A data processing method and apparatus for incomplete multi-source information

Abstract

This application relates to a data processing method and apparatus for incomplete multi-source information. The method includes: performing quality assessment on multi-source data channels to generate topological fingerprints characterizing the availability status of each data channel; dynamically routing data in a pre-constructed semantic topology graph to determine the target node corresponding to the current input data and the statistical parameter library of the target node; calculating the class-conditional inconsistency score of the current sample in the feature space; calibrating the information entropy of the inconsistency score distribution of the sample; adaptively weighting the sample for risk; and determining a dynamic safety threshold based on the weighted quantiles; comparing the inconsistency score of the current sample in each candidate category with the dynamic safety threshold to generate a prediction set containing categories that meet the threshold conditions as output. This method can ensure the stability and maintenance efficiency of the system during multi-source information fusion decision-making without the need for manual annotation.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and big data processing technology, and in particular to a data processing method and apparatus for incomplete multi-source information. Background Technology

[0002] In modern intelligent systems (such as multimodal content understanding, autonomous driving perception, intelligent voice interaction, and industrial IoT monitoring), decision-making often relies on the collaborative analysis of multiple data channels, including images, text, audio, vibration, and temperature. With the development of deep learning technology, multimodal fusion models have been widely applied in these fields.

[0003] However, in real-world high-concurrency data stream processing or complex physical environments, the integrity of data channels is difficult to guarantee continuously. Due to network jitter, sensor hardware failures, upstream interface timeouts, or privacy policy restrictions, partial data loss (i.e., modal missing) frequently occurs. Currently, industry solutions for handling such incomplete data mainly fall into two categories: one is interpolation completion, which uses generative models or statistical means to fill in missing values; the other is directly discarding samples containing missing values.

[0004] For incomplete data input, existing mainstream technical solutions typically have the following significant drawbacks: 1. Model Illusion and Risk of Misjudgment: Interpolation completion is essentially a "guess" of unknown information based on known information, inevitably introducing feature noise. Subsequent classification models often make "overconfident" erroneous judgments based on these completed features and lack an effective assessment of the reliability of the prediction results; in safety-critical fields such as healthcare and security, this risk is unacceptable.

[0005] 2. Ignoring the heterogeneity of missing data distribution: Existing systems often set a uniform confidence threshold (e.g., Softmax Score > 0.8) for all samples, ignoring the significant differences in feature distribution between "complete data" and "missing data" states. The result is that the false positive rate increases significantly when data is severely missing, while being overly conservative when data is complete may lead to missed detections, making it difficult to balance security and efficiency.

[0006] 3. Lack of adaptive decision boundary adjustment capability: Existing technologies struggle to quantify the "ambiguity" of samples in the feature space, making it impossible to automatically adjust the decision boundary based on the current information completeness under unsupervised conditions. For "difficult samples" on the classification edge, traditional methods often adopt a one-size-fits-all approach, lacking a mechanism to dynamically expand the candidate range to reduce the risk of missed classification. Summary of the Invention

[0007] Therefore, it is necessary to provide a data processing method and apparatus for incomplete multi-source information that can ensure the stability and maintenance efficiency of system operation when making decisions based on multi-source information fusion without the need for manual annotation, in order to address the above-mentioned technical problems.

[0008] A data processing method for incomplete multi-source information, the method comprising: The quality of multi-source data channels is assessed, and a topological fingerprint characterizing the availability status of each data channel is generated.

[0009] Based on the topological fingerprint, dynamic routing is performed in the pre-constructed semantic topology graph to determine the target node corresponding to the current input data and the statistical parameter library of the target node.

[0010] Based on the statistical parameter library of the target node, calculate the class conditional inconsistency score of the current sample in the feature space.

[0011] Based on the information entropy of the non-consistent score distribution of calibration samples within the target node, the samples are adaptively risk-weighted, and a dynamic safety threshold is determined based on the weighted quantile.

[0012] The non-consistency scores of the current sample in each candidate category are compared with dynamic safety thresholds, and a prediction set containing the categories that meet the threshold conditions is generated as the output.

[0013] Based on the sliding window caching and incremental update mechanism, the statistical parameter library is self-evolved online.

[0014] A data processing apparatus for incomplete multi-source information, the apparatus comprising: The topology fingerprint generation module is used to perform quality assessment on multi-source data channels and generate topology fingerprints that characterize the availability status of each data channel.

[0015] The statistical parameter library determination module is used to perform dynamic routing in a pre-built semantic topology graph based on topological fingerprints, and to determine the target node and the statistical parameter library of the target node corresponding to the current input data.

[0016] The score calculation module is used to calculate the class conditional inconsistency score of the current sample in the feature space based on the statistical parameter library of the target node.

[0017] The safety threshold setting module is used to adaptively risk-weight the samples based on the information entropy of the non-consistent score distribution of the calibration samples within the target node, and determine the dynamic safety threshold based on the weighted quantile.

[0018] The decision module compares the inconsistency scores of the current sample in each candidate category with a dynamic safety threshold, and generates a prediction set containing the categories that meet the threshold conditions as output.

[0019] The self-evolution module is used to perform online self-evolution of the statistical parameter library based on the sliding window caching and incremental update mechanism.

[0020] The aforementioned data processing method and apparatus for incomplete multi-source information first discretizes and encodes the missing states of multi-source channels using topological fingerprints to accurately characterize the heterogeneity of the input data distribution. Then, it constructs a semantic topological graph and performs dynamic routing based on Hamming distance, matching a dedicated statistical parameter library for each missing mode, fundamentally solving the problem of traditional uniform thresholds ignoring distribution differences. On this basis, it constructs class-conditional inconsistency scores by fusing Mahalanobis distance and local manifold density, achieving effective evaluation of sample reliability in a quantitative form, avoiding model overconfidence when information is missing. Simultaneously, it introduces an information entropy-driven risk weighting mechanism, adaptively adjusting sample weights according to the disorder of the inconsistency score distribution under the current missing mode, and dynamically determining the safety threshold through weighted quantiles, enabling the decision boundary to automatically adjust with environmental uncertainty. Finally, without manual annotation, it replaces empirical thresholds with statistical coverage constraints to achieve credibility calibration, and adapts to concept drift online through a sliding window incremental update mechanism, significantly reducing computation and annotation costs. In long-term deployment, during multi-source information fusion decision-making, it ensures system stability and maintenance efficiency, achieving a complete closed loop from data quality perception to adaptive decision-making. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a data processing method for incomplete multi-source information in one embodiment; Figure 2 This is a structural block diagram of a data processing device for incomplete multi-source information in one embodiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] In one embodiment, such as Figure 1 As shown, a data processing method for incomplete multi-source information is provided, including the following steps: Step 102: Perform quality assessment on the multi-source data channels and generate topological fingerprints that characterize the availability status of each data channel.

[0024] Multi-source data can be data from multiple sensors of the same type (such as a microphone array composed of multiple microphones), with the same features but different spatial locations; it can also be data of different types (such as image + text + sound), with large differences in feature space, providing complementary information; or it can be data from multiple independent sources measuring the same physical quantity (such as GPS + IMU + wheel speedometer for positioning), which improves the reliability and accuracy of the fused redundant information.

[0025] Specifically, the system contains three data channels: an image channel (v=1), a text channel (v=2), and an audio channel (v=3), with a total of 10 categories and a uniform feature dimension d=128. The system receives a sample in real time, which is the raw data from the three channels. First, the signal-to-noise ratio (SNR) of each channel signal is detected. The preset availability threshold δ_threshold=20dB. The calculated SNR for the image channel is 25dB, the SNR for the text channel is 15dB (due to the large amount of noise in the text), and the SNR for the audio channel is 30dB. This is based on the following formula: ; The topological fingerprint of the sample is obtained as m_target=[1,0,1].

[0026] Furthermore, simultaneously, for the labeled valid image and audio channels, pre-trained ResNet and audio neural networks are invoked to extract features, which are then mapped to a 128-dimensional unified space via a projection head. Since the text channel is unavailable, its feature vector is padded with zeros. Finally, the features from the three channels are fused to obtain the joint feature vector Z_i for the sample.

[0027] Step 104: Based on the topological fingerprint, perform dynamic routing in the pre-constructed semantic topology graph to determine the target node corresponding to the current input data and the statistical parameter library of the target node.

[0028] Specifically, the system has pre-constructed a semantic topology graph G=(V, E). Each node in the graph represents a historically occurring topological fingerprint pattern (such as [1,1,1], [1,1,0], [1,0,1], etc.). Edges connect nodes with a Hamming distance of 1.

[0029] Furthermore, the system checks if a complete node (i.e., a node that has accumulated a sufficient number of samples) exists in the graph corresponding to m_target=[1,0,1]. The query finds that the node ν_target exists and is complete. Therefore, the system performs direct hit routing, routing the sample and its feature vector Z_i to the node. and obtain its statistical parameter library. This contains the feature centroids of each category under that node. Shared covariance matrix and the set of inconsistencies in the scores of existing calibration samples .

[0030] Step 106: Calculate the class conditional inconsistency score of the current sample in the feature space based on the statistical parameter library of the target node.

[0031] Specifically, at the node Internally, the local sparsity factor ρ_i of the sample is first calculated. The HNSW index is used to retrieve K=10 nearest neighbor samples in the node reference library, and the average Euclidean distance between them and Z_i is calculated, resulting in ρ_i=0.35 (a small value, indicating that the sample is located in a high-density region).

[0032] Furthermore, the Ledoit-Wolf shrinkage algorithm is used to estimate the corrected covariance matrix Σ_reg within the node. The shrinkage coefficient λ_lw = 0.2 is set, and Σ_reg is calculated based on the sample covariance Σ_sample within the node and the feature dimension d = 128.

[0033] Furthermore, the Mahalanobis distance D(Z_i,k) from sample Z_i to each class k is calculated. For example, the Mahalanobis distance to the centroid μ_3 of the true class y=3 is calculated to be 2.1, and the Mahalanobis distance to the nearest competing class j=5 is calculated to be 2.8.

[0034] Furthermore, the composite class conditional inconsistency score is calculated. The manifold sensitivity hyperparameter γ is set to 1.0. The score for the true class k=3 is: ; Similar calculations were performed on other categories (e.g., k=5), yielding S(Z_i, 5) ≈ 1.34.

[0035] Step 108: Based on the information entropy of the non-consistent score distribution of the calibration samples within the target node, adaptive risk weighting is applied to the samples, and a dynamic safety threshold is determined based on the weighted quantiles.

[0036] Specifically, at the node Within the given set, there are N=500 calibration samples and their set of inconsistent scores S. This set is discretized into Bins=50 histogram intervals, and the Shannon entropy H(S) is calculated to be 2.3 (a moderately high entropy value, indicating that the classification boundary has some ambiguity in this missing data pattern).

[0037] Furthermore, risk weights are assigned to each calibration sample based on the entropy value. The base risk coefficient β = 2.0 and the entropy gain coefficient λ_ent = 0.5 are set. For a difficult sample in the set (whose score ranking r_i = 0.9), its unnormalized weight is: ; The weight of a simple sample (r_i=0.1) is only exp(0.43)≈1.54. After normalization, the weight w_i of the difficult sample is much higher than that of the simple sample.

[0038] Furthermore, a preset global false alarm rate control parameter α = 0.05 is set (i.e., the desired coverage rate is 95%). On the weighted calibration samples, the weighted quantiles are calculated to find the dynamic safety threshold Q_safe(ν_target) such that the cumulative weight reaches 95%. Q_safe = 1.20 is calculated.

[0039] Step 110: Compare the inconsistency scores of the current sample under each candidate category with the dynamic safety threshold, and generate a prediction set containing the categories that meet the threshold conditions as output.

[0040] Specifically, for the current test sample Z_i, its inconsistency scores under each candidate category have been calculated (e.g., S(Z_i,3)=1.00, S(Z_i,5)=1.34, ...). These scores are then compared with the dynamic safety threshold Q_safe=1.20.

[0041] Furthermore, categories satisfying S(Z_i,k) ≤ 1.20 are selected: Category 3 (1.00 ≤ 1.20) satisfies the condition, while Category 5 (1.34 > 1.20) does not. Finally, the generated prediction set is C(Z_test) = {3}.

[0042] Furthermore, since there is only one category in the prediction set, the system determines it to be in "high-precision mode" and directly outputs the unique category 3 and its corresponding inconsistency score of 1.00 (or the mapped confidence level), triggering the corresponding execution action.

[0043] Step 112: Based on the sliding window caching and incremental update mechanism, the statistical parameter library is self-evolved online.

[0044] Specifically, the system consists of nodes. Maintain a FIFO buffer queue with a capacity of 1000. The current sample Z_i and its future true feedback label y_i=3 are stored in the queue. When the queue is full, the oldest sample is removed.

[0045] Furthermore, after receiving the feedback tag, the system triggers an incremental update for the node. The centroids of category 3 are updated using an exponential moving average (EMA). Assuming the momentum coefficient η = 0.1, the eigenmean of the latest batch of 32 labeled samples is... If _batch is used, then the updated centroid is: ; Furthermore, the system periodically re-estimates the covariance matrix within each node at regular intervals (e.g., every 200 new samples received). Histogram statistics are performed on the inconsistent score set S, and the HNSW nearest neighbor index is reconstructed. Through this mechanism, the statistical parameter library can smoothly adapt to changes in data distribution, achieving online self-evolution of the system.

[0046] The aforementioned data processing method for incomplete multi-source information first uses topological fingerprinting to discretize and encode the missing states of multi-source channels, accurately characterizing the heterogeneity of the input data distribution. Then, a semantic topological graph is constructed and dynamic routing is performed based on Hamming distance, matching a dedicated statistical parameter library for each missing mode, fundamentally solving the problem of traditional unified thresholds ignoring distribution differences. On this basis, a class-conditional inconsistency score is constructed by fusing Mahalanobis distance and local manifold density, achieving effective evaluation of sample reliability in a quantitative form, avoiding overconfidence in the model when information is missing. Simultaneously, an information entropy-driven risk weighting mechanism is introduced, adaptively adjusting sample weights according to the disorder of the inconsistency score distribution under the current missing mode, and dynamically determining the safety threshold through weighted quantiles, enabling the decision boundary to automatically adjust with environmental uncertainty. Finally, without manual annotation, statistical coverage constraints replace empirical thresholds to achieve credibility calibration, and a sliding window incremental update mechanism adapts to concept drift online, significantly reducing computation and annotation costs. In long-term deployment, this ensures system stability and maintenance efficiency during multi-source information fusion decision-making, achieving a complete closed loop from data quality perception to adaptive decision-making.

[0047] In one embodiment, the signal-to-noise ratio (SNR) of the input signals of each data channel is detected, and data channels with an SNR lower than a preset threshold are marked as unavailable at the logical level. A binary topological fingerprint is generated based on the availability status of the marked data channels. ; in, For the first The sample at the th A binary topological fingerprint of the availability status of each channel. For channel indexing, To preset the industrial availability threshold, For the first The sample at the th The signal-to-noise ratio or confidence score on each channel.

[0048] In one embodiment, a semantic topology graph is constructed using the topological fingerprint as nodes and Hamming distances not exceeding a threshold as edges: ; ; in, For semantic topology graph, For topological fingerprint set, As an edge, For nodes that have appeared in the past, topological fingerprint patterns For statistical parameter library, For nodes Inner Class feature centroid, For nodes Internal shared covariance statistics For nodes Statistics on the set of inconsistencies in the internal calibration samples and their corresponding histograms. This is a category index. It checks if a complete node exists for the topological fingerprint of the current sample. If so, it routes directly to the current node; otherwise, or if the node is incomplete, it searches the set of complete nodes in the current node's neighborhood and synthesizes the current node's statistical parameters using inverse distance weighting. ; in, These are the statistical parameters for the current node. For a complete set of nodes, Let Hamming distance function be used. For the target topological fingerprint, These are the statistical parameters corresponding to the topological fingerprint of the current sample. The topological fingerprint of the current sample. This is a smoothing factor.

[0049] In one embodiment, based on the statistical parameter library of the target node, within the target node, a graph index-accelerated nearest neighbor search algorithm is used to retrieve a preset number of nearest neighbor samples of the current sample, and the local sparsity factor is calculated: ; in, For local sparsity factor, The preset number of neighboring samples. For the current sample The joint eigenvectors, For the first The feature vectors of the neighboring samples As a smoothing factor, The distance is Euclidean. The Ledoit-Wolf shrinkage algorithm is used to estimate the covariance matrix within the target node, resulting in the corrected covariance matrix: ; in, To correct the covariance matrix, For the target node, The shrinkage coefficient, For feature dimension, For in-node sample covariance estimation, The identity matrix is ​​used. The Mahalanobis distance from the current sample to each category is calculated based on the modified covariance matrix: ; in, For the current sample To Category Mahalanobis distance, To the target node Internal Category Characteristic centroid, For category indexing, To correct the inverse of the covariance matrix, the class conditional inconsistency score is generated based on the local sparsity factor and the Mahalanobis distance: ; in, For the current sample Category Class conditional non-consistency scores, For the current sample For the most recent competing categories Mahalanobis distance, As a smoothing factor, For manifold sensitivity hyperparameters, For local sparsity factor, It is the hyperbolic tangent function.

[0050] In one embodiment, the Shannon entropy is calculated after discretizing the set of inconsistencies in the calibration samples within the target node into a histogram: ; in, For the target node Shannon entropy of an internally inconsistent fractional distribution For the target node The set of inconsistent scores of the internal calibration samples For histogram interval indexing, For non-consistent scores falling in the th The sample probability of each interval The total number of intervals divided by the histogram. Based on the Shannon entropy and the normalized rank of each score in the set of inconsistent scores, an exponential risk weight is set: ; in, For unnormalized risk weights, Basic risk coefficient, This is the entropy gain coefficient. For distribution entropy, For the target node The set of inconsistent scores of the internal calibration samples This represents the normalized ranking of inconsistent scores within a node. To normalize risk weights, This is the index of all calibration samples within the node; Dynamic security thresholds are determined based on weighted quantiles.

[0051] In one embodiment, a dynamic safety threshold that makes the cumulative sample weights reach a preset coverage requirement is obtained based on the weighted quantiles: ; in, For dynamic security thresholds, For the target node, For the indeterminate boundary, For candidate threshold variables, For the set of real numbers, To normalize risk weights, To calibrate the inconsistency scores of the samples, This is the preset global false alarm rate control parameter.

[0052] In one embodiment, a class conditional inconsistency score is calculated for the current test sample under each candidate category, and the categories that satisfy the conditional inconsistency score not being greater than the dynamic security threshold are selected to form a prediction set as output.

[0053] In one embodiment, the system is assumed to include There are 1 data channel (modal / sensor), and the number of categories is 1. The unified feature dimension is For any sample Its channel availability vector is denoted as the topological fingerprint. The unified representation after merging is denoted as The semantic topology graph is denoted as... , where nodes This corresponds to a historically observed topological fingerprint pattern. Each node... Maintain the statistical parameter library: ; in, For nodes Inner Class feature centroid, For shared covariance statistics within nodes (which can also be extended to conditional covariance form when needed). This section describes the set of inconsistencies in the scores of calibration samples within a node, along with their histogram statistics, used for entropy and threshold estimation. To avoid dependence on the true labels during the inference phase, class-conditional inconsistencies are defined. Used for any candidate class Scoring is performed; during the calibration phase, data is constructed using real labels. During the reasoning stage Class-by-class calculation And form a set output.

[0054] Step 1: Refined quality assessment and topological fingerprinting of multi-channel signals The system first receives multimodal raw data objects from heterogeneous sources (such as vision sensors, text streams, acoustic arrays, and time-series signals). To adapt to complex industrial sites or network transmission environments, this step completes the mapping from "physical signal" to "logical state": it not only determines whether the data exists but also evaluates its signal-to-noise ratio quality to avoid low-quality inputs contaminating the feature space and provides a unified input pair for subsequent routing and calibration. .

[0055] 1.1 Signal-to-noise ratio sensing and hard thresholding: Traverse each data channel of the input object and calculate the sample... In the passage Signal-to-noise ratio (SNR) or confidence score The system has a preset industrial availability threshold. ,when At the same time, data transmission occurs in the technical channel, and it is also marked as "unavailable" at the logical level to prevent low-quality noise from entering the subsequent feature representation and decision-making process.

[0056] 1.2 Topological Fingerprint Generation: The system generates a binary vector as the "topological fingerprint" of the sample, used to accurately identify the modal combination state at the current moment. For each channel, if the signal quality meets the standard, it is marked as valid (1); otherwise, it is marked as missing (0). Assume the system contains... The data channel, then the... The first sample The state of each channel is defined as follows: ; This process transforms complex sensor failures, network packet loss, or modal loss scenarios into computable discrete state codes, providing an index for subsequent dynamic routing and policy switching.

[0057] 1.3 Multimodal Feature Encoding and Manifold Alignment: For channels labeled as valid, the system calls the corresponding pre-trained deep neural network to extract high-dimensional features (e.g., ResNet for images, BERT for text). Then, a projection head is used to map the different modal features to a unified-dimensional representation space (e.g., normalized to a hypersphere), and a joint feature vector for the object is generated through feature concatenation or weighted attention mechanisms. This yields a unified input pair for the samples. This serves as the basis for subsequent STG routing, deviation calculation, and set decision-making.

[0058] Step 2: Construct a dynamic routing mechanism based on semantic topology graphs To address the problem of long-tailed distribution of massive missing patterns caused by multi-channel combination explosion (some missing combinations are extremely rare, resulting in insufficient statistical samples), this invention does not employ isolated static grouping, but instead constructs a dynamic semantic topology graph (STG) to uniformly manage the node-level statistical parameter library, and enables the borrowing of parameters across adjacent missing patterns and accelerates cold starts.

[0059] 2.1 Definition of Graph Structure: STG is a weighted undirected graph. The set of nodes V Each node Represents a historically occurring topological fingerprint (i.e., a missing pattern); edge set Connecting any two nodes whose Hamming distance is no more than 1 indicates that the two modes differ by only one channel, possessing statistical distribution similarity and parameter transfer feasibility. The Hamming distance is defined as: ; 2.2 Complete Node Determination and Dynamic Routing Strategy: A node ν is called a "complete node" when its cumulative sample count satisfies the following condition. (Optional: satisfy each class) ), to ensure and The estimate is statistically stable. For a new sample... : Direct route hit: If there is a route in the graph that is... Corresponding complete nodes Then the sample will be routed to And directly use its parameter library .

[0060] Neighborhood Soft Routing: If the corresponding node is empty or incomplete (cold start), a breadth-first search (BFS) is triggered to search for it in the graph. Complete set of nodes in the order neighborhood As a "parent node set", it is used for parameter borrowing and initialization.

[0061] 2.3 Parameter Synthesis Formula (Inverse Distance Weighted IDW): When a soft route is triggered, the system borrows the statistical distribution parameters (such as the feature mean vector) of the parent node through Inverse Distance Weighted (IDW). Covariance matrix (etc.), to synthesize the initialization parameters for the current missing pattern. To avoid division by zero and numerical instability, introduce... : ; in, This represents a statistical parameter (mean or covariance). When there exists... When a complete node is found, the system directly executes the hit route without weighted synthesis. This mechanism enables rapid startup and stable calibration in zero-sample or few-sample scenarios. The route output is the target node. and its parameter library This serves as a benchmark for subsequent deviation calculations and threshold estimations.

[0062] Step 3: Calculation of characteristic deviation based on local manifold density Selected routing node Internally, it is necessary to accurately quantify the difficulty and degree of deviation of sample identification. To overcome the problem of reduced discriminative power of Euclidean distance in high-dimensional feature spaces due to the "curse of dimensionality", this invention introduces local manifold density to perform a second-order correction on Mahalanobis distance, thereby obtaining a more robust deviation measure and forming a class conditional inconsistency score that can be used for calibration and set prediction.

[0063] 3.1 Local manifold density factor (sparseness surrogate) estimation: Nearest neighbor search is accelerated using the HNSW (Hierarchical Navigable Small World) graph index. For the current sample representation... At the node Retrieve its feature space from the reference library The local sparsity factor is obtained by calculating the average distance to the nearest neighbors from the nearest neighbor samples. (It is monotonically inversely proportional to density): ; in, Smaller values ​​indicate that the sample is located in the core region of a high-density cluster (sample within the distribution). Larger values ​​indicate that the sample is located at the edge of a sparse manifold or in a noisy region (out-of-distribution sample).

[0064] 3.2 Regularized Mahalanobis Distance Calculation (Ledoit–Wolf Shrinkage): To eliminate the influence of feature dimension correlation, the system performs regularized Mahalanobis distance calculation at nodes. The Mahalanobis distance from the sample to the centroid of each class is calculated internally. Considering that the covariance matrix may be non-invertible under small sample conditions, this invention uses Ledoit-Wolf shrinkage estimation to obtain the corrected covariance matrix: ; in ， For in-node sample covariance estimation, I It is the identity matrix. The shrinkage coefficient, This is the feature dimension. This processing ensures numerical stability and improves the reliability of the distance metric. It is further defined to categories. Mahalanobis distance: ; 3.3 Class Conditional Inconsistency Score Synthesis: To balance "inter-class separability" and "manifold deviation", the system synthesizes scores for each candidate class. Calculate conditional non-consistency scores : ; The first term is the ratio of the relative distance from the sample to the candidate class center and the nearest competing class center; the second term is the density penalty term; and γ is the manifold sensitivity hyperparameter. This definition not only characterizes "which class the sample is closer to," but also further measures "whether the sample deviates from the normal data manifold." During the calibration phase, the true labels are used. Pick Write node statistics set ; during the reasoning phase, for all calculate Used for ensemble prediction.

[0065] Step 4: Information Entropy-Driven Adaptive Risk Gain Adjustment To achieve high-precision confidence calibration and prevent models from being blindly confident when data features are chaotic, this invention proposes a risk gain algorithm based on information entropy. This algorithm is used to perceive the "degree of chaos" of the non-consistent score distribution under the current node (missing mode) and dynamically adjust the penalty weight for hard samples accordingly, making threshold estimation more prudent.

[0066] 4.1 Distribution Entropy Calculation: Calculate the current node's entropy. Inconsistent score set of internal calibration samples Discretized From the histogram intervals, we obtain the probability vector. To avoid Do the histogram - Smooth and normalize the data, then calculate the Shannon entropy: ; The higher the entropy value, the more blurred the classification boundary is under the missing pattern, and the greater the overall uncertainty faced by the system.

[0067] 4.2 Adaptive Weight Allocation: The system calculates a risk weight for each calibration sample. Unlike traditional methods that rely solely on sorting position, this method introduces "distribution entropy" as a gain coefficient into the exponential weighting. Let... For sample scores The normalized ranking (the more difficult the sample size) is, then: ; in, This is the entropy gain coefficient. It is the basic risk coefficient. The technical principle is that when an increase in environmental entropy is detected (highly disordered feature distribution), the exponential term is significantly amplified, and the tail hard samples are given higher weights, thereby forcing subsequent threshold estimations to be more conservative, and offsetting environmental uncertainties and reducing the risk of missed detections by automatically expanding the range of the output candidate set.

[0068] Step 5: Construction and Output of Dynamic Elastic Decision Boundaries The system controls the global false alarm rate based on preset parameters (e.g., That is, guarantee (coverage), at the node It internally constructs a dynamic and flexible decision boundary and generates the final output.

[0069] 5.1 Weighted quantile threshold calculation: at node On the calibration samples, the system calculates the dynamic safety threshold. This ensures that the weights of samples with inconsistency scores not exceeding this threshold accumulate to reach the target confidence level. ; 5.2 Elastic Candidate Set Generation: For online test samples At the node Internal to all categories Calculate conditional scores The prediction set is then constructed by filtering all categories whose scores do not exceed a dynamic threshold. ; 5.3 Multi-level output protocol: High-precision mode: If If the result is high confidence, a unique category will be output directly and the action will be triggered.

[0070] Safe Mode: If If ambiguity is found, the candidate list and its corresponding inconsistency score (or relative confidence weight obtained through monotonic mapping) are output, triggering the host computer's "manual review" or "downgraded operation" logic.

[0071] Abnormal circuit breaker: If or If this occurs, an error alarm will be triggered, indicating "Insufficient valid information in the input data or significant abnormal distribution".

[0072] Step 6: Online parameter updates and self-evolution based on sliding window To adapt to concept drift (such as slow data distribution shifts caused by sensor aging), this system has online self-evolution capabilities, enabling the node parameter library to evolve. Update over time and continuously meet coverage constraints.

[0073] 6.1 Sample Buffer Queue (FIFO): The system uses a FIFO queue on each node. Maintain a fixed-capacity first-in-first-out queue Store inference samples within the most recent time window and the subsequent feedback tags (Ground Truth). This cache is used for updates. , as well as Histogram statistics.

[0074] 6.2 Incremental Update (EMA): After receiving feedback labels, the system updates the centroids of the corresponding nodes and categories using an exponential moving average. For nodes... With category Let the mean of the latest batch of labeled samples be... ,but: ; in, This is the momentum coefficient. The covariance statistic can be periodically reestimated over a sliding window and obtained through Ledoit-Wolf contraction. Simultaneously update the nearest neighbor index (e.g., HNSW) and the set of inconsistent scores. Histograms are used to maintain the timeliness of density estimation and entropy-driven thresholds. This mechanism ensures that the model can smoothly adapt to environmental changes, always remain sensitive to the latest data distribution, and avoid the cost and interruption risks of frequent full retraining.

[0075] It's worth noting that the existence of data in a channel is not used as the sole criterion. Instead, signal-to-noise ratio (SNR) or confidence score is introduced to evaluate channel quality online, and low-quality channels are logically considered "unavailable" through hard threshold truncation. Based on this, the availability status of each channel is encoded as a binary topological fingerprint. This invention achieves a mapping from complex physical faults, network packet loss, and privacy breaches to a computable discrete state space. This key point provides a unified entry point and index base for subsequent dynamic routing, parameter borrowing, and risk compensation. Secondly, addressing the missing pattern explosion and long-tail distribution caused by multi-channel combinations, this invention constructs a semantic topology graph using topological fingerprints as nodes. A borrowable statistical similarity structure is established based on adjacency relationships with Hamming distances not exceeding 1. During inference or calibration, if there are insufficient samples of the target missing pattern, a BFS search is triggered to find it. "Complete nodes" within the order neighborhood are synthesized using inverse distance weighting (IDW) to obtain statistical parameters (such as...). , This mechanism enables rapid startup. It avoids the problems of unpredictable parameters and calibration failures in traditional static binning under scarce modes, structurally improving the system's usable coverage for any missing mode. Furthermore, Mahalanobis distance is used to characterize the relative deviation of samples from class centroids, and Ledoit-Wolf contraction estimation is introduced to construct... This ensures the covariance matrix is ​​invertible and numerically stable under small sample conditions. Furthermore, HNSW is used to accelerate nearest neighbor retrieval and calculate the local density factor. The distance score is then modified in a second-order manner as a manifold penalty term, resulting in a non-uniform score that combines "inter-class discriminativeness" with "manifold edge / noise sensitivity". This key point effectively alleviates the curse of dimensionality problem of Euclidean distance and significantly enhances the sensitivity to identifying out-of-distribution samples and boundary-hard samples. Furthermore, under conditions of different missing patterns and varying feature distribution ambiguity, a uniform threshold will inevitably lead to an imbalance of "high false positives when missing data is severe and overly conservative when data is complete." This invention discretizes the non-consistent score distribution within nodes and calculates the Shannon entropy. H(S) This is used to characterize the overall uncertainty under the current missing patterns, and it is embedded as a risk gain coefficient into the sample weight exponential function to dynamically amplify the weight of tail-difficult samples, making the threshold calculation automatically conservative under high entropy conditions. This key point enables the system to complete confidence calibration without manual annotation and replaces empirical thresholds with statistically significant coverage constraints, significantly improving reliability in safety-critical scenarios. Then, the global false alarm rate is used to control parameters. To constrain this, a dynamic safety threshold is obtained using weighted quantiles. And generate a prediction set. As an output, unlike single-point classification, this invention provides a multi-level output protocol: when the set is a single class, it enters high-precision mode; when the set is multi-class, it outputs a candidate list to trigger manual review or downgraded operation; when the set is empty or contains all classes, it triggers an abnormal circuit breaker prompting "insufficient effective information." This key point achieves a controllable switch between "accuracy priority" and "safety priority," significantly reducing the risk of missed and false positives without sacrificing system availability. Finally, to address the slow drift in data distribution caused by sensor aging and scene changes, this invention maintains a fixed-capacity FIFO sample cache and uses exponential moving average (EMA) to incrementally update key statistics such as class centroids after obtaining feedback labels, thereby continuously keeping pace with the latest distribution without frequent full retraining. This key point ensures that the system has engineering-feasible online adaptability, improving the robustness of long-term deployment and maintenance efficiency.

[0076] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0077] In one embodiment, such as Figure 2 As shown, a data processing device for incomplete multi-source information is provided, including: a topological fingerprint generation module 202, a statistical parameter library determination module 204, a score calculation module 206, a security threshold setting module 208, a decision module 210, and a self-evolution module 212, wherein: The topology fingerprint generation module 202 is used to perform quality assessment on multi-source data channels and generate topology fingerprints that characterize the availability status of each data channel.

[0078] The statistical parameter library determination module 204 is used to perform dynamic routing in the pre-constructed semantic topology graph based on the topological fingerprint, and determine the target node and the statistical parameter library of the target node corresponding to the current input data.

[0079] The score calculation module 206 is used to calculate the class conditional inconsistency score of the current sample in the feature space based on the statistical parameter library of the target node.

[0080] The safety threshold setting module 208 is used to adaptively risk-weight the samples based on the information entropy of the non-consistent score distribution of the calibration samples within the target node, and determine the dynamic safety threshold based on the weighted quantile.

[0081] The decision module 210 is used to compare the inconsistency scores of the current sample under each candidate category with the dynamic safety threshold, and generate a prediction set containing the categories that meet the threshold conditions as output.

[0082] The self-evolution module 212 is used to perform online self-evolution of the statistical parameter library based on the sliding window caching and incremental update mechanism.

[0083] For specific limitations regarding a data processing device for incomplete multi-source information, please refer to the limitations of a data processing method for incomplete multi-source information described above, which will not be repeated here. Each module in the aforementioned data processing device for incomplete multi-source information can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0084] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0085] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchlink, DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data processing method for incomplete multi-source information, characterized in that, The method includes: A quality assessment is performed on the multi-source data channels to generate a topological fingerprint characterizing the availability status of each data channel; Based on the topological fingerprint, dynamic routing is performed in the pre-constructed semantic topology graph to determine the target node corresponding to the current input data and the statistical parameter library of the target node; Based on the statistical parameter library of the target node, calculate the class conditional inconsistency score of the current sample in the feature space; Based on the information entropy of the non-consistent score distribution of the calibration samples within the target node, the samples are adaptively risk-weighted, and a dynamic safety threshold is determined based on the weighted quantiles. The non-consistency scores of the current sample under each candidate category are compared with the dynamic safety threshold, and a prediction set containing the categories that meet the threshold conditions is generated as the output. The statistical parameter library is self-evolved online based on the sliding window caching and incremental update mechanism.

2. The method according to claim 1, characterized in that, A quality assessment is performed on the multi-source data channels to generate a topological fingerprint characterizing the availability status of each data channel, including: The signal-to-noise ratio (SNR) of the input signals of each data channel is detected, and the data channels with an SNR lower than a preset threshold are marked as unusable at the logic level. Generate a binary topological fingerprint based on the availability status of the marked data channels: in, For the first The sample at the th A binary topological fingerprint of the availability status of each channel. For channel indexing, To preset the industrial availability threshold, For the first The sample at the th The signal-to-noise ratio or confidence score on each channel.

3. The method according to claim 1, characterized in that, Based on the topological fingerprint, dynamic routing is performed in the pre-constructed semantic topology graph to determine the target node corresponding to the current input data and the statistical parameter library of the target node, including: Using the topological fingerprint as nodes, construct a semantic topological graph with edges where the Hamming distance does not exceed a threshold: in, For semantic topology graph, For topological fingerprint set, As an edge, For nodes that have appeared in the past, topological fingerprint patterns For statistical parameter library, For nodes Inner Class feature centroid, For nodes Internal shared covariance statistics For nodes Statistics on the set of inconsistencies in the internal calibration samples and their corresponding histograms. Indexed by category; Determine if a complete node exists for the topological fingerprint of the current sample. If so, route directly to the current node; otherwise, search the set of complete nodes in the neighborhood of the current node, and synthesize the statistical parameters of the current node using inverse distance weighting. in, These are the statistical parameters for the current node. For a complete set of nodes, Let Hamming distance function be used. For the target topological fingerprint, These are the statistical parameters corresponding to the topological fingerprint of the current sample. The topological fingerprint of the current sample. This is a smoothing factor.

4. The method according to claim 3, characterized in that, Based on the statistical parameter library of the target node, calculate the class conditional inconsistency score of the current sample in the feature space, including: Based on the statistical parameter library of the target node, within the target node, a graph index-accelerated nearest neighbor search algorithm is used to retrieve a preset number of nearest neighbor samples for the current sample, and the local sparsity factor is calculated: in, For local sparsity factor, The preset number of neighboring samples. For the current sample The joint eigenvectors, For the first The feature vectors of the neighboring samples As a smoothing factor, The distance is Euclidean. The Ledoit–Wolf shrinkage algorithm is used to estimate the covariance matrix within the target node, resulting in the corrected covariance matrix: in, To correct the covariance matrix, For the target node, The shrinkage coefficient, For feature dimension, For in-node sample covariance estimation, It is the identity matrix; Calculate the Mahalanobis distance from the current sample to each category based on the modified covariance matrix: in, For the current sample To Category Mahalanobis distance, To the target node Internal Category Characteristic centroid, For category indexing, To correct the inverse of the covariance matrix; Based on the local sparsity factor and the Mahalanobis distance, generate class conditional inconsistency scores: in, For the current sample Category Class conditional non-consistency scores, For the current sample For the most recent competing categories Mahalanobis distance, As a smoothing factor, For manifold sensitivity hyperparameters, For local sparsity factor, It is the hyperbolic tangent function.

5. The method according to claim 4, characterized in that, Based on the information entropy of the inconsistent score distribution of calibration samples within the target node, the samples are adaptively risk-weighted, and a dynamic safety threshold is determined based on the weighted quantiles, including: After discretizing the set of inconsistent scores of calibration samples within the target node into a histogram, the Shannon entropy is calculated: in, For the target node Shannon entropy of an internally inconsistent fractional distribution For the target node The set of inconsistent scores of the internal calibration samples For histogram interval indexing, For non-consistent scores falling in the th The sample probability of each interval This represents the total number of intervals into which the histogram is divided. Based on the Shannon entropy and the normalized ranking of each score in the set of inconsistent scores, an exponential risk weight is set: in, For unnormalized risk weights, Basic risk coefficient, This is the entropy gain coefficient. For distribution entropy, For the target node The set of inconsistent scores of the internal calibration samples This represents the normalized ranking of inconsistent scores within a node. To normalize risk weights, This is the index of all calibration samples within the node; Dynamic security thresholds are determined based on weighted quantiles.

6. The method according to claim 5, characterized in that, Determining dynamic security thresholds based on weighted quantiles includes: Based on the weighted quantiles, the dynamic safety threshold that makes the cumulative sample weights reach the preset coverage requirement is calculated: in, For dynamic security thresholds, For the target node, For the indeterminate boundary, For candidate threshold variables, For the set of real numbers, To normalize risk weights, To calibrate the inconsistency scores of the samples, This is the preset global false alarm rate control parameter.

7. The method according to claim 6, characterized in that, The inconsistency scores of the current sample in each candidate category are compared with the dynamic safety threshold, and a prediction set containing the categories that meet the threshold conditions is generated as output, including: For the current test sample, calculate the class conditional inconsistency score under each candidate category, and select the categories that satisfy the conditional inconsistency score not greater than the dynamic security threshold to form a prediction set as the output.

8. The method according to claim 7, characterized in that, Based on the sliding window caching and incremental update mechanism, the statistical parameter library undergoes online self-evolution, including: A fixed-capacity first-in-first-out (FIFO) buffer queue is maintained for each target node to store recent samples and their feedback labels. Based on the feedback labels, the feature centroids of the categories within the target node are iteratively updated using the exponential moving average algorithm: in, Category within the target node The updated feature centroid at the current moment. The momentum coefficient, Category within the target node The characteristic centroid at the previous moment, This is the feature mean vector of the latest batch of labeled samples. For time steps; The covariance matrix and histogram of the inconsistency score set within the target node are periodically re-estimated, and the nearest neighbor index is updated to complete the online self-evolution of the statistical parameter library.

9. A data processing device for incomplete multi-source information, characterized in that, The device includes: The topology fingerprint generation module is used to evaluate the quality of multi-source data channels and generate topology fingerprints that characterize the availability status of each data channel. The statistical parameter library determination module is used to perform dynamic routing in a pre-constructed semantic topology graph based on the topological fingerprint, and determine the target node corresponding to the current input data and the statistical parameter library of the target node; The score calculation module is used to calculate the class conditional inconsistency score of the current sample in the feature space based on the statistical parameter library of the target node. The safety threshold setting module is used to adaptively risk-weight the samples based on the information entropy of the non-consistent score distribution of the calibration samples within the target node, and determine the dynamic safety threshold based on the weighted quantiles. The decision module is used to compare the inconsistency score of the current sample under each candidate category with the dynamic safety threshold, and generate a prediction set containing the categories that meet the threshold conditions as output; The self-evolution module is used to perform online self-evolution of the statistical parameter library based on the sliding window caching and incremental update mechanism.

Images