A double-flow feature engineering method and system for stable coin risk early warning scenarios
By employing a dual-stream feature engineering approach and utilizing LSTM-VAE and TabNet models to process stablecoin market and transaction data, the problem of feature representation and threshold adaptability in stablecoin de-pegging risk monitoring is solved, enabling efficient risk warning and handling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-06-04
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies have limited feature representation capabilities and insufficient threshold adaptability in monitoring stablecoin de-pegging risks, making it difficult to capture implicit change patterns and nonlinear correlations, and the early warning results lack verifiability.
A dual-stream feature engineering approach is adopted, which extracts temporal latent features and reconstruction errors through the LSTM-VAE model, and generates dynamic thresholds by combining the TabNet model. Market data stream and transaction data stream are processed separately, and outputs a global market risk warning stream and a single transaction risk warning stream.
It has improved the ability to express market risk characteristics, enhanced the dynamic adaptability of transaction risk thresholds, improved the verifiability of early warning results, reduced false alarms and missed alarms, and improved the accuracy and efficiency of risk management.
Smart Images
Figure CN122492348A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of blockchain financial risk control, time series feature engineering and machine learning, and in particular to a dual-stream feature engineering method and system for stablecoin risk early warning scenarios. Background Technology
[0002] Stablecoins are important mediums of account, trading, and liquidity in the digital asset market, widely used for transaction settlement and on-chain fund transfers. They maintain value stability by maintaining a peg to fiat currency or other assets. Once they break away from this peg, it will trigger market panic, concentrated redemptions, on-chain liquidation, and a contraction in trading liquidity, thereby affecting the stable operation of related trading pairs, DeFi protocols, and the entire digital asset market.
[0003] To address the risk monitoring of stablecoin de-pegging, existing technologies typically construct risk indicators such as price deviation, short-term yield, volatility, trading volume, trading frequency, and liquidity changes from market and trading data, and use fixed thresholds, anomaly detection models, or machine learning classification models to identify the risk status. However, due to the characteristics of stablecoin market data, such as high-frequency volatility, high noise levels, low proportion of risk samples, and the time-series nature of risk evolution, existing methods have the following limitations: First, the feature expression capability is limited, relying heavily on manual statistical features, which can only depict explicit changes such as price, return, and volatility, and cannot fully express the implicit change patterns in continuous time windows and the nonlinear correlation between multiple indicators. Second, the threshold adaptability is insufficient, often using fixed thresholds or static discrimination rules, which cannot adapt to fluctuations under different market conditions, and are prone to false alarms when normal volatility increases, or false alarms when risks accumulate slowly. Third, the feature fusion and interpretation are insufficient, with a low degree of integration between market and trading features, and risk outputs are mostly single scores or classification results, lacking a correspondence with key risk features, which is not conducive to subsequent risk assessment and manual review. Summary of the Invention
[0004] The purpose of this invention is to provide a dual-stream feature engineering method and system for stablecoin risk warning scenarios, thereby solving the aforementioned technical problems.
[0005] To achieve the above objectives, this invention provides a dual-stream feature engineering method for stablecoin risk warning scenarios, comprising the following steps: S1: Receive stablecoin market data streams that enter at preset time intervals and stablecoin transaction data streams that enter based on transaction events; S2: For market data stream, calculate market statistical features, extract LSTM-VAE temporal latent features and calculate reconstruction error, concatenate market statistical features and LSTM-VAE features in time alignment and input them into TabNet model, output global market risk results, and generate dynamic threshold for trading risk based on global market risk results; S3: For the transaction data stream, calculate the risk characteristics of a single transaction and determine the risk of a single transaction based on the dynamic threshold of transaction risk. S4: Outputs global market risk warning stream and single transaction risk warning stream respectively.
[0006] Preferably, the calculation of market statistical characteristics in S2 is specifically as follows: Let the first... The stablecoin price at each sampling point is The anchor price is The sampling interval is The preset statistical window duration is Then the price deviation value Defined as: ; Corresponding window rate of return Defined as: ; in, , for window The corresponding number of sampling points; Further calculate the long-term price deviation. Market sentiment indicators and 5-minute deviation volatility This forms a set of market statistical characteristics, defined as: ; in, For the 30-minute window yield, The yield for a 60-minute window.
[0007] Preferably, the extraction of LSTM-VAE time series latent features in S2 specifically involves: using the market sequence within a continuous time window as the input to the LSTM-VAE model, with input features including price, trading volume, price deviation index, return rate, candlestick body length, candlestick amplitude, and market sentiment index; and outputting the latent distribution mean through the LSTM-VAE encoder. and the mean of the potential distribution As a temporal latent feature, the input sequence is calculated. With decoder reconstruction sequence The mean square error is used as the reconstruction error. LSTM-VAE feature set is formed. Defined as: ; in, As a preset latent feature dimension, For the first Each sampling point corresponds to Dimensional temporal latent features.
[0008] Preferably, the output of the global market risk result in S2 specifically involves: setting the market statistical characteristics set... With LSTM-VAE feature set Consolidate the data by time alignment to obtain the market risk feature vector. Defined as: ; Will Input a TabNet model and output the market risk probability. Market risk levels are classified according to preset probability ranges. .
[0009] Preferably, the dynamic threshold for generating transaction risk in S2 is specifically as follows: Let the first... The basic threshold for transaction risk characteristics is The preset lower threshold is The upper limit of the threshold is Then the first The corresponding sampling point of the th sampling point Dynamic threshold for transaction risk Defined as: ; in, This is an adjustment coefficient corresponding to the market risk level, and it decreases as the market risk level increases.
[0010] Preferably, the calculation of the risk characteristics of a single transaction in S3 is as follows: Let the current transaction be... The preset historical transaction window duration is Calculate the current transaction amount Transaction address in history window Number of transactions within Historical Window The cumulative transaction amount within Deviation from historical trading patterns This forms a set of risk characteristics for a single transaction. Defined as: .
[0011] Preferably, in S3, the risk assessment based on dynamic thresholds specifically involves: for the first... Risk characteristics of similar transactions If satisfied ≥ If any key transaction risk feature triggers a risk warning, or if multiple transaction risk features meet a preset combination judgment rule, the current transaction will be... Marked as a risky transaction.
[0012] Preferably, the dual early warning stream output in S4 is specifically: a global market risk early warning stream. Output based on market data sampling time, including stablecoin identifiers. Sampling time Market risk probability Market risk level and key features Global market risk early warning flow Defined as: ; Single Transaction Risk Warning Flow Output triggered by a transaction event, including the transaction hash. Trading hours The set of transaction risk characteristics that trigger risk rules Corresponding dynamic threshold set Risk assessment results Market risk level at the time of the transaction Single transaction risk warning flow Defined as: .
[0013] Preferably, it also includes offline model training, specifically: constructing training samples based on historical stablecoin market data, and standardizing price, trading volume, price deviation index, yield, candlestick body length, candlestick amplitude, and market sentiment index, assuming the th... The standardized market input vector for each sampling point is: The preset market time window length is Then continuous time window sequence samples for: ; Training the LSTM-VAE model, the encoder pairs Encode output latent distribution parameters and The latent vector is obtained after reparameterization. The decoder is based on Reconstructing the input sequence yields Training objectives for: ; in, Let be the mean square error function. For the first KL divergence values for each latent dimension This is the lower bound of the KL divergence term. The KL divergence weights increase with each training round; After training, take the mean of the latent distribution. As a latent temporal feature, the reconstruction error is calculated. Defined as: ; Market statistical characteristics LSTM-VAE temporal latent features and reconstruction error Consolidate the data by time alignment to obtain the market risk feature vector. Defined as: ; Historical risk label To train a TabNet model for monitoring signals, the model outputs the market risk probability. After training, save the model and standardized parameters; during inference, adjust the market risk probability according to the preset probability range. Mapped to market risk level Among them, market risk probability Defined as: .
[0014] A system for a dual-stream feature engineering method for stablecoin risk warning scenarios includes: The stablecoin market risk characteristic calculation and identification module is used to process stablecoin market data streams that enter at preset time intervals, calculate market statistical characteristics, extract LSTM-VAE time series potential characteristics, output market risk results through the TabNet model, and generate dynamic thresholds for trading risks based on the market risk results. The transaction risk characteristic calculation and judgment module is used to process transaction data streams that enter by transaction events, calculate transaction risk characteristics for a single transaction, and output transaction risk results based on dynamic transaction risk thresholds. The dual-early warning output module is connected to the stablecoin market risk characteristic calculation and identification module and the transaction risk characteristic calculation and judgment module, respectively. It is used to output a global market risk warning stream and a single transaction risk warning stream. The global market risk warning stream is output according to the market data sampling time, and the single transaction risk warning stream is output according to the transaction event.
[0015] Therefore, the present invention employs the above-mentioned dual-stream feature engineering method and system for stablecoin risk warning scenarios, which has the following beneficial effects: 1. Enhance the ability to express market risk characteristics: Based on traditional manual statistical features such as price deviation, yield, and volatility, the LSTM-VAE model is introduced to extract potential time-series features and reconstruction errors. This can effectively supplement the continuous time window change information and nonlinear correlation between multiple indicators that are difficult to capture by manual statistical features, and comprehensively depict the risk evolution characteristics of the stablecoin market.
[0016] 2. Enhance the dynamic adaptability of transaction risk thresholds: Establish a linkage mechanism between global market risk and individual transaction risk, and adaptively adjust the risk assessment threshold for individual transactions according to the global market risk level. This will make transaction risk assessment no longer dependent on fixed thresholds, and will automatically tighten or loosen the assessment criteria as the market risk status changes, thereby reducing false alarms and omissions.
[0017] 3. Improve the verifiability of early warning results: Adopt a dual early warning flow output mechanism to generate a global market risk early warning flow and a single transaction risk early warning flow, and clearly mark the key feature basis, trigger indicators and corresponding thresholds in the early warning results, so as to provide clear and specific data support for subsequent risk assessment and manual review, and improve the accuracy and efficiency of risk handling.
[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall structure of a dual-stream feature engineering method for stablecoin risk warning based on fusion temporal representation learning. Figure 2 This is a flowchart illustrating the stablecoin market risk characteristic calculation and identification module of the present invention; Figure 3 This is a flowchart illustrating the transaction risk characteristic calculation and determination module of the present invention; Figure 4 This is a flowchart illustrating the model training and deployment method of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.
[0021] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.
[0022] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0023] like Figures 1-4 As shown, this invention provides a dual-stream feature engineering method for stablecoin risk warning based on fusion temporal representation learning. By separating the parallel processing mechanism of market data stream and transaction data stream, and combining the fusion representation of statistical features and temporal potential features, it achieves linked early warning of global market risk and individual transaction risk. This effectively solves the problems of insufficient feature expression ability, poor threshold adaptability, and weak verifiability of early warning results in existing methods. The method includes the following steps: S1: Receive stablecoin market data streams entering at preset time intervals and stablecoin transaction data streams triggered by transaction events, respectively; S2: For the market data streams, calculate market statistical features, extract LSTM-VAE temporal latent features, and calculate reconstruction error. The reconstruction error reflects the degree of deviation between the current market state and historical normal market patterns. The larger the deviation, the higher the probability of market anomalies. This can effectively capture hidden risks that are difficult to detect with manual statistical features. Then, the market statistical features and LSTM-VAE features are concatenated in time alignment and input into the TabNet model to output global market risk results. Based on the global market risk results, a dynamic threshold for transaction risk is generated; S3: For the transaction data streams, calculate the risk features of individual transactions and determine the risk of individual transactions based on the dynamic threshold for transaction risk; S4: Output a global market risk warning stream and a single transaction risk warning stream, respectively. In the method of this invention, the input stablecoin market data stream... Transaction data stream anchor price (Generally set to 1), Time window length Transaction history window , No. The basic threshold for transaction risk characteristics is Threshold upper and lower limits Risk level classification rules, trained LSTM-VAE and TabNet models, and finally output a global market risk warning stream. Risk warning flow for single transactions .
[0024] The system first completes initialization, loading the market feature standardizer, pre-trained LSTM-VAE and TabNet models, basic transaction risk thresholds, and risk level classification rules. Then, it initiates the parallel processing of market data streams and transaction data streams. During the market data stream processing phase, for each market sampling point entering the system at preset time intervals... It reads price, volume, price deviation indicator, yield, candlestick body length, candlestick amplitude, and market sentiment indicator, and calculates multi-dimensional market statistical characteristics based on the current sampling point and its historical data to form a set of market statistical characteristics. Specifically, this includes: Let the first... The stablecoin price at each sampling point is The anchor price is The sampling interval is The preset statistical window duration is Then the price deviation value Defined as: ; Corresponding window rate of return Defined as: ; in, , for window The corresponding number of sampling points; Further calculate the long-term price deviation. Market sentiment indicators and 5-minute deviation volatility This forms a set of market statistical characteristics, defined as: ; in, For the 30-minute window yield, The yield for a 60-minute window.
[0025] Simultaneously, market sampling points are written in chronological order to a length of [length missing]. A sliding window, if the window length is insufficient The buffered sampling points wait for the next data input to ensure the continuity of temporal feature extraction; when the window length reaches... Construct a time window sequence Defined as: ; The time window sequence was processed in the same standardized manner as during the training phase. To ensure consistency in feature distribution between online inference and offline training, the standardized time window sequence is used. Input the LSTM-VAE model to obtain the latent distribution mean. and decoder reconstruction sequence and the mean of the potential distribution As a temporal latent feature, the input sequence is calculated. With decoder reconstruction sequence The mean square error is used as the reconstruction error. LSTM-VAE feature set is formed. Reconstruction error Defined as: ; LSTM-VAE Feature Set Defined as: ; in, As a preset latent feature dimension, For the first Each sampling point corresponds to Dimensional temporal latent features.
[0026] Set of market statistical characteristics With LSTM-VAE feature set Consolidate the data by time alignment to obtain the market risk feature vector. Defined as: ; Will Input a TabNet model and output the market risk probability. Market risk levels are classified according to preset probability ranges. .
[0027] Based on this, according to the market risk level If the transaction risk assessment threshold is dynamically updated, then the first... The corresponding sampling point of the th sampling point Dynamic threshold for transaction risk Defined as: ; in, This is an adjustment coefficient corresponding to the market risk level, which decreases as the market risk level increases, thus dynamically tightening the criteria for determining transaction risk. Finally, it outputs the global market risk warning flow corresponding to the current sampling point. .
[0028] For transaction data streams, execution is triggered by transaction events. When a transaction event occurs... Upon entering the system, the system first reads the basic information of the current transaction and its related entities in the history window. Calculate the current transaction amount from all transaction records within the current account. Transaction address in history window Number of transactions within Historical Window The cumulative transaction amount within Deviation from historical trading patterns This forms a set of risk characteristics for a single transaction. Defined as: .
[0029] Then, the market risk level corresponding to the time of the transaction is read. and dynamic threshold For each type of transaction risk characteristics Perform threshold checks one by one; if the conditions are met... If the trigger feature set is not empty, or if multiple transaction risk features meet the preset combination judgment rules, then the current transaction will be added to the trigger feature set. Mark transactions as risky and output a single transaction risk warning stream. Otherwise, it is recorded as no risk has been triggered. The system continuously executes the above market and transaction data flow processing flow in a loop to achieve 24 / 7 uninterrupted stablecoin risk monitoring.
[0030] This method employs a dual-early warning output mechanism, capable of simultaneously covering systemic risks at the macro market level and individual transaction risks at the micro transaction level. The global market risk early warning stream... Output based on market data sampling time, including stablecoin identifiers. Sampling time Market risk probability Market risk level and key features Global market risk early warning flow Defined as: .
[0031] Single Transaction Risk Warning Flow Output triggered by a transaction event, including the transaction hash. Trading hours The set of transaction risk characteristics that trigger risk rules Corresponding dynamic threshold set Risk assessment results Market risk level at the time of the transaction Single transaction risk warning flow Defined as: ; The warning results clearly indicate the characteristic basis and judgment threshold, which facilitates subsequent risk assessment and manual review, and greatly improves the interpretability and operability of the warning results.
[0032] To ensure the accuracy and consistency of the model, this method includes a complete offline model training and deployment process. The offline training phase includes: constructing training samples based on historical stablecoin market data; standardizing price, trading volume, price deviation indicators, yield, candlestick body length, candlestick amplitude, and market sentiment indicators to eliminate the influence of different units on model training; and setting the... The standardized market input vector for each sampling point is: The preset market time window length is , with a preset length Sliding window construction of continuous time window sequence samples This sequence is then used as the input sequence for the LSTM-VAE model. In one embodiment, the input fields include price, volume, depeg_abs, ret, candle_body, candle_range, and sentiment, along with the window length. .
[0033] Training an LSTM-VAE model involves the encoder processing the input sequence. Encode output latent distribution parameters and The latent vector is obtained after reparameterization. The decoder reconstructs the input sequence from the latent vectors to obtain... In one embodiment, the LSTM hidden dimension is 128 and the latent dimension is 32. The training objective consists of reconstruction error and KL divergence constraints, and KL weights are increased and lower bounds for KL terms are set. Defined as: ; in, Let be the mean square error function. For the first KL divergence values for each latent dimension This is the lower bound of the KL divergence term. The KL divergence weights increase with each training epoch; through the training objective This effectively balances reconstruction accuracy and latent space regularization, reducing the risk of overfitting or latent space collapse in the model. After training, the mean of the latent distribution is extracted. As a latent temporal feature, the reconstruction error is calculated. In one embodiment, 32-dimensional latent features and 1-dimensional reconstruction error are obtained.
[0034] The latent distribution mean Reconstruction error Market statistical characteristics Consolidate the data by time alignment to obtain the market risk feature vector. Defined as: ; Then use historical risk labels To train a TabNet model for monitoring signals, the model outputs the market risk probability. After training is complete, save the trained model and related standardized parameters.
[0035] During inference, the system loads a pre-trained model and parameters, and assigns market risk probabilities according to a preset probability range. Mapped to market risk level Among them, market risk probability Defined as: .
[0036] During model training, the online inference process strictly follows the field order, window length, standardization method, and feature splicing interface as used in the offline training phase to ensure consistency in feature interfaces between training and inference.
[0037] Corresponding to the above method, this invention also provides a dual-stream feature engineering system for stablecoin risk warning scenarios. This system mainly consists of three interconnected functional modules. The stablecoin market risk feature calculation and identification module is responsible for processing stablecoin market data streams entering at preset time intervals, completing core operations such as market statistical feature calculation, LSTM-VAE temporal potential feature extraction and reconstruction error calculation, feature fusion and splicing, TabNet market risk identification, and dynamic threshold generation for transaction risks. The transaction risk feature calculation and judgment module is responsible for processing transaction data streams entering according to transaction events, completing operations such as single transaction risk feature calculation, dynamic threshold reading, and risk judgment. The dual warning stream output module is communicatively connected to the first two modules, responsible for outputting a global market risk warning stream and a single transaction risk warning stream. The global market risk warning stream is output according to the market data sampling time, and the single transaction risk warning stream is output according to the transaction event.
[0038] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A double-flow feature engineering method for stable coin risk early warning scenarios, characterized in that: Includes the following steps: S1: Receive stablecoin market data streams that enter at preset time intervals and stablecoin transaction data streams that enter based on transaction events; S2: For market data stream, calculate market statistical features, extract LSTM-VAE temporal latent features and calculate reconstruction error, concatenate market statistical features and LSTM-VAE features in time alignment and input them into TabNet model, output global market risk results, and generate dynamic threshold for trading risk based on global market risk results; S3: For the transaction data stream, calculate the risk characteristics of a single transaction and determine the risk of a single transaction based on the dynamic threshold of transaction risk. S4: Outputs global market risk warning stream and single transaction risk warning stream respectively.
2. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 1, characterized in that, The specific calculation of the market statistical feature in S2 is: assuming that the stable coin price of the i th sampling point is , the anchor price is , the sampling interval is , and the preset statistical window length is , then the price deviation value is defined as: ； Corresponding window rate of return Defined as: ； in, , for window The corresponding number of sampling points; Further calculate the long-term price deviation. Market sentiment indicators and 5-minute deviation volatility This forms a set of market statistical characteristics, defined as: ； in, For the 30-minute window yield, The yield for a 60-minute window.
3. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 1, characterized in that, S2 extracts the latent time series features for LSTM-VAE as follows: The market sequence within a continuous time window is used as the input to the LSTM-VAE model. Input features include price, volume, price deviation indicators, returns, candlestick body length, candlestick amplitude, and market sentiment indicators. The latent distribution mean is output through the LSTM-VAE encoder. and the mean of the potential distribution As a temporal latent feature, the input sequence is calculated. With decoder reconstruction sequence The mean square error is used as the reconstruction error. LSTM-VAE feature set is formed. Defined as: ； in, As a preset latent feature dimension, For the first Each sampling point corresponds to Dimensional temporal latent features.
4. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 3, characterized in that, Specifically, S2 outputs the global market risk result by: setting the market statistical characteristics set. With LSTM-VAE feature set Consolidate the data by time to obtain the market risk feature vector. Defined as: ； Will Input a TabNet model and output the market risk probability. Market risk levels are classified according to preset probability ranges. .
5. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 4, characterized in that, The specific method for generating dynamic thresholds for trading risk in S2 is as follows: Let the first... The basic threshold for transaction risk characteristics is The preset lower threshold is The upper limit of the threshold is Then the first The corresponding sampling point of the th sampling point Dynamic threshold for transaction risk Defined as: ； in, This is an adjustment coefficient corresponding to the market risk level, and it decreases as the market risk level increases.
6. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 1, characterized in that, The specific steps for calculating the risk characteristics of a single transaction in S3 are as follows: Let the current transaction be... The preset historical transaction window duration is Calculate the current transaction amount Transaction address in history window Number of transactions within Historical Window The cumulative transaction amount within Deviation from historical trading patterns This forms a set of risk characteristics for a single transaction. Defined as: 。 7. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 6, characterized in that, In S3, risk assessment based on dynamic thresholds specifically involves: for the... Risk characteristics of similar transactions If satisfied ≥ If so, this feature will trigger a risk warning; When any key transaction risk feature triggers a risk rule, or when multiple transaction risk features meet a preset combination judgment rule, the current transaction will be... Marked as a risky transaction.
8. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 1, characterized in that: The dual early warning stream output in S4 is specifically: Global Market Risk Early Warning Stream. Output based on market data sampling time, including stablecoin identifiers. Sampling time Market risk probability Market risk level and key features Global market risk early warning flow Defined as: ； Single Transaction Risk Warning Flow Output triggered by a transaction event, including the transaction hash. Trading hours The set of transaction risk characteristics that trigger risk rules Corresponding dynamic threshold set Risk assessment results Market risk level at the time of the transaction Single transaction risk warning flow Defined as: 。 9. The dual-stream feature engineering method for stablecoin risk warning scenarios according to claim 1, characterized in that: This also includes offline model training, specifically: constructing training samples based on historical stablecoin market data, and standardizing price, trading volume, price deviation index, yield, candlestick body length, candlestick amplitude, and market sentiment index. Let the... The standardized market input vector for each sampling point is: The preset market time window length is Then continuous time window sequence samples for: ； Training the LSTM-VAE model, the encoder pairs Encode output latent distribution parameters and The latent vector is obtained after reparameterization. The decoder is based on Reconstructing the input sequence yields Training objectives for: ； in, Let be the mean square error function. For the first KL divergence values for each latent dimension This is the lower bound of the KL divergence term. The KL divergence weights increase with each training round; After training, take the mean of the latent distribution. As a latent temporal feature, the reconstruction error is calculated. Defined as: ； Market statistical characteristics LSTM-VAE temporal latent features and reconstruction error Consolidate the data by time alignment to obtain the market risk feature vector. Defined as: ； Historical risk label To train a TabNet model for monitoring signals, the model outputs the market risk probability. After training, save the model and standardized parameters; during inference, adjust the market risk probability according to the preset probability range. Mapped to market risk level Among them, market risk probability Defined as: 。 10. A system for a dual-stream feature engineering method for stablecoin risk warning scenarios according to any one of claims 1-9, characterized in that, include: The stablecoin market risk characteristic calculation and identification module is used to process stablecoin market data streams that enter at preset time intervals, calculate market statistical characteristics, extract LSTM-VAE time series potential characteristics, output market risk results through the TabNet model, and generate dynamic thresholds for trading risks based on the market risk results. The transaction risk characteristic calculation and judgment module is used to process transaction data streams that enter by transaction events, calculate transaction risk characteristics for a single transaction, and output transaction risk results based on dynamic transaction risk thresholds. The dual-early warning output module is connected to the stablecoin market risk characteristic calculation and identification module and the transaction risk characteristic calculation and judgment module, respectively. It is used to output a global market risk warning stream and a single transaction risk warning stream. The global market risk warning stream is output according to the market data sampling time, and the single transaction risk warning stream is output according to the transaction event.