CN-121981827-A - Financial transaction abnormal data real-time monitoring method and system based on deep learning

CN121981827ACN 121981827 ACN121981827 ACN 121981827ACN-121981827-A

Abstract

The invention relates to the technical field of financial risk control, in particular to a method and a system for monitoring abnormal data of financial transactions in real time based on deep learning. The method comprises the steps of collecting external macroscopic variables in real time, identifying macroscopic states at the current market, generating macroscopic state control signals, transmitting the macroscopic state control signals to a preset dynamic baseline management unit, retrieving a situation baseline model matched with the macroscopic states from a preset situation baseline model library, retrieving situation baseline parameters corresponding to the situation baseline model, encoding financial transaction flow data by using a deep neural network, generating microscopic behavior feature vectors, calculating semantic distances between the microscopic behavior feature vectors and normal feature spaces, comparing the semantic distances with a preset judging threshold value, and generating abnormal risk alarms in response to the semantic distances being larger than the judging threshold value. The invention obviously reduces the false alarm rate when the market fluctuates severely and the false alarm rate under the stable market, and improves the monitoring accuracy.

Inventors

ZHAO CHEN
DU CHENJIE
Chai Bencheng

Assignees

浙江万里学院

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (8)

1. The real-time monitoring method for abnormal data of financial transactions based on deep learning is characterized by comprising the following steps: collecting external macroscopic variables in real time; collecting financial transaction flow data in real time; Inputting an external macroscopic variable into a preset time sequence state model; Carrying out multi-source fusion analysis on external macroscopic variables by using a time sequence state model; Identifying a macroscopic state at the current market; generating a macro state control signal based on the macro state; Transmitting the macro state control signal to a preset dynamic baseline management unit; in response to the macro state control signal, retrieving a context baseline model matching the macro state from a preset context baseline model library; calling a situation baseline parameter corresponding to the situation baseline model; encoding the financial transaction streaming data by using a deep neural network; Generating microscopic behavior feature vectors; defining a normal feature space based on the contextual baseline parameters; calculating the semantic distance between the microscopic behavior feature vector and the normal feature space; Comparing the semantic distance with a preset judging threshold value; an abnormal risk alert is generated in response to the semantic distance being greater than the discrimination threshold.
2. The method for monitoring abnormal data of financial transactions based on deep learning according to claim 1 wherein said external macroscopic variables include panic index, industry specific ETF fund flow rate and financial news emotion heat index; the identifying a macroscopic state at the current market place includes: Clustering the current market environment into preset macroscopic state categories; The macroscopic state categories include stationary low wave dynamics, high frequency concussions, and extreme panic unilateral states.
3. The method for monitoring abnormal data of financial transactions based on deep learning according to claim 1, wherein the construction process of the context baseline model library comprises: Acquiring historical transaction data with a time stamp and a historical macro index; Dividing historical transaction data into different situation slices according to the historical macro state corresponding to the historical macro index; Independently training an unsupervised feature extraction model for each context slice; extracting a distribution center and boundary parameters of normal transaction behaviors in a feature space under a situation slice by using an unsupervised feature extraction model; and storing the distribution center and the boundary parameters into a situation baseline model library.
4. The method for monitoring abnormal data of financial transactions based on deep learning according to claim 1, wherein the generating microscopic behavior feature vectors comprises: Converting discrete financial transaction flow data into continuous high-dimensional vectors by using a deep neural network; generating a multidimensional numerical matrix qualitatively expressing deviation degree of financial transaction flow data relative to historical account behaviors, rhythm characteristics of operation frequency and topological form of fund flow; the multi-dimensional numerical matrix is determined as a microscopic behavioral feature vector.
5. The method for monitoring abnormal data of financial transactions based on deep learning according to claim 1, wherein calculating semantic distance between microscopic behavior feature vector and normal feature space comprises: taking the situation baseline parameter as a centroid or a distribution density function of a normal feature space; Calculating the mahalanobis distance of the microscopic behavior feature vector relative to the centroid; or calculating the log likelihood probability of the microscopic behavior feature vector under the distribution density function; the mahalanobis distance or log likelihood probability is determined as the semantic distance.
6. The method for monitoring abnormal data of financial transactions based on deep learning in real time according to claim 1, further comprising: collecting a discrimination result obtained by comparing the semantic distance with a discrimination threshold value and subsequent business rechecking feedback; Calculating a loss function based on business rechecking feedback; generating a gradient update signal; Responding to the fact that the false alarm rate in a specific macroscopic state is continuously higher than a preset level, and sending a correction instruction to a dynamic baseline management unit; adjusting boundaries of context baseline parameters corresponding to the specific macroscopic state based on the correction instruction; responding to macroscopic state switching hysteresis, and sending a correction instruction to the time sequence state model; the state partitioning sensitivity of the time series state model is optimized based on the correction instructions.
7. The financial transaction abnormal data real-time monitoring system based on deep learning, which is based on the financial transaction abnormal data real-time monitoring method based on deep learning as claimed in any one of claims 1 to 6, is characterized by comprising the following steps: The macro state sensing unit is used for collecting external macro variables in real time, inputting the external macro variables into a preset time sequence state model, identifying the macro state of the current market by using the time sequence state model, and generating a macro state control signal based on the macro state; The dynamic baseline management unit is used for storing a preset situation baseline model library, receiving a macro state control signal, responding to the macro state control signal, retrieving a situation baseline model matched with the macro state from the situation baseline model library, and retrieving situation baseline parameters corresponding to the situation baseline model; the micro-feature extraction unit is used for collecting financial transaction flow data in real time, encoding the financial transaction flow data by using the deep neural network and generating micro-behavior feature vectors; The context coupling discriminating unit is used for defining a normal feature space based on the context baseline parameters, calculating the semantic distance between the microscopic behavior feature vector and the normal feature space, comparing the semantic distance with a preset discriminating threshold value, and generating an abnormal risk alarm in response to the semantic distance being larger than the discriminating threshold value.
8. The deep learning based financial transaction anomaly data real-time monitoring system of claim 7, further comprising: The self-adaptive closed-loop feedback unit is used for collecting the discrimination result of the situation coupling discrimination unit and service rechecking feedback, calculating a loss function based on the service rechecking feedback and generating a gradient update signal; the self-adaptive closed-loop feedback unit is also used for sending a correction instruction to the dynamic baseline management unit to adjust the boundary of the situation baseline parameter in response to the false alarm rate in a specific macroscopic state being continuously higher than a preset level, and sending the correction instruction to the macroscopic state sensing unit to optimize the state division sensitivity in response to the macroscopic state switching hysteresis.

Description

Financial transaction abnormal data real-time monitoring method and system based on deep learning Technical Field The invention relates to the technical field of financial risk control, in particular to a method and a system for monitoring abnormal data of financial transactions in real time based on deep learning. Background Real-time risk monitoring of financial transactions is a key link for maintaining market stability, and the core technical problem is how to accurately identify abnormal transactions in complex and changeable market environments and balance false alarm rate and false alarm rate; In the prior art, the judgment is mostly carried out by depending on a static threshold value or a single model, the deep influence of a macroscopic market environment on microscopic transaction behaviors is ignored, and the macroscopic state is not used as a dynamic pre-condition for anomaly judgment, so that the judgment standard cannot be adaptively adjusted along with market fluctuation, namely, when the market fluctuates severely, a benign outlier with panic is easily misreported as anomaly, and when the market is stable, the high sensitivity to fine malicious attacks is difficult to maintain; In addition, the static system cannot cope with the conceptual drift of the market mode, and the performance is reduced when facing the continuously evolving attack means, so that a dynamic monitoring scheme capable of fusing macroscopic situation and microscopic characteristics is needed. The above information disclosed in the above background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to those of ordinary skill in the art. Disclosure of Invention In order to solve the technical problems, the invention discloses a financial transaction abnormal data real-time monitoring method and system based on deep learning, and specifically, the technical scheme of the invention is as follows: the real-time monitoring method for abnormal data of financial transactions based on deep learning comprises the following steps: collecting external macroscopic variables in real time; collecting financial transaction flow data in real time; Inputting an external macroscopic variable into a preset time sequence state model; Carrying out multi-source fusion analysis on external macroscopic variables by using a time sequence state model; Identifying a macroscopic state at the current market; generating a macro state control signal based on the macro state; Transmitting the macro state control signal to a preset dynamic baseline management unit; in response to the macro state control signal, retrieving a context baseline model matching the macro state from a preset context baseline model library; calling a situation baseline parameter corresponding to the situation baseline model; encoding the financial transaction streaming data by using a deep neural network; Generating microscopic behavior feature vectors; defining a normal feature space based on the contextual baseline parameters; calculating the semantic distance between the microscopic behavior feature vector and the normal feature space; Comparing the semantic distance with a preset judging threshold value; an abnormal risk alert is generated in response to the semantic distance being greater than the discrimination threshold. Optionally, the external macroscopic variables include panic index, industry-specific ETF funds flow rate, and financial news emotion heat index; the identifying a macroscopic state at the current market place includes: Clustering the current market environment into preset macroscopic state categories; The macroscopic state categories include stationary low wave dynamics, high frequency concussions, and extreme panic unilateral states. Optionally, the building process of the context baseline model library includes: Acquiring historical transaction data with a time stamp and a historical macro index; Dividing historical transaction data into different situation slices according to the historical macro state corresponding to the historical macro index; Independently training an unsupervised feature extraction model for each context slice; extracting a distribution center and boundary parameters of normal transaction behaviors in a feature space under a situation slice by using an unsupervised feature extraction model; and storing the distribution center and the boundary parameters into a situation baseline model library. Optionally, the generating the micro-behavior feature vector includes: Converting discrete financial transaction flow data into continuous high-dimensional vectors by using a deep neural network; generating a multidimensional numerical matrix qualitatively expressing deviation degree of financial transaction flow data relative to historical account behaviors, rhythm characteristics of operation frequency and top