomposition of supply chain time series data, which is difficult to adapt to different data characteristics, resulting in inaccurate feature extraction. In addition, the existing methods do not fully consider the reconstruction error during signal reconstruction, affecting the accuracy of data analysis. At the same time, a single-objective strategy is often adopted during the optimization process, failing to take into account the energy concentration, information entropy characteristics, and reconstruction error, limiting the optimization effect of the wavelet basis function. This application uses a multi-objective optimization method to dynamically select the optimal wavelet basis function from the wavelet basis function candidate set to adapt to the characteristic changes of supply chain data. By calculating the energy distribution probability, it is ensured that the selected basis function can accurately capture the main features. Calculate the relative entropy characteristics to reduce the information loss during data conversion. Further introduce the reconstruction error to keep the wavelet basis function with a high fitting degree during signal reconstruction, and optimize and iterate through gradient descent to improve the stability and accuracy of the selection. Compared with the prior art, this solution realizes the adaptive optimization of the wavelet basis function, makes the supply chain data decomposition more accurate, reduces the data reconstruction error, improves the reliability of supply chain prediction, and enhances the adaptability and generalization ability in complex environments.

[0116] In an alternative embodiment, the node fluctuation characteristic index value is used as the input, and the knowledge graph-enhanced graph attention network is used for feature aggregation to obtain the fluctuation feature vector, and the association strength between different nodes is calculated based on the multi-head self-attention mechanism. The state matrix reflecting the overall wave trend of the supply chain constructed based on the fluctuation feature vector and the association strength includes:

[0117] Obtain the node fluctuation characteristic index value of the supply chain node, and perform feature embedding on the fluctuation characteristic index value and the position encoding information to obtain the node representation vector;

[0118] Extract the node association information from the supply chain knowledge graph, and construct a knowledge-enhanced graph attention network based on the node association information. The knowledge-enhanced graph attention network fuses and calculates the node association information and the node representation vector to obtain the attention weight between nodes, and aggregates the features of the node representation vector according to the attention weight to generate a node-level fluctuation feature vector;

[0119] Input the fluctuation feature vector into the multi-head self-attention network. The multi-head self-attention network performs feature transformation on the fluctuation feature vector through multiple groups of feature mapping matrices to obtain the query vector, key vector, and value vector, calculates the association strength between different nodes based on the query vector and the key vector, and performs weighted fusion on the value vector using the association strength to generate the association feature vector of the node;

[0120] Construct a node state vector based on the fluctuation feature vector and the correlation feature vector, combine the node state vectors to form a state matrix, and optimize the state matrix by imposing multi-scale consistency constraints and temporal consistency constraints to obtain a state matrix that reflects the overall fluctuation trend of the supply chain.

[0121] Exemplarily, first obtain the fluctuation feature index values of the supply chain nodes. After obtaining the fluctuation feature index values of the supply chain nodes, it is necessary to perform feature embedding on these fluctuation feature index values and the corresponding location information to generate node representation vectors. The fluctuation feature index values refer to quantitative parameters used to measure the changes of supply chain nodes at different time points, such as inventory levels, transportation delays, order volume fluctuations, etc. The location encoding information is used to represent the position of the node in the supply chain network so that the model can utilize the location information for learning.

[0122] Extract the association information between nodes from the supply chain knowledge graph and construct a knowledge-enhanced graph attention network based on this association information. In this network, the knowledge-enhanced graph attention network utilizes the prior information contained in the knowledge graph to calculate the attention weights between nodes. The attention weights reflect the relative importance between different nodes, and the node representation vectors are feature aggregated through these weights to obtain a node-level fluctuation feature vector that can reflect the fluctuation of supply chain nodes.

[0123] Input the node-level fluctuation feature vector into a multi-head self-attention network, perform feature transformation on the fluctuation feature vector using multiple groups of feature mapping matrices to generate query vectors, key vectors, and value vectors. The query vectors, key vectors, and value vectors respectively represent the feature representations of different nodes in the supply chain network, where the query vectors are used to retrieve information of other nodes, the key vectors are used to calculate similarity, and the value vectors are used for feature fusion. Calculate the association strength between different nodes using the query vectors and the key vectors. The association strength is used to measure the closeness of the relationship between two nodes. Weightedly fuse the value vectors according to the calculated association strength, and finally generate the correlation feature vector of the node.

[0124] Combining the fluctuation eigenvector and the correlation eigenvector, a node state vector containing rich information is constructed, and the state vectors of multiple nodes are combined to form a state matrix. The state matrix is ​​used to represent the overall fluctuation trend of the supply chain. In order to optimize the state matrix so that it can more accurately reflect the fluctuation of the supply chain network, multi-scale consistency constraints and temporal consistency constraints are imposed on the state matrix. The multi-scale consistency constraint ensures that the fluctuation pattern at different time scales remains stable, and the temporal consistency constraint ensures that the change trend of the supply chain node is consistent with the actual business logic. These constraints ensure that the model maintains consistency at different time scales and can handle the continuity of node states at different time points. By optimizing these constraints, a state matrix that reflects the overall fluctuation trend of the supply chain can be obtained.

[0125] For example, in a fast-moving consumer goods supply chain management, you can obtain information such as inventory changes, transportation timeliness data, and order demand fluctuation data for each storage center. Input this data into the supply chain knowledge graph, combine it with the existing supply chain link association information in the knowledge graph, and use the graph attention network to calculate the influence relationship between storage centers. Subsequently, use the multi-head self-attention mechanism to calculate the strength of mutual correlation between different storage centers, identify key fluctuation points in the supply chain, and use the optimized state matrix to predict the risks of inventory shortages or logistics delays that may occur in the future, providing companies with reasonable response strategies.

[0126] like Figure 3 As shown, Figure 3 The distribution diagram of the prediction accuracy of the fluctuation trend of this embodiment shows the prediction performance under different fluctuation levels. The size of the bubble indicates the amount of data, and the larger the fluctuation coefficient, the more severe the supply chain fluctuation. Even in the case of high fluctuation (fluctuation coefficient 0.6), this solution can still maintain 86.1% prediction accuracy, while the accuracy of the traditional time series prediction method drops to 73.2%. This proves the robustness of this solution in dealing with complex fluctuation scenarios.

[0127] In this embodiment, by combining the fluctuation feature index, the knowledge graph, and the graph attention network, the limitations of the existing supply chain management technology in node dynamic fluctuation analysis are solved. The existing technology usually relies on static data or traditional statistical methods, lacking in-depth mining and dynamic updating of the complex relationships between nodes, resulting in the analysis results being unable to accurately reflect the real-time changes of the supply chain. Different from this, this solution dynamically captures the fluctuation features of nodes through feature embedding and the graph attention network, and uses the correlation information between nodes for in-depth fusion, which can more accurately reflect the interaction and influence between nodes. In addition, with the help of the knowledge graph, the solution fully mines the correlation information between nodes, enhancing the understanding of the supply chain structure, while traditional methods usually ignore these complex relationships. The multi-head self-attention network further optimizes the feature aggregation of nodes, can more precisely process the correlation features between nodes, and avoids the overly simple aggregation method of traditional methods. Compared with the existing technology, the improvement of this application lies in combining various feature information with deep learning methods, improving the prediction accuracy of node fluctuation features and the assessment ability of supply chain risks. Through this intelligent fusion method, this solution can achieve more efficient and accurate supply chain dynamic monitoring and management.

[0128] In an alternative embodiment, the historical fluctuation pattern is temporally matched with the current state matrix, and the output risk probability distribution includes:

[0129] The historical fluctuation data is decomposed into a trend component, a periodic component, and a random component through data decomposition, and the trend feature, the periodic feature, and the random feature are respectively calculated for the trend component, the periodic component, and the random component to construct a historical fluctuation pattern feature set;

[0130] The feature weights are determined by calculating the information entropy of each feature in the historical fluctuation pattern feature set, and the historical fluctuation pattern feature set is weighted and fused using the feature weights to obtain a historical fluctuation pattern matrix;

[0131] A multi-layer progressive similarity calculation network is constructed. In each layer of the network, the similarity between the current state matrix and the historical fluctuation pattern matrix is calculated, and the similarity result of each layer is used as the calculation weight of the next layer to obtain the temporally progressive matching weights between layers;

[0132] The time window size is adaptively adjusted according to the distribution characteristics of the temporally matching weights, the current state matrix is resampled based on the time window size, and the multi-layer progressive similarity calculation is performed at different scales. The calculation stops when the difference between the temporally matching weights obtained from two adjacent calculations is less than a preset weight threshold;

[0133] Non-linearly combine the temporal matching weights at different scales to obtain a comprehensive matching score, and use the comprehensive matching score to perform probability mapping on the risk indicators in the historical volatility pattern matrix, and output a risk probability distribution.

[0134] Exemplarily, the historical volatility data is split through data decomposition to obtain a trend component, a periodic component, and a random component. These components reflect different change trends in the data. The trend component represents the long-term change direction, the periodic component represents the periodic fluctuations, and the random component represents unpredictable noise or random fluctuations. Feature calculations are performed on each of the components obtained through the decomposition, so as to extract trend features, periodic features, and random features. These features can effectively represent different components of the historical volatility data, thereby constructing a historical volatility pattern feature set.

[0135] Calculate the information entropy of each feature in the historical volatility pattern feature set. Information entropy is an index that measures the uncertainty or amount of information of a feature. Through the calculation of information entropy, a weight can be assigned to each feature, and these weights reflect the importance of each feature to the volatility pattern. Using these feature weights, weighted fusion is performed on the historical volatility pattern feature set to generate a historical volatility pattern matrix. This matrix contains the comprehensive performance of each feature under different weights and is a comprehensive description of the historical volatility pattern.

[0136] Then construct a multi-layer progressive similarity calculation network. Each layer of this network calculates the similarity between the current state matrix and the historical volatility pattern matrix. In each layer of calculation, the matching result of the historical volatility pattern matrix is used as the weight of the current layer and passed to the next layer. This way of passing layer by layer can gradually optimize the similarity calculation and combine the information of each layer to obtain more accurate temporal matching weights.

[0137] Based on the distribution characteristics of the temporal matching weights, the system will adaptively adjust the size of the time window. The time window determines the range of historical data used in the calculation. By adjusting the window size, it is possible to adapt to different time scales in the data. When the difference between the temporal matching weights obtained from two adjacent calculations is less than a preset threshold, the calculation is stopped to ensure the stability and accuracy of the calculation results.

[0138] Non-linearly combine the temporal matching weights obtained at different scales to obtain a comprehensive matching score. This score reflects the matching degree between the current state and the historical volatility pattern. Use this score to perform probability mapping on the risk indicators in the historical volatility pattern matrix, thereby outputting a risk probability distribution. This distribution can help predict possible future risks and provide decision-making support.

[0139] In this embodiment, through data decomposition and feature extraction, the trends, cycles, and stochastic characteristics of historical fluctuation patterns are comprehensively characterized, improving the accuracy and robustness of pattern recognition. By adopting a multi-layer progressive similarity calculation network and an adaptive time window, time series matching at different scales is achieved, enhancing the adaptability of the algorithm to market fluctuations. Based on the feature weight calculation using information entropy and the matching score of non-linear combination, the risk probability distribution is made more objective and reasonable, providing a reliable basis for risk warning.

[0140] In an alternative embodiment, the risk probability distribution is input into a graph convolutional network. Based on the supply chain knowledge graph, feature propagation is performed. Through multi-layer spatial convolutional operations, a local structure feature matrix in the supply chain network is extracted. Combining with the expert rules in the supply chain knowledge graph, the key path of risk conduction is identified, and the risk warning signal including the conduction probability is generated, including:

[0141] The risk probability distribution is input into a graph convolutional network. An adjacency matrix is constructed based on the entity relationships in the supply chain knowledge graph. The adjacency matrix dynamically adjusts the node relationships using two-way adaptive weights. Through the adjacency matrix, multi-layer spatial convolutional operations are performed on the node features. Each layer of convolution introduces a residual structure and a skip connection with a gating mechanism. After being processed by a weight matrix and a non-linear activation function, a local structure feature matrix is obtained;

[0142] The structural similarity between nodes is calculated using the local structure feature matrix. The importance of historical conduction rules is attenuated based on a time series decay function. The rule confidence is obtained by combining the structural similarity and the decayed expert rule score through dynamic weight combination;

[0143] Based on the local structure feature matrix, a multi-hop conduction link between nodes is established. A set of conduction paths is constructed from the risk source node to the target node. The risk diffusion coefficient of each conduction path is calculated. The risk diffusion coefficient is calculated based on the risk probability magnitudes of the nodes on the path, the structural similarity between nodes, and the rule confidence. The conduction path with the largest risk diffusion coefficient is selected as the key path of risk conduction;

[0144] For adjacent nodes on the key path of risk conduction, the conduction intensity is calculated by combining the time series conduction mode and the spatial conduction mode between nodes. The product of the conduction intensity and the risk probability distribution of the nodes is used as the path conduction probability, and a risk warning signal including the key path of risk conduction and the path conduction probability is generated.

[0145] First, the risk probability distribution is input into the graph convolutional network, and an adjacency matrix is constructed based on the entity relationships in the supply chain knowledge graph. The adjacency matrix is a matrix that reflects the connection relationships between nodes. In this process, bidirectional adaptive weights are used to dynamically adjust the node relationships in the adjacency matrix. This means that the connection strength between nodes will be automatically optimized according to the current data and historical information to better reflect the actual influence between nodes.

[0146] Perform multi-layer spatial convolution operations on the node features through the adjacency matrix. The graph convolutional network is a neural network model that can process graph-structured data and extracts the local structural features of nodes through convolution operations. In each layer of convolution, a residual structure and a skip connection with a gating mechanism are introduced. The residual structure helps alleviate the vanishing gradient problem in deep neural networks and ensures the effective transmission of features in multi-layer networks. The gating mechanism controls the flow of information and avoids irrelevant information interfering with network learning. After convolution, it is processed through a weight matrix and a non-linear activation function to obtain the local structural feature matrix.

[0147] Calculate the structural similarity between nodes through the local structural feature matrix. The structural similarity is an index that measures the similarity degree between two nodes in the graph and reflects their structural similarity. Based on the time-series decay function, the influence of historical conduction rules is attenuated in importance. As time goes by, the influence of historical rules gradually weakens. The structural similarity and the decayed expert rule scores are combined through dynamic weights to obtain the rule confidence, which represents the credibility of the rule at the current moment.

[0148] Establish multi-hop conduction links between nodes. A multi-hop conduction link refers to a multi-level path from one node to another node, and the influence of each node on the final target node will be gradually transmitted along the path. Construct a set of conduction paths from the risk source node to the target node and calculate the risk diffusion coefficient of each path. The risk diffusion coefficient is jointly determined by the risk probabilities of the nodes on the path, the structural similarity between nodes, and the rule confidence. Select the path with the largest risk diffusion coefficient as the key risk conduction path, and this path has the greatest impact on the entire risk conduction process.

[0149] For adjacent nodes on the key risk conduction path, calculate the conduction strength by combining the time-series conduction mode and the spatial conduction mode between nodes. The time-series conduction mode describes the law of risk transmission between nodes over time, and the spatial conduction mode focuses on the geographical or structural relationships between nodes. The calculated conduction strength is combined with the product of the risk probability distributions of the nodes as the path conduction probability. The path conduction probability represents the possibility of transmitting risk on a specific path.

[0150] Finally, a risk warning signal including the key path of risk conduction and the path conduction probability is generated. This signal can reflect the potential risk conduction paths in the supply chain in real time, providing warning information for decision-makers so that corresponding preventive measures can be taken in advance.

[0151] In this embodiment, by combining graph convolutional network, knowledge graph, temporal decay function and multi-layer conduction calculation, the accuracy of risk prediction and management in the supply chain is significantly improved. Existing technologies usually rely on static data or a single rule engine, which are difficult to comprehensively capture the complex spatio-temporal relationships between nodes and cannot dynamically adjust the risk conduction paths. Different from this, this solution can reflect the actual changes and dependencies between nodes by dynamically adjusting the node relationships in the adjacency matrix. In addition, by using a multi-layer graph convolutional network combined with a residual structure and a gating mechanism, the expressive ability of the model is enhanced, the problem of gradient disappearance in information transmission is effectively solved, and the efficiency of feature extraction is improved. The innovation of the solution lies in effectively integrating time, space and structural information by adaptively adjusting the weights and the importance decay of historical conduction rules, enhancing the ability to capture multi-dimensional risk conduction paths. Most existing technologies are static paths or information transmission based on a single level. This technology dynamically identifies and optimizes risk conduction paths through progressive similarity calculation and multi-hop conduction links, further improving the prediction accuracy and flexibility. In this way, it can better cope with the complex node interactions and potential risk diffusion in the supply chain. In addition, the generation of the comprehensive matching score and the path conduction probability makes the risk warning signal more accurate and real-time, and can provide accurate warning information for decision-makers, thus effectively avoiding or reducing the impact of potential risks. Compared with traditional methods, this solution can handle a more complex supply chain environment, reduce manual intervention and prediction errors, and improve the overall stability and risk resistance of the supply chain.

[0152] In an alternative embodiment, calculating the conduction intensity for adjacent nodes on the key path of risk conduction by combining the temporal conduction mode and the spatial conduction mode between the nodes includes:

[0153] For adjacent nodes on the key path of risk conduction, extract the historical fluctuation patterns of the adjacent nodes based on a sliding time window, perform multi-dimensional temporal matching of the historical fluctuation patterns with the current state matrix through fluctuation frequency matching, fluctuation amplitude matching and fluctuation trend matching, and introduce fluctuation mutation detection to obtain a fluctuation mutation coefficient, and weight-combine the multi-dimensional temporal matching results with the fluctuation mutation coefficient to obtain a temporal conduction feature;

[0154] Extract the business association intensity, capital flow intensity, and information transfer intensity of adjacent nodes based on the current state matrix, and calculate the conduction impedance coefficient in combination with the time-series conduction characteristics. The conduction impedance coefficient characterizes the risk resistance ability of the node. Modify the business association intensity, capital flow intensity, and information transfer intensity through the conduction impedance coefficient to obtain the spatial conduction characteristics;

[0155] Perform an adaptive weighted combination of the time-series conduction characteristics and the spatial conduction characteristics to obtain the basic conduction intensity. Calculate the conduction critical value based on the basic conduction intensity. When the basic conduction intensity exceeds the conduction critical value, calculate the cascade impact coefficient;

[0156] Combine the basic conduction intensity and the cascade impact coefficient to obtain the conduction intensity between adjacent nodes.

[0157] Exemplarily, for adjacent nodes on the critical path of risk conduction, adopt the method of a sliding time window to extract the historical fluctuation patterns of each adjacent node. A sliding time window is a method of analyzing data by selecting a range of data in chronological order. Through this window, the fluctuation characteristics of the node in different time periods can be extracted. These fluctuation characteristics include the fluctuation frequency, fluctuation amplitude, and fluctuation trend, which respectively reflect the periodicity, intensity, and development direction of the node's fluctuations. Then, perform multi-dimensional time-series matching between these historical fluctuation patterns and the current state matrix. In this process, first perform fluctuation frequency matching to compare the periodicity of the historical fluctuation and the periodicity of the current state; then perform fluctuation amplitude matching to compare the intensity of the historical fluctuation and the intensity of the current state; finally, perform fluctuation trend matching to compare the direction of the historical fluctuation and the change direction of the current state. In addition, it is also necessary to introduce fluctuation mutation detection to identify the mutation points of the node's fluctuations. These mutation points are often the key points where risks occur. By calculating the fluctuation mutation coefficient, the information of these mutation points can be effectively incorporated into the time-series matching result.

[0158] Combine the obtained multi-dimensional time-series matching result and the fluctuation mutation coefficient, and through weighted combination, obtain the time-series conduction characteristics. The time-series conduction characteristics reflect the risk transfer characteristics of the node at different time scales. Through these characteristics, the propagation path and intensity of risks can be predicted more accurately.

[0159] Based on the current state matrix, extract its business association intensity, capital flow intensity, and information transfer intensity from adjacent nodes. These intensities respectively represent the interaction relationships between nodes at the business level, capital flow, and information flow transfer. By combining the time-series conduction characteristics, calculate the conduction impedance coefficient. The conduction impedance coefficient represents the resistance ability of the node when facing external risks, similar to the resistance in an electrical circuit. The larger the value, the stronger the risk resistance ability of the node and the smaller the risk conducted.

[0160] By correcting the business association strength, capital flow strength, and information transmission strength through the conduction impedance coefficient, the spatial conduction characteristics can be obtained. The spatial conduction characteristics represent the risk transmission characteristics between nodes in terms of space or structure, and these characteristics can reflect how risks between different nodes are affected by factors such as business and capital flow.

[0161] The temporal conduction characteristics and spatial conduction characteristics are adaptively weighted and combined to obtain the basic conduction strength. The basic conduction strength comprehensively considers the influence of the time dimension and the space dimension on risk transmission and can comprehensively evaluate the conduction situation of risks. Based on the basic conduction strength, the conduction critical value is calculated. When the basic conduction strength exceeds the conduction critical value, it indicates that the risk transmission has reached a dangerous level, and it is necessary to further calculate the cascade impact coefficient. The cascade impact coefficient is used to evaluate the degree of mutual influence between nodes during the risk conduction process. When the conduction strength reaches a certain critical value, a series of chain reactions may be triggered, thereby expanding the scope of risk spread.

[0162] Combining the basic conduction strength and the cascade impact coefficient, the conduction strength between adjacent nodes is obtained. This conduction strength reflects the mutual influence and synergy effect between adjacent nodes under a specific risk conduction path, provides a quantitative basis for risk management, and helps identify potential risk transmission chains.

[0163] In this embodiment, through the multi-dimensional analysis combining the sliding time window, fluctuation matching, temporal conduction characteristics, and spatial conduction characteristics, the ability to identify and predict supply chain risks is significantly improved. Existing technologies usually rely on static analysis and single-dimensional risk assessment and are difficult to cope with complex and dynamically changing supply chain environments. Traditional methods often ignore the linkage effect of time and space factors, while this solution effectively overcomes these deficiencies by dynamically matching historical fluctuation patterns, introducing mechanisms such as fluctuation mutation detection and conduction impedance coefficient. Through more comprehensive feature extraction and adaptive weighted combination, the identification process of the risk conduction path is optimized. Traditional technologies often have difficulty accurately capturing the multi-level and multi-dimensional spread of risks, while this solution can more accurately evaluate the diffusion process of potential risks by comprehensively considering the interaction between nodes in the time and space dimensions and dynamically adjusting the conduction strength. Especially through the introduction of the cascade impact coefficient, the ability to control the risk transmission chain is further enhanced, thus avoiding the limitations of single-path risk assessment.

[0164] In the second aspect of the embodiment of the present invention,

[0165] A kind of electronic device is provided, including:

[0166] A processor;

[0167] A memory for storing instructions executable by the processor;

[0168] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0169] In a third aspect of the embodiments of the present invention,

[0170] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0171] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for performing various aspects of the present invention.

[0172] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A supply chain multi-dimensional fluctuation prediction and inventory optimization method based on artificial intelligence, characterized by: include: Acquire and preprocess multi-dimensional data in the supply chain system, use piecewise linear interpolation to supplement missing data, generate standardized time series data sets, and build a supply chain knowledge graph based on the standardized time series data sets; Based on the standardized time series data set, the wavelet transform method is used to decompose the data at multiple scales, extract the trend component, periodic component and random component in the time series data, and calculate the node fluctuation characteristic index value; The node fluctuation characteristic index value is used as input, and the feature aggregation is performed through the graph attention network enhanced by the knowledge graph to obtain the fluctuation characteristic vector. The correlation strength between different nodes is calculated based on the multi-head self-attention mechanism, and a state matrix reflecting the overall fluctuation trend of the supply chain is constructed based on the fluctuation characteristic vector and the correlation strength. Match the historical fluctuation pattern with the current state matrix in time series and output the risk probability distribution; The risk probability distribution is input into the graph convolutional network, and feature propagation is performed based on the supply chain knowledge graph. The local structural feature matrix in the supply chain network is extracted through multi-layer spatial convolution operations. Combined with the expert rules in the supply chain knowledge graph, the key path of risk transmission is identified, and a risk warning signal containing the transmission probability is generated. Call the preset emergency response strategy library according to the risk warning signal to generate the corresponding risk prevention and control plan; The node fluctuation characteristic index value is used as input, and the fluctuation characteristic vector is obtained by feature aggregation through the graph attention network enhanced by the knowledge graph. The correlation strength between different nodes is calculated based on the multi-head self-attention mechanism. The state matrix reflecting the overall fluctuation trend of the supply chain is constructed based on the fluctuation characteristic vector and the correlation strength, including: Obtaining a fluctuation characteristic index value of a supply chain node, embedding the fluctuation characteristic index value and position coding information to obtain a node representation vector; Extracting node association information from the supply chain knowledge graph, building a knowledge-enhanced graph attention network based on the node association information, the knowledge-enhanced graph attention network fuses the node association information with the node representation vector to calculate the attention weights between nodes, and performing feature aggregation on the node representation vector according to the attention weights to generate a node-level fluctuation feature vector; The fluctuation feature vector is input into a multi-head self-attention network, and the multi-head self-attention network performs feature transformation on the fluctuation feature vector through multiple sets of feature mapping matrices to obtain a query vector, a key vector and a value vector, calculates the association strength between different nodes based on the query vector and the key vector, and performs weighted fusion on the value vector using the association strength to generate an association feature vector of the node; A node state vector is constructed based on the fluctuation eigenvector and the associated eigenvector, and the node state vectors are combined to form a state matrix. Multi-scale consistency constraints and temporal consistency constraints are imposed on the state matrix for optimization to obtain a state matrix that reflects the overall fluctuation trend of the supply chain.

2. The method according to claim 1, characterized in that Acquire and preprocess multi-dimensional data in the supply chain system, use piecewise linear interpolation to supplement missing data, generate standardized time series data sets, and build a supply chain knowledge graph based on the standardized time series data sets, including: Acquire multi-dimensional data in the supply chain system, including supplier production data, logistics and transportation data, warehouse inventory data, and market demand data; Establish a feature distribution matrix according to the multi-dimensional data, obtain a density feature vector using a local density calculation method, calculate a local deviation value of a data point based on the density feature vector, and compare it with an adaptive threshold to identify the first round of abnormal data, and record the location information of the first round of abnormal data in an abnormal index list; Extract the time series data segments around the first round of abnormal data according to the abnormal index list, calculate the change trend coefficients between the data points in the time series data segments, construct a data change feature matrix, use the data change feature matrix to perform secondary verification on the first round of abnormal data, and generate a final abnormal data set; Based on the location information of the final abnormal data set, a data interval to be supplemented is divided in the multi-dimensional data, the numerical difference and time span of the data points at both ends of the data interval to be supplemented are calculated, and the supplementary data weight is adaptively generated according to the numerical difference and time span, and the supplementary data weight is used to perform piecewise linear interpolation calculation to obtain a supplementary data set; Filling the supplementary data set into the multi-dimensional data, normalizing the filled data to generate a standardized data matrix; Based on the standardized data matrix, supply chain entity features are extracted, the multidimensional correlation between entities is calculated to obtain a correlation strength matrix, and a supply chain knowledge graph is constructed according to the correlation strength matrix, wherein nodes represent supply chain entities and the weights of connecting edges are determined by the correlation strength matrix.

3. The method according to claim 1, characterized in that Based on the standardized time series data set, the wavelet transform method is used to decompose the data at multiple scales, extract the trend component, periodic component and random component in the time series data, and calculate the node fluctuation characteristic index values ​​including: Calculate the energy concentration and information entropy of different wavelet basis functions based on the standardized supply chain time series data set, construct a multi-objective optimization function according to the energy concentration and information entropy, and obtain the optimal wavelet basis function by solving the multi-objective optimization function; The standardized supply chain time series data set is subjected to an adaptive multi-level decomposition, and the scale coefficient and wavelet coefficient are calculated at each decomposition level. The trend component is reconstructed according to the scale coefficient, and the periodic component and random component are reconstructed according to the wavelet coefficient. The energy distribution characteristics of the trend component, periodic component and random component at different scales are used as fluctuation characteristic indicator values.

4. The method according to claim 3, characterized in that: Based on the standardized supply chain time series data set, the energy concentration and information entropy of different wavelet basis functions are calculated, and a multi-objective optimization function is constructed according to the energy concentration and information entropy. The optimal wavelet basis function is obtained by solving the multi-objective optimization function, including: Constructing a wavelet basis function candidate set, and using each wavelet basis function in the wavelet basis function candidate set to perform wavelet transform on the supply chain time series data in the standardized supply chain time series data set to obtain corresponding wavelet coefficients; Calculating the energy value of each wavelet basis function based on the wavelet coefficient and performing normalization processing to obtain an energy distribution probability, and calculating the energy concentration of each wavelet basis function according to the energy distribution probability; Calculating the information entropy of each wavelet basis function based on the energy distribution probability, and obtaining the relative entropy characteristics of each wavelet basis function using the ratio of the information entropy to the theoretical maximum entropy; Reconstructing a signal using the wavelet coefficients and corresponding wavelet basis functions to obtain a reconstructed signal, and calculating a reconstruction error between the reconstructed signal and the supply chain time series data; The energy concentration, relative entropy characteristics and reconstruction error are weightedly combined to construct a multi-objective optimization function, and the multi-objective optimization function is iteratively optimized using the gradient descent method. When the adjacent iteration difference of the multi-objective optimization function is less than a preset convergence threshold and the function variance meets the stability requirement, the optimal wavelet basis function is selected from the wavelet basis function candidate set.

5. The method according to claim 1, characterized in that Match the historical fluctuation pattern with the current state matrix in time series, and output the risk probability distribution including: The historical fluctuation data is split into a trend component, a period component and a random component by data decomposition, and the trend component, the period component and the random component are respectively calculated to obtain trend features, period features and random features, so as to construct a historical fluctuation pattern feature set; The feature weight is determined by calculating the information entropy of each feature in the historical fluctuation pattern feature set, and the historical fluctuation pattern feature set is weightedly fused using the feature weight to obtain a historical fluctuation pattern matrix; Construct a multi-layer progressive similarity calculation network, calculate the similarity between the current state matrix and the historical fluctuation pattern matrix in each layer of the network, and use the similarity result of each layer as the calculation weight of the next layer to obtain the progressive time series matching weight between layers; Adaptively adjust the time window size according to the distribution characteristics of the timing matching weight, resample the current state matrix based on the time window size, and perform multi-layer progressive similarity calculation at different scales, and stop the calculation when the difference between the timing matching weights obtained by two adjacent calculations is less than a preset weight threshold; The time series matching weights at different scales are nonlinearly combined to obtain a comprehensive matching score, which is used to perform probability mapping on the risk indicators in the historical volatility pattern matrix to output the risk probability distribution.

6. The method according to claim 1, characterized in that The risk probability distribution is input into the graph convolutional network, and feature propagation is performed based on the supply chain knowledge graph. The local structural feature matrix in the supply chain network is extracted through multi-layer spatial convolution operations. Combined with the expert rules in the supply chain knowledge graph, the key path of risk transmission is identified, and risk warning signals containing transmission probability are generated, including: The risk probability distribution is input into the graph convolutional network, and an adjacency matrix is ​​constructed based on the entity relationship in the supply chain knowledge graph. The adjacency matrix uses bidirectional adaptive weights to dynamically adjust the node relationship. Multi-layer spatial convolution operations are performed on the node features through the adjacency matrix. Each layer of convolution introduces a residual structure and a skip-layer connection with a gating mechanism. The local structure feature matrix is ​​obtained through weight matrix and nonlinear activation function processing; The structural similarity between nodes is calculated using the local structural feature matrix, the importance of historical conduction rules is decayed based on the time series decay function, and the structural similarity and the decayed expert rule score are combined through dynamic weights to obtain the rule confidence; Based on the local structural feature matrix, a multi-hop transmission link between nodes is established, a transmission path set is constructed from the risk source node to the target node, and the risk diffusion coefficient of each transmission path is calculated. The risk diffusion coefficient is calculated by the risk probability size of the nodes on the path, the structural similarity between the nodes, and the rule confidence. The transmission path with the largest risk diffusion coefficient is selected as the key risk transmission path; For adjacent nodes on the critical path of risk conduction, the conduction strength is calculated by combining the temporal conduction mode and the spatial conduction mode between the nodes, and the multiplication value of the conduction strength and the risk probability distribution of the node is used as the path conduction probability to generate a risk warning signal including the critical path of risk conduction and the path conduction probability.

7. The method according to claim 6, characterized in that For the adjacent nodes on the critical risk transmission path, the transmission intensity is calculated by combining the temporal transmission mode and the spatial transmission mode between the nodes, including: For adjacent nodes on the critical path of risk transmission, the historical fluctuation pattern of the adjacent nodes is extracted based on the sliding time window, and the historical fluctuation pattern is matched with the current state matrix in a multi-dimensional time series through fluctuation frequency matching, fluctuation amplitude matching and fluctuation trend matching. The fluctuation mutation detection is introduced to obtain the fluctuation mutation coefficient, and the multi-dimensional time series matching result and the fluctuation mutation coefficient are weightedly combined to obtain the time series transmission characteristics; Based on the current state matrix, the business association strength, capital flow strength and information transmission strength of the adjacent nodes are extracted, and the conduction impedance coefficient is calculated in combination with the time series conduction characteristics. The conduction impedance coefficient represents the risk resistance of the node. The business association strength, capital flow strength and information transmission strength are corrected by the conduction impedance coefficient to obtain the spatial conduction characteristics; Adaptively weighting and combining the temporal conduction feature and the spatial conduction feature to obtain a basic conduction strength, calculating a conduction threshold value based on the basic conduction strength, and calculating a cascade influence coefficient when the basic conduction strength exceeds the conduction threshold value; The basic conduction strength and the cascade influence coefficient are combined to obtain the conduction strength between adjacent nodes.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images