Supply chain financing post-loan risk detection method and device based on time sequence prediction

Abstract

The invention discloses a supply chain financing post-loan risk detection method and device based on time sequence prediction. The method comprises the steps of obtaining a time sequence; sequentially coding the time sequence by using a one-dimensional convolutional neural network to obtain a time sequence mode feature of a historical adjacent time period, the historical adjacent time period being a historical time period directly adjacent to the to-be-predicted time period; performing jump convolution on the time sequence by using a plurality of convolution kernels of different sizes to obtain time sequence mode characteristics of a plurality of historical cycle periods of different scales, the historical cycle periods being historical periods which are separated from the to-be-predicted period by a predetermined time period; fusing the time sequence mode characteristics of the historical adjacent time periods and the time sequence mode characteristics of the multiple historical periodic time periods of different scales to obtain fused time sequence mode characteristics; and re-calibrating the fused time sequence mode features to obtain a post-loan risk detection result of the target merchant in the to-be-predicted time period.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and device for detecting post-loan risks of supply chain financing based on time series prediction. Background Art

[0002] In the field of supply chain financing financial services, regular testing of customer-related loan risks is an important part of post-loan management. However, the current common supply chain financing post-loan risk management methods have the following difficulties.

[0003] First, supply chain financing involves multiple links and multiple participants, and the data of each participant is difficult to integrate and coordinate, resulting in low detection efficiency and inaccurate detection results; second, traditional detection methods only rely on directly related information to build models, which makes it impossible to deeply analyze the complex relationships and potential risks in the supply chain; third, in the supply chain scenario, how to build a model that can be quickly trained and dynamically predicted to ensure the real-time nature of post-loan risk prediction. Summary of the invention

[0004] The present application provides a method and device for detecting post-loan risk of supply chain financing based on time series prediction, which helps to improve the accuracy and efficiency of post-loan detection of supply chain financing. The technical solution is as follows.

[0005] In a first aspect, a method for detecting post-loan risk of supply chain financing based on time series prediction is provided, characterized in that the method comprises:

[0006] Acquire a time series, wherein the time series indicates a corresponding relationship between multiple time series variables related to the supply chain and time, wherein the multiple time series variables include supply chain data of a target merchant, supply chain data of an upstream supplier of the target merchant, supply chain data of a downstream distributor of the target merchant, logistics data of the target merchant, quality control data of the target merchant, and contract data of the target merchant, wherein the supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow;

[0007] Sequentially encoding the time series using a one-dimensional convolutional neural network to obtain temporal pattern features of historical adjacent time periods, where the historical adjacent time periods are historical time periods directly adjacent to the time period to be predicted;

[0008] Using a plurality of convolution kernels of different sizes to perform jump convolution on the time series, so as to obtain time series pattern features of a plurality of historical periodic time periods of different scales, wherein the historical periodic time period is a historical period of time separated from the time period to be predicted by a predetermined time period;

[0009] The time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales are fused to obtain fused time series pattern features;

[0010] The fused time series pattern features are recalibrated to obtain the post-loan risk detection result of the target merchant in the time period to be predicted.

[0011] In some embodiments, the one-dimensional convolutional neural network includes M convolution kernels with a width of w and a height of c, the height of the convolution kernel is the number of dimensions of the time series, the width of the convolution kernel represents the range of data of historical adjacent time periods that determine the post-loan risk of the target merchant in a time period, and the one-dimensional convolutional neural network is used to sequentially encode the time series to obtain the temporal pattern characteristics of the historical adjacent time periods, including:

[0012] The time series is convolved using M convolution kernels to obtain the temporal pattern characteristics of historical adjacent time periods.

[0013] In some embodiments, the multiple historical periodic time periods of different scales include a first time period, a second time period, and a third time period, the multiple convolution kernels of different sizes include a first convolution kernel, a second convolution kernel, and a third convolution kernel, the height of the first convolution kernel, the height of the second convolution kernel, and the height of the third convolution kernel are all the number of dimensions of the time series, the width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period, and the use of multiple convolution kernels of different sizes to perform jump convolution on the time series to obtain time series pattern features of multiple historical periodic time periods of different scales includes:

[0014] sequentially encoding the time series using the first convolution kernel to obtain a time series pattern feature of a first time period;

[0015] sequentially encoding the time series using the second convolution kernel to obtain a time series pattern feature of a second time period;

[0016] The time series is sequentially encoded using the third convolution kernel to obtain the time series pattern characteristics of the third time period.

[0017] In some implementations, the fusing of the temporal pattern features of the historical adjacent time periods and the temporal pattern features of the historical periodic time periods of multiple different scales includes:

[0018] Performing global mean pooling on the temporal pattern features of the historical adjacent time periods and the temporal pattern features of the historical periodic time periods of multiple different scales;

[0019] The time series pattern features of the historical adjacent time periods after global mean pooling and the time series pattern features of the historical periodic time periods of multiple different scales are spliced ​​to obtain fused time series pattern features.

[0020] In some embodiments, recalibrating the fused temporal pattern features comprises:

[0021] Using a global mean pooling method, compressing the fused temporal pattern features to obtain compressed temporal pattern features;

[0022] Performing an excitation operation on the compressed temporal pattern features to obtain attention scores of multiple convolution channels;

[0023] The attention scores of the multiple convolution channels are multiplied by the fused temporal pattern features.

[0024] In a second aspect, a supply chain financing post-loan risk detection device based on time series prediction is provided, the device comprising:

[0025] an acquisition unit, configured to acquire a time series, wherein the time series indicates a corresponding relationship between a plurality of time series variables related to a supply chain and time, wherein the plurality of time series variables include supply chain data of a target merchant, supply chain data of an upstream supplier of the target merchant, supply chain data of a downstream distributor of the target merchant, logistics data of the target merchant, quality control data of the target merchant, and contract data of the target merchant, wherein the supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow;

[0026] A convolution unit is used to sequentially encode the time series using a one-dimensional convolutional neural network to obtain time series pattern features of historical adjacent time periods, where the historical adjacent time period is a historical time period directly adjacent to the time period to be predicted; and to perform jump convolution on the time series using a plurality of convolution kernels of different sizes to obtain time series pattern features of historical periodic time periods of a plurality of different scales, where the historical periodic time period is a historical period separated from the time period to be predicted by a predetermined time period;

[0027] A fusion unit, used to fuse the time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales to obtain fused time series pattern features;

[0028] The recalibration unit is used to recalibrate the fused time series pattern features to obtain the post-loan risk detection result of the target merchant in the time period to be predicted.

[0029] In some embodiments, the one-dimensional convolutional neural network includes M convolution kernels with a width of w and a height of c, the height of the convolution kernel is the number of dimensions of the time series, the width of the convolution kernel represents the range of data of historical adjacent time periods that determines the post-loan risk of the target merchant in a time period, and the convolution unit is used to convolve the time series using M convolution kernels to obtain the temporal pattern characteristics of the historical adjacent time periods.

[0030] In some embodiments, the multiple historical period time periods of different scales include a first time period, a second time period and a third time period, the multiple convolution kernels of different sizes include a first convolution kernel, a second convolution kernel and a third convolution kernel, the height of the first convolution kernel, the height of the second convolution kernel and the height of the third convolution kernel are all the number of dimensions of the time series, the width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period, and the convolution unit is used to sequentially encode the time series using the first convolution kernel to obtain the time series pattern characteristics of the first time period; sequentially encode the time series using the second convolution kernel to obtain the time series pattern characteristics of the second time period; and sequentially encode the time series using the third convolution kernel to obtain the time series pattern characteristics of the third time period.

[0031] In some embodiments, the fusion unit is used to perform global mean pooling on the timing pattern features of the historical adjacent time periods and the timing pattern features of the historical period periods of multiple different scales, respectively; and to splice the timing pattern features of the historical adjacent time periods after global mean pooling and the timing pattern features of the historical period periods of multiple different scales to obtain fused timing pattern features.

[0032] In some embodiments, the recalibration unit is used to compress the fused temporal pattern features by adopting a global mean pooling method to obtain compressed temporal pattern features; perform an excitation operation on the compressed temporal pattern features to obtain attention scores of multiple convolution channels; and multiply the attention scores of the multiple convolution channels by the fused temporal pattern features.

[0033] The embodiments of the present application have the following beneficial effects:

[0034] The embodiment of the present application provides a supply chain financing post-loan risk detection method based on time series prediction. Compared with the traditional supply chain financing post-loan risk detection method, the embodiment of the present application integrates the scattered supply chain data into a unified time series, which includes various data elements related to the supply chain, such as industrial and commercial information, equity registration, transaction flow, suppliers, logistics, quality control, contracts, etc., so as to achieve comprehensive supply chain quantitative data preparation, which is helpful to integrate more comprehensive information during risk detection. On this basis, the embodiment of the present application uses a one-dimensional convolutional neural network to obtain the time series pattern characteristics of the historical adjacent time periods in the time series, and uses skip convolution to obtain the time series pattern characteristics of the multi-scale historical cycle time periods in the time series. The time series pattern characteristics of the historical adjacent time periods and the time series pattern characteristics of the multi-scale historical cycle time periods are fused and recalibrated for real-time prediction of post-loan risks of multi-source supply chain data, thereby helping to deeply analyze the complex relationships and potential risks in the supply chain, improve the accuracy and efficiency of post-loan detection of supply chain financing, and provide financial institutions with more accurate supply chain post-loan risk assessment, so that they can better manage risks and provide post-loan services. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is an architecture diagram of a supply chain financing post-loan risk detection system based on time series prediction provided by an embodiment of the present application;

[0036] Figure 2 It is a flow chart of a method for detecting post-loan risk of supply chain financing based on time series prediction provided in an embodiment of the present application;

[0037] Figure 3 It is a structural diagram of a short-term dependency mode component provided in an embodiment of the present application;

[0038] Figure 4 It is a structural diagram of a long-term periodic mode component provided in an embodiment of the present application;

[0039] Figure 5 It is a flow chart of the structure of a recalibration component, feature fusion and feature recalibration provided in an embodiment of the present application;

[0040] Figure 6 It is a structural schematic diagram of a supply chain financing post-loan risk detection device based on time series prediction provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0042] The embodiment of the present application is aimed at the post-loan risk detection scenario of supply chain financing, and proposes to use time series to construct complex information and relationships after the supply chain loan, and adopts a deep learning method based on convolutional neural network (CNN) and skip convolution to predict post-loan risks, so as to fully fit the supply chain financing scenario in model establishment. Specifically, the detection algorithm based on time series can not only integrate the scattered supply chain data into a multi-source time series, which includes industrial and commercial information, loan transactions, transaction flow, etc., but also find the complex relationship and implicit association of the data series with time as the axis, so as to integrate more comprehensive information for modeling during risk detection, and predict whether the merchant will not be able to repay the loan in multiple periods in the future. At the same time, it can also regularly evaluate the risk of multi-target merchant loan groups based on the upstream and downstream of the supply chain, which fully fits the specific nature of the supply chain financing scenario. This method improves and improves the accuracy and efficiency of post-loan detection and risk control of traditional supply chain financing, and can provide financial institutions with a higher level of decision support and risk management in the current complex supply chain environment. The embodiment of the present application is expected to provide financial institutions with more accurate supply chain post-loan risk assessment, so that they can better manage risks and provide post-loan services, and further bring more precise control to loans in the field of supply chain finance.

[0043] The embodiment of the present application uses time series to organize the complex relationships of the entire chain of customer supply chains, providing a new idea for organizing post-loan data for supply chain financing. In addition, a post-loan risk detection algorithm based on deep learning of convolutional neural networks (CNN) and skip convolution is provided, which helps to provide more accurate and stable monitoring and early warning for the entire chain of supply chain financing loans.

[0044] The following is an explanation of some terminology concepts involved in the embodiments of the present application.

[0045] Supply chain financing post-loan risks: refers to potential problems and threats that may arise after the supply chain loan is issued, including but not limited to delayed payment, quality issues, market fluctuations, legal disputes and credit risks. For financial institutions and enterprises, it is crucial to effectively manage these risks to ensure the security and sustainability of loans.

[0046] Attention Network: It is a special network structure in neural network. It draws on the idea of ​​human attention mechanism to enable the model to focus on the most important part of the input information to improve the performance of the model. AN calculates the weights of different parts of the input information so that the model can focus on the most important information.

[0047] Convolutional Neural Network: Convolutional neural networks use convolutional layers and pooling layers to reduce the number of parameters through local connections and weight sharing. The convolutional layer is responsible for feature extraction, and the pooling layer achieves dimensionality reduction and reduces the amount of calculation. Through layer-by-layer abstraction, the network can automatically learn and extract deep features of input data, which can then be used for various downstream tasks.

[0048] Skip convolution (skip connection convolution): skip convolution means to establish a skip connection between the encoder and decoder of the convolutional neural network, connecting the feature map in the encoder with the corresponding feature map in the decoder. The purpose of this is to help the decoder learn to better reconstruct details while avoiding information loss and improving segmentation accuracy. Through skip connections, the network can utilize semantic information at different levels to better retain detail information.

[0049] The following is an example of an application scenario of the embodiment of the present application.

[0050] In the multivariate forecasting scenario, modeling complex nonlinear time series relationships, the length patterns of time series relationships, and the dependencies between variables are all key issues. Many existing statistical studies can only model linear correlations in time series variables, and such methods often cannot predict multiple time series variables at the same time, so they cannot be directly applied to supply chain financing post-loan forecasting. At the same time, these traditional supply chain financing post-loan detections usually rely on isolated databases and static data sources to detect the risks of a single merchant. Therefore, the following challenges are faced:

[0051] (1) Since the supply chain involves multiple links and multiple participants, each participant has independent data. These data are usually stored in different systems and are difficult to integrate and coordinate. Existing technologies do not fully explore the potential connections between these multi-source data, making it difficult to effectively capture the short-term and long-term temporal pattern characteristics in the supply chain time series.

[0052] (2) Traditional prediction methods only rely on directly related information to build models, which makes it difficult to integrate multi-hop potential information related to the merchants to be tested, and thus cannot deeply analyze the complex relationships and potential risks in the supply chain.

[0053] (3) In the supply chain scenario, a specific merchant usually has upstream suppliers and downstream distributors. When predicting that a supplier is at risk of not being able to repay the loan in the future, traditional detection methods cannot dynamically and in real time predict whether the merchant is also affected. Therefore, it is very important to build a model that can be quickly trained and dynamically predicted to ensure the real-time nature of post-loan risk prediction.

[0054] At present, most existing models use recurrent neural networks or their variant models to model complex time series relationships, which cannot solve the above challenges well. In combination with the current research status, the embodiment of the present application integrates the scattered supply chain data into a unified time series, which includes various data elements related to the supply chain, such as supplier data, logistics data, quality control data, contract data, etc., so as to achieve comprehensive supply chain quantitative analysis, and proposes a supply chain merchant post-loan risk prediction model based on convolutional neural networks and multi-scale time series pattern features. At the same time, the time series pattern features of historical adjacent time periods and the time series pattern features of historical cycle time periods in the time series are captured for modeling.

[0055] The embodiment of the present application first uses a standard one-dimensional convolutional neural network to sequentially encode the input time series to obtain the temporal pattern characteristics of historical adjacent periods. Secondly, based on the dilated convolution, the skip convolution is designed, and the skip convolution kernels of different sizes are used to sequentially encode the input time series to obtain the temporal pattern characteristics of multiple historical period periods of different scales. At the same time, for the above two types of convolution operations, the convolution kernel height is set to be the same as the number of dimensions of the time series to be predicted, so as to capture the local dependencies between regions in the time series. Finally, for the captured temporal pattern characteristics of multiple scales and the features mapped by multiple convolution kernels, the temporal pattern characteristics of multiple scales are further fused and recalibrated, and the fused and recalibrated features are input into the fully connected layer for real-time prediction of post-loan risks of supply chain data. Compared with the traditional post-loan detection method of supply chain financing, the embodiment of the present application improves the efficiency and accuracy of lending financial institutions in post-loan risk detection and promotes the transparency of supply chain loans. This will bring safer and more accurate risk detection to the field of supply chain finance.

[0056] Please refer to the attached Figure 1 , attached Figure 1 This is a diagram of the architecture of a supply chain financing post-loan risk detection system based on time series prediction provided by the embodiment of the present application, Figure 1The system shown includes a short-term dependency pattern component, a long-term periodic pattern component, and a time series feature fusion and recalibration component. The short-term dependency pattern component is used to obtain the time series pattern features of historical adjacent time periods using a standard convolutional neural network. The long-term periodic pattern component is used to obtain the time series pattern features of historical periodic time periods using skip convolution. The time series feature fusion and recalibration component is used to further process the time series pattern features of historical adjacent time periods obtained by the short-term dependency pattern component and the time series pattern features of historical periodic time periods of different scales obtained by the long-term periodic pattern component. For example, the time series feature fusion and recalibration component obtains the attention score at the convolution channel level through mean pooling and fully connected layer conversion, and uses the attention score to recalibrate the time series features to enhance the features that are beneficial to the prediction target and suppress the features that are less useful. The short-term dependency pattern component is connected in parallel with the long-term periodic pattern component, and the short-term dependency pattern component and the long-term periodic pattern component are respectively connected in series with the time series feature fusion and recalibration component. Appendix Figure 1 The system shown is applied to a computer device, for example, Figure 1 Each component in the system shown is software running on a computer device.

[0057] Please refer to the attached Figure 2 , attached Figure 2 Attached is a flow chart of a method for detecting post-loan risk of supply chain financing based on time series prediction provided by an embodiment of the present application. Figure 2 The method shown is applied to the Figure 1 In the system architecture shown, Figure 2 The method is executed by a computer device and includes the following steps.

[0058] S210, obtaining a time series.

[0059] The time series is, for example, a multi-source data series of post-loan risk. The time series indicates the corresponding relationship between multiple time series variables related to the supply chain and time. The multiple time series variables include the supply chain data of the target merchant, the supply chain data of the upstream supplier of the target merchant, the supply chain data of the downstream distributor of the target merchant, the logistics data of the target merchant, the quality control data of the target merchant, and the contract data of the target merchant. The supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow.

[0060] In some embodiments, the horizontal axis of the time series includes multiple time points arranged in chronological order, and the multiple time points include the current time point and the historical time points. The vertical axis of the time series includes time series variables of multiple dimensions related to the supply chain, for example, the vertical axis includes supplier data, logistics data, quality control data and contract data corresponding to each time point. Since the time series includes time series variables of multiple dimensions such as supplier data, logistics data, quality control data and contract data related to the supply chain, it is possible to integrate time series variables of multiple dimensions for post-loan risk detection.

[0061] In some embodiments, the time series contains data from multiple participants in the supply chain. For example, the time series contains data from multiple participants with upstream and downstream relationships in the supply chain at the same time point, such as the time series contains data from the first merchant at the first time point and data from the second merchant at the first time point, and the first merchant is the upstream supplier or downstream distributor of the second merchant. For example, the time series contains logistics data between Merchant 1 and Enterprise A, logistics information between Merchant 1 and Enterprise A, and the identification of the contract for sale between Merchant 1 and Enterprise A, logistics data between Merchant 2 and Enterprise B, logistics information between Merchant 2 and Enterprise B, and the identification of the contract for sale between Merchant 2 and Enterprise B on November 15, 2021. Since the time series contains data from multiple participants in the supply chain, the subsequent acquisition of features through the model and feature fusion will help to explore the potential connections between multi-source data.

[0062] S220, using a one-dimensional convolutional neural network to sequentially encode the time series to obtain the temporal pattern characteristics of historical adjacent time periods.

[0063] The historical adjacent period is the historical period in the time series that is directly adjacent to the time period to be predicted. The historical adjacent period can also be called a short-term time series segment. For example, the time period to be predicted is 10:00 on November 15, 2024, and the historical adjacent period is 9:00 on November 15, 2024. The time series pattern characteristics of the historical adjacent period are also called the time series pattern characteristics of the short-term dependency pattern. The time series pattern characteristics of the historical adjacent period characterize the impact of the data of the historical adjacent period in the time series on the post-loan risk of the target merchant in the time period to be predicted.

[0064] In some embodiments, standard CNN convolution operations are performed on time series to capture potential time series patterns in short time series segments and local dependencies between different merchants in the supply chain to obtain time series pattern characteristics of historical adjacent time periods.

[0065] In some embodiments, a one-dimensional convolutional neural network includes M convolution kernels. The M convolution kernels have the same width and the same height. For example, the width of the M convolution kernels is w, and the height of the M convolution kernels is c. M is a positive integer greater than or equal to 1. The height of the convolution kernel is the number of dimensions of the time series so as to capture the potential impact of changes in post-loan risks among multiple merchants. The width of the convolution kernel reflects the size of the convolution kernel mapped on the input time series. The width of the convolution kernel characterizes the range of historical data of adjacent time periods (such as industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flows) that determine the post-loan risk of the target merchant in a time period.

[0066] In some embodiments, the short-term dependency pattern component uses each of the M convolution kernels to convolve the time series separately to obtain M convolution results, and the M convolution results are the temporal pattern features of the historical adjacent time periods.

[0067] For example, please refer to the attached Figure 3 , attached Figure 3 The structure of the short-term dependency pattern component, i.e., the network architecture of the convolutional neural network, is shown. Figure 3 The short-term dependency pattern component shown in (a) contains 3 convolution kernels, each of which has a width of 3 and a height of 6. Figure 3 (a) shows that after the convolution kernel 1 with a width of 3 and a height of 6 is convolved with the region 1 in the time series, the convolution result 1 is obtained, and after the convolution kernel 3 with a width of 3 and a height of 6 is convolved with the region 3 in the time series, the convolution result 2 is obtained. Figure 3 (b) shows the data connection relationship between the input parameters, the convolution kernel in the short-term dependency pattern component, and the output results.

[0068] S230, using multiple convolution kernels of different sizes to perform jump convolution on the time series to obtain time series pattern features of historical period periods of multiple different scales.

[0069] In some embodiments, the historical cycle period is a historical period separated from the time period to be predicted by a predetermined time period. The historical cycle period and the time period to be predicted are periodic. For example, the historical cycle period and the time period to be predicted are the same period on different days. Alternatively, the historical cycle period and the time period to be predicted are the same period in different weeks but have the same week attribute. Alternatively, the historical cycle period and the time period to be predicted are the same period on the same day in different months. The historical cycle period can also be referred to as a long-term time series segment. The time series pattern characteristics of the historical cycle period characterize the impact of the data of the historical cycle period in the input time series on the post-loan risk of the target merchant in the time period to be predicted.

[0070] Some embodiments of the present application are based on dilated convolution design skip convolution operation to achieve the acquisition of long-term time series pattern features. When acquiring long-term time series pattern features, the factors of mutual dependence between different business circles, platforms, products, and customers are also considered.

[0071] In some embodiments, M convolution kernels are set to obtain the timing pattern features of historical periodic time periods. The M convolution kernels correspond to the M historical periodic time periods. The heights of the M convolution kernels are all the same. In some embodiments, the height of each of the M convolution kernels used to obtain the timing pattern features of the historical periodic time periods is set the same as the height of the convolution kernel used to obtain the timing pattern features of the historical adjacent time periods in S220. In some embodiments, the height of each of the M convolution kernels used to obtain the timing pattern features of the historical periodic time periods is the number of dimensions of the input time series. In some embodiments, the width of each of the M convolution kernels used to obtain the timing pattern features of the historical periodic time periods is different,

[0072] In some embodiments, a plurality of convolution kernels of different sizes for performing jump convolution include a first convolution kernel, a second convolution kernel, and a third convolution kernel. The acquired time series pattern features of the historical period time periods of multiple different scales include the time series pattern features of the first time period, the time series pattern features of the second time period, and the time series pattern features of the third time period. The height of the first convolution kernel, the height of the second convolution kernel, and the height of the third convolution kernel are all the number of dimensions of the time series. The width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period. The first convolution kernel is used to sequentially encode the time series to obtain the time series pattern features of the first time period; the second convolution kernel is used to sequentially encode the time series to obtain the time series pattern features of the second time period; the third convolution kernel is used to sequentially encode the time series to obtain the time series pattern features of the third time period. For example, the first convolution kernel among the M convolution kernels is used to capture the temporal pattern features with a daily cycle, the second convolution kernel is used to capture the temporal pattern features with a weekly cycle, and the third convolution kernel is used to capture the temporal pattern features with a monthly cycle.

[0073] For example, please refer to the attached Figure 4 , attached Figure 4 The structure of the long-term recurrent pattern component, i.e., the network architecture of skip convolution, is shown.

[0074] S240, fusing the time series pattern features of historical adjacent time periods and the time series pattern features of historical periodic time periods of multiple different scales to obtain fused time series pattern features;

[0075] The time series pattern features of historical adjacent time periods and the time series pattern features of historical period periods of multiple different scales are globally mean pooled respectively; the time series pattern features of historical adjacent time periods and the time series pattern features of historical period periods of multiple different scales after global mean pooling are concatenated, and the concatenated features are the fused multi-scale time series features.

[0076] Due to the fusion of short-term dependency pattern features and long-term cycle pattern features, the fused features can reflect both the impact of data from historical adjacent time periods on merchants and the impact of data from historical cycle time periods on merchants, thereby improving the accuracy of the prediction results.

[0077] S250, recalibrating the fused time series pattern features to obtain the post-loan risk detection results of the target merchant in the time period to be predicted.

[0078] By recalibrating the time series features, more accurate post-loan risk detection results can be obtained.

[0079] In a possible implementation, a global mean pooling method is used to compress the fused temporal pattern features to obtain compressed temporal pattern features; the compressed multi-scale temporal features are, for example, 1-dimensional vectors. Excitation operations are performed on the compressed temporal pattern features to obtain attention scores of multiple convolution channels; the attention scores of multiple convolution channels are multiplied with the fused temporal pattern features to obtain recalibrated multi-scale temporal features.

[0080] For example, see Figure 5 , Figure 5 The structure of the recalibration component, feature fusion and the process of feature recalibration are shown. The feature output by the one-dimensional convolutional neural network is a tensor of M*6*D dimensions. The features obtained after jump convolution include tensors of M*2*D dimensions, M*3*D dimensions and M*5*D dimensions. After global mean pooling and concatenation of the tensors of M*2*D dimensions, M*3*D dimensions and M*5*D dimensions, a tensor of M*4*1 dimensions is obtained. After compression of the tensor of M*4*1 dimensions, a tensor of M*1*1 dimensions is obtained. The tensor of M*1*1 dimensions is sequentially input into two layers of fully connected layers, and the tensor output by the fully connected layer is stimulated by the sigmoid function to obtain a tensor of M*1*1 dimensions. The tensor of M*1*1 dimensions is scaled to obtain a tensor of M*4*1 dimensions. The tensor of M*4*1 dimensions is the result of post-loan risk detection.

[0081] The post-loan risk detection method provided in this embodiment can effectively combine the risk of a single merchant with the risk of the entire supply chain financial system, thereby more comprehensively evaluating the risk in the supply chain. Finally, through the three components of the short-term dependency pattern component, the long-term cycle pattern component, and the time series feature fusion and recalibration component, the time series is predicted to obtain the post-loan risk of the target merchant at the current time point, so as to bring safer and more accurate risk detection services to the field of supply chain finance.

[0082] Figure 6 600 is a schematic diagram of a supply chain financing post-loan risk detection device based on time series prediction provided in an embodiment of the present application. The device 600 includes:

[0083] An acquisition unit 610 is used to acquire a time series, wherein the time series indicates a corresponding relationship between multiple time series variables related to the supply chain and time, wherein the multiple time series variables include supply chain data of a target merchant, supply chain data of an upstream supplier of the target merchant, supply chain data of a downstream distributor of the target merchant, logistics data of the target merchant, quality control data of the target merchant, and contract data of the target merchant, wherein the supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow;

[0084] The convolution unit 620 is used to sequentially encode the time series using a one-dimensional convolutional neural network to obtain time series pattern features of historical adjacent time periods, where the historical adjacent time period is a historical time period directly adjacent to the time period to be predicted; and to perform jump convolution on the time series using a plurality of convolution kernels of different sizes to obtain time series pattern features of historical periodic time periods of a plurality of different scales, where the historical periodic time period is a historical period separated from the time period to be predicted by a predetermined time period;

[0085] A fusion unit 630, configured to fuse the time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of different scales to obtain a fused time series pattern feature;

[0086] The recalibration unit 640 is used to recalibrate the fused time series pattern features to obtain the post-loan risk detection result of the target merchant in the time period to be predicted.

[0087] In some embodiments, the one-dimensional convolutional neural network includes M convolution kernels with a width of w and a height of c, the height of the convolution kernel is the number of dimensions of the time series, the width of the convolution kernel represents the range of data of historical adjacent time periods that determines the post-loan risk of the target merchant in a time period, and the convolution unit 620 is used to convolve the time series using M convolution kernels to obtain the temporal pattern characteristics of the historical adjacent time periods.

[0088] In some embodiments, the multiple historical period time periods of different scales include a first time period, a second time period and a third time period, the multiple convolution kernels of different sizes include a first convolution kernel, a second convolution kernel and a third convolution kernel, the height of the first convolution kernel, the height of the second convolution kernel and the height of the third convolution kernel are all the number of dimensions of the time series, the width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period, and the convolution unit 620 is used to use the first convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the first time period; use the second convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the second time period; use the third convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the third time period.

[0089] In some embodiments, the fusion unit 630 is used to perform global mean pooling on the timing pattern features of the historical adjacent time periods and the timing pattern features of the historical period periods of multiple different scales, respectively; and to splice the timing pattern features of the historical adjacent time periods after global mean pooling and the timing pattern features of the historical period periods of multiple different scales to obtain fused timing pattern features.

[0090] In some embodiments, the recalibration unit 640 is used to compress the fused temporal pattern features by adopting a global mean pooling method to obtain compressed temporal pattern features; perform an excitation operation on the compressed temporal pattern features to obtain attention scores of multiple convolution channels; and multiply the attention scores of the multiple convolution channels by the fused temporal pattern features.

[0091] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0092] A refers to B, which means that A is the same as B or A is a simple variant of B.

[0093] The terms "first" and "second" in the description and claims of the embodiments of the present application are used to distinguish different objects, rather than to describe a specific order of objects, and cannot be understood as indicating or implying relative importance. For example, the first convolution kernel and the second convolution kernel are used to distinguish different convolution kernels, rather than to describe a specific order of convolution kernels, and cannot be understood as the first convolution kernel being more important than the second convolution kernel.

[0094] In the embodiments of the present application, unless otherwise specified, "at least one" means one or more, and "a plurality" means two or more. For example, a plurality of convolution kernels means two or more convolution kernels.

[0095] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loading and executing computer program instructions on a computer, the process or function described in accordance with the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrations. Available media can be magnetic media, (e.g., floppy disk, hard disk, tape), optical media (e.g., DVD), or semiconductor media (e.g., solid-state hard disk Solid State Disk (SSD)), etc.

[0096] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting post-loan risk of supply chain financing based on time series prediction, characterized in that: The method comprises: Acquire a time series, wherein the time series indicates a corresponding relationship between multiple time series variables related to the supply chain and time, wherein the multiple time series variables include supply chain data of a target merchant, supply chain data of an upstream supplier of the target merchant, supply chain data of a downstream distributor of the target merchant, logistics data of the target merchant, quality control data of the target merchant, and contract data of the target merchant, wherein the supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow; Sequentially encoding the time series using a one-dimensional convolutional neural network to obtain temporal pattern features of historical adjacent time periods, where the historical adjacent time periods are historical time periods directly adjacent to the time period to be predicted; Using a plurality of convolution kernels of different sizes to perform jump convolution on the time series, so as to obtain time series pattern features of a plurality of historical periodic time periods of different scales, wherein the historical periodic time period is a historical period of time separated from the time period to be predicted by a predetermined time period; The time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales are fused to obtain fused time series pattern features; The fused time series pattern features are recalibrated to obtain the post-loan risk detection result of the target merchant in the time period to be predicted.

2. The method according to claim 1, characterized in that The one-dimensional convolutional neural network includes M convolution kernels with a width of w and a height of c. The height of the convolution kernel is the number of dimensions of the time series. The width of the convolution kernel represents the range of data of historical adjacent time periods that determine the post-loan risk of the target merchant in a time period. The one-dimensional convolutional neural network is used to sequentially encode the time series to obtain the temporal pattern characteristics of the historical adjacent time periods, including: The time series is convolved using M convolution kernels to obtain the temporal pattern characteristics of historical adjacent time periods.

3. The method according to claim 1, characterized in that The multiple historical periodic time periods of different scales include a first time period, a second time period and a third time period, the multiple convolution kernels of different sizes include a first convolution kernel, a second convolution kernel and a third convolution kernel, the height of the first convolution kernel, the height of the second convolution kernel and the height of the third convolution kernel are all the number of dimensions of the time series, the width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period, and the use of multiple convolution kernels of different sizes to perform jump convolution on the time series to obtain time series pattern features of multiple historical periodic time periods of different scales includes: sequentially encoding the time series using the first convolution kernel to obtain a time series pattern feature of a first time period; sequentially encoding the time series using the second convolution kernel to obtain a time series pattern feature of a second time period; The time series is sequentially encoded using the third convolution kernel to obtain the time series pattern characteristics of the third time period.

4. The method according to claim 1, characterized in that The fusing of the time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales includes: Performing global mean pooling on the temporal pattern features of the historical adjacent time periods and the temporal pattern features of the historical periodic time periods of multiple different scales; The time series pattern features of the historical adjacent time periods after global mean pooling and the time series pattern features of the historical periodic time periods of multiple different scales are spliced ​​to obtain fused time series pattern features.

5. The method according to claim 1, characterized in that The recalibrating the fused temporal pattern features includes: Using a global mean pooling method, compressing the fused temporal pattern features to obtain compressed temporal pattern features; Performing an excitation operation on the compressed temporal pattern features to obtain attention scores of multiple convolution channels; The attention scores of the multiple convolution channels are multiplied by the fused temporal pattern features.

6. A supply chain financing post-loan risk detection device based on time series prediction, characterized in that: The device comprises: an acquisition unit, configured to acquire a time series, wherein the time series indicates a corresponding relationship between a plurality of time series variables related to a supply chain and time, wherein the plurality of time series variables include supply chain data of a target merchant, supply chain data of an upstream supplier of the target merchant, supply chain data of a downstream distributor of the target merchant, logistics data of the target merchant, quality control data of the target merchant, and contract data of the target merchant, wherein the supply chain data includes industrial and commercial information, equity registration, loan transactions, administrative penalties, and transaction flow; A convolution unit is used to sequentially encode the time series using a one-dimensional convolutional neural network to obtain time series pattern features of historical adjacent time periods, where the historical adjacent time period is a historical time period directly adjacent to the time period to be predicted; and to perform jump convolution on the time series using a plurality of convolution kernels of different sizes to obtain time series pattern features of historical periodic time periods of a plurality of different scales, where the historical periodic time period is a historical period separated from the time period to be predicted by a predetermined time period; A fusion unit, used to fuse the time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales to obtain fused time series pattern features; The recalibration unit is used to recalibrate the fused time series pattern features to obtain the post-loan risk detection result of the target merchant in the time period to be predicted.

7. The device according to claim 6, characterized in that The one-dimensional convolutional neural network includes M convolution kernels with a width of w and a height of c. The height of the convolution kernel is the number of dimensions of the time series. The width of the convolution kernel represents the range of data from historical adjacent time periods that determines the post-loan risk of the target merchant in a time period. The convolution unit is used to convolve the time series using M convolution kernels to obtain the temporal pattern characteristics of the historical adjacent time periods.

8. The device according to claim 6, characterized in that The multiple historical period time periods of different scales include a first time period, a second time period and a third time period, the multiple convolution kernels of different sizes include a first convolution kernel, a second convolution kernel and a third convolution kernel, the height of the first convolution kernel, the height of the second convolution kernel and the height of the third convolution kernel are all the number of dimensions of the time series, the width of the first convolution kernel represents the first time period, the width of the second convolution kernel represents the second time period, and the width of the third convolution kernel represents the third time period, and the convolution unit is used to use the first convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the first time period; use the second convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the second time period; use the third convolution kernel to sequentially encode the time series to obtain the time series pattern characteristics of the third time period.

9. The device according to claim 6, characterized in that The fusion unit is used to perform global mean pooling on the time series pattern features of the historical adjacent time periods and the time series pattern features of the historical periodic time periods of multiple different scales; The time series pattern features of the historical adjacent time periods after global mean pooling and the time series pattern features of the historical periodic time periods of multiple different scales are spliced ​​to obtain fused time series pattern features.

10. The device according to claim 6, characterized in that The recalibration unit is used to compress the fused temporal pattern features by adopting a global mean pooling method to obtain compressed temporal pattern features; and perform an excitation operation on the compressed temporal pattern features to obtain attention scores of multiple convolution channels; The attention scores of the multiple convolution channels are multiplied by the fused temporal pattern features.

Images