Rwa risk identification method and system based on artificial intelligence

Abstract

The present application relates to the technical field of commercial risk identification, and particularly relates to a RWA risk identification method and system based on artificial intelligence. Based on the entity category of the target real world asset, corresponding historical asset state time series data is acquired; according to the historical asset state time series data, an asset state prediction model is constructed by combining meta-learning and ensemble learning methods; data acquisition parameters are configured according to the asset state prediction model, and real-time state time series data segments of the target real world asset are correspondingly acquired; the real-time state time series data segments are input into the asset state prediction model to acquire an initial prediction set, and a prediction risk boundary is correspondingly constructed; real-time asset state data of the target real world asset is continuously monitored, and the real-time asset state data is compared with the prediction risk boundary, and when the real-time asset state data exceeds the prediction risk boundary, a risk warning instruction is generated. The present application can improve the accuracy, timeliness and reliability of RWA risk identification.

Description

Technical Field

[0001] This invention relates to the field of business risk identification technology, and in particular to an AI-based risk identification method and system for risk identification using the RWA (Risk Assessment and Risk Assessment) framework. Background Technology

[0002] In the current field of Risk Assessment (RWA), traditional methods rely heavily on human experience and simple statistical analysis. Data collection lacks category-specificity, and the use of uniform standards easily leads to the collection of large amounts of irrelevant historical data. Furthermore, risk identification models often employ a single machine learning model, making it difficult to adapt to different asset classes and resulting in low predictive accuracy. Risk warnings depend on periodic sampling checks, lacking quantitative risk boundaries. These problems lead to low data processing efficiency and poor model adaptability in RWA risk identification, resulting in delayed risk detection, strong subjectivity in judgment, and an increased probability of asset risk and potential loss. Summary of the Invention

[0003] This invention addresses the technical problems of low data processing efficiency, poor model adaptability, untimely risk detection, and strong subjectivity in judgment in existing technologies by providing an AI-based RWA risk identification method and system.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides an AI-based RWA risk identification method, comprising: obtaining corresponding historical asset status time series data based on the entity category of the target real-world asset; Based on the historical asset status time-series data, and combining meta-learning and ensemble learning methods, an AI-based asset status prediction model is constructed and trained. Data acquisition parameters are configured according to the asset status prediction model, and corresponding real-time status time-series data fragments of the target real-world asset are acquired. These real-time status time-series data fragments are input into the asset status prediction model to obtain an initial prediction set, and a corresponding prediction risk boundary is constructed. The real-time asset status data of the target real-world asset is continuously monitored, and the real-time asset status data is compared with the prediction risk boundary. When the real-time asset status data exceeds the prediction risk boundary, a risk warning instruction is generated.

[0005] Optionally, based on the entity category of the target real-world asset, the corresponding historical asset status time series data is obtained, including: obtaining risk identification requirement information of the target real-world asset and extracting the corresponding look-ahead window length; defining a historical data collection window for the target real-world asset according to the look-ahead window length, wherein the historical data collection window is K times the look-ahead window length, and K is greater than or equal to 2; and combining the entity category of the target real-world asset with the historical data collection window to obtain the corresponding historical asset status time series data.

[0006] The process involves constructing and training an AI-based asset state prediction model based on the historical asset state time-series data, combined with meta-learning and ensemble learning methods. This includes: acquiring the entity volatility characteristics of the target real-world asset; adaptively configuring prediction model training parameters using a pre-trained meta-learning model based on the entity category of the target real-world asset, the risk identification requirement information, and the entity volatility characteristics, wherein the prediction model training parameters include at least the input segment length, the prediction segment length, and the prediction confidence threshold; applying the prediction model training parameters, combined with the sliding window method, to segment the historical asset state time-series data, generating a model training sample set; and constructing and training the asset state prediction model based on the model training sample set using ensemble learning methods.

[0007] The construction steps of the meta-learning model include: obtaining the full entity category of the target manufacturer corresponding to the target real-world asset, and determining multiple reference real-world assets accordingly; traversing the multiple reference real-world assets to obtain the original asset state sequence dataset within a preset time interval, wherein the original asset state sequence data includes at least the reference entity category, reference risk identification requirement information, and reference entity volatility characteristics; obtaining the original model training parameter data corresponding to the original asset state sequence data, wherein the original model training parameter data is determined based on historical performance evaluation and includes at least the reference input segment length, reference prediction segment length, and reference prediction confidence threshold; based on the original asset state sequence data and the original model training parameter data, data reconstruction is performed using a meta-learning training paradigm to construct multiple meta-training tasks, wherein each meta-training task includes a support set and a query set; based on the prediction results of multiple query sets, a joint loss objective is constructed and calculated, wherein the joint loss objective includes at least the residuals between the query input segment length, query prediction segment length, and query prediction confidence threshold in the prediction results and the original model training parameter data; and training and validating the meta-learning model with the goal of minimizing the joint loss objective.

[0008] Specifically, the process involves applying the training parameters of the prediction model, using a sliding window method to segment the historical asset status time-series data, generating a model training sample set, and then constructing and training an asset status prediction model based on the model training sample set using an ensemble learning method. This includes: Based on the input segment length and the prediction segment length, the sliding window length is determined; according to the sliding window length, the historical asset status time series data is randomly slid-cut using the sliding window method, and each sliding cut result is used as a model training sample to generate a model training sample set; multiple basic prediction models are initialized, and the multiple basic prediction models are trained in parallel and integrated using the model training sample set to obtain the asset status prediction model, wherein the multiple basic prediction models are configured differently.

[0009] The process of configuring data acquisition parameters according to the asset status prediction model and acquiring real-time status time-series data segments of the target real-world asset includes: defining the data acquisition parameters with the input segment length of the asset status prediction model as the acquisition window length; and performing backtracking acquisition based on the data acquisition parameters, starting from the current time point, to acquire the real-time status time-series data segments of the target real-world asset in real time.

[0010] The process of inputting the real-time state time-series data segment into the asset state prediction model to obtain an initial prediction set and construct a corresponding prediction risk boundary includes: using the real-time state time-series data segment as input, obtaining initial prediction results of multiple basic prediction models through the asset state prediction model, and storing them as the initial prediction set; performing confidence filtering on the initial prediction set according to the prediction confidence threshold to obtain a confidence prediction set; and calculating and constructing the prediction risk boundary within the future prediction segment length based on the confidence prediction set.

[0011] Secondly, the present invention provides an AI-based RWA risk identification system, comprising: The historical asset status acquisition module is used to acquire the corresponding historical asset status time series data based on the entity category of the target real-world asset. The asset status prediction model training module is used to construct and train an artificial intelligence-based asset status prediction model based on the historical asset status time series data and by combining meta-learning and ensemble learning methods. The real-time status time series data acquisition module is used to configure data acquisition parameters according to the asset status prediction model and acquire real-time status time series data fragments of the target real-world asset accordingly. The prediction risk boundary construction module is used to input the real-time status time series data fragments into the asset status prediction model, obtain an initial prediction set, and construct the prediction risk boundary accordingly. The risk warning instruction generation module is used to continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning instruction when the real-time asset status data exceeds the predicted risk boundary.

[0012] By implementing this invention, it is possible to obtain the corresponding historical asset status time series data based on the entity category of the target real-world asset, avoid the blindness of data collection, reduce the interference of irrelevant data on the subsequent analysis process, reduce data processing costs, and improve data utilization efficiency. By implementing this invention, it is possible to construct and train an AI-based asset status prediction model based on the historical asset status time series data, combined with meta-learning and ensemble learning methods. This can improve the accuracy and reliability of asset status prediction. Ensemble learning effectively avoids the biases and errors that may exist in a single model, while meta-learning accelerates the model's adaptation to new scenarios. The combination of the two allows the trained model to predict asset status more accurately in practical applications. By implementing this invention, it is possible to configure data acquisition parameters according to the asset status prediction model and obtain real-time status time series data fragments of the target real-world assets, thereby improving the targeting and effectiveness of real-time data acquisition, avoiding the collection of too much redundant data or the omission of key data, reducing the pressure of data transmission and storage, and ensuring the data quality input to the model, thus improving the accuracy of subsequent prediction results. By implementing this invention, the real-time status time series data fragments can be input into the asset status prediction model to obtain an initial prediction set and construct a corresponding prediction risk boundary, providing a clear basis for risk identification. The establishment of the prediction risk boundary transforms risk identification from a vague qualitative judgment to a precise quantitative judgment, avoiding the subjectivity and uncertainty of human judgment. By implementing this invention, it is possible to continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning command when the real-time asset status data exceeds the predicted risk boundary. This minimizes the time for risk detection and response, preventing further escalation of the risk. Through real-time monitoring and automatic warnings, relevant personnel can be informed of the asset risk situation immediately, quickly activate emergency plans, and reduce the damage caused to the asset.

[0013] In summary, by implementing this invention, the accuracy, timeliness, and reliability of RWA risk identification can be effectively improved, the probability of asset risks and losses can be reduced, and strong technical support can be provided for the secure operation of RWA. Attached Figure Description

[0014] Figure 1 A flowchart illustrating an AI-based RWA risk identification method provided by this invention; Figure 2 This is a schematic diagram of the structure of an AI-based RWA risk identification system provided by the present invention.

[0015] In the attached diagram, the components represented by each number are as follows: Module 11 for acquiring historical asset status, module 12 for training asset status prediction model, module 13 for acquiring real-time status time series data, module 14 for constructing prediction risk boundaries, and module 15 for generating risk warning instructions. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0019] Example 1, as Figure 1 As shown, this embodiment of the invention provides an AI-based RWA risk identification method, including: S100: Based on the entity category of the target real-world asset, obtain the corresponding historical asset status time series data; S200: Based on the historical asset status time series data, and combining meta-learning and ensemble learning methods, construct and train an artificial intelligence-based asset status prediction model; S300: Configure data acquisition parameters according to the asset status prediction model, and obtain the real-time status time series data fragments of the target real-world asset accordingly; S400: Input the real-time status time series data fragment into the asset status prediction model to obtain an initial prediction set and construct the corresponding prediction risk boundary; S500: Continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning instruction when the real-time asset status data exceeds the predicted risk boundary.

[0020] In step S100 of this application embodiment, based on the entity category of the target real-world asset, the corresponding historical asset status time series data is obtained, including: Obtain risk identification requirements for the target real-world assets and extract the corresponding forward window length; Based on the aforementioned look-ahead window length, a historical data acquisition window for the target real-world asset is defined, wherein the historical data acquisition window is K times the look-ahead window length, and K is greater than or equal to 2; By combining the entity category of the target real-world asset with the historical data acquisition window, the corresponding historical asset status time series data is obtained.

[0021] In this embodiment of the application, step S100 is a basic data preparation step for the RWA risk identification method based on artificial intelligence. The purpose is to provide high-quality historical data support that is adapted to the risk identification needs for the subsequent construction and training of the asset status prediction model.

[0022] To achieve the above objectives, it is first necessary to obtain the risk identification requirements information of the target real-world assets and extract the corresponding forward window length. That is, to clarify the specific risk identification requirements of the target RWA, such as "early warning of equipment failure risk in the next 30 days" or "monitoring of accounts receivable overdue risk in the next 20 days," and then extract the core time parameter from the requirements, namely the forward window length. The forward window length is the length of the future time interval that the asset status prediction model needs to predict, such as 30 days and 20 days for the above requirements.

[0023] Next, the historical data collection window for the target real-world asset needs to be defined based on the aforementioned look-ahead window length. This is calculated according to the rule: "Historical data collection window length = K × Look-ahead window length (K≥2)". For example, if the look-ahead window length is 30 days and K=3, then the historical data collection window length is 90 days, meaning asset state data for the past 90 days needs to be collected. The purpose of this design is to cover the asset's state changes over multiple periods with sufficiently long historical data, ensuring that the subsequent asset state prediction model can learn the complete state evolution pattern and avoid prediction bias due to insufficient historical data.

[0024] Next, it is necessary to combine the entity category of the target real-world asset with the historical data collection window to obtain the corresponding historical asset status time-series data. That is, to determine the entity category of the target real-world asset, such as "real estate", "machinery and equipment", "financial assets", etc., and to clarify the core status indicators of the asset category. For example, the core status indicators for real estate are "market price, vacancy rate, and location and supporting infrastructure maturity", and the core status indicators for machinery and equipment are "operating temperature, failure rate, and maintenance records".

[0025] Then, based on the historical data collection window defined in the aforementioned steps, such as the past 90 days, time-series data of the core status indicators corresponding to this type of asset are collected to obtain continuous data arranged in chronological order, such as the daily price of real estate and the hourly operating temperature of machinery and equipment, forming the "historical asset status time-series data".

[0026] In step S200 of this application embodiment, based on the historical asset status time-series data, and combining meta-learning and ensemble learning methods, an artificial intelligence-based asset status prediction model is constructed and trained, including: To obtain the physical volatility characteristics of the target real-world asset; Based on the entity category of the target real-world asset, the risk identification requirement information, and the entity volatility characteristics, the prediction model training parameters are adaptively configured through a pre-trained meta-learning model, wherein the prediction model training parameters include at least the input segment length, the prediction segment length, and the prediction confidence threshold. The prediction model training parameters are applied, and the historical asset status time series data is segmented using the sliding window method to generate a model training sample set. Based on the model training sample set, the asset status prediction model is constructed and trained using an ensemble learning method.

[0027] In this embodiment of the application, the purpose of step S100 is to construct a high-precision asset status prediction model that adapts to the characteristics of the target asset and meets the requirements of risk identification by adaptive configuration of parameters through meta-learning and improvement of prediction stability through ensemble learning.

[0028] First, it is necessary to obtain the physical volatility characteristics of the target real-world asset. Specifically, it is necessary to determine the core state indicators of the target asset, such as the failure rate and operating temperature fluctuations of machinery and equipment, and the daily price fluctuations and trading volume fluctuations of commodities. Then, the volatility parameters of the indicator are calculated statistically or by algorithms. For example, the standard deviation is used to measure the magnitude of price fluctuations, the coefficient of variation is used to measure the relative degree of price fluctuations, volatility clustering analysis is used to determine whether there are concentrated bursts of volatility, and so on, ultimately forming the "entity volatility characteristics".

[0029] Furthermore, based on the entity category of the target real-world asset, the risk identification requirement information, and the entity volatility characteristics, the prediction model training parameters need to be adaptively configured through a pre-trained meta-learning model. That is, the entity category, the risk identification requirement information, and the entity volatility characteristics are input into the meta-learning model, and the meta-learning model adaptively outputs adapted prediction model training parameters based on the input information. These parameters include the input segment length, the prediction segment length, and the prediction confidence threshold.

[0030] The input segment length is the length of historical data used by the prediction model for prediction, such as "30 days" for high volatility assets and "15 days" for low volatility assets. The forecast segment length is the time interval of the future asset state that the forecast model needs to predict, and it is specifically matched with the risk identification requirements, such as a forecast segment length of 15 days. The prediction confidence threshold is a criterion for selecting reliable prediction results. For example, for high-risk demand scenarios, the prediction confidence threshold of 90% can be output to reduce the false alarm rate.

[0031] In step S200 of this application embodiment, the construction step of the meta-learning model includes: Obtain the full entity category of the target manufacturer corresponding to the target real-world asset, and identify multiple reference real-world assets accordingly; Traverse multiple real-world reference assets to obtain a dataset of original asset state sequences within a preset time interval. The original asset state sequence data includes at least reference entity categories, reference risk identification requirements, and reference entity volatility characteristics. Obtain the original model training parameter data corresponding to the original asset state sequence data, wherein the original model training parameter data is determined based on historical performance evaluation and includes at least the reference input segment length, the reference prediction segment length, and the reference prediction confidence threshold. Based on the original asset state sequence data and the original model training parameter data, data reconstruction is performed using a meta-learning training paradigm to construct multiple meta-training tasks, wherein each meta-training task includes a support set and a query set. Based on the prediction results of multiple query sets, a joint loss objective is constructed and calculated, wherein the joint loss objective includes at least the residuals between the query input segment length, the query prediction segment length, the query prediction confidence threshold, and the original model training parameter data in the prediction results; The meta-learning model is trained and validated with the goal of minimizing the joint loss.

[0032] In this embodiment of the application, the purpose of constructing the meta-learning model is to train a meta-learning model that can quickly adapt the optimal training parameters according to the characteristics of the target asset by transferring the parameter configuration experience of multiple reference real-world assets, so as to provide intelligent decision-making basis for the subsequent prediction model parameter configuration of the target asset.

[0033] The first step is to obtain the full entity category of the target manufacturer corresponding to the target real-world asset, and then identify multiple reference real-world assets.

[0034] The target real-world assets are categorized into "all entity categories" of the manufacturer, such as "production equipment," "warehousing facilities," and "raw material inventory" for a manufacturing company. Then, assets with sufficient historical data and clear risk identification cases are selected from each entity category as "reference real-world assets," such as selecting three machine tools that have been running for three years and have a history of malfunctions from the "production equipment" category.

[0035] The second step requires iterating through multiple real-world reference assets to obtain a dataset of the original asset state sequence within a preset time interval.

[0036] First, a "preset time interval" needs to be set for each reference real-world asset, such as the past two years, to ensure that there is enough time to cover the various state changes of the asset.

[0037] Next, it is necessary to collect "original asset state sequence data" of reference real-world assets within the preset time interval. The original asset state sequence data includes reference entity categories, such as "production equipment - machine tool"; reference risk identification requirement information, such as "the machine tool's past risk requirement was 'predicting the failure risk in the next 7 days'"; and reference entity volatility characteristics, such as "the monthly standard deviation of the failure rate of this machine tool over the past 2 years = 0.6".

[0038] Finally, the above information is organized in chronological order to form a structured sequence dataset.

[0039] The third step is to obtain the original model training parameter data corresponding to the original asset state sequence data.

[0040] First, for each reference real-world asset, it is necessary to review its past history of training models using different training parameters, including input segment length, prediction segment length, and prediction confidence threshold. Then, based on historical performance evaluation indicators such as prediction accuracy and risk warning false alarm rate, the "optimal training parameters" of the corresponding real-world assets under specific risk requirements are selected as "original model training parameter data". For example, for a certain machine tool under the requirement of "7-day fault warning", the optimal input segment length is 21 days, the optimal prediction segment length is 7 days, and the optimal confidence threshold is 85%. Finally, a one-to-one correspondence is established between the "original asset state sequence data" and the "original model training parameter data".

[0041] The fourth step involves reconstructing the data based on the original asset state sequence data and the original model training parameter data, combined with a meta-learning training paradigm, to construct multiple meta-training tasks.

[0042] First, based on the data corresponding to "asset characteristics (original asset state sequence data) - optimal parameters (original model training parameter data)" established in step three, the data needs to be split according to the meta-learning training paradigm. For example, model-independent meta-learning can be selected for data splitting.

[0043] Specifically, for each dataset referencing real-world assets, it is necessary to randomly divide it into a support set and a query set; The support set contains a small number of "asset characteristics - optimal parameters" samples, which are used to allow the meta-learning model to initially learn the parameter matching rules of this type of asset. The query set includes another part of the "asset characteristics - optimal parameters" sample, which is used to verify whether the rules learned by the meta-learning model are general. Then, repeat the above process to construct multiple independent "meta-training tasks" for all reference real-world assets, each meta-training task containing one support set and one query set.

[0044] The fifth step requires constructing and calculating a joint loss target based on the prediction results of multiple query sets.

[0045] First, the asset characteristics of the "support set" in each meta-training task are input into the meta-learning model to be trained. After learning based on the support set, the meta-learning model outputs the "prediction parameters" for the asset characteristics of the "query set". The prediction parameters are the predicted input segment length, the predicted segment length, and the confidence threshold. Then, construct the "joint loss objective": calculate the residuals between the "predicted parameters" output by the meta-learning model and the "original model training parameter data (optimal parameters)" in the query set, such as mean squared error, absolute error, etc., and sum the residuals of all meta-training tasks to form the joint loss; The joint loss must cover all key parameters, including at least the "residual between predicted input segment length and actual input segment length", the "residual between predicted segment length and actual predicted segment length", and the "residual between predicted confidence threshold and actual confidence threshold".

[0046] The sixth step is to train and validate the meta-learning model with the goal of minimizing the joint loss.

[0047] First, with the goal of "minimizing the joint loss" as the optimization direction, optimization algorithms such as Adam and SGD gradient descent are used to update the weight parameters of the meta-learning model; Then, during training, the performance of the meta-learning model is periodically tested using independent "validation meta-tasks," where each validation meta-task is a validation set separately partitioned from the reference real-world assets. If the joint loss of the validation set decreases and tends to stabilize, it indicates that the meta-learning model has learned general parameter matching rules. If the validation loss increases, adjustments are made through regularization, increasing the number of reference assets, etc., until the meta-learning model performs stably on both the training and validation sets, thus completing the training of the meta-learning model.

[0048] Furthermore, it is necessary to apply the training parameters of the prediction model, combine the sliding window method to segment the historical asset status time series data, generate a model training sample set, and construct and train the asset status prediction model based on the model training sample set and the ensemble learning method.

[0049] By inputting the entity category, risk identification requirements, and volatility characteristics of the target real-world asset into the meta-learning model, the model can predict and output adaptively configured training parameters. For example, the predicted output might have a historical data length of 30 days, a prediction segment length of 15 days, and a prediction confidence threshold of 90%.

[0050] In step S200 of this application embodiment, the prediction model training parameters are applied, and the historical asset status time series data is segmented using the sliding window method to generate a model training sample set. Based on the model training sample set, an asset status prediction model is constructed and trained using an ensemble learning method, including: The sliding window length is determined based on the input segment length and the predicted segment length. Based on the sliding window length, the historical asset status time series data is randomly slid-cut using the sliding window method, and each sliding cut result is used as a model training sample to generate a model training sample set. Multiple basic prediction models are initialized, and the multiple basic prediction models are trained in parallel and integrated using the model training sample set to obtain the asset status prediction model, wherein the multiple basic prediction models are configured differently.

[0051] In this embodiment of the application, the purpose of the above steps is to transform the training parameters of the prediction model configured by meta-learning into an asset state prediction model that can accurately predict the asset state through scientific data segmentation and differential model integration.

[0052] The first step is to determine the sliding window length based on the input segment length and the prediction segment length. First, it's necessary to clarify two core parameters of the meta-learning configuration: one is the input segment length, which is the length of historical data used by the asset status prediction model for prediction, such as "30 days," meaning predicting the future based on the asset status of the past 30 days; the other is the prediction segment length, which is the future time interval that the asset status prediction model needs to predict, such as "7 days," meaning predicting the asset status of the next 7 days. Optionally, the total window length can be calculated as "sliding window length = input segment length + prediction segment length". For example, in the above example, the input segment length = 30 days and the prediction segment length = 7 days, then the sliding window length = 37 days.

[0053] The second step is to randomly slide and cut the historical asset status time series data according to the sliding window length and the sliding window method, and use each sliding cut result as a model training sample to generate a model training sample set.

[0054] First, we need to prepare the "historical asset status time series data" obtained in step S100, which includes daily asset status data for the past year, totaling 365 records; Then, using the sliding window length determined in the first step as the cutting unit, the data is randomly cut starting from the beginning of the historical data. For example, here the sliding window length is set to 37 days: For example, in the first cut, data from days 1 to 37 are taken, with the first 30 days (input segment length) as "sample input X1" and the last 7 days (prediction segment length) as "sample output Y1", forming the first model training sample (X1, Y1); in the second cut, the sliding step size is randomly selected, such as 3 days, that is, data from days 5 to 41 are taken, with the first 30 days as "sample input X2" and the last 7 days as "sample output Y2", forming the second model training sample (X2, Y2). Next, repeat the above random sliding cut until all historical asset state time series data are traversed, generating a "model training sample set" containing a large number of (X,Y) pairs.

[0055] The third step requires initializing multiple basic prediction models, using the model training sample set to train and integrate the multiple basic prediction models in parallel, and obtaining the asset status prediction model.

[0056] First, it's necessary to select multiple types of base prediction models, such as Recurrent Neural Networks (LSTM), Gradient Boosting Trees (XGBoost), and Logistic Regression (LR), or to set different hyperparameters for the same type of model, for example, setting two LSTMs: LSTM1 with 64 hidden layer nodes and LSTM2 with 128 hidden layer nodes. This ensures that there are "structural differences" or "parameter differences" between the models, avoiding redundancy after ensemble integration. Optionally, three basic prediction models can be initialized: LSTM for capturing temporal dependencies, XGBoost for capturing nonlinear features, and LR for ensuring model stability.

[0057] Then, the "model training sample set" generated in the second step is divided into a training set and a validation set in an 8:2 ratio. Multiple basic prediction models are then trained in parallel using the training set, i.e., simultaneously, to improve efficiency. Mean squared error (MSE) is used as the loss function, with the goal of minimizing the error between the predicted output and the true output (sample Y). Simultaneously, the validation set is used to adjust the learning rate, number of iterations, and other hyperparameters of each basic prediction model to ensure that each model achieves optimal performance.

[0058] Next, weighted voting, stacking, and other integration strategies are used to integrate the prediction results of multiple basic prediction models.

[0059] The weighted voting method assigns weights based on the accuracy of each basic prediction model on the validation set. The weights are positively correlated with the accuracy. For example, if the LSTM accuracy is 92%, the XGBoost accuracy is 90%, and the LR accuracy is 88%, then the final prediction result is (LSTM prediction value × 0.35) + (XGBoost prediction value × 0.33) + (LR prediction value × 0.32). Stacking involves using the prediction results of each basic prediction model as "new features" and inputting them into a secondary model, such as a logistic regression model. The secondary model then outputs the final prediction result, further improving the integration accuracy. The basic prediction model integrated through the above methods is the "asset status prediction model".

[0060] In step S300 of this application embodiment, data acquisition parameters are configured according to the asset status prediction model, and real-time status time-series data fragments of the target real-world asset are obtained accordingly, including: The data acquisition parameters are defined with the length of the input segment of the asset status prediction model as the acquisition window length. Based on the data acquisition parameters, backtracking data is collected starting from the current time point to obtain real-time state time-series data fragments of the target real-world assets.

[0061] In this embodiment of the application, the purpose of step S300 is to provide the asset status prediction model with real-time input data that is format-adapted and time-synchronized, so as to ensure that the model can make predictions based on the latest asset status.

[0062] To achieve the above objectives, the data acquisition parameters must first be defined with the length of the input segment of the asset status prediction model as the acquisition window length.

[0063] Specifically, it is necessary to extract the core training parameters associated with the asset status prediction model, namely the "input segment length" of the meta-learning model adaptive configuration in the aforementioned step S200, such as "30 days", which means that the asset status prediction model uses 30 days of historical data as input during training; The "input segment length" is directly mapped to the "collection window length," and this serves as the core to define complete data collection parameters. The data collection parameters should include at least the following: collection duration, which is equal to the input segment length, such as 30 days; data frequency, which is consistent with the frequency of historical asset status time-series data. For example, if the historical data is daily, then the real-time data will also collect daily data; if the historical data is hourly, then the real-time data will collect hourly data to ensure data granularity matching; and data dimension, which is consistent with the core indicators of historical asset status time-series data. For example, if the historical data includes "equipment operating temperature and failure rate," then the real-time data will also collect these two types of indicators to avoid incomplete model input features due to missing dimensions.

[0064] Then, based on the data acquisition parameters, backtracking data collection needs to be performed starting from the current time point to obtain the real-time state time series data fragments of the target real-world assets.

[0065] Specifically, the "current time" needs to be determined, such as the time when the system performed data collection on May 30, 2024 at 12:00. Next, "backtracking data collection" is performed according to the collection parameters defined in the aforementioned steps. For example, if the collection window length is 30 days and the data frequency is daily, then backtracking for 30 days from May 30, 2024, daily asset status data from May 1, 2024 to May 30, 2024 is collected, such as the maximum daily equipment operating temperature and the daily failure rate. If the collection window length is 72 hours and the data frequency is hourly, then the asset status data for each hour will be collected from 12:00 on May 30, 2024 to 12:00 on May 30, 2024, a 72-hour backtracking process. Finally, the collected real-time data is organized into a continuous sequence according to time order to form the "real-time status time series data segment".

[0066] In step S400 of this application embodiment, the real-time status time-series data fragment is input into the asset status prediction model to obtain an initial prediction set, and a corresponding prediction risk boundary is constructed, including: Using the real-time status time series data segment as input, the initial prediction results of multiple basic prediction models are obtained through the asset status prediction model and stored as the initial prediction set; The initial prediction set is filtered based on the prediction confidence threshold to obtain a confidence prediction set; Based on the confidence prediction set, the predicted risk boundary within the predicted segment length in the future is calculated and constructed.

[0067] In this embodiment of the application, the purpose of step S400 is to transform the output of the asset status prediction model into a "predicted risk boundary" that can be used for risk assessment by integrating the prediction results of multiple models and screening the confidence level, so as to provide a clear judgment standard for subsequent real-time risk monitoring.

[0068] To achieve the above objectives, the first step is to take the real-time status time series data segment as input, obtain the initial prediction results of multiple basic prediction models through the asset status prediction model, and store them as the initial prediction set.

[0069] Specifically, the "real-time status time series data fragments" obtained in step S300, such as equipment operating temperature and failure rate data over the past 30 days, need to be input into the asset status prediction model; The asset status prediction model consists of multiple differentiated underlying prediction models that operate in parallel, each outputting a prediction of the asset status within a future prediction period. Then, the prediction results of all the basic prediction models are integrated according to the "time dimension" to form an initial prediction set.

[0070] The second step is to perform confidence filtering on the initial prediction set based on the prediction confidence threshold to obtain a confidence prediction set.

[0071] Specifically, it is necessary to extract the "prediction confidence threshold" configured in the meta-learning model in step S200, such as 90%, which means that the reliability of the prediction results of the basic prediction model must be ≥90% to be accepted. For the prediction results of each base model in the initial prediction set, determine whether its confidence level reaches the threshold. Keep the prediction results of the base model with a confidence level greater than or equal to the prediction confidence threshold, and remove the results with a confidence level less than the prediction confidence threshold. For example, if the LR model has a prediction confidence of only 78% < 90% due to poor fit of the training samples, its prediction results will be discarded; if the prediction confidence of LSTM and XGBoost is greater than or equal to the prediction confidence threshold, their prediction results will be retained. Finally, all predictions that reached the prediction confidence threshold were re-integrated to form a "confidence prediction set". The third step is to calculate and construct the predicted risk boundary within the predicted segment length in the future, based on the confidence prediction set.

[0072] First, the “prediction segment length” needs to be determined. This prediction segment length is configured by meta-learning, such as 7 days, meaning the risk boundary needs to cover the next 7 days. Then, statistical calculations are performed on the "daily forecast values" in the confidence forecast set to determine the "normal fluctuation range" for that day, i.e., the daily upper and lower limits of the risk boundary: Common calculation method: Determine the normal fluctuation range using "mean ± standard deviation × coefficient". The coefficient is typically set to 1.5-2 to balance sensitivity and false alarm rate. For example: if the predicted result of a certain indicator on a certain day in the confidence prediction set is [8%, 10%] (excluding the prediction results of the LR model), with a mean of 9% and a standard deviation of ≈1.414, and a coefficient of 2, then the upper and lower limits of the risk boundary for Day 1 are 9% ± 2 × 1.414% ≈ [6.17%, 11.83%]; similarly, the predicted risk boundaries for other days within the prediction period can be calculated. Finally, the predicted risk boundaries for each day of the next 7 days are integrated in chronological order to form the "predicted risk boundaries within the length of the future prediction period". For example, Day 1: [6.2%, 11.8%], Day 2: [9.0%, 18.0%], ..., Day 7: [12.5%, 16.5%].

[0073] In step S500 of this application embodiment, it is also necessary to continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning instruction when the real-time asset status data exceeds the predicted risk boundary.

[0074] That is, according to the data collection parameters of the asset, the real-time data of the target asset is continuously collected; then the collected real-time data is compared with the corresponding time period prediction risk boundary constructed by S400 in terms of time dimension, such as comparing the equipment failure probability of the day with the prediction risk boundary of the day [6.2%, 11.8%]; If real-time data exceeds the predicted risk boundary, the system will automatically generate a risk warning instruction and complete the warning.

[0075] Example 2, as Figure 2As shown, based on the same inventive concept as the AI-based RWA risk identification method provided in Embodiment 1, this embodiment of the invention also provides an AI-based RWA risk identification system, including: The historical asset status acquisition module 11 is used to acquire the corresponding historical asset status time series data based on the entity category of the target real-world asset. The asset status prediction model training module 12 is used to construct and train an artificial intelligence-based asset status prediction model based on the historical asset status time series data and by combining meta-learning and ensemble learning methods. The real-time status time series data acquisition module 13 is used to configure data acquisition parameters according to the asset status prediction model and acquire real-time status time series data fragments of the target real-world asset accordingly. The prediction risk boundary construction module 14 is used to input the real-time status time series data fragments into the asset status prediction model, obtain an initial prediction set, and construct the prediction risk boundary accordingly. The risk warning instruction generation module 15 is used to continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning instruction when the real-time asset status data exceeds the predicted risk boundary.

[0076] Furthermore, the historical asset status acquisition module 11 includes the following execution steps: Obtain risk identification requirements for the target real-world assets and extract the corresponding forward window length; Based on the aforementioned look-ahead window length, a historical data acquisition window for the target real-world asset is defined, wherein the historical data acquisition window is K times the look-ahead window length, and K is greater than or equal to 2; By combining the entity category of the target real-world asset with the historical data acquisition window, the corresponding historical asset status time series data is obtained.

[0077] Furthermore, the asset status prediction model training module 12 includes the following execution steps: To obtain the physical volatility characteristics of the target real-world asset; Based on the entity category of the target real-world asset, the risk identification requirement information, and the entity volatility characteristics, the prediction model training parameters are adaptively configured through a pre-trained meta-learning model, wherein the prediction model training parameters include at least the input segment length, the prediction segment length, and the prediction confidence threshold. The prediction model training parameters are applied, and the historical asset status time series data is segmented using the sliding window method to generate a model training sample set. Based on the model training sample set, the asset status prediction model is constructed and trained using an ensemble learning method.

[0078] The steps for constructing the meta-learning model include: Obtain the full entity category of the target manufacturer corresponding to the target real-world asset, and identify multiple reference real-world assets accordingly; Traverse multiple real-world reference assets to obtain a dataset of original asset state sequences within a preset time interval. The original asset state sequence data includes at least reference entity categories, reference risk identification requirements, and reference entity volatility characteristics. Obtain the original model training parameter data corresponding to the original asset state sequence data, wherein the original model training parameter data is determined based on historical performance evaluation and includes at least the reference input segment length, the reference prediction segment length, and the reference prediction confidence threshold. Based on the original asset state sequence data and the original model training parameter data, data reconstruction is performed using a meta-learning training paradigm to construct multiple meta-training tasks, wherein each meta-training task includes a support set and a query set. Based on the prediction results of multiple query sets, a joint loss objective is constructed and calculated, wherein the joint loss objective includes at least the residuals between the query input segment length, the query prediction segment length, the query prediction confidence threshold, and the original model training parameter data in the prediction results; The meta-learning model is trained and validated with the goal of minimizing the joint loss.

[0079] Specifically, the process involves applying the training parameters of the prediction model, using a sliding window method to segment the historical asset status time-series data, generating a model training sample set, and then constructing and training an asset status prediction model based on the model training sample set using an ensemble learning method. This includes: The sliding window length is determined based on the input segment length and the predicted segment length. Based on the sliding window length, the historical asset status time series data is randomly slid-cut using the sliding window method, and each sliding cut result is used as a model training sample to generate a model training sample set. Multiple basic prediction models are initialized, and the multiple basic prediction models are trained in parallel and integrated using the model training sample set to obtain the asset status prediction model, wherein the multiple basic prediction models are configured differently.

[0080] Furthermore, the real-time status time series data acquisition module 13 includes the following execution steps: The data acquisition parameters are defined with the length of the input segment of the asset status prediction model as the acquisition window length. Based on the data acquisition parameters, backtracking data is collected starting from the current time point to obtain real-time state time-series data fragments of the target real-world assets.

[0081] Furthermore, the risk boundary prediction construction module 14 includes the following execution steps: Using the real-time status time series data segment as input, the initial prediction results of multiple basic prediction models are obtained through the asset status prediction model and stored as the initial prediction set; The initial prediction set is filtered based on the prediction confidence threshold to obtain a confidence prediction set; Based on the confidence prediction set, the predicted risk boundary within the predicted segment length in the future is calculated and constructed.

[0082] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0083] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based RWA risk identification method, characterized in that, include: Based on the entity category of the target real-world asset, obtain the corresponding historical asset status time series data. The entity category includes machinery and equipment, and the core status indicators corresponding to machinery and equipment are operating temperature, failure rate, and maintenance records. Based on the historical asset status time series data, and combining meta-learning and ensemble learning methods, an artificial intelligence-based asset status prediction model is constructed and trained. Configure data acquisition parameters according to the asset status prediction model, and obtain real-time status time series data fragments of the target real-world assets accordingly. The real-time status time series data fragments are input into the asset status prediction model to obtain an initial prediction set, and a corresponding prediction risk boundary is constructed. The system continuously monitors the real-time asset status data of the target real-world asset and compares the real-time asset status data with the predicted risk boundary. When the real-time asset status data exceeds the predicted risk boundary, a risk warning instruction is generated. Based on the historical asset status time-series data, and combining meta-learning and ensemble learning methods, an artificial intelligence-based asset status prediction model is constructed and trained, including: To obtain the physical volatility characteristics of the target real-world asset; Based on the entity category, risk identification requirements, and volatility characteristics of the target real-world asset, the prediction model training parameters are adaptively configured through a pre-trained meta-learning model, wherein the prediction model training parameters include at least the input segment length, the prediction segment length, and the prediction confidence threshold. The prediction model training parameters are applied, and the historical asset status time series data is segmented using the sliding window method to generate a model training sample set. Based on the model training sample set, the asset status prediction model is constructed and trained using an ensemble learning method. The prediction model training parameters are applied, and the historical asset status time series data is segmented using the sliding window method to generate a model training sample set. Based on the model training sample set, an asset status prediction model is constructed and trained using an ensemble learning method, including: The sliding window length is determined based on the input segment length and the predicted segment length. Based on the sliding window length, the historical asset status time series data is randomly slid-cut using the sliding window method, and each sliding cut result is used as a model training sample to generate a model training sample set. Multiple basic prediction models are initialized, and the multiple basic prediction models are trained in parallel and integrated using the model training sample set to obtain the asset status prediction model, wherein the multiple basic prediction models are configured differently. The real-time status time-series data fragments are input into the asset status prediction model to obtain an initial prediction set, and a corresponding prediction risk boundary is constructed, including: Using the real-time status time series data segment as input, the initial prediction results of multiple basic prediction models are obtained through the asset status prediction model and stored as the initial prediction set; The initial prediction set is filtered based on the prediction confidence threshold to obtain a confidence prediction set; Based on the confidence prediction set, the predicted risk boundary within the predicted segment length in the future is calculated and constructed.

2. The AI-based RWA risk identification method as described in claim 1, characterized in that, Based on the entity category of the target real-world asset, obtain the corresponding historical asset status time series data, including: Obtain risk identification requirements for the target real-world assets and extract the corresponding forward window length; Based on the aforementioned look-ahead window length, a historical data acquisition window for the target real-world asset is defined, wherein the historical data acquisition window is K times the look-ahead window length, and K is greater than or equal to 2; By combining the entity category of the target real-world asset with the historical data acquisition window, the corresponding historical asset status time series data is obtained.

3. The AI-based RWA risk identification method as described in claim 1, characterized in that, The steps for constructing the meta-learning model include: Obtain the full entity category of the target manufacturer corresponding to the target real-world asset, and identify multiple reference real-world assets accordingly; Traverse multiple real-world reference assets to obtain a dataset of original asset state sequences within a preset time interval. The original asset state sequence data includes at least reference entity categories, reference risk identification requirements, and reference entity volatility characteristics. Obtain the original model training parameter data corresponding to the original asset state sequence data, wherein the original model training parameter data is determined based on historical performance evaluation and includes at least the reference input segment length, the reference prediction segment length, and the reference prediction confidence threshold. Based on the original asset state sequence data and the original model training parameter data, data reconstruction is performed using a meta-learning training paradigm to construct multiple meta-training tasks, wherein each meta-training task includes a support set and a query set. Based on the prediction results of multiple query sets, a joint loss objective is constructed and calculated, wherein the joint loss objective includes at least the residuals between the query input segment length, the query prediction segment length, the query prediction confidence threshold, and the original model training parameter data in the prediction results; The meta-learning model is trained and validated with the goal of minimizing the joint loss.

4. The AI-based RWA risk identification method as described in claim 1, characterized in that, Based on the asset status prediction model, configure the data acquisition parameters and obtain corresponding real-time status time-series data segments of the target real-world asset, including: The data acquisition parameters are defined with the length of the input segment of the asset status prediction model as the acquisition window length. Based on the data acquisition parameters, backtracking data is collected starting from the current time point to obtain real-time state time-series data fragments of the target real-world assets.

5. An AI-based RWA risk identification system, characterized in that, The system is used to implement the AI-based RWA risk identification method as described in any one of claims 1-4, and the system comprises: The historical asset status acquisition module is used to acquire the corresponding historical asset status time series data based on the entity category of the target real-world asset. The asset status prediction model training module is used to construct and train an artificial intelligence-based asset status prediction model based on the historical asset status time series data and by combining meta-learning and ensemble learning methods. The real-time status time series data acquisition module is used to configure data acquisition parameters according to the asset status prediction model and acquire real-time status time series data fragments of the target real-world asset accordingly. The prediction risk boundary construction module is used to input the real-time status time series data fragments into the asset status prediction model, obtain an initial prediction set, and construct the prediction risk boundary accordingly. The risk warning instruction generation module is used to continuously monitor the real-time asset status data of the target real-world asset, compare the real-time asset status data with the predicted risk boundary, and generate a risk warning instruction when the real-time asset status data exceeds the predicted risk boundary.

Images