A Visualized Risk Detection Method for Settlement Business Based on Intelligent Risk Identification
By combining multimodal data fusion and hybrid learning models with visualization and incremental learning, the problem of insufficient multi-dimensional data fusion in settlement business risk detection has been solved, enabling real-time risk identification and automatic handling, improving the accuracy and comprehensiveness of risk identification, and meeting the requirements for real-time response.
Patent Information
- Application Number
- CN202611288688.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing risk detection methods for settlement operations fail to achieve deep integration of multi-dimensional data, resulting in an inability to fully explore the complex correlations and complementary values between different modalities of data, an inability to effectively identify unknown emerging frauds, and a lack of real-time response capabilities and intuitive risk visualization interfaces, thus affecting the comprehensiveness and accuracy of risk identification.
Multimodal data acquisition and preprocessing are employed, and a hybrid model combining supervised learning, unsupervised learning, and reinforcement learning is used for risk identification. Cross-modal feature fusion is performed through an attention mechanism to generate a unified fused feature vector, thereby achieving risk scoring and level classification. Combined with visualization and incremental learning mechanisms, differentiated treatment strategies are implemented.
It enables millisecond-level real-time detection and hierarchical automatic handling of settlement business risks, improves the accuracy and comprehensiveness of risk identification, enhances the efficiency of risk decision-making and model adaptability, reduces the rate of missed judgments and false judgments, meets the requirements of real-time response, and provides an intuitive risk visualization interface.
Smart Images

Figure CN123117786A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of settlement business risk detection technology, and in particular to a method for visualizing settlement business risks based on intelligent risk identification. Background Technology
[0002] Settlement services, as a core infrastructure of the modern financial system, directly impact financial security and stability through their risk control capabilities. With the widespread adoption of digital finance, diversified transaction scenarios, fragmented channels, and the intelligent evolution of fraud methods have led to the emergence of new and complex models such as cross-channel collaborative fraud and account-based group fraud. Traditional risk control systems based on rule engines and manual review are no longer sufficient to cope. Related technologies have gradually shifted from single-rule-driven to multimodal data fusion and AI-driven approaches. These include continuous fraud detection through autoencoders, anomaly identification by integrating transaction and behavioral data, building intelligent risk control systems by combining blockchain and multimodal AI, and implementing tiered risk control by introducing machine learning to adjust rule weights. This encompasses multiple key stages from data collection and feature extraction to risk assessment.
[0003] However, existing risk detection methods for settlement operations often rely on single models or shallow feature fusion, failing to achieve deep integration of multi-dimensional data such as transactions, behaviors, devices, biometrics, and context. This can lead to an inability to fully exploit the complex relationships and complementary value between different modalities, or a severe deficiency in detecting only known fraud patterns while neglecting emerging frauds, thus impacting the comprehensiveness and accuracy of risk identification. Furthermore, current technologies generally lack a complete business loop from data collection, identification, and display to processing and optimization. Complex models are slow inference speeds, failing to meet the millisecond-level real-time response requirements of settlement operations, and lack intuitive and comprehensive risk visualization interfaces, hindering risk management personnel from conducting global monitoring and making rapid decisions. Summary of the Invention
[0004] The present invention aims to at least partially solve one of the technical problems in the related art.
[0005] Therefore, the first objective of this invention is to propose a method for visually detecting settlement business risks based on intelligent risk identification.
[0006] Another objective of this invention is to propose a settlement business risk visualization detection device based on intelligent risk identification.
[0007] The third objective of this invention is to provide a computer device.
[0008] The fourth objective of this invention is to provide a non-transitory computer-readable storage medium.
[0009] To achieve the above objectives, a first aspect of the present invention proposes a method for visually detecting settlement business risks based on intelligent risk identification, comprising: S1. Collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. Then, perform data cleaning, normalization, encoding and timestamp alignment processing on the collected data in sequence to obtain preprocessed multimodal data. S2, Modality-specific features are extracted from the preprocessed multimodal data to obtain feature data for each modality, and the feature data for each modality are processed by cross-modal fusion based on an attention mechanism to generate a unified fusion feature vector; S3, the unified fusion feature vector is input into the supervised learning model, the unsupervised learning model and the reinforcement learning model respectively for risk identification, and the identification results of the three models are weighted and fused to obtain the risk score of the transaction, and the risk level is divided according to the risk score; S4. Implement differentiated handling strategies according to the risk level, and incrementally update the supervised learning model, unsupervised learning model, and reinforcement learning model based on the handling results and newly added risk data. At the same time, generate visualized risk monitoring information to display the risk status.
[0010] In one embodiment of the present invention, S1 includes: Duplicate records are removed using the transaction serial number as the unique primary key, and outliers in numerical fields are identified using the 3σ criterion. First, the sample mean μ and standard deviation σ of the field are calculated. Data that satisfy |x-μ|>3σ are identified as outliers. Outliers are filled with the median of the same field to obtain cleaned multimodal data. The numerical features in the cleaned multimodal data are mapped to the [0,1] interval using the Min-Max normalization method. The normalization calculation formula is as follows: ,in , The minimum and maximum values of the corresponding fields are used to obtain the normalized multimodal data; Differential encoding is used for categorical features in the normalized multimodal data. Low-cardinality categorical features are converted into 0-1 vectors of the corresponding dimension using one-hot encoding. High-cardinality categorical features are mapped to low-dimensional dense vectors using a learnable embedding matrix. The embedding encoding calculation formula is as follows: ,in The index vector corresponding to the category. To embed the weight matrix, the encoded multimodal data is obtained; Based on the transaction occurrence time, a sliding window with a fixed time step is used to perform time-series alignment on the encoded multimodal data. Missing sampling points within the window are filled using linear interpolation, with the interpolation formula being: ,in , For adjacent sampling times, , The values at the corresponding times are used to obtain the preprocessed multimodal data.
[0011] In one embodiment of the present invention, it further includes: Low-cardinality category features of transaction channels and document types are encoded using one-hot encoding and converted into 0-1 vectors of the corresponding dimensions; Embedding encoding is used for high cardinality category features of account identifiers and device fingerprints. Discrete categories are mapped to low-dimensional dense vectors through a learnable embedding matrix.
[0012] In one embodiment of the present invention, S2 includes: For structured transaction data and device environment data, a three-layer deep neural network with residual connections is used for feature extraction. The three layers consist of an input layer, a 256-dimensional hidden layer, a 128-dimensional hidden layer, and a 64-dimensional hidden layer. Each hidden layer uses the ReLU activation function, and residual connections are added between adjacent hidden layers. The forward propagation core is calculated as follows: The residual output is The final output is a 64-dimensional dense feature. For sequential operation behavior data, a bidirectional long short-term memory network with temporal attention is used for feature extraction. An attention layer is added after the output of BiLSTM to automatically weight key temporal segments. After the bidirectional outputs are concatenated, attention weighting is applied to obtain 128-dimensional temporal features. For biometric data, a lightweight convolutional neural network based on depthwise separable convolution is used for feature extraction. The depthwise separable convolution is divided into two steps: depthwise convolution and pointwise convolution. The depthwise convolution independently convolves each input channel, and the pointwise convolution completes channel fusion through 1×1 convolution. Finally, 96-dimensional features are output through global average pooling. For text-based contextual data, a Chinese BERT-base pre-trained model is used to extract text features. The output corresponding to the delimiter is taken as the sentence-level semantic vector, and then the dimensionality is reduced to 224 dimensions through a fully connected layer. The dimensionality reduction calculation formula is as follows: .
[0013] In one embodiment of the present invention, it further includes: Through the Gate of Oblivion Input gate Cell state update Output gate and hidden state Complete the gated computation of unidirectional LSTM; The outputs of the forward and backward LSTMs are concatenated to obtain a bidirectional output; An attention layer is added after the bidirectional output, and 128-dimensional temporal features are obtained by automatically weighting key time segments.
[0014] In one embodiment of the present invention, S3 includes: The unified fused feature vector is input into the supervised learning model. The supervised learning model trains XGBoost and a deep neural network using labeled historical fraud data. XGBoost uses log loss with L1 and L2 regularization terms as the objective function to iteratively construct a regression tree to fit the residuals. The deep neural network uses cross-entropy loss as the optimization objective. The outputs of the two models are fused by weighted average to identify known fraud patterns and obtain the supervised learning risk identification result. The unified fused feature vector is input into an unsupervised learning model, which combines an isolated forest and autoencoder algorithms. The isolated forest constructs multiple isolated trees by randomly dividing the feature space and calculates the anomaly score based on the average path length of the samples. The autoencoder compresses the input into a low-dimensional latent vector through an encoder and then reconstructs the input through a decoder, using the reconstruction error as the anomaly score. The two algorithms detect global anomalies and local anomalies respectively. After the results are fused, unknown and emerging fraud patterns are identified, resulting in unsupervised learning risk identification results. The unified fused feature vector is input into a reinforcement learning model, which employs a deep Q-network algorithm. The model uses the fused features and risk assessment result of the current transaction as the state, the optional actions as the actions, and the final outcome as the reward. The training process is stabilized through experience replay and a target network mechanism. The target Q-value is calculated as follows: ,in As a discount factor, The target network parameters are defined as follows: the network is updated by minimizing the mean square error between the Q value and the target value; the risk judgment strategy is continuously optimized through trial and error learning to obtain the reinforcement learning risk identification result. The results of supervised learning risk identification, unsupervised learning risk identification, and reinforcement learning risk identification are weighted and fused with weights of 0.5, 0.3, and 0.2, respectively, to obtain the final risk score of the transaction, which is then divided into four risk levels: low, medium, high, and extremely high.
[0015] In one embodiment of the present invention, S4 includes: When the risk level is low, the transaction is automatically approved and the risk information is recorded, which is then used as a training sample for subsequent model iterations. When the risk level is medium risk, intelligent handling suggestions are pushed to the reviewers, and an investigation report containing transaction details, related account information and historical risk records is automatically generated, along with auxiliary tools for correlation analysis and behavior trajectory analysis. When the risk level is high, the account's withdrawal and transfer functions will be automatically frozen, and the request will be sent to senior reviewers for priority review. When the risk level is extremely high, the transaction will be directly intercepted and all funds in the associated account will be frozen. The relevant departments will be notified and a fund recovery procedure will be initiated in accordance with regulatory requirements.
[0016] To achieve the above objectives, a second aspect of the present invention provides a settlement business risk visualization detection device based on intelligent risk identification, comprising: The multimodal preprocessing module is used to collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. The module performs data cleaning, normalization, encoding and timestamp alignment on the collected data in sequence to obtain preprocessed multimodal data. The cross-modal feature fusion module is used to extract modality-specific features from the preprocessed multimodal data to obtain feature data for each modality, and to generate a unified fused feature vector by performing cross-modal fusion processing based on an attention mechanism on the feature data of each modality. The multi-model risk discrimination module is used to input the unified fused feature vector into the supervised learning model, the unsupervised learning model and the reinforcement learning model respectively for risk identification, and to perform weighted fusion of the identification results of the three models to obtain a risk score for the transaction, and to classify the risk level according to the risk score; The disposal iteration monitoring module is used to execute differentiated disposal strategies according to the risk level, and to incrementally update the supervised learning model, unsupervised learning model and reinforcement learning model based on the disposal results and newly added risk data, while generating visualized risk monitoring information to display the risk status.
[0017] The present invention discloses a method and apparatus for visualizing and detecting settlement business risks based on intelligent risk identification. This method and apparatus can achieve millisecond-level real-time detection and hierarchical automatic handling of settlement business risks, improve the accuracy and comprehensiveness of risk identification, and enhance the efficiency of risk decision-making and the adaptive capability of the model through multi-dimensional visualization and incremental learning mechanisms.
[0018] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing the method described in the first aspect embodiment.
[0019] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the first aspect.
[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0021] Figure 1 This is a flowchart of a settlement business risk visualization detection method based on intelligent risk identification according to an embodiment of the present invention; Figure 2 This is a structural diagram of a settlement business risk visualization detection device based on intelligent risk identification according to an embodiment of the present invention; Figure 3 It is a computer device according to an embodiment of the present invention. Detailed Implementation
[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.
[0024] The following description, with reference to the accompanying drawings, describes a method and apparatus for visually detecting settlement business risks based on intelligent risk identification, according to an embodiment of the present invention.
[0025] Figure 1 This is a flowchart of a settlement business risk visualization detection method based on intelligent risk identification according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1. Collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. Then, perform data cleaning, normalization, encoding and timestamp alignment processing on the collected data in sequence to obtain preprocessed multimodal data. S2, Modality-specific features are extracted from the preprocessed multimodal data to obtain feature data for each modality, and the feature data for each modality are processed by cross-modal fusion based on an attention mechanism to generate a unified fusion feature vector; S3, the unified fusion feature vector is input into the supervised learning model, the unsupervised learning model and the reinforcement learning model respectively for risk identification, and the identification results of the three models are weighted and fused to obtain the risk score of the transaction, and the risk level is divided according to the risk score; S4. Implement differentiated handling strategies according to the risk level, and incrementally update the supervised learning model, unsupervised learning model, and reinforcement learning model based on the handling results and newly added risk data. At the same time, generate visualized risk monitoring information to display the risk status.
[0026] The settlement business visualization risk detection platform provided by this invention adopts a modular design and includes six core functional modules. The modules interact with each other and transmit instructions through standardized interfaces.
[0027] Furthermore, the multimodal data acquisition and preprocessing module serves as the platform's data entry point. Through a unified interface, it connects to the bank's core system, payment system, online banking, mobile banking, third-party payment platforms, and external data service providers, collecting five major categories of data in real time: transaction logs, operational behaviors, device environment, biometrics, and contextual information. After acquisition, the module sequentially performs data cleaning, normalization, encoding, and timestamp alignment operations. In the data cleaning phase, duplicate records are first removed using the transaction log number as the unique primary key, and then... The criteria for identifying outliers in numeric fields first calculate the sample mean of the field. with standard deviation , will satisfy Data identified as outliers were filled using the median of the same field to avoid missing time-series information due to direct deletion. The normalization stage employed the Min-Max normalization method, mapping all numerical features to the [0,1] interval. The calculation formula is as follows: .in , To standardize the value ranges of features with different dimensions for the minimum and maximum values of the corresponding fields, the encoding stage employs differentiated encoding for categorical features. Low-cardinality categorical features such as transaction channels and document types are encoded using one-hot encoding, converting them into 0-1 vectors of the corresponding dimensions. High-cardinality categorical features such as account identifiers and device fingerprints are encoded using embedding, mapping discrete categories to low-dimensional dense vectors through a learnable embedding matrix. The calculation formula is as follows: ,in The index vector corresponding to the category. To embed the weight matrix, the timestamp alignment stage uses the transaction occurrence time as the benchmark and employs a sliding window with a fixed time step to perform temporal alignment of the multimodal data. Missing sampling points within the window are filled using linear interpolation, with the interpolation formula being: ,in , For adjacent sampling times, , To ensure the consistency of data across different modalities in the time dimension, the values are provided at the corresponding time points.
[0028] Further, after data preprocessing, the system sends the data to the multimodal data fusion module for feature-level fusion. This module adopts a two-stage fusion strategy, with each stage's network structure optimized for financial risk scenarios, rather than being a simple concatenation of general models. The first stage is modality-specific feature extraction, selecting appropriate improved neural network models based on the characteristics of different data types. For structured transaction and device data, a three-layer deep neural network with residual connections is used. The network consists of an input layer, a 256-dimensional hidden layer (ReLU activation), a 128-dimensional hidden layer (ReLU activation), and a 64-dimensional hidden layer (ReLU activation). Residual connections are added between adjacent hidden layers to alleviate the gradient vanishing problem in deep networks. The forward propagation core computation is... The residual output is The final output is a 64-dimensional dense feature. For sequential operational behavior data, a bidirectional long short-term memory network with temporal attention is used. An attention layer is added after the BiLSTM output to automatically weight key temporal segments. The core gating computation of the unidirectional LSTM includes the forget gate. Input gate Cell state update Output gate Hidden state The bidirectional outputs are concatenated and then weighted by attention to obtain 128-dimensional temporal features. For biometric data, a lightweight convolutional neural network based on depthwise separable convolution is used. Depthwise separable convolution replaces standard convolution to compress the number of parameters. Depthwise separable convolution consists of two steps: depthwise convolution and pointwise convolution. Depthwise convolution independently convolves each input channel, while pointwise convolution fuses channels through 1×1 convolution. Finally, global average pooling outputs 96-dimensional features. For textual contextual data, a Chinese BERT-based pre-trained model is used to extract text features. The output corresponding to the delimiter is taken as the sentence-level semantic vector, and then the dimensionality is reduced to 224 dimensions through a fully connected layer. The dimensionality reduction calculation formula is as follows: The second stage is cross-modal feature fusion, employing an 8-head attention mechanism that incorporates modality embeddings. Learnable modality embedding vectors are added to the standard multi-head attention mechanism, and unique identifiers are added to different modality features to mitigate cross-modal semantic differences. First, each modality feature is added to its corresponding modality embedding, then mapped to a query Q, key K, and value V matrix. The single-head attention is calculated as follows: ,in The outputs of the eight attention heads are concatenated and linearly transformed to generate a unified fusion feature vector with a dimension of 512, which is the key vector dimension.
[0029] Furthermore, after receiving the fused feature vector, the intelligent risk identification engine performs risk assessment using a hybrid model combining supervised learning, unsupervised learning, and reinforcement learning. The supervised learning sub-model trains XGBoost and a deep neural network separately using labeled historical fraud data. XGBoost uses a logarithmic loss with L1 and L2 regularization terms as the objective function, iteratively constructing a regression tree to fit the residuals. The deep neural network uses cross-entropy loss as the optimization objective. The outputs of the two models are fused through a weighted average to identify known fraud patterns such as account theft, identity fraud, and money laundering. The unsupervised learning sub-model combines isolated forest and autoencoder algorithms. Isolated forest constructs multiple isolated trees by randomly partitioning the feature space, calculating anomaly scores based on the average path length of the samples; higher scores indicate a greater probability of anomaly. The autoencoder compresses the input into a low-dimensional latent vector through an encoder, then reconstructs the input through a decoder, using the reconstruction error as the anomaly score. The two algorithms detect global and local anomalies respectively, and the fused results are used to identify unknown and emerging fraud patterns. The reinforcement learning sub-model employs a deep Q-network algorithm, using the current transaction's fusion features and risk assessment results as the state, optional actions as actions, and the final outcome as the reward. The training process is stabilized through experience replay and a target network mechanism, with the target Q-value calculated as follows: ,in As a discount factor, The target network parameters are defined as follows: the network is updated by minimizing the mean squared error between the Q-value and the target value, and the risk judgment strategy is continuously optimized through trial and error learning. The outputs of the three sub-models are weighted and fused with weights of 0.5, 0.3, and 0.2 to obtain the final risk score of the transaction, which is divided into four risk levels: low, medium, high, and very high.
[0030] Furthermore, the visualized risk monitoring and early warning module utilizes web front-end technology to construct a multi-dimensional dashboard, providing risk management personnel with real-time transaction risk heatmaps, risk transaction trend analysis charts, high-risk account and device distribution maps, fraud pattern analysis and evolution charts, and a risk handling progress tracking interface. It supports data drill-down from global statistics to individual transaction details and multi-dimensional correlation queries. The module establishes a tiered early warning mechanism based on different risk levels, sending corresponding level early warning notifications through system pop-ups, SMS, email, telephone, and other channels. Extremely high-risk transactions directly trigger an emergency alarm mechanism, ensuring that risk events reach management personnel promptly.
[0031] Furthermore, the intelligent guidance and automatic handling module implements differentiated handling strategies based on the transaction risk level. Low-risk transactions are automatically released and risk information is recorded as training samples for subsequent model iterations; medium-risk transactions push intelligent handling suggestions to reviewers, automatically generate an investigation report containing transaction details, related account information, and historical risk records, and are accompanied by auxiliary tools such as correlation analysis and behavioral trajectory analysis to improve review efficiency; high-risk transactions automatically freeze account withdrawal and transfer functions and push them to senior reviewers for priority review; extremely high-risk transactions are directly intercepted and all funds in related accounts are frozen, and the relevant departments are notified and a fund recovery procedure is initiated in accordance with regulatory requirements. The module also provides a customizable configuration interface, allowing financial institutions to adjust risk thresholds and handling logic according to their own business characteristics.
[0032] Furthermore, the model self-optimization module employs an incremental learning mechanism, continuously updating model parameters based on newly generated risk data and response results. During training, the parameters of the underlying feature extraction network are fixed, and only the top-level classification and decision layers of the model are fine-tuned, eliminating the need for full retraining and reducing training time and computational resource consumption. The module automatically calculates the model's accuracy, precision, recall, and F1 score daily. When the F1 score falls below a preset threshold of 0.85, a full retraining process is automatically triggered, retraining and validating the model using historical data from the past three months to ensure the long-term effectiveness of risk identification capabilities.
[0033] The embodiments of the present invention also have the following technical effects: This invention offers more comprehensive multimodal data fusion, fully covering five major categories of core risk data. Through a deep fusion method based on a multi-head attention mechanism, it fully explores the complex correlations and complementary information between different modalities, resulting in more comprehensive risk feature extraction and effectively reducing false negative and false positive rates. Regarding risk identification capabilities, the hybrid AI model architecture combines supervised learning's ability to accurately identify known patterns, unsupervised learning's ability to detect unknown patterns, and reinforcement learning's ability to continuously optimize. This enables the simultaneous and effective identification of various complex fraud patterns, significantly improving accuracy. In terms of system design, a complete business closed-loop system has been established, enabling real-time risk monitoring, intelligent early warning, automatic handling, and continuous model evolution, addressing the limitation of existing technologies that only focus on risk identification. In terms of performance and user experience, through model lightweighting and inference optimization, millisecond-level risk detection and automatic handling are achieved, reducing risk response time from hours to seconds. A multi-dimensional visualization dashboard intuitively displays the overall risk status, facilitating rapid decision-making by risk management personnel. Simultaneously, the intelligent guidance and automatic handling mechanisms significantly reduce manual review, substantially lowering labor costs.
[0034] To achieve the above embodiments, such as Figure 2 As shown, this embodiment also provides a settlement business risk visualization detection device 10 based on intelligent risk identification, including: The multimodal preprocessing module 100 is used to collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. The collected data is then subjected to data cleaning, normalization, encoding and timestamp alignment processing in sequence to obtain preprocessed multimodal data. The cross-modal feature fusion module 200 is used to extract modality-specific features from the preprocessed multimodal data to obtain feature data for each modality, and to generate a unified fusion feature vector by performing cross-modal fusion processing based on an attention mechanism on the feature data of each modality. The multi-model risk discrimination module 300 is used to input the unified fusion feature vector into a supervised learning model, an unsupervised learning model, and a reinforcement learning model respectively for risk identification, and to perform weighted fusion of the identification results of the three models to obtain a risk score for the transaction, and to classify the risk level according to the risk score; The disposal iteration monitoring module 400 is used to execute differentiated disposal strategies according to the risk level, and to incrementally update the supervised learning model, unsupervised learning model and reinforcement learning model based on the disposal results and newly added risk data, while generating visualized risk monitoring information to display the risk status.
[0035] This invention provides a settlement business risk visualization detection device based on intelligent risk identification, which can realize millisecond-level real-time detection and hierarchical automatic handling of settlement business risks, improve the accuracy and comprehensiveness of risk identification, and enhance risk decision-making efficiency and model adaptability through multi-dimensional visualization and incremental learning mechanisms.
[0036] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 3 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.
[0037] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.
[0038] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0039] Furthermore, 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 technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
Claims
1. A settlement business risk visualization detection method based on intelligent risk identification, characterized in that, include: S1. Collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. Then, perform data cleaning, normalization, encoding and timestamp alignment processing on the collected data in sequence to obtain preprocessed multimodal data. S2, Modality-specific features are extracted from the preprocessed multimodal data to obtain feature data for each modality, and the feature data for each modality are processed by cross-modal fusion based on an attention mechanism to generate a unified fusion feature vector; S3, the unified fusion feature vector is input into the supervised learning model, the unsupervised learning model and the reinforcement learning model respectively for risk identification, and the identification results of the three models are weighted and fused to obtain the risk score of the transaction, and the risk level is divided according to the risk score; S4. Implement differentiated handling strategies according to the risk level, and incrementally update the supervised learning model, unsupervised learning model, and reinforcement learning model based on the handling results and newly added risk data. At the same time, generate visualized risk monitoring information to display the risk status.
2. The method of claim 1, wherein, S1 includes: Duplicate records are removed using the transaction serial number as the unique primary key, and outliers in numerical fields are identified using the 3σ criterion. First, the sample mean μ and standard deviation σ of the field are calculated. Data that satisfy |x-μ|>3σ are identified as outliers. Outliers are filled with the median of the same field to obtain cleaned multimodal data. The Min-Max normalization method is used for numerical features in the cleaned multi-modal data to map to the interval [0, 1], and the normalization calculation formula is Wherein , is the minimum and maximum value of the corresponding field, and the normalized multi-modal data is obtained; Differential encoding is adopted for the normalized multi-modal data, low-base number category features are converted into 0-1 vectors of corresponding dimensions through one-hot encoding, and high-base number category features are mapped into low-dimensional dense vectors through a learnable embedding matrix, and the embedding encoding formula is wherein is an index vector corresponding to the category, is an embedding weight matrix, and the encoded multi-modal data is obtained. Based on the transaction occurrence time, a sliding window with a fixed time step is used to perform time-series alignment on the encoded multimodal data. Missing sampling points within the window are filled using linear interpolation, with the interpolation formula being: ,in , For adjacent sampling times, , The values at the corresponding times are used to obtain the preprocessed multimodal data.
3. The method as described in claim 2, characterized in that, The method further includes: Low-cardinality category features of transaction channels and document types are encoded using one-hot encoding and converted into 0-1 vectors of the corresponding dimensions; Embedding encoding is used for high cardinality category features of account identifiers and device fingerprints. Discrete categories are mapped to low-dimensional dense vectors through a learnable embedding matrix.
4. The method as described in claim 1, characterized in that, S2 includes: For structured transaction data and device environment data, a three-layer deep neural network with residual connections is used for feature extraction. The three layers consist of an input layer, a 256-dimensional hidden layer, a 128-dimensional hidden layer, and a 64-dimensional hidden layer. Each hidden layer uses the ReLU activation function, and residual connections are added between adjacent hidden layers. The forward propagation core is calculated as follows: The residual output is The final output is a 64-dimensional dense feature. For sequential operation behavior data, a bidirectional long short-term memory network with temporal attention is used for feature extraction. An attention layer is added after the output of BiLSTM to automatically weight key temporal segments. After the bidirectional outputs are concatenated, attention weighting is applied to obtain 128-dimensional temporal features. For biometric data, a lightweight convolutional neural network based on depthwise separable convolution is used for feature extraction. The depthwise separable convolution is divided into two steps: depthwise convolution and pointwise convolution. The depthwise convolution independently convolves each input channel, and the pointwise convolution completes channel fusion through 1×1 convolution. Finally, 96-dimensional features are output through global average pooling. For text-based contextual data, a Chinese BERT-base pre-trained model is used to extract text features. The output corresponding to the delimiter is taken as the sentence-level semantic vector, and then the dimensionality is reduced to 224 dimensions through a fully connected layer. The dimensionality reduction calculation formula is as follows: .
5. The method as described in claim 4, characterized in that, The method further includes: Through the Gate of Oblivion Input gate Cell state update Output gate and hidden state Complete the gated computation of unidirectional LSTM; The outputs of the forward and backward LSTMs are concatenated to obtain a bidirectional output; An attention layer is added after the bidirectional output, and 128-dimensional temporal features are obtained by automatically weighting key time segments.
6. The method as described in claim 1, characterized in that, The S3 includes: The unified fused feature vector is input into the supervised learning model. The supervised learning model trains XGBoost and a deep neural network using labeled historical fraud data. XGBoost uses log loss with L1 and L2 regularization terms as the objective function to iteratively construct a regression tree to fit the residuals. The deep neural network uses cross-entropy loss as the optimization objective. The outputs of the two models are fused by weighted average to identify known fraud patterns and obtain the supervised learning risk identification result. The unified fused feature vector is input into an unsupervised learning model, which combines an isolated forest and autoencoder algorithms. The isolated forest constructs multiple isolated trees by randomly dividing the feature space and calculates the anomaly score based on the average path length of the samples. The autoencoder compresses the input into a low-dimensional latent vector through an encoder and then reconstructs the input through a decoder, using the reconstruction error as the anomaly score. The two algorithms detect global anomalies and local anomalies respectively. After the results are fused, unknown and emerging fraud patterns are identified, resulting in unsupervised learning risk identification results. The unified fused feature vector is input into a reinforcement learning model, which employs a deep Q-network algorithm. The model uses the fused features and risk assessment result of the current transaction as the state, the optional actions as the actions, and the final outcome as the reward. The training process is stabilized through experience replay and a target network mechanism. The target Q-value is calculated as follows: ,in As a discount factor, The target network parameters are defined as follows: the network is updated by minimizing the mean square error between the Q value and the target value; the risk judgment strategy is continuously optimized through trial and error learning to obtain the reinforcement learning risk identification result. The results of supervised learning risk identification, unsupervised learning risk identification, and reinforcement learning risk identification are weighted and fused with weights of 0.5, 0.3, and 0.2, respectively, to obtain the final risk score of the transaction, which is then divided into four risk levels: low, medium, high, and extremely high.
7. The method as described in claim 1, characterized in that, The S4 includes: When the risk level is low, the transaction is automatically approved and the risk information is recorded, which is then used as a training sample for subsequent model iterations. When the risk level is medium risk, intelligent handling suggestions are pushed to the reviewers, and an investigation report containing transaction details, related account information and historical risk records is automatically generated, along with auxiliary tools for correlation analysis and behavior trajectory analysis. When the risk level is high, the account's withdrawal and transfer functions will be automatically frozen, and the request will be sent to senior reviewers for priority review. When the risk level is extremely high, the transaction will be directly intercepted and all funds in the associated account will be frozen. The relevant departments will be notified and a fund recovery procedure will be initiated in accordance with regulatory requirements.
8. A settlement business risk visualization detection device based on intelligent risk identification, characterized in that, include: The multimodal preprocessing module is used to collect multimodal settlement business data, which includes transaction data, operation behavior data, equipment environment data, biometric data and context data. The module performs data cleaning, normalization, encoding and timestamp alignment on the collected data in sequence to obtain preprocessed multimodal data. The cross-modal feature fusion module is used to extract modality-specific features from the preprocessed multimodal data to obtain feature data for each modality, and to generate a unified fused feature vector by performing cross-modal fusion processing based on an attention mechanism on the feature data of each modality. The multi-model risk discrimination module is used to input the unified fused feature vector into the supervised learning model, the unsupervised learning model and the reinforcement learning model respectively for risk identification, and to perform weighted fusion of the identification results of the three models to obtain a risk score for the transaction, and to classify the risk level according to the risk score; The disposal iteration monitoring module is used to execute differentiated disposal strategies according to the risk level, and to incrementally update the supervised learning model, unsupervised learning model and reinforcement learning model based on the disposal results and newly added risk data, while generating visualized risk monitoring information to display the risk status.
9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the method as described in any one of claims 1-7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.
