High-robustness decision support system based on real-time reasoning

Abstract

The invention discloses a high-robustness decision support system based on real-time reasoning, and belongs to the technical field of artificial intelligence decision. The system comprises a data perception and acquisition module, a data processing and fusion module, a real-time reasoning engine, a robustness enhancement module and a decision presentation interface. And the real-time inference engine adopts a hybrid model fusing symbolic inference and sub-symbolic inference, and outputs a preliminary decision through dynamic weighting of a self-adaptive arbiter. And the robustness enhancement module performs multi-dimensional evaluation and correction on the preliminary decision through a triple mechanism of uncertainty quantization (such as Monte Carlo Dropout), resistance sample detection and situation consistency verification, and finally generates a reliable decision with high robustness. The method effectively solves the problems of poor reliability and low stability of an existing decision-making system caused by noise, abnormity and adversarial attacks in a complex real environment, has real-time performance, high robustness and interpretability, and is suitable for key fields of smart cities, industrial control and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and decision science, and particularly relates to a decision support system (DSS) combining mixed reasoning, real-time computing and robustness enhancement technology. BACKGROUND

[0002] With the development of Internet of Things (IoT), big data and artificial intelligence technology, decision support systems are widely used in smart cities, industrial automation, emergency management and other key fields. These systems usually rely on deep learning models for real-time perception and reasoning.

[0003] However, existing decision support systems have obvious limitations: Lack of robustness: Most systems are based on a single data-driven model (such as deep learning), which is very sensitive to noise, outliers and carefully constructed adversarial attacks in input data, which may lead to decision collapse and difficulty in stable operation in complex and unpredictable real environments; Single reasoning mechanism: Systems often only use "black box" sub-symbolic reasoning (such as neural networks), lack of explainability, and cannot utilize existing domain knowledge, rules and logic for collaborative reasoning, and perform poorly when encountering scenarios not covered by training data; Real-time and accuracy contradiction: In order to ensure real-time, light-weight models are often used, but this sacrifices accuracy; while complex models have high accuracy but large reasoning delay, making it difficult to meet the requirements of high real-time scenarios; Lack of self-evolution ability: Once the system is deployed, its performance is fixed and cannot continuously learn and optimize from subsequent interactions and feedback, making it difficult to cope with new challenges.

[0004] Therefore, there is an urgent need for a comprehensive solution that can achieve high robustness, real-time and explainable decision-making in complex and uncertain environments. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a high-robustness decision support system based on real-time reasoning. The system effectively improves the decision reliability and stability in noisy, disturbed and adversarial environments by integrating symbolic reasoning and sub-symbolic reasoning and introducing a multi-dimensional robustness enhancement mechanism.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: A high-robustness decision support system based on real-time reasoning, characterized by the following core modules: Data sensing and acquisition module: responsible for continuously and in real time acquiring multimodal raw data (such as temperature, pressure, images, and text streams) from heterogeneous data sources such as IoT sensors, video surveillance, databases, and third-party APIs located in various locations. Data Processing and Fusion Module: This module receives raw data and performs preprocessing operations such as cleaning, denoising, and normalization. Then, using the Dynamic Time Warping (DTW) algorithm or the Kalman filter algorithm, it fuses data from different sources and time series into a unified, spatiotemporally aligned feature vector, providing high-quality input for subsequent inference. Real-time Inference Engine: This is the core computing unit of the system. It receives the processed feature vectors and inputs them into its built-in hybrid inference model. This model integrates: Symbolic Reasoning Unit: Based on a predefined domain knowledge graph and rule base, this unit utilizes rule engines such as Drools for deterministic logical deductive reasoning. The unit outputs decision suggestions that conform to logical rules and possess high interpretability. Sub-symbolic reasoning unit: Based on a deep neural network optimized using model lightweighting techniques such as pruning and quantization, this unit performs perceptual and pattern recognition reasoning. It excels at processing unstructured data and outputs data-driven decision recommendations. Adaptive Arbitrator: This component uses an attention-based weighted fusion algorithm or ensemble learning algorithm to dynamically analyze the characteristics of the current input data (such as uncertainty and modal integrity), and assign appropriate weights to the results of symbolic reasoning and sub-symbolic reasoning accordingly, and finally fuses them to generate a preliminary decision suggestion and its confidence level.

[0007] Robustness Enhancement Module: This module is key to achieving high robustness in this invention. It performs three lines of defense to verify the initial decision output by the real-time inference engine: Uncertainty quantification unit: Using Monte Carlo Dropout (MC Dropout) or Bayesian approximation methods, the uncertainty score of the decision is calculated through multiple sampling inferences. If the score exceeds a threshold, the decision is deemed unreliable and correction is initiated. Adversarial Sample Detection Unit: This unit determines whether the input is a maliciously constructed adversarial sample by calculating the gradient sensitivity of the input features or utilizing a Generative Adversarial Network (GAN). Once an attack is detected, a defense strategy is immediately activated, such as performing local smoothing filtering to denoise the input data, or ignoring sub-symbolic inference results and relying entirely on the symbolic inference unit for security decisions. Contextual consistency verification unit: By constructing a dynamic contextual graph and using graph matching algorithms or temporal consistency detection algorithms, the current decision is compared with the historical decision trajectory and domain constraint rules to verify its logical rationality and consistency, and to avoid abnormal deviations.

[0008] After the above verification and correction, the module outputs a high-robustness final decision.

[0009] Decision presentation and execution interface: convert the final decision into an operational instruction or a visual interface element, and provide it to a human decision maker or directly issue it to an execution mechanism (such as a controller, actuator); System resilience maintenance unit (optional): contains a simulation and stress test environment that can generate various types of faults and attack scenarios on a regular or on-demand basis, conduct "fire drills" on the system, and continuously exercise and enhance the resilience of the entire system.

[0010] The beneficial effects of the present application are: Significantly improve robustness: through the three lines of defense of uncertainty quantification, adversarial sample detection and context consistency verification, the fault tolerance of the system to data noise, anomalies and malicious attacks is greatly enhanced, and the decision is more reliable and safe; Ensure real-time and accuracy: hybrid reasoning model combines the accuracy of symbolic reasoning and the efficiency of sub-symbolic reasoning, and the adaptive arbitrator realizes intelligent fusion; at the same time, model lightweight technology ensures real-time performance; Enhance explainability: the introduction of symbolic reasoning unit makes the decision-making process no longer a "black box", users can understand the logic rules behind the system, and increase the trust degree; Possess continuous evolution ability: through the feedback learning loop, the system can continuously optimize from actual feedback using online learning or incremental learning algorithm, adapt to new environment and new threats, and has long-term vitality. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0012] Figure 1 The overall architecture module diagram of the system provided in the present application; Figure 2 The workflow diagram of the hybrid reasoning model in the real-time reasoning engine provided in the present application; Figure 3 The multi-dimensional evaluation and correction flowchart of the robustness enhancement module provided in the present application. DETAILED DESCRIPTION

[0013] The present application will be further described in detail below in combination with the drawings and embodiments.

[0014] This system can be deployed as software on server clusters or edge computing devices.

[0015] Example: Taking intelligent traffic signal control in urban areas as an example.

[0016] Data perception and collection: Multimodal data is collected through intersection cameras (to acquire traffic flow images), inductive loops (to acquire vehicle numbers), radar (to acquire vehicle speeds), and traffic management databases (to acquire historical traffic patterns).

[0017] Data processing and fusion: Target detection is performed on the images to extract vehicle counts, and inductive loop and radar data are filtered. All data are then fused using a Kalman filter algorithm to generate a unified feature vector representing "current traffic flow, average vehicle speed, and queue length in each direction at the intersection."

[0018] 3. Real-time reasoning: Subsymbolic reasoning unit: Input a feature vector of a time series into a lightweight LSTM neural network, predict the traffic flow in each direction in the next minute, and output a suggestion to "extend the green light by X seconds"; Symbolic reasoning unit: Simultaneously, the rule engine checks rules in the knowledge base, such as "IF (emergency vehicles pass) THEN (force green light to open)". If this condition does not exist, it may output the suggestion of "maintain current phase"; Adaptive Arbitrator: Given a stable current traffic flow and high confidence in the neural network, the arbitrator assigns it a high weight and adopts its suggestion to "extend the green light" as a preliminary decision.

[0019] 4. Enhanced robustness: Uncertainty quantification: MC Dropout calculations show that due to sudden weather changes (heavy rain), the uncertainty of image recognition increases, and the confidence of the initial decision decreases; Adversarial detection: (Assuming) no adversarial attacks were detected; Contextual verification: Contextual consistency verification revealed that extending the green light time for this direction would conflict with the historical pattern of "this should be the peak time for pedestrian crossings" during the same period. The system determined that this decision might be risky.

[0020] Final Decision and Execution: The robustness enhancement module integrates the above information and modifies the initial decision, potentially generating a more conservative decision (e.g., "only slightly extend the green light by 5 seconds and prepare to switch phases"). This instruction is sent to the traffic light controller via the execution interface.

[0021] Feedback learning: The system records the actual traffic efficiency feedback at the intersection after each decision. If the feedback is unsatisfactory, the system will fine-tune the parameters of the LSTM model or adjust the weight strategy of the arbitrator through incremental learning algorithms.

[0022] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A high robustness decision support system based on real-time inference, characterized in that, The system comprises: a data perception and collection module for acquiring multi-modal raw data from a plurality of heterogeneous data sources in real time; a data processing and fusion module connected to the data perception and collection module, for cleaning, denoising and normalizing the multi-modal raw data, and fusing them into a unified spatio-temporal alignment feature vector based on a dynamic time warping algorithm or a Kalman filtering algorithm; a real-time inference engine connected to the data processing and fusion module, for receiving the spatio-temporal alignment feature vector and performing real-time calculation based on a built-in hybrid inference model, and outputting a preliminary decision suggestion and its confidence level; a robustness enhancement module connected to the real-time inference engine, for multi-dimensional evaluation and correction of the preliminary decision suggestion, including uncertainty quantification, adversarial sample detection and context consistency verification, to generate a high-robustness final decision; a decision presentation and execution interface connected to the robustness enhancement module, for providing the final decision to the user in a visualized form or directly converting it into a control instruction and sending it to an execution mechanism; a system resilience maintenance unit for periodically calling a simulation and stress test environment to conduct resilience training on the system.

2. The system of claim 1, wherein, The hybrid inference model of the real-time inference engine integrates a symbolic inference unit and a sub-symbolic inference unit; The symbolic inference unit performs logical deductive reasoning based on a pre-defined knowledge graph and rule base, using a Drools or Jess rule engine; The sub-symbolic inference unit performs perception and pattern recognition reasoning based on a lightweight deep neural network, which adopts pruning or quantization techniques to meet real-time requirements; The hybrid inference model is configured with an adaptive arbitrator for dynamically weighted fusion of the outputs of the symbolic inference unit and the sub-symbolic inference unit according to the characteristics of the input data, to form the preliminary decision suggestion; wherein the adaptive arbitrator adopts a weighted fusion algorithm based on an attention mechanism or an ensemble learning algorithm, and the weights are dynamically updated through online learning.

3. The system of claim 2, wherein, The robustness enhancement module comprises: an uncertainty quantification unit for calculating the variance of the prediction result as an uncertainty score through multiple forward propagations using a Monte Carlo Dropout method, or calculating the uncertainty score of the preliminary decision suggestion using a Bayesian approximation method in Bayesian deep learning, and triggering a decision correction process when the uncertainty score exceeds a first preset threshold; an adversarial sample detection unit for identifying whether the spatio-temporal alignment feature vector is an adversarial attack sample by calculating the gradient sensitivity of the input features or using a generative adversarial network, and if so, activating a defense strategy; the defense strategy includes but is not limited to denoising the input data using local smoothing filtering, or activating the symbolic inference unit for rule-based logical decision; a context consistency verification unit for comparing the preliminary decision suggestion with historical decision records and domain constraint rules, detecting whether there is a logical conflict or abnormal deviation by constructing a dynamic context graph and using a graph matching algorithm or a temporal consistency detection algorithm, and correcting accordingly.

4. The system of claim 3, wherein, The robustness enhancement module further comprises a feedback learning loop configured to: Collect execution effect feedback data of the final decision in an actual application scenario; Based on the execution effect feedback data, optimize and adjust the evaluation threshold of the hybrid inference model and / or the robustness enhancement module using an online learning algorithm or an incremental learning algorithm; wherein the online learning algorithm uses a variant of stochastic gradient descent, and the incremental learning algorithm is used to learn new scene patterns without forgetting old knowledge, to realize self-evolution and improvement of system robustness.

5. The system of claim 1, wherein, The system resilience maintenance unit contains a simulation and stress test environment that can inject various types of noise data, abnormal data streams, and simulated attack samples according to preset strategies or adaptive generation algorithms to normalize stress testing and resilience training of the real-time inference engine and the robustness enhancement module.

6. A decision support method based on the system of any one of claims 1-5, characterized by, The method comprises: Real-time acquisition and fusion of multi-modal raw data to form a spatio-temporal alignment feature vector; Input the feature vector into the hybrid inference model, perform symbolic reasoning and sub-symbolic reasoning in parallel, and output through an adaptive arbitrator to generate a preliminary decision suggestion; Perform multi-dimensional robustness evaluation on the preliminary decision suggestion, including uncertainty quantification, adversarial sample detection, and context consistency verification; Based on the evaluation results, revise the preliminary decision suggestion to generate a final decision with high robustness; Output and execute the final decision.

Images