Multi-source data fusion method and system based on knowledge graph constraint

Abstract

The application discloses a multi-source data fusion method and system based on knowledge graph constraint, and relates to the technical field of data fusion. Multi-modal physiological signals are synchronously acquired and converted into structured vectors; signal quality scanning is performed on the multi-modal data vectors, dynamic confidence of each mode is calculated to generate a set of weighting coefficients; the multi-modal data vectors are input into a field knowledge graph model to output knowledge reasoning vectors; the multi-modal data vectors mapped by the set of weighting coefficients are arithmetically combined with the knowledge reasoning vectors to obtain an initial fusion state index; a state mapping package of a monitoring object is acquired, the initial fusion state index and the multi-modal data vectors are combined as a current state, a dynamic correction step is calculated by using a reinforcement learning calibration model; and finally, compensation operation is performed on the initial fusion state index to generate a final fusion state index. By fusing data driving and knowledge driving, the application improves the accuracy, robustness and individual adaptability of the fusion state index.

Description

Technical Field

[0001] This invention relates to the field of data fusion technology, specifically to a method and system for multi-source data fusion based on knowledge graph constraints. Background Technology

[0002] Multi-source data fusion technology is an important research direction in the field of intelligent information processing. It aims to obtain a more comprehensive and accurate estimate of the state of a target system by integrating heterogeneous observation data from different sensors. In recent years, with the popularization of sensor technology and the improvement of computing power, multi-source data fusion methods have been widely applied to complex system state assessment scenarios such as industrial process monitoring, intelligent transportation, and environmental perception. The accuracy and reliability of the fusion results directly affect the effectiveness of upper-level decision-making.

[0003] Current mainstream data fusion methods mostly employ Kalman filtering and its variants based on statistical estimation, DS inference based on evidence theory, or end-to-end learning models based on neural networks. When processing multi-source data, these methods typically weight and combine data based on the statistical characteristics of each data source (such as noise variance) or preset fusion weights, which can suppress random noise and improve the accuracy of state estimation to some extent. However, these data-driven fusion methods have a fundamental limitation: the fusion result is entirely determined by the numerical characteristics of the observed data, lacking awareness and adherence to the underlying domain logic. When the quality of a data source degrades due to environmental interference, sensor failure, or other reasons, traditional methods, although able to reduce its contribution by adjusting weights, still cannot determine whether the fusion result violates the basic physical laws of system operation or the causal logic of the domain.

[0004] Traditional multi-source data fusion methods treat data as the sole source of information, focusing only on statistical correlations at the data level, while failing to embed domain prior knowledge into the fusion process in a computable form. In real-world scenarios where data quality dynamically changes, fusion results relying solely on statistical characteristics are prone to producing anomalous outputs that deviate from domain logic when some data sources are disturbed. This logical inaccuracy is difficult to detect through conventional signal quality assessments, but it can lead to misjudgments in decision-making systems based on the fusion results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-source data fusion method and system based on knowledge graph constraints.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] In a first aspect, this invention discloses a multi-source data fusion method based on knowledge graph constraints, comprising the following steps:

[0008] Synchronously acquire multimodal monitoring signal sequences and convert them into structured multimodal data vectors, wherein the multimodal data vectors include at least feature components reflecting different dimensional states;

[0009] The multimodal data vector is subjected to real-time signal quality scanning, the dynamic confidence level of each modal signal corresponding to the acquired multimodal data vector is calculated, and a corresponding weighted coefficient set is generated according to the distribution ratio of the dynamic confidence level.

[0010] The multimodal data vector is input into a preset domain knowledge graph model. Based on the preset causal constraint logic in the knowledge graph model, the multimodal data vector is subjected to topological association analysis and graph convolution operation to output a knowledge reasoning vector representing the suppression state logic constraint.

[0011] The multimodal data vector is weighted and mapped using the weighted coefficient set, and then arithmetically synthesized with the knowledge reasoning vector to obtain the initial fusion state index.

[0012] Obtain the state mapping package of the current monitored object, use the initial fusion state index and the multimodal data vector as the current environmental state information, and use a preset reinforcement learning calibration model to calculate the dynamic correction step size for the initial fusion state index.

[0013] The initial fusion state index is compensated based on the dynamic correction step size to generate the final fusion state index.

[0014] Secondly, this invention discloses a multi-source data fusion system based on knowledge graph constraints, using the aforementioned multi-source data fusion method based on knowledge graph constraints, including:

[0015] The data acquisition module is used to synchronously acquire multimodal monitoring signal sequences and convert them into structured multimodal data vectors, wherein the multimodal data vectors include at least feature components reflecting different dimensional states;

[0016] The quality assessment module is used to perform real-time signal quality scanning on the multimodal data vector, calculate the dynamic confidence of each modal signal corresponding to the acquired multimodal data vector, and generate a corresponding weighted coefficient set according to the distribution ratio of the dynamic confidence.

[0017] The knowledge reasoning module is used to input the multimodal data vector into a preset domain knowledge graph model, perform topological association analysis and graph convolution operation on the multimodal data vector based on the preset causal constraint logic in the domain knowledge graph model, and output a knowledge reasoning vector representing the suppression state logic constraint.

[0018] The fusion computing module is used to perform arithmetic synthesis of the multimodal data vector with the knowledge reasoning vector after weighting the multimodal data vector through the weighted coefficient set, so as to obtain the initial fusion state index.

[0019] The reinforcement learning calibration module is used to obtain the state mapping package of the current monitored object, take the initial fusion state index and the multimodal data vector as the current environmental state information, and calculate the dynamic correction step size for the initial fusion state index using a preset reinforcement learning calibration model.

[0020] The output module is used to perform compensation calculations on the initial fusion state index according to the dynamic correction step size to generate the final fusion state index.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. This invention performs real-time quality scanning on multi-source data, calculates the dynamic confidence level of each data source, and generates a weighted coefficient set accordingly. This allows subsequent fusion calculations to adaptively adjust the contribution weights of different data sources. When a data source is disturbed or its quality degrades, its impact on the fusion process will be dynamically weakened, thereby effectively improving the robustness and anti-interference capability of multi-source data fusion.

[0023] 2. This invention utilizes multimodal data vectors while introducing a pre-defined domain knowledge graph model. Through topological association analysis and graph convolution operations, it outputs knowledge inference vectors that contain domain causal logic constraints. The data-driven weighted mapping results are arithmetically synthesized with the knowledge-driven inference vectors, resulting in a fusion result that incorporates both the dynamic changes of real-time data and conforms to the inherent logic of the domain. This achieves an effective balance between data fluctuations and prior knowledge, significantly enhancing the interpretability and logical consistency of the fusion result.

[0024] 3. This invention acquires a state mapping package containing the static features of the target object, and constructs the current environmental state by combining the initial fusion result with multimodal data vectors. It then uses a reinforcement learning model to calculate a targeted dynamic correction step size. This mechanism enables the fusion calculation to adapt to the individual differences of different objects and to adaptively compensate the fusion result based on the temporal evolution trend of the data, achieving personalized and precise calibration that surpasses general population models. Attached Figure Description

[0025] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0026] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0027] Figure 2 This is a flowchart illustrating the working principle of the present invention;

[0028] Figure 3 This is a system module architecture diagram of the present invention. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] Application Overview:

[0031] In the field of multi-source data fusion technology, existing methods mostly employ Kalman filtering based on statistical estimation, DS inference based on evidence theory, or end-to-end learning models based on neural networks. When processing multi-source observation data, these methods typically perform weighted combinations based on the statistical characteristics or preset weights of each data source. However, these data-driven fusion methods face inherent limitations in complex application scenarios: when different modal observation data have complementary representation capabilities of the target state—some modalities are sensitive to state changes but susceptible to environmental interference, while others are stable but less sensitive—and key causal input variables provide logical constraints on state evolution, existing methods struggle to effectively utilize this complementary characteristic between modalities. Especially under dynamically changing environmental conditions, sensitive modalities are easily contaminated by noise, leading to instantaneous jumps or misjudgments in the fusion results, while system response lag often causes a time misalignment between the fusion results and the true state, making it difficult to meet the requirements of high-reliability state assessment for complex systems.

[0032] To address the aforementioned issues, research has revealed the complementarity of different modal observation data in characterizing the state of the target system: high-sensitivity modes respond rapidly to state changes but are susceptible to environmental disturbances, while low-sensitivity modes offer good stability but have slow responses. Key causal input variables, on the other hand, provide logical constraints on state evolution. Therefore, a hybrid computational model combining multimodal data fusion and knowledge graph logical constraints is proposed. Anti-interference fusion is achieved through dynamic confidence assessment, and the causal constraint logic of the domain knowledge graph is used to verify the logical consistency of the fusion results, ensuring that the fusion results still conform to the system's inherent causal laws even when some observation data is disturbed. Further experimental verification involves incorporating the static attribute characteristics of the target system into the design of the reinforcement learning reward function, forming a dynamic compensation mechanism for system response lags, and achieving adaptive time-delay calibration of the fusion results.

[0033] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] Example 1:

[0035] After introducing the basic concept of the present invention, the specific implementation of the present invention will be described in detail below, taking anesthesia depth monitoring as a typical application scenario.

[0036] like Figure 1 As shown, the multi-source data fusion method based on knowledge graph constraints includes the following steps:

[0037] Simultaneously acquire multimodal monitoring signal sequences and convert them into structured multimodal data vectors. The multimodal data vectors include at least feature components reflecting different dimensional states. Specifically, EEG signals, ECG signals, blood pressure waveforms, and target-controlled infusion (TCI) data are acquired in real-time through the multi-sensor interface of the anesthesia monitor, and then preprocessed and converted into structured multimodal data vectors. The multimodal data vectors include at least:

[0038] (1) EEG characteristics: Based on the power spectrum characteristics (δ band: 0.5-4Hz, α band: 8-13Hz, β band: 13-30Hz power ratio), biphasic coherence index and burst inhibition ratio extracted from multi-lead EEG signals, the degree of consciousness inhibition in the cerebral cortex is reflected.

[0039] (2) Circulatory system characteristics: including heart rate variability (HRV) frequency domain index (LF / HF ratio), systolic blood pressure to diastolic blood pressure ratio and pulse wave conduction time, reflecting the autonomic nervous system's response to anesthetic drugs;

[0040] (3) Dosage concentration characteristics: Real-time plasma concentration prediction and dosing rate from the target-controlled infusion pump provide a baseline input for pharmacological causal analysis.

[0041] Real-time signal quality scanning is performed on the multimodal data vectors. The dynamic confidence level of each modal signal corresponding to the acquired multimodal data vectors is calculated, and a corresponding weighted coefficient set is generated based on the distribution ratio of the dynamic confidence level. The calculation process of the dynamic confidence level specifically includes:

[0042] The instantaneous envelope features of each modal signal in the multimodal data vector are extracted. The standard deviation to mean ratio of each modal signal in the time domain is calculated through a preset sliding window (window length 2 seconds, step size 0.5 seconds) to obtain the dispersion index.

[0043] The power spectrum distribution of each modal signal is obtained using a preset Fast Fourier Transform (FFT). Based on preset monitoring characteristic frequency bands, the signal power within the corresponding frequency band is extracted as the effective signal power. The ratio of the effective signal power to the noise power across the entire frequency band is calculated to obtain the real-time signal-to-noise ratio (SNR). The dispersion index and the real-time SNR are then weighted and summed to output the dynamic confidence level of each modal signal.

[0044] In a preferred embodiment of the present invention, the dynamic confidence level of each modal signal The following formula is used to dynamically calculate:

[0045] ;

[0046] in, represents the dynamic confidence level of the m-th modal signal, with a value range of [0,1].

[0047] α and β are preset perception weights that satisfy α+β=1. For example, the values ​​can be α=0.4 and β=0.6.

[0048] The normalized value of the dispersion index is calculated using the min-max normalization method: , where D is the real-time dispersion index (standard deviation / mean). and These are the minimum and maximum values ​​of the dispersion of this mode in historical data (which can be pre-defined statistically).

[0049] This is the normalized value of the real-time signal-to-noise ratio, and the normalization method is the same as... , The value range is mapped to [0,1].

[0050] The overall reliability of the current channel signal is quantified by weighted summation of the normalized dispersion index and the normalized signal-to-noise ratio for each mode. If the dynamic confidence level of any mode signal is lower than a preset threshold... (The Youden exponent is determined by analyzing the data distribution of historical high-quality and contaminated signals, using the point of maximum value on the receiver operating characteristic curve (ROC curve). The value typically ranges from 0.2 to 0.4, with the preferred value being...) If the modal data vector is set to zero in the weighted coefficient set, the signal channel is shielded from severe contamination, which fundamentally solves the problem of instability of the index caused by noise such as electrosurgical interference and electromyographic artifacts in the clinical surgical environment.

[0051] Multimodal data vectors are input into a pre-defined domain knowledge graph model. The construction and reasoning process of the domain knowledge graph model includes:

[0052] State variable entities refer to core variables affecting the system state, such as anesthetic drug concentration; observation feature entities refer to system features that can be directly or indirectly observed, such as EEG power spectrum; target state entities refer to the system state to be evaluated, such as the degree of consciousness inhibition. A pre-defined knowledge structure topology graph includes drug entities (anesthetic drug nodes such as propofol and remifentanil), physiological feature entities (EEG feature nodes, HRV nodes, etc.), and anesthetic state entities (state nodes such as awake, sedated, surgical anesthesia, and deep inhibition), and defines directed edge permissions between entities according to pre-defined pharmacological causal rules (e.g., increased propofol effect-room concentration → deeper consciousness inhibition).

[0053] The system acquires the current drug concentration characteristics and calculates the predicted effect-compartment concentration at the current moment based on a pre-defined three-compartment system dynamics model (Schnider model), which serves as the initial activation node for the knowledge structure topology graph. A three-layer graph convolutional network (GCN) is then used to update the initial activation node by treating multimodal data vectors as external observations. The influence weights of each entity node within the topology space are calculated to generate a knowledge inference vector.

[0054] The three-layer graph convolutional network was trained using the Adam optimizer with an initial learning rate of 0.001, 200 training epochs, and a batch size of 64. Training was stopped early when the validation set loss no longer decreased after 10 consecutive epochs. A cross-entropy loss function was used, combined with a graph structure regularization term to maintain topological consistency between nodes. The training dataset contained multimodal data from 10,000 anesthesia surgeries and corresponding anesthesia statuses labeled by anesthesiologists.

[0055] In a preferred embodiment of the present invention, the feature update formula for each layer of the graph convolutional network is as follows:

[0056] ;

[0057] Where N=312, it means that the knowledge graph contains a total of 312 nodes, including drug entities, physiological characteristic entities, and anesthesia state entities;

[0058] Let l be the node feature matrix of the l-th layer. As the feature dimension of this layer, this embodiment adopts a three-layer structure with dimensions of 64, 32, and 16 respectively;

[0059] As an adjacency matrix, directed edges between entities are defined according to pre-defined pharmacological causal rules (e.g., "increased propofol concentration leads to deeper inhibition of consciousness"). If a causal relationship exists, then... Otherwise, it is 0;

[0060] D is a degree matrix, which is a diagonal matrix. , where represents the number of edges connecting the i-th node;

[0061] It is the negative 1 / 2 power of the degree matrix, used to perform symmetric normalization on the adjacency matrix to balance the influence of nodes with different degrees;

[0062] Let be the learnable weight matrix of the l-th layer, which is optimized through training;

[0063] This is the ReLU activation function.

[0064] The meaning of this formula is: the node features of the (l+1)th layer. It is through the features of the l-th layer Perform neighbor aggregation (achieved by a normalized adjacency matrix) and linear transformation (multiply by) The features are obtained after nonlinear activation. Through multi-layer stacking, the features of each node can be fused with the pharmacological causal information of multi-hop neighbors, and the final generated node features are used to output the knowledge inference vector.

[0065] Domain knowledge graph models encode pharmacokinetic causal relationships into a graph structure, introducing pharmacological constraints lacking in pure data-driven models into an exponential calculation framework in a computable form, thus making up for the fundamental deficiency of existing methods in lacking causal logic at the mechanism level.

[0066] The multimodal data vectors are weighted and mapped using a weighted coefficient set, and then arithmetically synthesized with the knowledge reasoning vectors to obtain the initial fusion state index. Specific steps include:

[0067] The multimodal data vectors processed by the weighted coefficient set are mapped to a feature space of a preset dimension through a two-layer fully connected nonlinear transformation, and the data-driven observations are output. The knowledge inference vectors are input into a preset logistic regression model and the knowledge-driven reference values ​​are output. The Euclidean distance between the data-driven observations and the knowledge-driven reference values ​​is calculated.

[0068] Determine whether the Euclidean distance exceeds the preset logical consistency boundary threshold: if it does not exceed, then the data-driven observation value and the knowledge-driven reference value are fused with equal weight to obtain an intermediate fused value; if it exceeds, then the dynamic adjustment coefficient is calculated based on the ratio of the difference between the Euclidean distance and the boundary threshold, and the data-driven observation value is corrected using the dynamic adjustment coefficient, and then fused with the knowledge-driven reference value to obtain an intermediate fused value; the intermediate fused value is compressed to the [0,100] interval after linear mapping, and used as the initial anesthesia depth index.

[0069] In another preferred embodiment of the present invention, the formula for calculating the dynamic adjustment coefficient when the deviation between the data-driven observation and the knowledge-driven reference value exceeds the logical consistency boundary threshold is as follows:

[0070] ;

[0071] in, These are dynamically adjusted coefficients used to correct data-driven observations.

[0072] As a preset logical consistency boundary threshold, the normalized feature space in this embodiment ranges from 0.15 to 0.35. The specific value is determined through clinical pre-experiment statistics to balance the contributions of data-driven and knowledge-driven approaches.

[0073] The Euclidean distance between data-driven observations and knowledge-driven references is given by... Then take Avoid division by zero errors.

[0074] This mechanism ensures that when data signals are contaminated or model outputs deviate from pharmacological logic, knowledge graph constraints can proactively intervene to correct them, preventing abnormal data from causing exponential jumps, thereby solving the problem of existing methods lacking medical logic constraints.

[0075] Obtain the state mapping package of the currently monitored object. The state mapping package should at least contain the initial policy network weights, learning rate scaling factor, and reward function weight adjustment coefficients for this patient cluster. The specific process includes:

[0076] The static attribute features of the monitored objects (static attribute features refer to individual inherent attributes that do not change rapidly or change slowly over time, such as age, weight, ASA classification, and other pre-existing disease states in anesthesia monitoring; and equipment model, manufacturing date, and installation location in industrial equipment monitoring) are obtained and transformed into structured individual difference vectors. The individual difference vectors are then input into a preset feature clustering model to locate the cluster center to which the monitored object belongs within a preset sample sensitivity space. Based on the cluster center, the corresponding preset parameter configuration is obtained to form a state mapping package for the current monitored object, which is used to initialize the policy network of the reinforcement learning calibration model.

[0077] The feature clustering model employs the K-Means algorithm, pre-trained on 100,000 anesthesia samples. Through fine-grained clustering of individual characteristics of monitored subjects, it provides personalized initial policy parameters for different monitored subject groups (especially the elderly, children, and critically ill patients), addressing the problem of weak adaptability to individual differences in existing methods. The reinforcement learning calibration model uses the PPO algorithm, with both the policy network and value network being two-layer fully connected networks. The hidden layer has 128 neurons, and the activation function is ReLU. The optimizer is Adam, with a learning rate of 3e-4, a discount factor γ=0.99, GAE parameter λ=0.95, and clip parameter ε=0.2, updated 4 times per round.

[0078] The dynamic correction step size for the initial anesthesia depth index is calculated using a pre-defined reinforcement learning calibration model. The dynamic correction calculation of the reinforcement learning calibration model includes:

[0079] A pre-defined spatiotemporal attention mechanism (TSA) is used to mine features of the current environmental state information, extracting the evolution trend features of the multimodal monitoring signal sequence within a pre-defined historical period (30-second sliding window); a reinforcement learning reward function based on time delay compensation is constructed. The reinforcement learning reward function includes a causal response lag evaluation term and a state stationarity maintenance term. The causal response lag evaluation term is used to measure the time matching degree between the current fusion state and the historical control input. The system response time constant can be pre-determined according to the dynamic characteristics of the specific system; for example, it is the pharmacokinetic time constant in anesthesia monitoring and the system inertia time constant in industrial control.

[0080] In this embodiment, the reward function includes a pharmacological response lag evaluation term and a physiological homeostasis maintenance term:

[0081] (1) The pharmacological response lag assessment item quantifies the causal matching degree between the initial anesthesia depth index at the current moment and the historical drug administration pulse based on the preset pharmacokinetic time constant.

[0082] (2) The physiological homeostasis maintenance term penalizes the jitter rate of the initial anesthesia depth index output, suppressing non-physiological fluctuations.

[0083] In a preferred embodiment of the present invention, the reinforcement learning reward function The calculation formula is:

[0084] ;

[0085] in:

[0086] ;

[0087] ;

[0088] in, This is a lag assessment item for pharmacological response, reflecting the degree of matching between the current anesthesia depth index and historical drug administration effects;

[0089] The initial anesthesia depth index at time t;

[0090] The Sigmoid model is used for the preset concentration-effect mapping function: Where k is the steepness coefficient and EC50 is the half-maximal effective concentration, the specific value of which is preset according to the drug type (e.g., propofol k=0.8, EC50=3.2μg / mL).

[0091] for Predicted concentration values ​​in the effect room at any given time;

[0092] The preset pharmacokinetic time constant is dynamically calculated based on the system kinetic model (e.g., the reciprocal of the effect compartment equilibrium rate constant ke0 using the Schnider model), with an example range of 0.5 to 2 minutes.

[0093] This is a term for maintaining physiological homeostasis, used to penalize rapid fluctuations in the exponent;

[0094] This is the steady-state weighting coefficient, used to adjust the intensity of the penalty for exponential fluctuations; preferably, it is set to 0.2.

[0095] Var(·) represents the variance of the sliding window. Let T be the variance of the initial anesthesia depth index within the time window [tT, t], where the time window length T is 30 seconds.

[0096] The reinforcement learning calibration model uses the PPO (proximal policy optimization) algorithm. Based on the feedback guidance of the above reward function, it searches for the optimal dynamic correction step size within the preset action space (correction step size range [-3,+3]). The reinforcement learning model is pre-trained on an offline dataset, and only performs policy inference in the online stage without performing parameter updates.

[0097] Based on the dynamic correction step size, the initial anesthesia depth index Perform compensation calculations to generate the final anesthesia depth index. After the compensation calculation is completed, the data smoothing and safety filtering steps are also included: it is determined whether the value after the compensation calculation is within the preset valid state range [0,100]. If it exceeds the range, the value is forcibly corrected to the corresponding safety boundary value. The valid state range is preset according to the target state value range of the specific application field. For example, it is 0~100 in anesthesia monitoring and 0~1 in industrial equipment health assessment.

[0098] Mean filtering is performed on the continuous anesthesia depth index sequence with a preset step size to eliminate non-uniform jumps caused by instantaneous external disturbances, generating a smoothed final anesthesia depth index. .

[0099] In this invention, a logical mapping operator is used to infer the expected value of one type of feature based on one type of observed feature. Its construction methods include statistical regression based on historical data, analytical mapping based on physical models, and associative reasoning based on knowledge graphs. A physiological logical mapping operator is used to reverse-calculate the expected conscious inhibition feature value corresponding to the current circulatory state based on the evolution trend of circulatory system features. The construction of the physiological logical mapping operator is based on the prior knowledge in physiology that "there is a coupling relationship between the state of the circulatory system and the degree of inhibition of the central nervous system," which can be implemented in the following ways:

[0100] Method 1: Mapping model based on statistical regression:

[0101] A large amount of clinical anesthesia case data was collected in advance, including synchronously recorded circulatory system characteristics (such as heart rate variability, LF / HF ratio, and systolic blood pressure change rate) and corresponding EEG characteristics of consciousness inhibition (such as delta power ratio and bispectral index, BIS). An empirical mapping function from circulatory system characteristics to consciousness inhibition characteristics was established through multivariate nonlinear regression analysis.

[0102] ;

[0103] in:

[0104] The desired conscious inhibition characteristic (e.g., the desired δ power ratio);

[0105] This is a frequency domain index of heart rate variability (LF / HF ratio).

[0106] The rate of change in systolic blood pressure (current blood pressure / baseline blood pressure);

[0107] It is a non-linear mapping function, which can be obtained by fitting using support vector regression (SVR) or a shallow neural network.

[0108] In a preferred embodiment of the present invention, a three-layer BP neural network (input layer dimension 3, hidden layer number of neurons 16, output layer dimension 1) is used as the mapping operator and trained to convergence (root mean square error < 0.05) on 5,000 labeled cases.

[0109] Method 2: Reverse reasoning based on the pharmacology-physiology coupling model:

[0110] By combining pharmacokinetic / pharmacodynamic (PK / PD) models with circulatory system response models, a reasoning mechanism is constructed to infer the degree of central nervous system inhibition from circulatory effects:

[0111] ;

[0112] ;

[0113] ;

[0114] in, For effect room concentration;

[0115] For system dynamics models (such as the Schnider model);

[0116] For pharmacodynamic models, describing the relationship between concentration and circulatory system response (such as heart rate changes), the following methods can be used: Model fitting;

[0117] The inverse function of has a unique solution within the monotonic interval.

[0118] Method 3: Knowledge Graph-Based Association Reasoning

[0119] By utilizing the association edges between "circulatory system feature nodes" and "EEG feature nodes" in a pre-defined domain knowledge graph, the conditional probability distribution of each EEG feature node under given circulatory feature conditions is calculated through a graph attention mechanism, and the expected value is taken as the expected consciousness inhibition feature value.

[0120] ;

[0121] in, This is the set of EEG feature nodes connected to the currently looping feature node;

[0122] Attention weights are calculated based on the weights of edges and node attributes in the graph.

[0123] This represents the typical value of the j-th EEG feature node.

[0124] Regardless of the method used, the physiological logic mapping operator needs to be pre-trained or calibrated on a clinical dataset. In a preferred embodiment of the present invention, method one (neural network regression model) is used, which is trained on a database containing a large amount of clinical data, and the model output is offset and corrected according to the patient's static characteristics (age, weight, etc.) during deployment to improve individual adaptability.

[0125] During the surgery, the system extracts the evolution trend of circulatory system features in real time (such as the rate of change of the LF / HF ratio over the past 30 seconds) and inputs it into a pre-trained physiological logic mapping operator. Output the EEG consciousness inhibition characteristic value "expected" for the current loop state. Subsequently, this expected value was compared with the actual extracted EEG features. Compare and calculate the deviation measure. The formula is as follows:

[0126] ;

[0127] when Exceeding the preset causal conflict threshold (Based on clinical trial statistics, the typical value range for the normalized scale of EEG characteristics is 0.1-0.2. In this embodiment, we take...) When a logical conflict exists between multimodal signals, indicating that a certain modality may be contaminated, a data self-repair process is triggered. In this case, the state assessment derived from the knowledge graph based on the current drug concentration and physiological logic is given priority, and the results driven by potentially contaminated EEG feature data are covered or smoothed.

[0128] like Figure 2 The diagram shown illustrates the working principle of this invention. The working principle of this invention lies in constructing an anesthesia depth quantification framework with a three-layer progressive architecture of multimodal adaptive fusion, medical knowledge constraints, and individualized reinforcement calibration, and integrating it into a bedside anesthesia monitoring computing unit. For example, the computing platform can employ an embedded GPU module, supporting real-time graph neural network inference and online reinforcement learning updates.

[0129] Specifically, the system achieves synchronous and high-precision acquisition of multimodal physiological signals from patients through multiple digital interfaces (EEG acquisition chip, 12-bit ADC cyclic signal acquisition board). The sampling rate of the EEG channel is set to 256Hz, and the cyclic channel is set to 125Hz. After the raw signal is filtered by an on-chip FIR filter to remove power frequency noise (50Hz notch filter) and baseline drift (0.5Hz high-pass filter), the preprocessing module extracts the EEG power spectrum features (δ / α / β band power ratio), HRV time-frequency domain index, and TCI pump real-time concentration, and integrates them into a multimodal data vector.

[0130] The dynamic confidence assessment module calculates the SNR and dispersion of each modal signal corresponding to the multimodal data vector in real time based on the above features, and generates an adaptive weighting coefficient set. This effectively shields against electrosurgical interference (typical frequency band 200kHz~500kHz) and electromyographic artifacts generated by surgical operations (>100Hz), ensuring the signal quality foundation for subsequent fusion calculations. By dynamically clearing the weighting coefficients of severely contaminated channels to zero, this module enables the entire calculation framework to automatically degrade when signal quality deteriorates, ensuring the continuity and stability of the exponential output.

[0131] In one embodiment, the domain knowledge graph is pre-installed in the system firmware, containing 312 entity nodes (covering 18 commonly used anesthetic drugs, 47 physiological characteristic indicators, and 12 anesthetic state classifications) and 1548 weighted directed edges (based on causal relationships extracted from clinical pharmacology textbooks and anesthesia guidelines). The graph convolutional inference module receives effect-room concentration as activation input in real time, completes causal chain inference from drug molecule action to macroscopic state of consciousness through a three-layer GCN, and outputs a knowledge inference vector. Under the logical consistency verification framework, when the output deviation between data-driven and knowledge-driven approaches exceeds the boundary threshold, the system automatically initiates a dynamic adjustment coefficient correction mechanism to prevent abnormal exponential jumps caused by data noise, ensuring the self-consistency of the algorithm output within the scope of medical logic, thereby specifically solving the logical disconnect problem caused by the lack of pharmacological causal constraints in existing technologies.

[0132] The reinforcement learning calibration module initializes the policy network based on the state mapping package of the monitored object after the start of surgery. It then continuously receives environmental state feedback in 5-second increments and updates the policy parameters online. This mechanism enables the system to gradually perceive the personalized response curve of the monitored object to anesthetic drugs during surgery, achieving individualized and precise calibration that surpasses general population models. This effectively solves the problem that existing methods cannot adapt to individual differences in special patient groups such as the elderly and critically ill patients.

[0133] In addition, the present invention also includes a cross-modal trend verification step: real-time comparison of the evolution trend of the first type of monitoring feature and the evolution trend of the second type of monitoring feature in the multimodal data vector;

[0134] When the first type of monitoring feature indicates a deepening of the target state and the second type of monitoring feature indicates an enhanced correlation state, a co-confirmation signal is generated. Based on this co-confirmation signal, the gain of the final fusion state index is increased. The first type of monitoring feature refers to observation features sensitive to changes in the target state, and the second type of monitoring feature refers to correlation features that have a physiological or physical coupling relationship with the target state. When the first type of feature indicates a change in the target state in a certain direction, and the second type of feature indicates a synchronous change in the coupled feature, a co-confirmation signal can be generated to enhance the confidence level.

[0135] In this embodiment, the evolution trends of EEG features and circulatory system features are compared in real time. When the EEG features show a decrease in the depth of consciousness and the circulatory system features show an increase in heart rate variability, a co-confirmation signal is generated, and the gain of the final anesthesia depth index is increased based on the co-confirmation signal to prompt the anesthesiologist to adjust the dosing regimen in a timely manner. The invention also includes cross-modal semantic consistency verification and data self-repair steps: the evolution trend of circulatory system features is acquired in real time, input into a preset physiological logic mapping operator, and the expected consciousness inhibition feature value corresponding to the current circulatory state is obtained by reverse calculation; the actual extracted value of the EEG features and the expected consciousness inhibition feature value are semantically aligned and verified, and the deviation measure between the two is calculated; when the deviation measure exceeds a preset causal conflict threshold, it is determined that there is interference pollution in the multimodal monitoring signal sequence, and the inference component associated with the current drug concentration feature in the domain knowledge graph model is called to perform nonlinear smoothing correction based on medical logic on the initial anesthesia depth index.

[0136] This invention achieves high precision and reliability in quantifying anesthesia depth through adaptive fusion of multimodal signals and constraints based on medical knowledge. The dynamic confidence weighting mechanism effectively addresses the instability of the index caused by signal contamination in the clinical environment; the causal reasoning constraints based on knowledge graphs overcome the fundamental flaw of purely data-driven models that neglect pharmacological logic; and the individualized online calibration based on reinforcement learning enhances the algorithm's adaptability to different monitoring groups.

[0137] The following is a specific example of a tertiary hospital's general surgery department using the knowledge graph-constrained multi-source data fusion method of this invention to monitor the depth of anesthesia in a 65-year-old male patient (weighing 70kg, ASA II, undergoing laparoscopic cholecystectomy with propofol target-controlled infusion):

[0138] During the system initialization phase, individual difference vectors were constructed based on the patient's age (65 years), weight (70 kg), and ASA II level. The K-Means feature clustering model was used to locate the cluster center c3 (corresponding to the elderly intermediate-risk patient group), and the corresponding parameter configuration (ke0=0.456 min-1, policy network initial weight θ3) was retrieved to initialize the reinforcement learning calibration model.

[0139] Five minutes after anesthesia induction, multimodal data vector extraction yielded the following results: EEG delta power ratio 0.72, alpha power ratio 0.08, burst inhibition ratio 0; HRV LF / HF (autonomic balance index) ratio 0.85; systolic blood pressure 82 mmHg (baseline 100 mmHg); current propofol plasma concentration... =3.2μg / mL.

[0140] Dynamic confidence assessment: EEG channel Normalized value of dispersion index Circulation system channel , The confidence level for the dosing concentration channel (TCI pump direct digital interface) is 1.0. The dynamic confidence levels for the three channels are calculated as follows: =0.730, =0.812, =1.0, and the normalized weighted coefficient set is [0.292, 0.325, 0.383]. The confidence scores of each channel are all higher than the threshold. All of them are incorporated into fusion computing.

[0141] Knowledge Graph Reasoning: Calculating Effect-Room Concentration Based on the Schnider Model =2.91μg / mL, activates the moderate inhibition pathway in the knowledge graph corresponding to the propofol effect chamber concentration of 2.5~3.5μg / mL, and generates a knowledge inference vector through three-layer GCN inference, with the corresponding consciousness inhibition interval mapping value of [40,60].

[0142] Fusion computing: Data-driven observations Knowledge-driven reference values Euclidean distance It did not exceed the logical consistency boundary threshold. The initial anesthesia depth index is obtained by equal weight fusion. .

[0143] Reinforcement learning calibration: Spatiotemporal attention mechanism extracts trend features over the past 30 seconds ( The value decreased continuously from 58.2 to 49.7, a decrease rate of approximately 0.28 / s. The reward function evaluates the lag term of the pharmacological response. (Judging that the causal match between the current index and the drug administration pulse is too high), steady-state jitter penalty term The policy network outputs a dynamically adjusted step size. .

[0144] Final output: Anesthesia depth index after compensation calculation The final anesthesia depth index is obtained after the data is within the effective range [0,100] and undergoes 5 steps of mean filtering. Within the clinically appropriate depth of anesthesia (40-60), the anesthesiologist does not need to adjust the dosing regimen. In this embodiment, the system's dynamic confidence masking mechanism ensures the exponential continuity of the electromyography interference period during surgery; the synergistic effect of knowledge graph constraints and reinforcement learning individualized calibration ensures that the output results are consistent with the anesthesiologist's comprehensive clinical judgment.

[0145] This embodiment demonstrates the complete computational process of the method of the present invention in a typical clinical anesthesia scenario. Through the synergistic effect of multimodal signal dynamic confidence adaptive fusion, knowledge graph causal constraints, and reinforcement learning individualized calibration, the system achieves accurate quantification of the anesthesia depth state of the monitored subjects. The system's anesthesia depth index error is reduced by 38% compared to the single-modal method, and by 52% in patients with comorbid circulatory system diseases. This effectively solves the systematic inaccuracies caused by signal contamination, lack of pharmacological logic, and weak adaptability to individual differences in existing technologies.

[0146] Example 2:

[0147] like Figure 3 As shown, this invention provides a multi-source data fusion system based on knowledge graph constraints. This system is used to implement the multi-source data fusion method based on knowledge graph constraints described in Embodiment 1 above. The system is integrated into a bedside anesthesia monitoring computing unit and specifically includes:

[0148] The data acquisition module is used to synchronously collect multimodal monitoring signal sequences and convert them into structured multimodal data vectors. The multimodal data vectors include at least EEG features, circulatory system features, and drug concentration features.

[0149] The quality assessment module is used to perform real-time signal quality scanning on multimodal data vectors, calculate the dynamic confidence of each modal signal corresponding to the acquired multimodal data vectors, and generate the corresponding weighted coefficient set according to the distribution ratio of the dynamic confidence.

[0150] The knowledge reasoning module is used to input multimodal data vectors into a preset domain knowledge graph model, perform topological association analysis and graph convolution operations on the multimodal data vectors based on the preset causal constraint logic in the domain knowledge graph model, and output knowledge reasoning vectors that represent the suppression state logic constraints.

[0151] The fusion computing module is used to perform arithmetic synthesis of multimodal data vectors with knowledge reasoning vectors after weighted mapping through a set of weighted coefficients, to obtain the initial anesthesia depth index.

[0152] The reinforcement learning calibration module is used to obtain the state mapping package of the current monitored object, take the initial anesthesia depth index and multimodal data vector as the current environmental state information, and use the preset reinforcement learning calibration model to calculate the dynamic correction step size for the initial anesthesia depth index.

[0153] The output module is used to perform compensation calculations on the initial anesthesia depth index based on the dynamic correction step size, and after data smoothing and safety filtering, generate the final anesthesia depth index.

[0154] This invention constructs an anesthesia depth calculation architecture with real-time adaptive capabilities through deep hardware and software coupling. The hardware composition and functional cooperation relationship of each module are as follows:

[0155] The data acquisition module consists of a high-precision multi-channel biosignal acquisition board (core chip: ADS1299 EEG front-end, 12-bit ADC cyclic signal acquisition unit), a digital TCI pump communication interface (RS-232 / USB), and an embedded preprocessing DSP unit. This module achieves synchronous sampling of multimodal signals (clock synchronization accuracy <1ms) and real-time feature extraction at the physical layer, ensuring the time alignment of the multimodal feature vectors required for subsequent fusion calculations, directly supporting the accuracy of dynamic confidence assessment and fusion calculations.

[0156] The quality assessment module consists of on-chip signal analysis logic (based on FPGA to implement FFT accelerated computation) and a dynamic threshold comparator. When the dynamic confidence of any channel is lower than the preset threshold, the hardware flag automatically sets the corresponding weighting coefficient to zero, preventing low-quality signals from contaminating the fusion result, thus achieving a hardware-level solution to the signal robustness problem in the background technology.

[0157] The knowledge reasoning module consists of a graph neural network reasoning acceleration unit (based on GPU Tensor Cores to implement batch graph convolution operations) and a knowledge graph static storage area (Flash ROM). The pre-built domain knowledge graph is stored in the read-only storage area, and the graph convolution reasoning results are written to the high-speed SRAM cache in real time for the fusion computing module to read, ensuring that the pharmacological causal constraints remain effective throughout the exponential calculation process.

[0158] The fusion computing module consists of a fully connected neural network inference unit and a logical consistency check comparator. It is responsible for performing weighted mapping, Euclidean distance calculation and dynamic adjustment coefficient correction, and finally outputs the initial anesthesia depth index.

[0159] The reinforcement learning calibration module consists of a policy network computing unit (embedded GPU), a spatiotemporal attention feature extractor, and an online optimizer. The policy network parameters are stored in non-volatile memory (NAND Flash), supporting online rolling updates during the surgical procedure, thus achieving parameter persistence and individualized adaptive calibration.

[0160] The output module consists of a safety boundary hardware limiter, a mean filter DSP module, and a standardized display interface (RS-485 / Ethernet), and is responsible for outputting the final anesthesia depth index to the anesthesia monitor display and the operating room information system (AIMS) in real time.

[0161] Taking the cardiothoracic surgery operating room of a top-tier hospital as an example: The system is integrated into the anesthesia workstation, monitoring a 42-year-old female patient (weighing 58kg, ASA I, undergoing thoracoscopic lobectomy with sevoflurane inhalation anesthesia). The data acquisition module connects to the patient's scalp electrodes via a BNC interface to collect 8-lead EEG, simultaneously acquiring the pulse oximetry waveform and the digital output of the sevoflurane concentration from the anesthesia machine. The quality assessment module detects high-amplitude electromyography artifacts during endotracheal intubation (CEEG in the EEG channel drops to 0.21, below the threshold of 0.30), automatically setting the EEG weighting coefficient to zero, and the fusion calculation is now dominated by circulatory and drug concentration characteristics; 30 seconds after intubation, the EEG signal quality recovers, and the EEG channel is automatically reintegrated into the fusion calculation. Throughout the process... The output was smooth and stable, without any false spikes in the index caused by electromyography interference. As a result, the anesthesiologist avoided unnecessary procedures to deepen the anesthesia, and the surgery was completed smoothly.

[0162] This invention achieves full automation of the anesthesia depth index calculation process, from signal acquisition, quality assessment, knowledge reasoning, multimodal fusion to individualized calibration, through a multi-module collaborative systematic architecture. Real-time data interaction between modules and hardware acceleration design ensure reliable operation of the system in clinical anesthesia scenarios (index update cycle ≤ 1 second), significantly improving the accuracy and safety of anesthesia depth monitoring.

[0163] This embodiment presents a multi-source data fusion system based on knowledge graph constraints, used to implement the aforementioned multi-source data fusion method based on knowledge graph constraints. Therefore, the specific implementation of the multi-source data fusion system based on knowledge graph constraints can be found in the previous embodiment section of the multi-source data fusion method based on knowledge graph constraints. To avoid redundancy, it will not be repeated here.

[0164] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0165] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0166] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0167] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A multi-source data fusion method based on knowledge graph constraints, characterized in that: Includes the following steps: Synchronously acquire multimodal monitoring signal sequences and convert them into structured multimodal data vectors, wherein the multimodal data vectors include at least feature components reflecting different dimensional states; The multimodal data vector is subjected to real-time signal quality scanning, the dynamic confidence level of each modal signal corresponding to the acquired multimodal data vector is calculated, and a corresponding weighted coefficient set is generated according to the distribution ratio of the dynamic confidence level. The multimodal data vector is input into a preset domain knowledge graph model. Based on the preset causal constraint logic in the domain knowledge graph model, the multimodal data vector is subjected to topological association analysis and graph convolution operation to output a knowledge reasoning vector representing the suppression state logic constraint. The multimodal data vector is weighted and mapped using the weighted coefficient set, and then arithmetically synthesized with the knowledge reasoning vector to obtain the initial fusion state index. Obtain the state mapping package of the current monitored object, use the initial fusion state index and the multimodal data vector as the current environmental state information, and use a preset reinforcement learning calibration model to calculate the dynamic correction step size for the initial fusion state index. The initial fusion state index is compensated based on the dynamic correction step size to generate the final fusion state index.

2. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: The calculation process for the dynamic confidence level of each modal signal corresponding to the acquired multimodal data vector includes: Extract the instantaneous envelope features of each modal signal in the multimodal data vector, and calculate the ratio of the standard deviation to the mean of each modal signal in the time domain through a preset sliding window to obtain the dispersion index; The power spectrum distribution of each mode signal is obtained by using a preset fast Fourier transform. The signal power in the corresponding frequency band is extracted as the effective signal power according to the preset monitoring characteristic frequency band. The ratio of the effective signal power to the noise power of the whole frequency band is calculated to obtain the real-time signal-to-noise ratio. The discreteness index and the real-time signal-to-noise ratio are weighted and summed to output the dynamic confidence of each modal signal; if the dynamic confidence of any modal signal is lower than a preset threshold, the coefficient of that modal data vector in the weighted coefficient set is set to zero.

3. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: The multimodal data vector is input into a preset domain knowledge graph model. The construction and reasoning process of the domain knowledge graph model includes: The system pre-defines a knowledge structure topology graph containing state variable entities, observation feature entities, and target state entities, and defines directed edge permissions between entities according to pre-defined causal rules. At least one feature component is obtained from the multimodal data vector, and the core state variable value corresponding to the current moment is calculated based on the preset system dynamics model, and used as the initial activation node of the knowledge structure topology graph; Using a multi-layer graph convolutional network, the multimodal data vector is used as an external observation to update the initial activated node, the influence weight of each entity node in the topological space is calculated, and the knowledge inference vector is generated.

4. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: The specific steps for weighting and mapping the multimodal data vector via the weighting coefficient set include: The multimodal data vector, after being weighted by the set of weighted coefficients, is transformed nonlinearly and mapped to a feature space of a preset dimension, and the output data drives the observation values. The knowledge reasoning vector is input into a preset logistic regression model, and the knowledge-driven reference value is output. Calculate the Euclidean distance between the data-driven observations and the knowledge-driven references; Determine whether the Euclidean distance exceeds a preset logical consistency boundary threshold: If the value is not exceeded, the data-driven observation value and the knowledge-driven reference value are fused to obtain an intermediate fused value. If the value exceeds the limit, a dynamic adjustment coefficient is calculated based on the ratio of the difference between the Euclidean distance and the boundary threshold. The data-driven observation is then corrected using the dynamic adjustment coefficient, and the corrected data-driven observation is fused with the knowledge-driven reference value to obtain an intermediate fused value. The intermediate fusion value is used as the initial fusion state index.

5. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: The process of obtaining the state mapping packet of the currently monitored object includes: Obtain the static attribute features of the monitored object and convert them into a structured individual difference vector; The individual difference vector is input into a preset feature clustering model to locate the cluster center to which the monitored object belongs within a preset sample sensitivity space; Based on the cluster centers, the corresponding preset parameter configurations are obtained to form a state mapping package for the current monitored object, which is used to initialize the policy network of the reinforcement learning calibration model.

6. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: In calculating the dynamic correction step size for the initial fusion state index using a preset reinforcement learning calibration model, the dynamic correction calculation of the reinforcement learning calibration model includes: The current environmental state information is used to perform feature mining by utilizing a preset spatiotemporal attention mechanism to extract the evolution trend features of the multimodal monitoring signal sequence within a preset historical period; A reinforcement learning reward function based on time delay compensation is constructed; wherein, the reinforcement learning reward function includes a causal response lag evaluation term and a state stability maintenance term; wherein, the causal response lag evaluation term quantifies the causal matching degree between the initial fusion state index at the current moment and the historical control input pulse according to a preset system response time constant; The reinforcement learning calibration model, based on the evolution trend characteristics, searches for the optimal dynamic correction step size within a preset action space through feedback guidance from the reinforcement learning reward function.

7. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: After generating the final fusion state index, the method further includes data smoothing and security filtering steps, specifically including: Determine whether the value after the compensation operation is within a preset valid state range. If it exceeds the valid state range, then forcibly correct the value to a preset safety boundary value. The mean filtering process is performed on the continuous fusion state index sequence with a preset step size to eliminate the non-uniform jumps caused by instantaneous external disturbances and generate the smoothed final fusion state index.

8. The multi-source data fusion method based on knowledge graph constraints according to claim 1, characterized in that: After generating the final fusion state index, the method further includes a cross-modal trend verification step, specifically including: The evolution trends of the first type of monitoring features and the second type of monitoring features in the multimodal data vector are compared in real time. When the first type of monitoring feature shows that the target state is deepening and the second type of monitoring feature shows that the correlation state is strengthening, a collaborative confirmation signal is generated, and the gain of the final fusion state index is increased based on the collaborative confirmation signal.

9. The multi-source data fusion method based on knowledge graph constraints according to claim 8, characterized in that: The method further includes cross-modal semantic consistency verification and data self-repair steps, specifically including: The evolution trend of the first type of monitoring feature in the multimodal data vector is obtained in real time, and a preset logical mapping operator is input to calculate the expected value corresponding to the current first type of monitoring feature in reverse. The actual extracted values ​​of the second type of monitoring features are semantically aligned with the expected values, and the deviation measure between the two is calculated. When the deviation measure exceeds the preset causal conflict threshold, it is determined that there is interference pollution in the multimodal monitoring signal sequence. The inference component associated with the current key state variable in the domain knowledge graph model is called to perform nonlinear smoothing correction on the initial fusion state index.

10. A multi-source data fusion system based on knowledge graph constraints, characterized in that: The method for multi-source data fusion based on knowledge graph constraints as described in any one of claims 1 to 9 includes: The data acquisition module is used to synchronously acquire multimodal monitoring signal sequences and convert them into structured multimodal data vectors, wherein the multimodal data vectors include at least feature components reflecting different dimensional states; The quality assessment module is used to perform real-time signal quality scanning on the multimodal data vector, calculate the dynamic confidence of each modal signal corresponding to the acquired multimodal data vector, and generate a corresponding weighted coefficient set according to the distribution ratio of the dynamic confidence. The knowledge reasoning module is used to input the multimodal data vector into a preset domain knowledge graph model, perform topological association analysis and graph convolution operation on the multimodal data vector based on the preset causal constraint logic in the domain knowledge graph model, and output a knowledge reasoning vector representing the suppression state logic constraint. The fusion computing module is used to perform arithmetic synthesis of the multimodal data vector with the knowledge reasoning vector after weighting the multimodal data vector through the weighted coefficient set, so as to obtain the initial fusion state index. The reinforcement learning calibration module is used to obtain the state mapping package of the current monitored object, take the initial fusion state index and the multimodal data vector as the current environmental state information, and calculate the dynamic correction step size for the initial fusion state index using a preset reinforcement learning calibration model. The output module is used to perform compensation calculations on the initial fusion state index according to the dynamic correction step size to generate the final fusion state index.

Images