Visual prediction method and system for lower-layer supply chain risk based on visual technology

Abstract

The invention discloses a lower-layer supply chain risk visual prediction method and system based on a visual technology, and belongs to the technical field of supply chain risk assessment, and the method specifically comprises the steps: obtaining related data of a lower-layer supply chain, carrying out the preprocessing, obtaining first monitoring data, constructing a target recognition model, and carrying out the recognition of the first monitoring data; identifying and extracting spatio-temporal features and time sequence features of the first monitoring data, constructing a multi-level risk assessment model, assessing a comprehensive risk score of a lower-layer supply chain, and predicting future risks; the key elements in the supply chain can be accurately positioned by performing feature extraction and recognition on the collected multi-modal data, and the potential risk factors and the comprehensive risk value can be effectively recognized and visually displayed by combining the risk assessment model.

Description

Technical Field

[0001] The present invention belongs to the technical field of supply chain risk assessment, and specifically discloses a visual prediction method and system for lower-level supply chain risks based on visual technology. Background Art

[0002] The supply chain has evolved from a simple linear chain into a complex network structure with many levels and links. The lower-level supply chain, as the basic level that directly supports the production and operation of core enterprises, covers many key nodes such as raw material suppliers, parts manufacturers, and logistics subcontractors. The stability of its operation plays a vital role in the stable operation of the entire supply chain.

[0003] Traditional supply chain risk prediction methods mainly focus on statistical analysis of historical transaction data, and predict common risks such as raw material price fluctuations and supplier delivery delays by building regression models and time series models. However, this method that relies on historical data has many disadvantages. On the one hand, historical data reflects the past market environment and operating conditions; on the other hand, questionnaires and manual experience judgment are also commonly used risk assessment methods. The questionnaire process is cumbersome, time-consuming, and the responses are highly subjective.

[0004] Supply chain risk management based on visual technology is still in the exploratory stage. Although there have been some attempts, most applications are still concentrated on the static visual analysis and object identification level, and are unable to fully integrate multimodal data and dynamically changing information for efficient risk prediction and visualization. Therefore, the proposal of a lower-level supply chain risk visualization prediction method based on visual technology aims to overcome the limitations of traditional supply chain risk management methods, and combine modern visual technology with big data analysis to provide a risk assessment and prediction method based on multimodal fusion, dynamic update, and real-time feedback of visual data. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a method and system for visual prediction of lower-level supply chain risks based on visual technology, which can conduct comprehensive analysis, real-time update and dynamic feedback of multimodal data and intelligent decision support, and can improve the risk prediction accuracy and response speed of supply chain finance.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The visual prediction method of lower-level supply chain risk based on visual technology includes:

[0008] Obtain relevant data of the lower supply chain and perform preprocessing to obtain first monitoring data;

[0009] Constructing a target recognition model to recognize and extract the spatiotemporal features and time series features of the first monitoring data;

[0010] Construct a multi-level risk assessment model to evaluate the comprehensive risk score of the lower-level supply chain and predict future risks.

[0011] Specifically, the target recognition model is constructed to recognize and extract the spatiotemporal features and time series features of the first monitoring data, including:

[0012] Constructing a target recognition model and training the target recognition model to obtain a trained target recognition model;

[0013] The first monitoring data is input into the trained target recognition model to extract the spatiotemporal features and time series features of the image in the first monitoring data.

[0014] Specifically, the building of the target recognition model and training the target recognition model include:

[0015] According to the characteristics and scenarios of the lower-level supply chain, select CNN and LSTM networks to build the target recognition model, and configure the number, order, and parameters of convolutional layers, pooling layers, and fully connected layers;

[0016] The labeled public data set is divided into a training set, a validation set, and a test set. The target recognition model is trained using the training set. The parameters of the target recognition model are adjusted through the back propagation algorithm and the optimizer. The loss function is set. The target recognition model is trained until the loss function converges and remains unchanged, and the training is stopped.

[0017] After the training is completed, the target recognition model is finally evaluated using the test set. The accuracy, recall rate, and F1 value are calculated to measure the performance of the target recognition model. If the model performance meets the expected requirements, a trained target recognition model is obtained. If the model performance does not meet the expected requirements, the target recognition model is optimized to obtain a trained target recognition model.

[0018] Specifically, the extracting of the spatiotemporal features and the time series features of the image in the first monitoring data includes:

[0019] The first monitoring data is input into the trained target recognition model, and the pixel coordinates in a window in any image in the first monitoring data are set as , at time and The grayscale values ​​are and , Represents a unit time interval, It represents the change of the horizontal axis within a unit time interval. Represent the change of the ordinate within a unit time interval and establish the optical flow constraint equation;

[0020] Establish an optical flow constraint equation for multiple pixel points in a window in any image in the first monitoring data, solve the overdetermined equation group, that is, multiple optical flow constraint equations, and calculate the optical flow vector , determine the movement speed and direction of an object in any image in the first monitoring data, and obtain the spatiotemporal characteristics of the image in the first monitoring data;

[0021] The long-term dependency and dynamic changes of the video in the first monitoring data are captured, and the temporal features in the video are extracted.

[0022] Specifically, the multi-level risk assessment model is constructed to evaluate the comprehensive risk score of the lower supply chain and predict future risks, including:

[0023] Construct a multi-level risk assessment model, including first-level assessment, second-level assessment and third-level assessment;

[0024] Calculate the comprehensive risk score of the lower supply chain based on the results of the first, second and third level assessments;

[0025] Predict future risks in the lower-level supply chain and identify potential high-risk areas.

[0026] Specifically, the multi-level risk assessment model is constructed, including first-level assessment, second-level assessment and third-level assessment, including:

[0027] In the first level of evaluation, the extracted spatiotemporal features and time series features are preliminarily identified to obtain the risk prediction label of each data sample in the first monitoring data, including: no risk and risky;

[0028] In the second-level assessment, the spatiotemporal features and time series features corresponding to the risky data samples in the first-level assessment are detected and classified for anomalies, potential abnormal behaviors or problem areas are identified, and the abnormal behaviors or problem areas are classified;

[0029] In the third assessment, the outputs of the first and second level assessments are integrated to quantify the risk using a multi-level risk scoring model;

[0030] By locally training a multi-level risk assessment model, introducing differential privacy technology, sharing model parameters through a federal aggregation mechanism, and utilizing adaptive multi-model weighted aggregation, the first weight is dynamically assigned according to the performance indicators of each model on the local data set. Based on the differences in the models, the same data nodes or adjacent data nodes are used for training to obtain model performance indicators, and the second weight is dynamically assigned. The model parameters corresponding to the first weight and the model parameters corresponding to the second weight are weightedly summed to obtain the model shared parameters.

[0031] Specifically, the future risks of the lower supply chain are predicted and potential high-risk areas are identified, including:

[0032] Using clustering algorithms, different links of the lower supply chain are clustered according to risk scores. The center point of each cluster represents the average risk level of the area corresponding to the center point.

[0033] Based on the average risk level of each area, a risk heat map is generated to identify potential high-risk areas.

[0034] Specifically, the relevant data of the lower supply chain includes: production environment monitoring data, warehouse site monitoring data, financial data and logistics and transportation monitoring data.

[0035] A visual prediction system for lower-level supply chain risks based on visual technology, used to implement the visual prediction method for lower-level supply chain risks based on visual technology, comprising: a data processing module, a feature extraction module and a risk assessment and prediction module;

[0036] The data processing module is used to obtain relevant data of the lower supply chain and perform preprocessing to obtain first monitoring data;

[0037] The feature extraction module is used to construct a target recognition model to identify and extract the spatiotemporal features and time series features of the first monitoring data;

[0038] The risk assessment and prediction module is used to construct a multi-level risk assessment model, evaluate the comprehensive risk score of the lower-level supply chain, and predict future risks.

[0039] Specifically, the risk assessment and prediction module includes: a first assessment unit, a second assessment unit, a third assessment unit, a comprehensive assessment unit and a prediction unit;

[0040] The first evaluation unit is used to preliminarily identify a risk prediction label for each data sample in the first monitoring data;

[0041] The second evaluation unit is used to perform anomaly detection and classification on the spatiotemporal features and time series features corresponding to the risky data samples in the first layer evaluation;

[0042] The third assessment unit is used to integrate the outputs of the first-level assessment and the second-level assessment and use a multi-level risk scoring model to quantitatively score the risk;

[0043] The comprehensive assessment unit is used to comprehensively calculate the risks of the lower supply chain;

[0044] The prediction unit is used to predict future risks of the lower supply chain and identify potential high-risk areas.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention proposes a visual prediction method for lower-level supply chain risks based on visual technology. By extracting and identifying features from the collected multimodal data, it can accurately locate the key elements in the supply chain. Combined with the risk assessment model, it can effectively identify potential risk factors and assess the comprehensive risk value.

[0047] 2. The present invention proposes a visual prediction method for lower-level supply chain risks based on visual technology. By combining data from different modalities, it can conduct a comprehensive analysis of supply chain risks from multiple dimensions, so that the risk assessment model can more comprehensively assess risks and identify potential correlation factors, further improving the accuracy of risk prediction.

[0048] 3. The present invention proposes a visual prediction method and system for lower-level supply chain risks based on visual technology, which can present complex risk assessment results to decision makers in a graphical and visual manner. Through intuitive methods such as charts, heat maps, and risk maps, potential risk areas in the supply chain can be quickly understood, providing strong support for decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 A flow chart of a visual prediction method for lower-level supply chain risks based on visual technology provided by the present invention;

[0050] Figure 2 A schematic diagram of model parameter sharing provided by the present invention;

[0051] Figure 3 This is an architecture diagram of the lower-level supply chain risk visualization prediction system based on visual technology provided by the present invention. DETAILED DESCRIPTION

[0052] The present application is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements can also be made without departing from the concept of the present application. These all belong to the protection scope of the present application.

[0053] In the present invention, unless otherwise specified, all parts and percentages are weight units, and the equipment and raw materials used can be purchased from the market or are commonly used in the art. The methods in the following embodiments, unless otherwise specified, are conventional methods in the art. The components or equipment in the following embodiments, unless otherwise specified, are universal standard parts or components known to those skilled in the art, and their structures and principles are known to those skilled in the art through technical manuals or conventional experimental methods.

[0054] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present invention. However, one or more embodiments may also be implemented by those skilled in the art without these specific details.

[0055] Example 1

[0056] In this application, the visual prediction method for the lower-level supply chain risk based on visual technology is specifically applied to the visual prediction of the lower-level supply chain financial risk. The lower-level supply chain is composed of core enterprises, core agents and distributors. The core enterprises sell to core agents through financing. The core agents need to pay advance payments before delivery. After the core agents receive the goods, they sell them to distributors, and there are financial risks in the process.

[0057] See also Figure 1-Figure 2 , an embodiment provided by the present invention: a visual prediction method for lower-level supply chain risk based on visual technology, comprising the following specific steps:

[0058] Step S1: Obtain relevant data of the lower supply chain, including: production environment monitoring data, storage site monitoring data, financial data and logistics transportation monitoring data, and perform preprocessing to obtain first monitoring data;

[0059] The production environment monitoring data in step S1 includes: by deploying high-definition cameras and industrial cameras, real-time capture of the status of the production line, collection of the operation of production equipment, images of production items, abnormal stoppages or failures of the production line, storage environment, physical conditions in the production environment, etc.;

[0060] Warehouse site monitoring data includes obtaining information such as the type, quantity, and storage location of goods in the warehouse by deploying RFID (radio frequency identification) devices and cameras;

[0061] Financial data includes: transaction data, invoice information, credit rating, etc. Transaction data includes order records, payment records, credit records of suppliers and customers, historical data of product purchases and deliveries, income, expenditure, balance sheet, cash flow, etc.;

[0062] Logistics and transportation monitoring data include: vehicle location, transportation route, driver status, and vehicle status through vehicle-mounted cameras and GPS positioning systems;

[0063] The preprocessing includes: data cleaning, removing noise data and redundant data; image denoising, using an adaptive median filter algorithm to remove noise caused by light interference, equipment electronic noise, etc., to improve image clarity; image enhancement, using an improved algorithm based on histogram equalization to stretch the image contrast, making the detail features in the image more obvious, and performing statistical analysis on the image grayscale histogram to redistribute pixel grayscale values ​​and enhance the distinction between different areas; multimodal data fusion, including time synchronization and space synchronization, using Kalman filtering, particle filtering and other methods to fuse data from different sources;

[0064] The principle of the adaptive median filtering algorithm is: if the minimum value of a pixel at a point in the window is less than the median value and the median value is less than the maximum value, then the pixel value of the output image at that point is the median value of the pixel values ​​in the window; otherwise, the pixel value of the output image at that point is the pixel value of the input image at that point.

[0065] Time synchronization: Since different types of data may have different collection times, all data need to be synchronized through timestamps. By using high-precision clocks (such as GPS synchronized clocks) to accurately timestamp various devices, the time consistency between data sources is ensured, so that multimodal data can be accurately connected on the time axis; Spatial synchronization:

[0066] Step S2: constructing a target recognition model to identify and extract the spatiotemporal features and time series features of the first monitoring data;

[0067] The specific steps of step S2 are:

[0068] Step S201: constructing a target recognition model and training the target recognition model to obtain a trained target recognition model;

[0069] Step S202: input the first monitoring data into the trained target recognition model to extract the spatiotemporal features and time series features of the image in the first monitoring data.

[0070] The specific steps of step S201 include:

[0071] Step S2011: According to the characteristics and scenarios of the lower supply chain, select CNN and LSTM networks to build a target recognition model, and configure the number, order, and parameters of convolutional layers, pooling layers, and fully connected layers;

[0072] In this embodiment, for example, for a more complex task of identifying the operating status of equipment, a deeper network structure may need to be designed to fully learn the subtle features and complex patterns of the equipment; while for the identification of the inventory quantity of raw materials, a relatively shallow network structure may be sufficient. In the design process, the computational complexity and training time of the model also need to be considered to ensure efficient operation in practical applications.

[0073] Step S2012: Divide the annotated public data set into a training set, a validation set, and a test set, use the training set to train the target recognition model, adjust the parameters of the target recognition model through the back propagation algorithm and the optimizer, set the loss function, train the target recognition model until the loss function converges and remains unchanged, and stop training;

[0074] In this embodiment, during the training process, the validation set is used to monitor the performance of the model to prevent overfitting. Overfitting means that the model performs well on the training set, but performs poorly on new, unseen data (test set). By evaluating the performance on the validation set, the model's hyperparameters (such as learning rate, regularization parameters, etc.) can be adjusted in time to optimize the model's generalization ability;

[0075] Step S2013: After the training is completed, the target recognition model is finally evaluated using the test set, and the accuracy, recall rate, and F1 value are calculated to measure the performance of the target recognition model. If the model performance meets the expected requirements, a trained target recognition model is obtained. If the model performance does not meet the expected requirements, the target recognition model is optimized to obtain a trained target recognition model.

[0076] Preferably, the optimization method includes adjusting the model architecture, increasing training data, adopting data enhancement techniques (such as rotating, flipping, and scaling images), adjusting hyperparameters, etc., so that the model can accurately and stably identify key elements of the supply chain through continuous iterative optimization.

[0077] The specific steps of step S202 include:

[0078] Step S2021: Input the first monitoring data into the trained target recognition model, and set the pixel coordinates in a window in any image in the first monitoring data as , at time and The grayscale values ​​are and , Represents a unit time interval, It represents the change of the horizontal axis within a unit time interval. Represents the change in the ordinate within a unit time interval, and establishes the optical flow constraint equation. The specific formula is:

[0079] ;

[0080] in, represents the grayscale gradient of any image in the first monitoring data in the horizontal direction, represents the grayscale gradient of any image in the first monitoring data in the vertical direction, u represents the component of the optical flow in the horizontal direction, v represents the component of the optical flow in the vertical direction, represents the time gradient;

[0081] In this embodiment, optical flow refers to the motion vector of a pixel point in an image on the image plane over time. The optical flow constraint equation is derived based on several basic assumptions: the constant brightness assumption and the small motion assumption. The constant brightness assumption is specifically: within a short time interval, the brightness of a point on the surface of an object remains unchanged during the motion process, that is, the pixel grayscale values ​​corresponding to the same object point in images at different times are the same. For example, in a continuously shot video, the grayscale value of a specific point on a moving car in two adjacent frames of images will not change immediately due to the movement of the car; the small motion assumption is specifically: the movement of the object between adjacent frames is small, that is, , and are all very small quantities, and using Taylor series Expand it and set the brightness constant assumption to get the optical flow constraint equation;

[0082] Step S2022: Establish an optical flow constraint equation for multiple pixel points in a window in any image in the first monitoring data, solve the overdetermined equation group, that is, multiple optical flow constraint equations, and calculate the optical flow vector , determine the movement speed and direction of an object in any image in the first monitoring data, and obtain the spatiotemporal characteristics of the image in the first monitoring data;

[0083] In this embodiment, the optical flow constraint equation is essentially a binary linear equation. It is impossible to uniquely determine the two unknowns u and v by only one equation. Multiple optical flow constraint equations are required to solve the optical flow equation. The solution method of the optical flow equation is: 1) Lucas-Kanade method: Assume that the optical flow is constant in a small spatial neighborhood, that is, in a small window centered on a certain pixel point, all pixels have the same optical flow vector , within this small window, an optical flow constraint equation can be established for each pixel point, thus obtaining an overdetermined set of equations (because the number of pixels in the window is usually much larger than the number of unknowns); 2) Horn-Schunck method: Based on the optical flow constraint equation, a smoothness constraint condition is introduced, assuming that the optical flows of adjacent pixels should be similar, that is, the optical flow field should be smooth, and the optical flow is solved by minimizing an energy function containing optical flow constraints and smoothness constraints; these solutions are all based on the optical flow constraint equation, combined with different assumptions and constraints, to more accurately calculate the optical flow vector of the pixel points in the image, thereby obtaining the motion information of the object.

[0084] Step S2023: Capture the long-term dependency and dynamic changes of the video in the first monitoring data, and extract the timing features in the video.

[0085] For example, abnormal changes in the transportation process are learned from consecutive video frames, or abnormal pauses in the production line are identified.

[0086] Step S3: Construct a multi-level risk assessment model to evaluate the comprehensive risk score of the lower-level supply chain and predict future risks.

[0087] The specific steps of step S3 are:

[0088] Step S301: construct a multi-level risk assessment model, including first-level assessment, second-level assessment and third-level assessment, and train the multi-level risk assessment model locally;

[0089] Step S302: Calculate the comprehensive risk score of the lower supply chain based on the results of the first-level assessment, the second-level assessment, and the third-level assessment. The specific formula is:

[0090] ;

[0091] Among them, R represents the comprehensive risk score of the lower supply chain, n represents the total number of risks, represents the score of the ith risk, represents the weight of the ith risk;

[0092] Step S303: predict the future risks of the lower supply chain and identify potential high-risk areas.

[0093] The specific steps of step S301 are:

[0094] Step S3011: In the first level evaluation, the extracted spatiotemporal features and time series features are preliminarily identified to obtain a risk prediction label for each data sample in the first monitoring data, including: no risk and risk;

[0095] In this embodiment, the extracted spatiotemporal features and time series features are preliminarily classified by support vector machine (SVM), random forest and other algorithms. The output of the classification model is usually a preliminary identification of risk, such as no risk and risk; no risk in the risk prediction label: the area or behavior is judged to be normal and no measures are required; risk in the risk prediction label: potential risks are identified and further processing is required. The specific processing method is to enter the next stage of risk identification and problem location, that is, the second-level assessment;

[0096] Step S3012: In the second-level assessment, anomaly detection and classification are performed on the spatiotemporal features and time series features corresponding to the risky data samples in the first-level assessment, potential abnormal behaviors or problem areas are identified, and the abnormal behaviors or problem areas are classified;

[0097] In this embodiment, the anomaly detection methods include: Isolation Forest, DBSCAN, One-Class SVM, etc., and the classification models usually include: Support Vector Machine, Random Forest, etc. Through the mid-level risk identification and classification models, the potential risk areas in the supply chain can be preliminarily determined. These potential risk areas may include: irregular stacking of items in the warehouse, abnormal transportation routes, supplier credit risks, etc.;

[0098] Step S3013: In the third assessment, the outputs of the first and second level assessments are integrated to quantitatively score the risks using a multi-level risk scoring model;

[0099] Step S3014: Through local training of a multi-level risk assessment model, differential privacy technology is introduced, model parameters are shared through a federal aggregation mechanism, and adaptive multi-model weighted aggregation is used to dynamically assign a first weight according to the performance indicators of each model on the local data set. Based on the differences in the models, the same data nodes or adjacent data nodes are used for training to obtain model performance indicators, and a second weight is dynamically assigned. The model parameters corresponding to the first weight and the model parameters corresponding to the second weight are weighted summed to obtain the model shared parameters.

[0100] like Figure 2 As shown, the same data is data contained in the first layer, the second layer, and the third layer, and the adjacent data is the common data points of the first layer and the second layer, and the same data is eliminated.

[0101] In this embodiment, in the supply chain risk assessment, the data used contains a large amount of sensitive information. The differential privacy technology can ensure that even if a data entry is deleted or changed in the data set, the final analysis result will not be affected, thus avoiding the risk of individual information leakage. The basic principle of differential privacy is that if a data entry is deleted or changed in a data set, the final analysis result should hardly be affected. Any data, whether it appears in the data set or not, should not significantly change the result. This design ensures the protection of individual privacy.

[0102] In the supply chain scenario, three layers of assessment are used for risk scoring. Due to different data distributions, different models may converge slowly or not converge during the aggregation process. The model parameters are shared through a federated aggregation mechanism. The specific steps are: 1) In the preparation stage, the multi-level risk assessment model and collected data are prepared; 2) In the upload stage, the data is uploaded to a local or central server using blockchain and differential privacy technology; 3) In the aggregation calculation stage, the uploaded data is verified, and an adaptive aggregation algorithm (an improved FedProx to solve the Non-IID problem) is used to aggregate and calculate the parameters of the multi-level risk assessment model, and abnormal nodes are processed (if the node times out and does not respond (>30s), the parameters of the previous round are used to participate in the aggregation); 4) In the distribution stage, a global model is generated based on the global model parameters obtained by the aggregation calculation, and the global model is transmitted to the designated server to obtain a trained multi-level risk assessment model, so that the model can learn a wider range of features and patterns.

[0103] The risk scoring model quantifies the risk based on the weighted average of each feature or a machine learning model. The risk score usually ranges from 0 to 1, where 0 indicates no risk and 1 indicates the highest risk. For example, for warehouse management, the risk score can be calculated based on multiple factors such as inventory change rate and inventory status.

[0104] The specific steps of step S303 are:

[0105] Step S3031: using a clustering algorithm, clustering different links of the lower supply chain according to risk scores, with the center point of each cluster representing the average risk level of the area corresponding to the center point;

[0106] Step S3032: Generate a risk heat map based on the average risk level of each area to identify potential high-risk areas.

[0107] In this embodiment, in the risk heat map, the darker the color, the higher the risk score. By analyzing the risk score distribution of each area, the area with a higher risk score, that is, the potential high-risk area, is identified. Specifically, the threshold of the risk score is set by those skilled in the art to identify the potential high-risk area.

[0108] Example 2

[0109] See also Figure 3 ,Another embodiment provided by the present invention: A visual prediction system for lower-level supply chain risks based on visual technology, comprising: a data processing module, a feature extraction module and a risk assessment and prediction module;

[0110] The data processing module is used to obtain relevant data of the lower supply chain and perform preprocessing to obtain first monitoring data;

[0111] The feature extraction module is used to construct a target recognition model to identify and extract the spatiotemporal features and time series features of the first monitoring data;

[0112] The risk assessment and prediction module is used to construct a multi-level risk assessment model, evaluate the comprehensive risk score of the lower-level supply chain, and predict future risks.

[0113] The risk assessment and prediction module includes: a first assessment unit, a second assessment unit, a third assessment unit, a comprehensive assessment unit and a prediction unit;

[0114] The first evaluation unit is used to preliminarily identify a risk prediction label for each data sample in the first monitoring data;

[0115] The second evaluation unit is used to perform anomaly detection and classification on the spatiotemporal features and time series features corresponding to the risky data samples in the first layer evaluation;

[0116] The third assessment unit is used to integrate the outputs of the first-level assessment and the second-level assessment and use a multi-level risk scoring model to quantitatively score the risk;

[0117] The comprehensive assessment unit is used to comprehensively calculate the risks of the lower supply chain;

[0118] The prediction unit is used to predict future risks of the lower supply chain and identify potential high-risk areas.

[0119] In addition, the parts of the above-mentioned technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0120] The specific implementation modes as described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation mode of the present invention and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A visual prediction method for lower-level supply chain risks based on visual technology, characterized in that: include: Obtain relevant data of the lower supply chain and perform preprocessing to obtain first monitoring data; Constructing a target recognition model to recognize and extract the spatiotemporal features and time series features of the first monitoring data; Construct a multi-level risk assessment model to evaluate the comprehensive risk score of the lower-level supply chain and predict future risks.

2. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 1, characterized in that: The target recognition model is constructed to recognize and extract the spatiotemporal features and time series features of the first monitoring data, including: Constructing a target recognition model and training the target recognition model to obtain a trained target recognition model; The first monitoring data is input into the trained target recognition model to extract the spatiotemporal features and time series features of the image in the first monitoring data.

3. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 2, characterized in that: The target recognition model is constructed and the target recognition model is trained, including: According to the characteristics and scenarios of the lower-level supply chain, select CNN and LSTM networks to build the target recognition model, and configure the number, order, and parameters of convolutional layers, pooling layers, and fully connected layers; The labeled public data set is divided into a training set, a validation set, and a test set. The target recognition model is trained using the training set. The parameters of the target recognition model are adjusted through the back propagation algorithm and the optimizer. The loss function is set. The target recognition model is trained until the loss function converges and remains unchanged, and the training is stopped. After the training is completed, the target recognition model is finally evaluated using the test set. The accuracy, recall rate, and F1 value are calculated to measure the performance of the target recognition model. If the model performance meets the expected requirements, a trained target recognition model is obtained. If the model performance does not meet the expected requirements, the target recognition model is optimized to obtain a trained target recognition model.

4. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 2, characterized in that: The step of extracting the spatiotemporal features and the temporal features of the image in the first monitoring data comprises: The first monitoring data is input into the trained target recognition model, and the pixel coordinates in a window in any image in the first monitoring data are set as , at time and The grayscale values ​​are and , Represents a unit time interval, It represents the change of the horizontal axis within a unit time interval. Represent the change of the ordinate within a unit time interval and establish the optical flow constraint equation; Establish an optical flow constraint equation for multiple pixel points in a window in any image in the first monitoring data, solve the overdetermined equation group, that is, multiple optical flow constraint equations, and calculate the optical flow vector , determine the movement speed and direction of an object in any image in the first monitoring data, and obtain the spatiotemporal characteristics of the image in the first monitoring data; The long-term dependency and dynamic changes of the video in the first monitoring data are captured, and the temporal features in the video are extracted.

5. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 1, characterized in that: The multi-level risk assessment model is constructed to evaluate the comprehensive risk score of the lower supply chain and predict future risks, including: Construct a multi-level risk assessment model, including first-level assessment, second-level assessment and third-level assessment; Calculate the comprehensive risk score of the lower supply chain based on the results of the first, second and third level assessments; Predict future risks in the lower-level supply chain and identify potential high-risk areas.

6. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 5, characterized in that: The multi-level risk assessment model is constructed, including first-level assessment, second-level assessment and third-level assessment, including: In the first level of evaluation, the extracted spatiotemporal features and time series features are preliminarily identified to obtain the risk prediction label of each data sample in the first monitoring data, including: no risk and risky; In the second-level assessment, the spatiotemporal features and time series features corresponding to the risky data samples in the first-level assessment are detected and classified for anomalies, potential abnormal behaviors or problem areas are identified, and the abnormal behaviors or problem areas are classified; In the third assessment, the outputs of the first and second level assessments are integrated to quantify the risk using a multi-level risk scoring model; By locally training a multi-level risk assessment model, introducing differential privacy technology, sharing model parameters through a federal aggregation mechanism, and utilizing adaptive multi-model weighted aggregation, the first weight is dynamically assigned according to the performance indicators of each model on the local data set. Based on the differences in the models, the same data nodes or adjacent data nodes are used for training to obtain model performance indicators, and the second weight is dynamically assigned. The model parameters corresponding to the first weight and the model parameters corresponding to the second weight are weightedly summed to obtain the model shared parameters.

7. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 5, characterized in that: The future risks of the lower supply chain are predicted and potential high-risk areas are identified, including: Using clustering algorithms, different links of the lower supply chain are clustered according to risk scores. The center point of each cluster represents the average risk level of the area corresponding to the center point. Based on the average risk level of each area, a risk heat map is generated to identify potential high-risk areas.

8. The visual prediction method for lower-level supply chain risk based on visual technology as claimed in claim 1, characterized in that: The relevant data of the lower supply chain includes: production environment monitoring data, warehouse site monitoring data, financial data and logistics and transportation monitoring data.

9. A visual prediction system for lower-level supply chain risks based on visual technology, used to implement the visual prediction method for lower-level supply chain risks based on visual technology as described in any one of claims 1 to 8, characterized in that: include: Data processing module, feature extraction module and risk assessment and prediction module; The data processing module is used to obtain relevant data of the lower supply chain and perform preprocessing to obtain first monitoring data; The feature extraction module is used to construct a target recognition model to identify and extract the spatiotemporal features and time series features of the first monitoring data; The risk assessment and prediction module is used to construct a multi-level risk assessment model, evaluate the comprehensive risk score of the lower-level supply chain, and predict future risks.

10. The visual prediction system for lower-level supply chain risks based on visual technology as claimed in claim 9, characterized in that: The risk assessment and prediction module includes: a first assessment unit, a second assessment unit, a third assessment unit, a comprehensive assessment unit and a prediction unit; The first evaluation unit is used to preliminarily identify a risk prediction label for each data sample in the first monitoring data; The second evaluation unit is used to perform anomaly detection and classification on the spatiotemporal features and time series features corresponding to the risky data samples in the first layer evaluation; The third assessment unit is used to integrate the outputs of the first-level assessment and the second-level assessment and use a multi-level risk scoring model to quantitatively score the risk; The comprehensive assessment unit is used to comprehensively calculate the risks of the lower supply chain; The prediction unit is used to predict future risks of the lower supply chain and identify potential high-risk areas.

Images