Investment decision-making system and method based on federated learning and multi-modal risk control

Abstract

The invention provides an investment decision-making system and method based on federated learning and multi-modal risk control, and the system comprises a data collection layer which is used for integrating satellite remote sensing data, supply chain Internet of Things data and financial market data; the federal processing layer is used for constructing a distributed training framework based on a FedAvg algorithm and realizing cross-mechanism data fusion and global model parameter updating through homomorphic encryption and differential privacy; the intelligent decision-making layer adopts a TD3 reinforcement learning engine and adopts a dynamic gating mechanism to adjust fusion weights of texts, time sequences and image modalities in real time; and the execution monitoring layer is used for detecting transaction execution deviation in real time through a digital twinning technology, realizing abnormal response by adopting a CUSUM algorithm, and dynamically adjusting a federal learning rate through Bayesian optimization. The key technical defects of data islands, tail risk response delay, insufficient cross-market linkage analysis and the like of a traditional investment system are overcome through an innovatively designed federated learning architecture and a time-varying risk control model.

Description

Technical Field

[0001] The present invention relates to the field of intelligent financial technology, and in particular to an investment decision-making system and method based on federated learning and multimodal risk control. Background Art

[0002] Traditional investment decision-making systems usually rely on a series of quantitative and qualitative analysis methods to evaluate potential investment opportunities and make corresponding investment decisions. The goal of these systems is to reduce investment risks and improve investment returns through scientific methods. With the development of science and technology, traditional investment decision-making systems are gradually moving towards a more intelligent direction, such as using machine learning and artificial intelligence technologies to make more accurate market forecasts and risk assessments.

[0003] The traditional investment decision-making system has the following key flaws: Data silos and response delays: Existing systems mostly rely on a single financial data source (such as market time series data) and are unable to integrate non-traditional data such as satellite remote sensing and supply chain IoT, resulting in limited risk monitoring dimensions.

[0004] Insufficient tail risk capture capability: The existing system monitors less than 50 indicators, making it difficult to identify new risk sources in real time (such as supply chain disruptions in port satellite images and sudden changes in social media sentiment). For example, the traditional method has a delay of more than 24 hours in early warning of supply chain disruptions and an accuracy rate of less than 70%.

[0005] Lack of cross-market linkage analysis: Traditional models rely on the historical correlation coefficient method and are unable to quantify the transmission effects of regional adjustment events or supply chain shocks on financial markets, resulting in weak cross-market risk early warning capabilities.

[0006] Patent application document CN116167833A discloses an Internet financial risk control system and method based on federated learning, which uses a neural network model based on deep learning to mine the implicit correlation feature information between the weight matrices of various Internet financial risk assessment models, so as to fuse the various Internet financial risk assessment models to improve the assessment accuracy of the fused Internet financial risk assessment model. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of the present invention. Summary of the invention

[0007] In view of the defects in the prior art, the purpose of the present invention is to provide an investment decision-making system and method based on federated learning and multimodal risk control.

[0008] The investment decision-making system based on federated learning and multimodal risk control provided by the present invention includes: The data collection layer is used to integrate satellite remote sensing data, supply chain IoT data, and financial market data; The federated processing layer builds a distributed training framework based on the FedAvg algorithm, and implements cross-institutional data fusion and global model parameter updates through homomorphic encryption and differential privacy; The intelligent decision-making layer uses the TD3 reinforcement learning engine. The state space includes the gradient of the surface standard deviation, market skewness, and the change in the normalized vegetation index. The action space is the dynamic adjustment of asset weights and leverage control. The optimization target is a hybrid model of risk assessment and Black-Litterman constraints. A dynamic gating mechanism is used to adjust the fusion weights of text, time series, and image modalities in real time. The execution monitoring layer uses digital twin technology to detect transaction execution deviations in real time, adopts the CUSUM algorithm to achieve abnormal response, and dynamically adjusts the federated learning rate through Bayesian optimization.

[0009] Preferably, the data collection layer includes: Satellite data channel, collects multispectral images through satellites and calculates the normalized vegetation index. The formula is: , where NIR is the reflectivity of the near infrared band, and Red is the reflectivity of the red light band. Then, the shadow changes of the oil storage tanks are detected by combining wavelet transform to predict the price fluctuations of agricultural product futures. The supply chain data channel collects temperature, humidity and vibration data in real time through a globally deployed RFID sensor network, and uses a composite model of LSTM and GARCH to calculate anomaly scores. The formula is:

[0010] in, is the abnormality score, is the long short-term memory network function, Indicates from Temperature, humidity and vibration data from time period to t, is the wavelet transform function, is the frequency spectrum of the vibration data; Financial data channel, collects the implied volatility surface data of the option market and calculates its gradient As an indicator of market microstructure mutation, the formula is:

[0011] in, It is a gradient operator, which is used to describe the rate of change of a physical quantity in space; is some physical quantity on the surface; Representing physical quantities Partial derivative with respect to temperature T; Representing physical quantities Partial derivative with respect to curvature K; It is a proportional coefficient used to adjust the weight of the influence of curvature on the physical quantity.

[0012] Preferably, the global model parameter updating formula is:

[0013] in, is the updated global model parameter at time step t+1, is the dynamic aggregation weight based on the Sharpe ratio, is the number of data samples on client k, K is the total number of clients participating in the training process, are the local model parameters of client k at time step t.

[0014] Preferably, the mixed model expression of risk assessment and Black-Litterman constraint is:

[0015] in, is the weight vector; is the weight vector based on the Black-Litterman model; t is the time unit.

[0016] Preferably, the federation processing layer further comprises: The cross-modal feature alignment module maps the satellite data coordinates to the geographic coordinate system of the supply chain nodes and fuses multi-source features through a multi-head attention mechanism. The formula is:

[0017] in, is the fused feature vector, is the multi-head attention mechanism function, is the query vector for satellite data, is the key vector of supply chain data, is the value vector of financial data; The dynamic risk profile module updates three types of risk indicators in real time: Regional adjustment risk: based on satellite image event detection model, updated according to preset period; Supply chain risk: It is calculated by compounding LSTM anomaly score and GARCH volatility, and the update frequency is real-time; Market liquidity risk: Through the detection of sudden changes in the volatility surface gradient, when it exceeds the preset range, a market structure warning is triggered, which is updated once a second.

[0018] Preferably, the intelligent decision-making layer further includes: Time-varying risk budget model, dynamically adjusting asset weights , the calculation formula is:

[0019] in, represents the weight of asset i at time t; is the adjustment parameter at time t, ; is the risk-adjusted return of asset i at time t; Hierarchical optimization pipeline, including: daily adjustment of all asset categories, hourly rebalancing of core asset pools, and minute-by-minute adjustment of liquidity assets.

[0020] Preferably, the execution monitoring layer further comprises: Adaptive threshold control module to dynamically adjust risk parameters , the calculation formula is:

[0021] in, is the adjusted parameter at time t, the basic threshold =5%, liquidity sensitivity coefficient =0.15, is the volatility at time t, is the market liquidity at time t; is the exponential decay term of liquidity to the parameter; The closed-loop feedback mechanism generates a heat map of risk factor contribution through SHAP value analysis, triggers feature extraction strategy optimization at the data collection layer, and calibrates model parameters in real time.

[0022] Preferably, the data collection layer further comprises: The edge computing preprocessing module deploys a lightweight feature extraction algorithm at the data source to achieve data cleaning, outlier filtering, and feature normalization; The dynamic data lake building module adopts the DataMesh architecture to achieve cross-domain data autonomous governance and federated access.

[0023] Preferably, the data fusion process of the federated processing layer further includes: Timing synchronization mechanism to establish a unified timestamp; In the feature encoding module, 3D convolution is used to extract spatial features of satellite data, bidirectional LSTM is used to extract temporal features of supply chain data, and the SVI model is used to fit volatility surface parameters of financial data.

[0024] The investment decision-making method based on federated learning and multimodal risk control provided by the present invention comprises the following steps: Multimodal data collection and preprocessing steps: Generate feature vectors through satellite image analysis, RFID sensor data stream and volatility surface analysis; Federated model training and aggregation step: Periodically perform a cross-institutional model update; Dynamic risk control and asset allocation steps: Periodically perform multiple asset weight optimizations through the TD3 algorithm; Trade execution and feedback calibration: real-time monitoring of slippage and triggering of model recalibration; The TD3 algorithm is:

[0025] in, is the state vector at time t; is the gradient of the surface standard deviation; is the skewness from time t-1 to t; is the change in the Normalized Difference Vegetation Index.

[0026] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention solves the core pain points of traditional systems in data integration, real-time risk control and cross-market analysis through multimodal federated learning, dynamic risk modeling and reinforcement learning optimization, and achieves generational breakthroughs in risk warning timeliness, model robustness and return stability; (2) Integrate 12 types of heterogeneous data sources (such as Sentinel satellite NDVI data, RFID supply chain trajectories, and option volatility surfaces), expand monitoring indicators to more than 300 items, and cover multi-dimensional risks such as regional adjustments, supply chains, and market liquidity; through satellite image analysis (NDVI calculation error <0.5 pixel) and supply chain sensor networks (RFID nodes >5000), early warning of supply chain disruptions can be issued 36-72 hours in advance, with an accuracy rate of 92.3%; (3) The CUSUM algorithm is used to achieve 800ms-level anomaly detection, with transaction slippage of <1.5bps; the hierarchical optimization pipeline (strategic / tactical / high frequency) compresses decision delays to seconds (high-frequency position adjustment delay <10 seconds), which is two orders of magnitude higher than the performance of traditional systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings: Figure 1 This is a system architecture diagram of the present invention; Figure 2 For the calibration process; Figure 3 Design for critical processes; Figure 4 For data processing procedures; Figure 5 Flowchart for modeling supply chain-to-stock market network shocks; Figure 6 Designed an architecture for a three-stage optimized pipeline; Figure 7 It is the threshold action logic; Figure 8 Contribute to the revenue at each stage; Fig. 9 Contribution to data value; Fig.10 It is a diagram of the relationship between layers. DETAILED DESCRIPTION

[0028] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0029] Example like Figure 1 ,The present invention provides an adaptive investment decision-making system based on federated learning and multimodal dynamic risk control, including: a data acquisition layer, a federated processing layer, an intelligent decision-making layer and an execution monitoring layer; 1. Data collection layer; Core role: Heterogeneous data fusion, integrating 12 types of structured and unstructured data sources, including: Market time series data: high-frequency trading data (10ms-level collection), industry index, exchange rate fluctuations; Non-traditional data: satellite imagery (port container density heatmap), social media sentiment index (multilingual BERT sentiment analysis); Related network data: supply chain map (covering node relationships of 3,000+ listed companies).

[0030] Technical features: Edge computing preprocessing: deploy lightweight feature extraction modules at the data source to complete data cleaning, outlier filtering, and feature normalization; Dynamic data lake construction: Adopt DataMesh architecture to achieve autonomous governance and federated access of cross-domain data products.

[0031] 2. Federal processing layer; Core role: Privacy and security modeling, building a distributed training framework based on the FedAvg algorithm, achieving:

[0032] in, is the updated global model parameter at time step t+1, is the dynamic aggregation weight based on the Sharpe ratio, is the number of data samples on client k, K is the total number of clients participating in the training process, are the local model parameters of client k at time step t.

[0033] Technical features: Double encryption channel: Combining homomorphic encryption and differential privacy (ε=2.3), the parameter transmission error is controlled within ±0.5%; Heterogeneous device adaptation: supports hybrid training of edge nodes (FPGA) and cloud servers, increasing model convergence speed by 37%.

[0034] 3. Intelligent decision-making layer; Core function: Multimodal risk control, using a dynamic gating mechanism to adjust the fusion weights of text, time series, and image modalities in real time. The expression is:

[0035] in, represents the fusion weight at time step t, is the activation function, is the weight matrix, is the hidden state vector of the text modality, is the hidden state vector of the temporal modality, is the hidden state vector of the image modality, is the bias term.

[0036] Technical features: Reinforcement learning optimization engine: integrated PPO algorithm, performs portfolio rebalancing every 15 seconds; Explainable decision support: Apply SHAP value analysis to generate a heat map of risk factor contribution.

[0037] 4.Execution monitoring layer; Core function: Closed-loop feedback control, building a dynamic monitoring matrix based on digital twin technology.

[0038] Technical features: Streaming anomaly detection: Using the CUSUM algorithm to achieve 800ms-level instruction latency; Dynamic parameter tuning: The federated learning rate is adjusted in real time through Bayesian optimization, and the adaptation efficiency of non-independent and identically distributed data is improved by 42%.

[0039] like Fig.10 , which is the data connection between levels.

[0040] Data flow details: 1. Forward data flow; The data collection layer transmits the preprocessed multimodal features (dimensional compression rate 85%) to the federated processing layer through the TLS1.3 encrypted channel; The federated processing layer performs cross-institutional model aggregation every 5 minutes and outputs the global risk feature matrix to the intelligent decision-making layer.

[0041] 2. Feedback control flow; The execution monitoring layer transmits transaction execution deviation (Δ<0.5%) and market impact response data in real time, triggering fine-tuning of the decision-making layer model; The intelligent decision-making layer identifies key risk factors through SHAP value analysis and guides the data collection layer to optimize feature extraction strategies.

[0042] 3. Cross-layer coordination mechanism; Adopt the DataMesh federal governance model to establish a three-layer consensus protocol: Data product level: feature vector schema standardization; Model level: dynamic negotiation of federation aggregation weights; Decision-making level: risk thresholds are synchronized across domains.

[0043] The data collection layer includes: Satellite data channel: The data source is the Sentinel-2 satellite multispectral imager (10 bands), focusing on collecting the Normalized Difference Vegetation Index (NDVI); the spatial resolution is 10 meters / pixel, and the collection frequency is once every 5 days for global coverage; the preprocessing process formula is:

[0044] Among them, the reflectivity of the near infrared band is NIR=842nm band, and the reflectivity of the red band is Red=665nm band; Application scenario: Predicting agricultural product futures price fluctuations through vegetation cover changes (correlation coefficient r=0.79).

[0045] Supply chain data channel: The data source is the RFID temperature sensor network (>5,000 nodes deployed globally), the sampling frequency is to record temperature / humidity / vibration data every minute, and the anomaly detection model formula is:

[0046] in, is the abnormality score, is the long short-term memory network function, Indicates from Temperature, humidity and vibration data from time period to t, is the wavelet transform function, is the frequency spectrum of the vibration data.

[0047] Data value: 24-72 hours in advance warning of supply chain disruption risks (accuracy rate 92.3%); Financial data channel: The data source is the implied volatility surface of the options market (CBOE / VIX related derivatives), the collection dimensions are term structure (1M / 3M / 6M), volatility skewness (Skew>0.15 is considered market panic), and the key indicator formula is:

[0048] in, It is a gradient operator, which is used to describe the rate of change of a physical quantity in space; is a physical quantity on the surface (such as surface stress, surface tension, etc.); Representing physical quantities Partial derivative with respect to temperature T; Representing physical quantities Partial derivative with respect to curvature K; It is a proportional coefficient used to adjust the weight of the influence of curvature on the physical quantity.

[0049] This formula describes the gradient of a physical quantity on the surface (which may be stress or other related variables), combining the effects of temperature and a parameter (such as curvature). The gradient value reflects the sudden change of the market microstructure.

[0050] The federation processing layer includes: Cross-modal feature fusion: Spatial alignment: Map satellite data coordinates to the geographic coordinate system of supply chain nodes (WGS84→UTM); Time series synchronization: Establish a unified timestamp (UNIX timestamp ±50ms error); Feature encoding: Satellite data is 3D convolution to extract spatial features (Conv3Dkernel5×5×5), supply chain data is bidirectional LSTM to extract temporal features (hidden layer 128 dimensions), financial data is volatility surface parameterization (SVI model fitting), and the fusion algorithm is:

[0051] in, is the fused feature vector, is the multi-head attention mechanism function, is the query vector for satellite data, is the key vector of supply chain data, It is the value vector of financial data.

[0052] Risk Dimensions:

[0053] The intelligent decision-making layer includes: Time-varying optimization model: dynamic parameter adjustment, the formula is:

[0054] The mixed optimization objective is:

[0055] TD3 decision engine: state space design, the formula is:

[0056] in, It is a dynamic parameter, which changes with time; is the Sigmoid function, which maps the input to the (0,1) interval. is the value of a variable at time t; is the weight vector; is the weight vector based on the Black-Litterman model; is the state vector at time t; is the gradient of the surface standard deviation; is the skewness from time t-1 to t; is the change in the Normalized Difference Vegetation Index.

[0057] Action space: The asset weight adjustment range is ±15%, and the leverage ratio is controlled to be dynamically adjusted at 1.2-2.0 times.

[0058] The execution monitoring layer includes: Real-time trading system, the execution strategy is:

[0059] Closed-loop feedback mechanisms, such as Figure 2 ,Calibration process: market response → performance evaluation → error analysis → model parameter update.

[0060] Key indicators: strategy drift rate is <2% / week, order execution slippage is <1.5bps, and model recalibration cycle is every 24 hours.

[0061] This architecture achieves generational improvement of the existing system through three technological breakthroughs: 1. Federated multimodal fusion: Under the premise of data privacy protection, the spatial and temporal alignment of satellite physical data and financial derivatives data is achieved; 2. Dynamic risk transmission modeling: Establish a cross-market shock transmission model of NDVI → agricultural product supply → futures price → volatility surface; 3. Millisecond-level adaptive portfolio adjustment: The TD3 algorithm achieves asset allocation optimization more than 100 times per second, which is two orders of magnitude higher than traditional systems.

[0062] Core technology innovation: 1. Multimodal federated learning architecture; Design a secure fusion channel for heterogeneous data, support federated feature alignment of 6 types of data sources, and develop a cross-modal attention mechanism:

[0063] in, is the satellite feature vector, is the supply chain timing characteristic, From the eigenvector to feature vector The attention weight, , is the weight matrix, are other feature vectors.

[0064] 2. Time-varying risk budget model; Innovative dynamic weight adjustment algorithm:

[0065] in, represents the weight of asset i at time t; is the adjustment parameter at time t, ; is the risk-adjusted return of asset i at time t.

[0066] 3. Hybrid optimization decision engine; Constructing the BL-RiskParity-TD3 triple optimization framework:

[0067] in, is the covariance matrix at time t, is the covariance matrix based on historical data, is the covariance matrix based on news or latest information; is the weight vector; is a tuning parameter used to balance the weight between the two objectives; is the weight vector based on the Black-Litterman model; is the policy function at time t, is the state vector at time t, is the gradient of the surface standard deviation, is the skewness at time t.

[0068] Key process design such as Figure 3 .

[0069] Data dimension subsystem: Satellite image analysis, data processing flow is as follows Figure 4 .

[0070] The key technical parameters are:

[0071] Logistics sensor processing, data processing chain: sensor node ->> edge computing gateway: raw data stream (10Hz), edge computing gateway ->> central server: feature vector (256 dimensions), central server ->> risk model: LSTM anomaly score; Core algorithm:

[0072] in, is the health index at time t, and the weight Dynamically adjust according to the importance of node i, n is the number of anomaly scores, is the anomaly score of the ith node.

[0073] Volatility Surface Analysis: Surface modeling method: Using the stochastic volatility model (SVI parameterization):

[0074] Key derivative calculations:

[0075] in, is the volatility squared as a function of k and t; is a constant term; , , , is the coefficient; is the logarithmic price; is the center point; is the gradient vector of volatility; is the partial derivative of volatility with respect to time T; It is the partial derivative of volatility with respect to the logarithmic price K; when ||∇σ||2>0.15, a market structure warning is triggered.

[0076] Algorithm innovation subsystem: Inter-market shock transmission model: Technical implementation: 1. Establish the transmission equation from satellite data to commodity futures:

[0077] in, is the change in commodity futures prices at time t, the coefficient of NDVI index on futures price changes α=0.32 (p<0.01), the coefficient of previous volatility gradient on futures price changes β=0.18 (p<0.05), is the normalized difference vegetation index at time t, is the volatility gradient at time t-1; is a random error term.

[0078] 2. Build a network impact model from supply chain to stock market. The process is as follows Figure 5 .

[0079] 3. Develop inter-market impact intensity indicators:

[0080] in, is the inter-market shock intensity index, n is the total number of markets, is the sensitivity of the value of market i to market j The partial derivative of is the volatility of market j.

[0081] The architecture design of the three-stage optimization pipeline is as follows Figure 6 .

[0082] Technical parameters are:

[0083] Adaptive Threshold Control: Dynamic adjustment mechanism:

[0084] in, is the adjusted parameter at time t, the basic threshold =5%, liquidity sensitivity coefficient =0.15, is the volatility at time t, is the market liquidity at time t; is the exponential decay term of the liquidity parameter.

[0085] Threshold action logic is as follows Figure 7 .

[0086] The revenue contribution of each stage is as follows: Figure 8 .

[0087] Innovation points: 1. Full-dimensional data fusion: for the first time, the spatiotemporal alignment and joint modeling of satellite physical data, supply chain operation data and financial derivatives data are realized; 2. Quantification of cross-market shocks: Establish a market shock transmission model under the computable general equilibrium (CGE) framework to break through the limitations of the traditional correlation coefficient method; 3. Hierarchical optimization system: Through the three-stage structure of strategy-tactics-high frequency, we can achieve enhanced returns while maintaining investment discipline; 4. Intelligent threshold control: Develop an adaptive threshold algorithm based on the linkage between market liquidity and volatility to solve the problem of lagging risk control.

[0088]

[0089] Data value contribution Fig. 9 .

[0090] Satellite data processing flow: 1. Image acquisition: Acquire Sentinel-2 10-meter resolution NDVI data every day; 2. Feature extraction: Detect the shadow change of the oil tank through wavelet transform (accuracy ±3%); 3. Capacity forecast: Establish NDVI-agricultural product futures price transmission model (R²=0.89).

[0091] Supply Chain Monitoring System: Real-time collection of temperature, humidity and RFID data from 1,200 logistics nodes around the world; Build an LSTM anomaly detection model (AUC=0.93); Develop supply chain disruption early warning indicators (lead time > 36 hours).

[0092] Those skilled in the art know that, in addition to implementing the system, device and its various modules provided by the present invention in a purely computer-readable program code, it is entirely possible to implement the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and its various modules provided by the present invention can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures within the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing the method and structures within the hardware component.

[0093] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An investment decision-making system based on federated learning and multimodal risk control, characterized in that: include: The data collection layer is used to integrate satellite remote sensing data, supply chain IoT data, and financial market data; The federated processing layer builds a distributed training framework based on the FedAvg algorithm, and implements cross-institutional data fusion and global model parameter updates through homomorphic encryption and differential privacy; The intelligent decision-making layer uses the TD3 reinforcement learning engine. The state space includes the gradient of the surface standard deviation, market skewness, and the change in the normalized vegetation index. The action space is the dynamic adjustment of asset weights and leverage control. The optimization target is a hybrid model of risk assessment and Black-Litterman constraints. A dynamic gating mechanism is used to adjust the fusion weights of text, time series, and image modalities in real time. The execution monitoring layer uses digital twin technology to detect transaction execution deviations in real time, adopts the CUSUM algorithm to achieve abnormal response, and dynamically adjusts the federated learning rate through Bayesian optimization.

2. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The data collection layer includes: Satellite data channel, collects multispectral images through satellites and calculates the normalized vegetation index. The formula is: , where NIR is the reflectivity of the near infrared band, and Red is the reflectivity of the red light band. Then, the shadow changes of the oil storage tanks are detected by combining wavelet transform to predict the price fluctuations of agricultural product futures. The supply chain data channel collects temperature, humidity and vibration data in real time through a globally deployed RFID sensor network, and uses a composite model of LSTM and GARCH to calculate anomaly scores. The formula is: in, is the abnormality score, is the long short-term memory network function, Indicates from Temperature, humidity and vibration data from time period to t, is the wavelet transform function, is the frequency spectrum of the vibration data; Financial data channel, collects the implied volatility surface data of the option market and calculates its gradient As an indicator of market microstructure mutation, the formula is: in, It is a gradient operator, which is used to describe the rate of change of a physical quantity in space; is some physical quantity on the surface; Representing physical quantities Partial derivative with respect to temperature T; Representing physical quantities Partial derivative with respect to curvature K; It is a proportional coefficient used to adjust the weight of the influence of curvature on the physical quantity.

3. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The global model parameter update formula is: in, is the updated global model parameter at time step t+1, is the dynamic aggregation weight based on the Sharpe ratio, is the number of data samples on client k, K is the total number of clients participating in the training process, are the local model parameters of client k at time step t.

4. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The mixed model expression of risk assessment and Black-Litterman constraint is: in, is the weight vector; is the weight vector based on the Black-Litterman model; t is the time unit.

5. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The federation processing layer further comprises: The cross-modal feature alignment module maps the satellite data coordinates to the geographic coordinate system of the supply chain nodes and fuses multi-source features through a multi-head attention mechanism. The formula is: in, is the fused feature vector, is the multi-head attention mechanism function, is the query vector for satellite data, is the key vector of supply chain data, is the value vector of financial data; The dynamic risk profile module updates three types of risk indicators in real time: Regional adjustment risk: based on satellite image event detection model, updated according to preset period; Supply chain risk: It is calculated by compounding LSTM anomaly score and GARCH volatility, and the update frequency is real-time; Market liquidity risk: Through the detection of sudden changes in the volatility surface gradient, when it exceeds the preset range, a market structure warning is triggered, which is updated once a second.

6. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The intelligent decision-making layer further includes: Time-varying risk budget model, dynamically adjusting asset weights , the calculation formula is: in, represents the weight of asset i at time t; is the adjustment parameter at time t, ; is the risk-adjusted return of asset i at time t; Hierarchical optimization pipeline, including: daily adjustment of all asset categories, hourly rebalancing of core asset pools, and minute-by-minute adjustment of liquidity assets.

7. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The execution monitoring layer further comprises: Adaptive threshold control module to dynamically adjust risk parameters , the calculation formula is: in, is the adjusted parameter at time t, the basic threshold =5%, liquidity sensitivity coefficient =0.15, is the volatility at time t, is the market liquidity at time t; is the exponential decay term of liquidity to the parameter; The closed-loop feedback mechanism generates a heat map of risk factor contribution through SHAP value analysis, triggers feature extraction strategy optimization at the data collection layer, and calibrates model parameters in real time.

8. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The data collection layer further includes: The edge computing preprocessing module deploys a lightweight feature extraction algorithm at the data source to achieve data cleaning, outlier filtering, and feature normalization; The dynamic data lake building module adopts the DataMesh architecture to achieve cross-domain data autonomous governance and federated access.

9. The investment decision-making system based on federated learning and multimodal risk control according to claim 1, characterized in that: The data fusion process of the federated processing layer further includes: Timing synchronization mechanism to establish a unified timestamp; In the feature encoding module, 3D convolution is used to extract spatial features of satellite data, bidirectional LSTM is used to extract temporal features of supply chain data, and the SVI model is used to fit volatility surface parameters of financial data.

10. An investment decision method based on the investment decision system based on federated learning and multimodal risk control according to any one of claims 1 to 9, characterized in that: The following steps are involved: Multimodal data collection and preprocessing steps: Generate feature vectors through satellite image analysis, RFID sensor data stream and volatility surface analysis; Federated model training and aggregation step: Periodically perform a cross-institutional model update; Dynamic risk control and asset allocation steps: Periodically perform multiple asset weight optimizations through the TD3 algorithm; Trade execution and feedback calibration: real-time monitoring of slippage and triggering of model recalibration; The TD3 algorithm is: in, is the state vector at time t; is the gradient of the surface standard deviation; is the skewness from time t-1 to t; is the change in the Normalized Difference Vegetation Index.

Images