An intelligent screening method based on multi-source data

Abstract

The application relates to the technical field of information retrieval, in particular to an intelligent screening method based on multi-source data, which comprises the following steps: acquiring high-frequency time sequence data of a first data source, dividing the high-frequency time sequence data into state change intervals and state stable intervals; extracting dynamic features from the state change intervals and extracting statistical features from the state stable intervals to generate a mode feature cluster sequence; acquiring low-frequency data of a second data source, taking a time stamp as an anchor point, combining the mode feature cluster sequence to acquire a precursor node and a successor node of the mode feature cluster, and generating an anchor point association graph; performing bidirectional prediction verification based on the anchor point association graph, calculating a bidirectional consistency score, and generating an alignment confidence label; monitoring the incremental update of the high-frequency time sequence data, re-performing reverse prediction verification, updating the alignment confidence label, screening anchor point association records with high confidence, fusing low-frequency data features and mode feature clusters, and outputting a screening result.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, specifically to an intelligent screening method based on multi-source data. Background Technology

[0002] With the rapid development of information technology, intelligent screening technology based on multi-source data has been widely applied in many fields. Intelligent screening systems collect, process, and analyze multi-dimensional feature data from different data sources to achieve automated assessment and identification of the status of target objects, playing a significant role in improving screening efficiency and reducing labor costs.

[0003] Existing intelligent screening systems typically employ a unidirectional processing architecture of "data acquisition—feature extraction—model inference—result output" in their data processing flow. Under this architecture, after the screening module completes the analysis of the input data, it provides the screening results or risk level as the final output to the user or downstream system, thus ending the screening process.

[0004] However, the aforementioned one-way processing architecture has significant technical shortcomings. Firstly, there is a lack of effective data interaction interfaces between the screening module and the downstream decision support module. Screening results cannot directly drive downstream modules to generate targeted decision support information, requiring manual intervention from users to transform screening results into specific decision solutions, thus reducing the system's automation level and response efficiency. Secondly, the existing system lacks a feedback path from downstream modules to the screening module. The decision execution effect cannot be fed back to the screening module for model optimization, making it difficult for the system to adaptively adjust based on actual application results, thus limiting the continuous iteration capability of the screening model.

[0005] Furthermore, in multi-source data scenarios, the update frequency and timeliness of data from different data sources vary. Existing systems lack a dynamic perception mechanism for the update status of multi-source data, making it difficult to trigger the re-execution of the screening process in a timely manner when data changes, which affects the timeliness and accuracy of screening results.

[0006] Therefore, there is a need for an intelligent screening method that can enable bidirectional linkage between the screening module and the decision support module and support feedback optimization, in order to solve the technical problems of the disconnect between screening results and downstream decision support and the lack of closed-loop optimization capability in the existing technology.

[0007] Therefore, an intelligent screening method based on multi-source data is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent screening method based on multi-source data. This method constructs a pattern feature cluster sequence from high-frequency time-series data and establishes a correlation map between different data sources using low-frequency data timestamps as anchors. It combines bidirectional prediction consistency verification and incremental update mechanisms to integrate high- and low-frequency features and output high-confidence screening results.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] An intelligent screening method based on multi-source data includes:

[0011] The system acquires high-frequency time-series data from the first data source, calculates the difference between adjacent data points, and divides the high-frequency time-series data into state-change intervals and state-stationary intervals. Dynamic features are extracted from the state-change intervals, and statistical features are extracted from the state-stationary intervals to generate a pattern feature cluster sequence.

[0012] Low-frequency data from a second data source is acquired, and the timestamps are used as anchor points to determine the forward background window and backward influence window of the anchor points. The predecessor and successor nodes of the pattern feature clusters are obtained by combining the pattern feature cluster sequence, and the anchor point association map is generated.

[0013] Bidirectional prediction verification is performed based on the anchor point association graph: the positive prediction bias is obtained by predicting the low-frequency data features of the anchor point using the feature sequence of the predecessor node, and the negative prediction bias is obtained by predicting the pattern type of the successor node based on the low-frequency data features; the bidirectional consistency score is calculated and the alignment confidence label is generated.

[0014] Monitor the incremental updates of high-frequency time series data. When the incremental data falls into the backward influence window of an existing anchor point, process the pattern feature cluster sequence and update the successor node set, re-predict and verify, update the alignment confidence label, filter the anchor point association records with high confidence, perform fusion processing on low-frequency data features and pattern feature clusters, and output the screening results.

[0015] Preferably, generating a pattern feature cluster sequence includes: traversing high-frequency time-series data in chronological order, calculating the difference value between adjacent data points, comparing the difference value with a preset change threshold; marking the state change start point when five consecutive difference values ​​exceed the change threshold, marking the state change end point when five consecutive difference values ​​are lower than the change threshold, marking the data interval between the start point and the end point as the state change interval, and marking the remaining data interval as the state stable interval.

[0016] The state change interval is divided by sliding segmentation using the first window size. Within each sub-window, the window start value, window end value, extreme point position within the window, change direction identifier, and average change rate are extracted as dynamic features.

[0017] For the stable interval, a second window size larger than the first window size is used for sliding segmentation, and the window mean, window variance and window duration are extracted as statistical features in each sub-window.

[0018] Dynamic features and statistical features are assembled into pattern feature clusters. Each pattern feature cluster contains a cluster identifier, a time start point, a time end point, an interval type identifier, and a feature vector. All pattern feature clusters are arranged in chronological order to form a pattern feature cluster sequence.

[0019] Preferably, determining the foreground window and the back influence window of the anchor point includes:

[0020] Extract the acquisition timestamp of each data record from the low-frequency data, and use the acquisition timestamp as the time reference of the anchor point; determine the time length parameters of the forward background window and the backward influence window based on the acquisition cycle of the low-frequency data and the time coverage of the high-frequency time series data; extend the time length of the forward background window in the negative direction of the time axis based on the acquisition timestamp of the anchor point, and determine the start time boundary and end time boundary of the forward background window; extend the time length of the backward influence window in the positive direction of the time axis, and determine the start time boundary and end time boundary of the backward influence window.

[0021] Preferably, generating an anchor point association graph includes: creating an anchor point record for each anchor point, the anchor point record containing an anchor point identifier, a collection timestamp, and the original feature vector of low-frequency data; retrieving pattern feature clusters whose time range falls within the forward background window from the pattern feature cluster sequence according to the time boundary of the anchor point's forward background window, as the set of predecessor nodes for the anchor point; retrieving pattern feature clusters whose time range falls within the backward influence window from the pattern feature cluster sequence according to the time boundary of the anchor point's backward influence window, as the set of successor nodes for the anchor point; using the anchor point as the central node of the graph, establishing directed edges of historical background type between the central node and each predecessor node, and establishing directed edges of subsequent response type between the central node and each successor node, the attributes of each directed edge including the time interval value between the node and the anchor point; associating and storing the original feature vector of the anchor point with the feature vector sequences of the predecessor nodes and the feature vector sequences of the successor nodes to form an anchor point association graph.

[0022] Preferably, the positive prediction bias acquisition process includes: acquiring the feature vectors of all predecessor nodes within the forward background window of the anchor point, arranging them in chronological order to form a predecessor feature vector sequence; inputting the predecessor feature vector sequence into a sequence encoding module to encode the variable-length feature vector sequence, generating a fixed-dimensional background representation vector; inputting the background representation vector into a feature mapping module to map the background representation vector to the same dimensional space as the low-frequency data features, outputting a predicted value vector for the low-frequency data features of the anchor point; comparing the predicted value vector with the actual low-frequency data feature vector of the anchor point, calculating the vector distance between the predicted value vector and the actual feature vector, and using the vector distance as the positive prediction bias.

[0023] Preferably, the reverse prediction bias acquisition process includes: acquiring the original feature vector of low-frequency data and the forward background representation vector of the anchor point, and concatenating the two to form a joint input vector; dividing the backward influence window of the anchor point into multiple time slots according to a preset time interval, each time slot corresponding to a pattern type label to be predicted, the pattern types including rising change type, falling change type, fluctuating change type and stationary type; inputting the joint input vector into the pattern type prediction module, and outputting the predicted probability distribution of the pattern type that should appear in each time slot; acquiring the interval type identifier of the actual pattern feature cluster corresponding to each time slot in the backward influence window as the actual pattern type distribution; comparing the predicted probability distribution with the actual pattern type distribution, calculating the distribution difference metric between the predicted distribution and the actual distribution, and using the distribution difference metric as the reverse prediction bias.

[0024] Preferably, the calculation of the bidirectional consistency score and the generation of alignment confidence labels include: normalizing the positive prediction bias to obtain a positive normalized bias value; normalizing the negative prediction bias to obtain a negative normalized bias value; and calculating the bidirectional consistency score based on the positive and negative normalized bias values.

[0025] Set a high confidence threshold and a low confidence threshold. Anchor point association records with bidirectional consistency scores higher than the high confidence threshold are marked as high confidence level, those with scores between the high and low confidence thresholds are marked as medium confidence level, and those with scores lower than the low confidence threshold are marked as low confidence level. The confidence level is then attached as an alignment confidence label to the corresponding anchor point association record.

[0026] Preferably, the screening result acquisition step includes: screening anchor point association records with a high confidence level alignment confidence label from all anchor point association records to form a set of records to be merged;

[0027] For each anchor-related record in the record set to be fused, extract the original low-frequency data feature vector of the anchor, the pattern feature cluster feature vector sequence of the predecessor node, and the pattern feature cluster feature vector sequence of the successor node;

[0028] The original feature vector of low-frequency data is used as the static feature component, and the feature vector sequences of predecessor and successor nodes are concatenated in time order to be used as the dynamic feature component.

[0029] The static and dynamic feature components are jointly processed at the feature level to generate a fused feature representation vector.

[0030] The fused feature representation vector is input into the screening and classification module, which outputs the screening conclusion category and the corresponding confidence value. The screening conclusion category and the confidence value are then organized into a screening result data structure and output.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] 1. This invention divides high-frequency time-series data into state-change intervals and state-stationary intervals, extracts dynamic and statistical features respectively, constructs a pattern feature cluster sequence, and then uses low-frequency data timestamps as anchor points, introducing forward background windows and backward influence windows to establish an anchor point correlation map. This achieves refined time alignment between data with different sampling frequencies and different semantic levels. This invention can characterize the complete state evolution background and response process before and after the occurrence of low-frequency events, so that low-frequency data is no longer an isolated point, but embedded in a continuous high-frequency pattern structure. This significantly reduces screening misjudgments caused by mismatches and random correlations, and improves the credibility and interpretability of screening results at the time-series level.

[0033] 2. This invention introduces a bidirectional prediction verification mechanism combining positive and negative prediction bias: on the one hand, it uses the characteristics of the precursor patterns of the anchor point to predict low-frequency data features and evaluates the explanatory power of historical behavior on the anchor point results; on the other hand, it uses the low-frequency data features to predict the distribution of pattern types after the anchor point and evaluates the constraint ability of the anchor point on the subsequent state evolution. By normalizing the two types of prediction biases and jointly calculating the bidirectional consistency score, a high-confidence label is assigned only when the history, result, and subsequent events are semantically and statistically self-consistent. This mechanism can effectively filter out spurious associations that are accidental co-occurrences, noise-driven associations, or well-fitted one-sided associations, significantly improving the stability and reliability of the screening conclusions.

[0034] 3. This invention addresses the practical application scenario of continuously growing high-frequency time-series data by employing a dynamic re-verification mechanism based on incremental updates. When new high-frequency data falls into the backward influence window of existing anchor points, instead of recalculating the entire dataset, it performs local updates on the affected pattern feature clusters and successor node sets, and re-executes reverse prediction verification and consistency scoring, thereby correcting the anchor point association relationships and their confidence labels in real time. This invention ensures that the screening results adaptively correct themselves over time while significantly reducing computational overhead, avoiding screening distortion caused by model obsolescence, and enabling the screening results to maintain long-term consistency with the real data distribution and operating status, making it more suitable for engineering and online deployment applications. Attached Figure Description

[0035] Figure 1 A flowchart of an intelligent screening method based on multi-source data provided by the present invention;

[0036] Figure 2 This is a flowchart of the anchor point association map construction process provided by the present invention;

[0037] Figure 3 The flowchart for obtaining screening results provided by this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0039] Example 1:

[0040] Please see Figure 1This invention provides an intelligent screening method based on multi-source data, the technical solution of which is as follows: High-frequency time-series data from a first data source is acquired, the difference between adjacent data points is calculated, and the high-frequency time-series data is divided into state change intervals and state stability intervals; dynamic features are extracted from the state change intervals, and statistical features are extracted from the state stability intervals to generate a pattern feature cluster sequence; low-frequency data from a second data source is acquired, with the collection timestamp as the anchor point, and the forward background window and backward influence window of the anchor point are determined; the predecessor and successor nodes of the pattern feature clusters are obtained by combining the pattern feature cluster sequence, and an anchor point association graph is generated; based on the anchor point association graph... The spectrum performs bidirectional prediction verification: positive prediction bias is obtained by predicting the low-frequency data features of anchor points using the feature sequences of predecessor nodes, and negative prediction bias is obtained by predicting the pattern type of successor nodes based on the low-frequency data features; bidirectional consistency score is calculated and alignment confidence labels are generated; incremental updates of high-frequency time series data are monitored, and when incremental data falls into the backward influence window of existing anchor points, the pattern feature cluster sequence is processed and the successor node set is updated, reverse prediction verification is performed again, alignment confidence labels are updated, high-confidence anchor point association records are filtered, low-frequency data features and pattern feature clusters are fused, and the screening results are output.

[0041] Preferably, generating a pattern feature cluster sequence includes: traversing high-frequency time-series data in chronological order, calculating the difference value between adjacent data points, comparing the difference value with a preset change threshold; marking the state change start point when five consecutive difference values ​​exceed the change threshold, marking the state change end point when five consecutive difference values ​​are lower than the change threshold, marking the data interval between the start point and the end point as the state change interval, and marking the remaining data interval as the state stable interval.

[0042] The state change interval is divided by sliding segmentation using the first window size. Within each sub-window, the window start value, window end value, extreme point position within the window, change direction identifier, and average change rate are extracted as dynamic features.

[0043] For the stable interval, a second window size larger than the first window size is used for sliding segmentation, and the window mean, window variance and window duration are extracted as statistical features in each sub-window.

[0044] Dynamic features and statistical features are assembled into pattern feature clusters. Each pattern feature cluster contains a cluster identifier, a time start point, a time end point, an interval type identifier, and a feature vector. All pattern feature clusters are arranged in chronological order to form a pattern feature cluster sequence.

[0045] Specifically, the process involves: acquiring high-frequency time-series data continuously collected from a first data source, which is a numerical sequence formed by continuous sampling at fixed time intervals, with the sampling interval being on the order of seconds or minutes; the first data source is a smart terminal device platform, and the high-frequency time-series data is multi-dimensional feature index data continuously collected by the platform at the order of minutes or seconds, including a sequence of continuously collected physiological feature parameters.

[0046] The high-frequency time series data is traversed in chronological order, and the difference between adjacent data points is calculated. For the i-th data point and the (i+1)-th data point in the time series, the difference between the latter and the former is calculated as the i-th difference value.

[0047] Each difference value is compared with a preset change threshold, and a sliding counting window mechanism is used for continuity determination. A continuous determination window length parameter is set. When all difference values ​​within the sliding window exceed the change threshold, the current position is marked as the state change start point; when all difference values ​​within the sliding window are below the change threshold, the current position is marked as the state change end point. The continuous determination window length can be configured according to the sampling frequency and data characteristics of the high-frequency time series data. The data interval between the state change start point and the state change end point is marked as the state change interval, and the remaining data intervals are marked as the state stability interval. The change threshold is determined through statistical analysis of historical samples of the high-frequency time series data. The mean and standard deviation of the difference values ​​of adjacent data points in the historical samples are calculated, and the sum of the mean and a preset multiple of the standard deviation is used as the change threshold.

[0048] For the state change interval, a sliding segmentation is performed using the first window size. Within each sub-window, the following dynamic features are extracted: window start value (the value of the first data point in the sub-window); window end value (the value of the last data point in the sub-window); extreme point position within the window (the relative position index of the data point with the largest or smallest value in the sub-window); change direction indicator, determined based on the relationship between the window end value and the window start value. If the window end value is greater than the window start value, it is marked as rising; if the window end value is less than the window start value, it is marked as falling; if the difference between the two is less than a preset fluctuation judgment threshold but the range within the window is greater than the threshold, it is marked as fluctuating; and the mean rate of change, calculated by taking the arithmetic mean of the absolute values ​​of the differences between adjacent data points within the window.

[0049] For the stable interval, a second window size is used for sliding segmentation, which is larger than the first window size. The following statistical features are extracted in each sub-window: window mean, which is the arithmetic mean of all data points in the sub-window; window variance, which is the arithmetic mean of the squares of the differences between each data point in the sub-window and the window mean; and window duration, which is the time span covered by the sub-window, equal to the number of data points in the sub-window multiplied by the sampling interval.

[0050] The features extracted from each sub-window are assembled into pattern feature clusters. The data structure of each pattern feature cluster includes the following fields: cluster identifier, which is a unique number assigned by the system to distinguish different pattern feature clusters; time start point, which is the start timestamp of the original data interval corresponding to the pattern feature cluster; time end point, which is the end timestamp of the original data interval corresponding to the pattern feature cluster; interval type identifier, which takes the value of variable type or stationary type; feature vector, which is a numerical vector composed of dynamic features or statistical features extracted in the sub-window arranged in a fixed order.

[0051] All pattern feature clusters are arranged in chronological order to form a pattern feature cluster sequence. If the length of the pattern feature cluster sequence exceeds the preset maximum length threshold, the stationary pattern feature clusters are merged and compressed. Specifically, adjacent stationary pattern feature clusters are merged into one, and the feature vector of the merged cluster is the average of the feature vectors of each merged cluster. Variational pattern feature clusters are retained first. If the sequence length is less than the minimum length threshold, the longer stationary intervals are subdivided and decomposed, and finally a fixed-length pattern feature cluster sequence is output.

[0052] Preferably, determining the foreground window and the back influence window of the anchor point includes:

[0053] Extract the acquisition timestamp of each data record from the low-frequency data, and use the acquisition timestamp as the time reference for the anchor point; determine the time length parameters of the forward background window and the backward influence window based on the acquisition cycle of the low-frequency data and the time coverage of the high-frequency time series data; extend the time length of the forward background window in the negative direction of the time axis based on the acquisition timestamp of the anchor point, and determine the start time boundary and end time boundary of the forward background window. The forward background window is used to associate the historical evolution pattern of high-frequency data in the period before the anchor point; extend the time length of the backward influence window in the positive direction of the time axis, and determine the start time boundary and end time boundary of the backward influence window. The backward influence window is used to associate the subsequent response pattern of high-frequency data in the period after the anchor point.

[0054] Preferably, an anchor point association map is generated, referring to... Figure 2 ,include:

[0055] An anchor record is created for each anchor point, and the anchor record contains the anchor point identifier, the collection timestamp, and the original feature vector of the low-frequency data;

[0056] Based on the time boundary of the forward background window of the anchor point, retrieve the pattern feature clusters whose time range falls within the forward background window from the pattern feature cluster sequence, and use them as the set of predecessor nodes of the anchor point.

[0057] Based on the time boundary of the backward influence window of the anchor point, retrieve the pattern feature clusters whose time range falls within the backward influence window from the pattern feature cluster sequence, and use them as the set of successor nodes of the anchor point.

[0058] The anchor point is used as the central node of the graph. Directed edges of the historical background type are established between the central node and each predecessor node, and directed edges of the subsequent response type are established between the central node and each successor node. The attributes of each directed edge include the time interval value between the node and the anchor point.

[0059] The original feature vector of the anchor point is associated with and stored in relation to the feature vector sequences of the predecessor node and the feature vector sequences of the successor node, forming an anchor point association graph data structure.

[0060] Specifically, low-frequency data from the second data source is acquired. In this embodiment, the second data source is a data management organization, and the low-frequency data is structured static data collected by the organization on an annual or quarterly basis, including periodically collected multi-dimensional feature indicators. In typical application scenarios of multi-source heterogeneous data fusion, the first data source can be a smart wearable device data collection platform, and the second data source can be a periodically collected data organization. The data from the two types of data sources correspond to high-frequency continuous data collection and low-frequency periodic data collection, respectively.

[0061] Extract the collection timestamp of each data record from the low-frequency data, and use this timestamp as the time base for the anchor point. Create an anchor point record for each anchor point. The data structure of the anchor point record contains the following three fields:

[0062] Anchor identifier: A string or numeric code that uniquely identifies the anchor point;

[0063] Data collection timestamp: The time when low-frequency data was collected;

[0064] Original feature vector of low-frequency data: A numerical array consisting of all feature values ​​of this low-frequency data record.

[0065] Based on the acquisition cycle of low-frequency data and the time coverage of high-frequency time-series data, the time length parameters of the forward background window and the backward influence window are determined. The time lengths of the forward background window and the backward influence window can be configured according to the specific application scenario.

[0066] Using the anchor point's acquisition timestamp as a reference, the forward background window is extended negatively along the time axis to determine its start and end time boundaries. The end time boundary of the forward background window is the anchor point's acquisition timestamp, and the start time boundary is the acquisition timestamp minus the forward background window's time length. The forward background window is used to correlate historical evolution patterns of high-frequency data within the time period preceding the anchor point.

[0067] Using the anchor point's acquisition timestamp as a reference, the backward influence window is extended in the positive direction of the time axis to determine its start and end time boundaries. The start time boundary of the backward influence window is the anchor point's acquisition timestamp, and the end time boundary is the acquisition timestamp plus the backward influence window's duration. The backward influence window is used to correlate subsequent response patterns of high-frequency data within the time period following the anchor point.

[0068] Based on the time boundary of the forward background window of the anchor point, retrieve the pattern feature clusters whose time range falls within the forward background window from the pattern feature cluster sequence. Specifically, traverse each pattern feature cluster in the pattern feature cluster sequence, and if the start and end times of the cluster are both located between the start and end time boundaries of the forward background window, then include the cluster in the predecessor node set.

[0069] Based on the time boundary of the backward influence window of the anchor point, retrieve pattern feature clusters whose time range falls within the backward influence window from the pattern feature cluster sequence. Specifically, traverse each pattern feature cluster in the pattern feature cluster sequence, and if the start and end times of the cluster are both located between the start and end time boundaries of the backward influence window, then include the cluster in the successor node set.

[0070] Using the anchor point as the central node of the graph, directed edges of historical background type are established between the central node and each predecessor node, with the edges pointing from the predecessor node to the central node. Similarly, directed edges of subsequent response type are established between the central node and each successor node, with the edges pointing from the central node to the successor node. The attributes of each directed edge include the time interval between the node and the anchor point, calculated as the absolute value of the difference between the node's midpoint time and the anchor point's data collection timestamp.

[0071] The original feature vector of the anchor point is associated with and stored in conjunction with the feature vector sequences of its predecessor and successor nodes, forming an anchor point association graph data structure. The anchor point association graph can be stored using an adjacency list structure, with the anchor point identifier as an index, storing the set of its associated predecessor and successor nodes and their corresponding edge attributes.

[0072] Preferably, the positive prediction bias is obtained by using the feature sequence of the predecessor node to predict the low-frequency data features of the anchor point, including:

[0073] Obtain the feature vectors of all predecessor nodes within the forward background window of the anchor point, and arrange them in chronological order to form a sequence of predecessor feature vectors.

[0074] The preceding feature vector sequence is input into the sequence encoding module to encode the variable-length feature vector sequence and generate a fixed-dimensional background representation vector.

[0075] The background representation vector is input into the feature mapping module, which maps the background representation vector to the same dimensional space as the low-frequency data features, and outputs the predicted value vector of the low-frequency data features of the anchor point.

[0076] The predicted value vector is compared with the actual low-frequency data feature vector of the anchor point, and the vector distance between the predicted value vector and the actual feature vector is calculated. The vector distance is used as the positive prediction bias.

[0077] The sequence encoding module is implemented using the following scheme: the preceding feature vector sequence is input into a two-layer fully connected network. The number of neurons in the first layer is 1.5 times the input dimension, and the ReLU activation function is used. The number of neurons in the second layer is the same as the dimension of the input feature vector, and no activation function is used. During the training phase, Dropout regularization is applied to the output of the first layer with a dropout rate of 0.2 to prevent overfitting. The design goal of this encoding structure is to achieve lossless conversion of variable-length sequences to fixed-dimensional vectors while maintaining feature expressive power.

[0078] The feature mapping module employs a three-layer fully connected network structure: the first layer's input dimension equals the dimension of the background representation vector, and its output dimension is twice that dimension, using the ReLU activation function; the second layer's input dimension is the output dimension of the previous layer, and its output dimension is 1.5 times that value, also using the ReLU activation function; the third layer's input dimension is the output dimension of the previous layer, and its output dimension is the dimension of the original feature vector of the low-frequency data, without using an activation function. Dropout regularization is applied to the first and second layers, with a dropout rate of 0.2 in both layers. This multi-layer structure can learn the non-linear mapping relationship from background information to low-frequency features, improving prediction accuracy.

[0079] During model training, the Adam optimizer was used with an initial learning rate of 0.001. In each training epoch, if the validation set loss function value did not improve within three epochs, the learning rate was multiplied by 0.8 to decay. The batch size was set to 32. For scenarios with fewer than 1000 training samples, L2 regularization was applied with a regularization coefficient of 0.001. A dynamic stopping strategy was used for training epochs; training stopped when the average prediction bias on the validation set did not improve for five consecutive epochs.

[0080] Preferably, the reverse prediction bias obtained by predicting the pattern type of the successor node based on low-frequency data features includes:

[0081] Obtain the original feature vector and forward background representation vector of the low-frequency data of the anchor point, and concatenate the two to form a joint input vector;

[0082] The backward influence window of the anchor point is divided into multiple time slots according to a preset time interval. Each time slot corresponds to a pattern type label to be predicted. The pattern types include rising change type, falling change type, fluctuating change type and stationary type.

[0083] The joint input vector is input into the pattern type prediction module, which outputs the predicted probability distribution of the pattern type that should appear in each time slot.

[0084] Obtain the interval type identifier of the actual mode feature cluster corresponding to each time slot within the backward influence window, as the actual mode type distribution;

[0085] The predicted probability distribution is compared with the actual pattern type distribution, and the distribution difference measure between the predicted distribution and the actual distribution is calculated. The distribution difference measure is used as the reverse prediction bias.

[0086] Specifically, the pattern type prediction module adopts the following architecture: The joint input vector is first fed into a three-layer fully connected network for feature transformation. The number of neurons in each layer is 1.5 times, 1.2 times, and 1 times the input dimension, respectively. The first two layers use the ReLU activation function, and the last two layers apply Dropout regularization (dropout rate of 0.2). The output vector after feature transformation is then fed into a time series decoder, which uses a single-layer unidirectional LSTM structure with 1.5 times the feature vector dimension of the hidden units. The output of the LSTM passes through a fully connected classification layer at each time step. This layer has 4 output neurons (corresponding to 4 pattern types) and uses the Softmax activation function to generate a four-dimensional probability distribution vector for that time step.

[0087] The output dimension of the pattern type prediction module is "number of time slots × 4", where each row is a four-dimensional probability distribution vector corresponding to a time slot. This design enables the module to predict the pattern type of each time slot sequentially through the hidden state of the LSTM, capturing the temporal dependency between adjacent time slots (e.g., a stationary pattern is more likely to follow a decreasing pattern than an increasing pattern).

[0088] The distribution dissimilarity metric is calculated using the weighted cross-entropy approach: for each time slot, the cross-entropy loss between the one-hot vector of the actual pattern type label and the predicted probability distribution vector is calculated. The cross-entropy loss imposes a larger penalty on low-confidence erroneous predictions and a relatively smaller penalty on high-confidence errors. The average of the cross-entropy losses across all time slots is used as the final distribution dissimilarity metric.

[0089] Preferably, the bidirectional consistency score is calculated to generate alignment confidence labels, including:

[0090] The positive prediction bias is normalized by mapping it to a numerical range of 0 to 1, thus obtaining the positive normalized bias value.

[0091] The back prediction bias is normalized by mapping it to a numerical range of 0 to 1 to obtain the back normalized bias value.

[0092] The bidirectional consistency score is calculated based on the positive normalization deviation value and the negative normalization deviation value. The bidirectional consistency score is negatively correlated with the positive normalization deviation value and the negative normalization deviation value. The higher the score, the higher the consistency between the bidirectional prediction and the actual value.

[0093] Set a high confidence threshold and a low confidence threshold. Anchor association records with bidirectional consistency scores higher than the high confidence threshold are marked as high confidence level, records with scores between the high confidence threshold and the low confidence threshold are marked as medium confidence level, and records with scores lower than the low confidence threshold are marked as low confidence level.

[0094] The confidence level is attached as an alignment confidence label to the corresponding anchor point associated record.

[0095] Specifically, bidirectional prediction verification is performed based on the anchor point association map, including two processes: forward prediction and reverse prediction.

[0096] The forward prediction process is used to verify whether the high-frequency data patterns within the forward background window can effectively predict the low-frequency data features of the anchor point. The specific steps are as follows:

[0097] Obtain the feature vectors of all predecessor nodes within the forward background window of the anchor point, and arrange them in chronological order to form a predecessor feature vector sequence. If the predecessor node set contains N pattern feature clusters, then the predecessor feature vector sequence is an ordered set of N feature vectors arranged in chronological order.

[0098] The preceding feature vector sequence is input into the sequence encoding module, which encodes the variable-length feature vector sequence to generate a fixed-dimensional background representation vector. In this embodiment, the sequence encoding module is specifically implemented by performing mean pooling on the feature vectors at each time step in the preceding feature vector sequence along the time dimension. That is, the arithmetic mean of the values ​​taken at all time steps is calculated for each feature dimension, ultimately generating a background representation vector with the same dimension as a single feature vector. This method can compress a variable-length sequence into a fixed-dimensional representation while preserving the overall statistical properties of the sequence.

[0099] The background representation vector is input into the feature mapping module, which maps the background representation vector to the same dimensional space as the low-frequency data features and outputs the predicted value vector of the anchor low-frequency data features. The feature mapping module uses a fully connected layer structure to realize dimensional transformation. The input dimension of the fully connected layer is equal to the dimension of the background representation vector, and the output dimension is equal to the dimension of the original feature vector of the low-frequency data.

[0100] The predicted value vector is compared with the actual low-frequency data feature vector of the anchor point, and the vector distance between the predicted value vector and the actual feature vector is calculated. The vector distance is calculated by: calculating the square of the difference between the corresponding elements of the two vectors, summing all the square values ​​and taking the square root to obtain the Euclidean distance value; this vector distance is used as the positive prediction bias; the smaller the positive prediction bias, the stronger the correlation between the high-frequency data pattern in the forward background window and the low-frequency data feature of the anchor point.

[0101] The reverse prediction process is used to verify whether the low-frequency data features of the anchor point can effectively predict the high-frequency data pattern types that should appear within the backward influence window. The specific steps are as follows:

[0102] The original feature vector of the low-frequency data at the anchor point and the forward background representation vector generated during the forward prediction process are obtained. These two vectors are then concatenated to form a joint input vector. The concatenation method involves linking each element of the forward background representation vector to each element of the original feature vector of the low-frequency data, forming a new vector with a dimension equal to the sum of the dimensions of the two vectors.

[0103] The backward influence window of the anchor point is divided into multiple time slots according to a preset time interval, which can be set to a value comparable to the average time span of the pattern feature cluster; each time slot corresponds to a pattern type label to be predicted. There are four types of patterns: rising change, falling change, fluctuating change, and stationary.

[0104] The joint input vector is input into the pattern type prediction module. After performing a nonlinear transformation on the joint input vector, the pattern type prediction module generates a four-dimensional probability vector for each time slot through the classification output layer. The four dimensions correspond to the prediction probabilities of four pattern types: rising change, falling change, fluctuating change, and stationary change. The sum of the four probability values ​​is one.

[0105] Obtain the interval type identifier of the actual pattern feature cluster corresponding to each time slot within the backward influence window, as the actual pattern type distribution. Specifically, based on the time range of each time slot, retrieve the pattern feature cluster falling into that slot from the successor node set, and read its interval type identifier as the actual pattern type of that slot.

[0106] The predicted probability distribution is compared with the actual pattern type distribution to calculate a distribution difference metric between the two. The specific calculation process is as follows: All time slots in the backward influence window are traversed. For each time slot, the pattern type with the highest probability value in the predicted probability distribution is taken as the predicted type and compared with the corresponding actual pattern type. The number of time slots where the predicted type and the actual type are inconsistent is counted, and this number is divided by the total number of time slots to obtain the distribution difference metric. This value ranges from zero to one; a smaller value indicates more accurate backpropagation. This distribution difference metric is used as the backpropagation bias.

[0107] The process of calculating the bidirectional consistency score is as follows:

[0108] The positive prediction deviation is normalized by setting a baseline value for the positive deviation, which can be determined based on the statistical distribution of the positive prediction deviation in historical data. The positive prediction deviation is divided by the baseline value to obtain the normalization result. If the normalization result is greater than 1, it is truncated to 1; if the normalization result is less than 0, it is truncated to 0. Finally, the positive normalized deviation value with a range of zero to one is obtained.

[0109] The range of the reverse prediction bias is already within the interval of zero to one, so it can be directly used as the reverse normalized bias value.

[0110] Calculate the arithmetic mean of the forward normalized deviation and the reverse normalized deviation to obtain the mean normalized deviation. Subtract the mean normalized deviation from the mean normalized deviation to obtain the bidirectional consistency score. The score ranges from 0 to 1 and is negatively correlated with both the forward and reverse normalized deviations. A higher score indicates a higher degree of consistency between the bidirectional prediction and the actual value, and better alignment quality.

[0111] The process of generating alignment confidence labels is as follows:

[0112] Set a high confidence threshold and a low confidence threshold. The high confidence threshold is greater than the low confidence threshold, and both thresholds are in the range of zero to one.

[0113] The bidirectional consistency score is compared with a threshold: if the bidirectional consistency score is higher than the high confidence threshold, the anchor-related record is marked as high confidence level; if the bidirectional consistency score is between the high confidence threshold and the low confidence threshold, it is marked as medium confidence level; if the bidirectional consistency score is lower than the low confidence threshold, it is marked as low confidence level.

[0114] The confidence level is used as an alignment confidence label and attached to the corresponding anchor link record to form an anchor link record with a confidence label.

[0115] Preferably, high-confidence anchor point association records are selected, and low-frequency data features and pattern feature clusters are fused to output the screening results, which are then referenced. Figure 3 ,include:

[0116] From all anchor-related records, select anchor-related records with a high alignment confidence level to form a set of records to be merged;

[0117] For each anchor-related record in the record set to be fused, extract the original low-frequency data feature vector of the anchor, the pattern feature cluster feature vector sequence of the predecessor node, and the pattern feature cluster feature vector sequence of the successor node;

[0118] The original feature vector of low-frequency data is used as the static feature component, and the feature vector sequences of predecessor and successor nodes are concatenated in time order to be used as the dynamic feature component.

[0119] The static and dynamic feature components are jointly processed at the feature level to generate a fused feature representation vector.

[0120] The fused feature representation vector is input into the screening and classification module, which outputs the screening conclusion category and the corresponding confidence value. The screening conclusion category and the confidence value are then organized into a screening result data structure and output.

[0121] Specifically, the incremental updates of high-frequency time-series data are monitored, and an incremental data listening mechanism is established. When the first data source generates new high-frequency time-series data, the listener obtains the timestamp range of the new data, including the earliest and latest timestamps of the new data.

[0122] In the anchor point association graph, query the set of anchor points whose backward influence window ends later than the earliest timestamp of the new data, and mark these anchor points as pending re-verification.

[0123] When the timestamp of incremental data falls within the backward influence window of an existing anchor point, pattern agglomeration processing is performed on the incremental data, including difference value calculation, state interval partitioning, and feature extraction, generating a new pattern feature cluster. The newly generated pattern feature cluster is then appended to the successor node set of the corresponding anchor point.

[0124] Re-execute the reverse prediction verification process: Based on the updated successor node set, re-obtain the actual pattern type distribution of each time slot, recalculate the distribution difference metric as the updated reverse prediction bias, and then recalculate the bidirectional consistency score and alignment confidence label. Update the confidence label field in the anchor point association record.

[0125] Filter anchor link records with high confidence: Traverse all anchor link records and filter those with a high confidence level alignment label to form a set of records to be merged. Anchor link records with medium and low confidence levels are not included in subsequent fusion processing.

[0126] The specific steps for fusing low-frequency data features and pattern feature clusters are as follows:

[0127] For each anchor-related record in the set of records to be fused, the following three types of data are extracted: the original feature vector of the low-frequency data of the anchor, the sequence of predecessor feature vectors composed of the feature vectors of each pattern feature cluster in the predecessor node set, and the sequence of successor feature vectors composed of the feature vectors of each pattern feature cluster in the successor node set.

[0128] The original feature vector of low-frequency data is used as a static feature component, which reflects the state snapshot of the target object at the time of low-frequency acquisition.

[0129] The preceding and succeeding feature vector sequences are concatenated chronologically to form a complete temporal feature vector sequence. This sequence is then aggregated to generate dynamic feature components. The aggregation operation is consistent with that of the sequence encoding module, performing mean pooling along the time dimension to generate dynamic feature component vectors with the same dimension as individual feature vectors. These components reflect the dynamic evolution characteristics of the target object before and after the anchor point.

[0130] Feature-level joint processing of static and dynamic feature components: The static feature component vector and the dynamic feature component vector are concatenated to form a fused feature representation vector with a dimension equal to the sum of the dimensions of the two.

[0131] The fused feature representation vector is input into the screening and classification module. This module employs a multi-classifier structure to classify the fused feature representation vector and outputs the screening conclusion category and its corresponding confidence score. The screening conclusion category includes several pre-defined target state categories, used to characterize the comprehensive state assessment result of the target object. The confidence score represents the reliability of the classification result, ranging from 0 to 1.

[0132] From the fused feature representation vector, extract the feature dimensions and their values ​​that contribute significantly to the classification result to form a list of key features. The key feature extraction method is as follows: record the degree of influence of each input feature dimension on the final classification result during the classifier's inference process, select several feature dimensions with the highest influence, and record their dimension indices and corresponding feature values.

[0133] The screening result data structure is organized and output based on the screening conclusion categories, confidence scores, and a list of key features. The screening result data structure can be stored in key-value pair format, containing a conclusion category field, a confidence score field, and a list of key features field.

[0134] This invention proposes an intelligent screening method based on multi-source data. By structurally associating and bidirectionally verifying high-frequency continuous time-series data with low-frequency periodic data, it achieves high-reliability screening of complex object states. The method first performs differential analysis on the high-frequency time-series data, automatically distinguishing between state-changing and state-stationary intervals, and extracting dynamic and statistical features to form a sequence of pattern feature clusters with temporal semantics. Then, using the low-frequency data acquisition timestamp as anchor points, a forward background window and a backward influence window are constructed, associating the anchor points with their preceding and following high-frequency patterns to generate an anchor point association map. Based on this, bidirectional consistency verification is performed through forward prediction ("high-frequency predicting low-frequency") and reverse prediction ("low-frequency inferring high-frequency") to quantify the reliability of the anchor point association and generate aligned confidence labels. Incremental updates of high-frequency data are supported, allowing dynamic correction of subsequent patterns and real-time updates of confidence. Finally, multi-source feature fusion is performed only on high-confidence anchor points, combining static low-frequency features and dynamic pattern features to output screening conclusions and confidence levels. This solution effectively reduces the risk of multi-source data mismatch, improves the accuracy, interpretability and stability of screening results, and is suitable for intelligent screening scenarios involving the fusion of complex time series and heterogeneous data.

[0135] Example 2:

[0136] Based on Embodiment 1, this embodiment further provides an optimized intelligent screening method. By introducing a pattern feature cluster redundancy detection and merging mechanism, a dynamic edge weight calculation mechanism, and an anomaly anchor point marking and isolation mechanism, the data processing efficiency and the reliability of the screening results are improved.

[0137] The method in this embodiment also includes four main steps: acquiring high-frequency time series data and generating pattern feature cluster sequences, acquiring low-frequency data and generating anchor point association maps, performing bidirectional prediction verification, incremental update and outputting screening results. The basic processing flow of each step is the same as in Embodiment 1. The following focuses on describing the optimized features of this embodiment.

[0138] After generating the pattern feature cluster sequence, this embodiment adds a redundancy detection and merging process to reduce redundant information in the sequence and reduce the computational overhead of subsequent processing.

[0139] After generating the pattern feature cluster sequence, the following steps are also included:

[0140] Redundancy detection is performed on adjacent pattern feature clusters in the pattern feature cluster sequence, and the similarity value between the feature vectors of two adjacent pattern feature clusters is calculated.

[0141] When the similarity value exceeds the preset merging threshold and the interval type identifiers of two adjacent pattern feature clusters are the same, the two adjacent pattern feature clusters are merged into one merged pattern feature cluster.

[0142] The starting point of the merged pattern feature cluster is the smaller value of the starting points of the two merged clusters, and the ending point is the larger value of the ending points of the two merged clusters. The interval type identifier remains unchanged, and the feature vector is the arithmetic mean of the corresponding elements of the feature vectors of the two merged clusters.

[0143] The merged pattern feature cluster replaces two adjacent pattern feature clusters in the original sequence, and the pattern feature cluster sequence is updated.

[0144] Specifically, the redundancy detection targets two adjacent pattern feature clusters in the pattern feature cluster sequence. For the i-th and (i+1)-th adjacent pattern feature clusters in the sequence, their feature vectors are extracted, and the similarity value between the two feature vectors is calculated.

[0145] The similarity value is calculated as follows: First, calculate the inner product of the two feature vectors, which is to multiply the corresponding elements of the two vectors and then sum them; then calculate the magnitude of each feature vector, which is the square root of the sum of the squares of the elements of each vector; finally, divide the inner product value by the product of the two magnitudes to obtain the cosine similarity value; the cosine similarity value ranges from -1 to 1, and the closer the value is to 1, the closer the directions of the two vectors are and the more similar their features are.

[0146] A merging threshold is set, which is a value between 0 and 1. When the calculated similarity value exceeds the merging threshold, and the interval type identifiers of two adjacent pattern feature clusters are the same, it is determined that the two pattern feature clusters are redundant, and a merging operation is performed.

[0147] The specific process of the merging operation is as follows: A new merged pattern feature cluster is created, with its cluster identifier being a newly generated unique identifier. The time start point of the merged pattern feature cluster is taken from the smaller value of the time start points of the two merged clusters to ensure coverage of the earlier time boundary; the time end point is taken from the larger value of the time end points of the two merged clusters to ensure coverage of the later time boundary. The interval type identifier remains unchanged, because only pattern feature clusters with the same interval type identifier will be merged. The feature vector is calculated as follows: the corresponding elements of the feature vectors of the two merged clusters are added together and then divided by two to obtain the arithmetic mean vector, which is used as the feature vector of the merged pattern feature cluster.

[0148] Replace the i-th and (i+1)-th pattern feature clusters in the original sequence with the merged pattern feature clusters, update the pattern feature cluster sequence, and reduce the sequence length by one unit after the replacement.

[0149] Redundancy detection and merging operations can be performed iteratively. After completing one round of merging, if there are still adjacent pattern feature clusters in the sequence that meet the merging conditions, the next round of detection and merging will continue until there are no more adjacent cluster pairs that meet the conditions in the sequence, or the sequence length reaches the preset minimum length threshold.

[0150] The aforementioned redundancy detection and merging mechanism effectively compresses the length of the pattern feature cluster sequence and reduces the number of nodes in the anchor point association graph, thereby lowering the computational complexity of bidirectional prediction verification and fusion processing. Furthermore, since the merging operation is only performed on adjacent clusters with highly similar features and of the same type, the merged feature vector can still effectively represent the data characteristics of the corresponding time period without causing significant loss of useful information.

[0151] By performing feature similarity detection and conditional merging on adjacent pattern feature clusters, redundant information in the pattern feature cluster sequence can be effectively reduced, the sequence length and computational overhead of subsequent processing can be reduced, and key pattern features with discriminative power in the sequence can be preserved, thereby improving data processing efficiency without losing effective information.

[0152] When generating the anchor point association graph and establishing directed edges, this embodiment calculates the edge weight value for each directed edge to quantify the difference in contribution of different nodes to the anchor point association.

[0153] When establishing directed edges, the process also includes calculating the edge weight for each directed edge:

[0154] The time decay factor is calculated based on the time interval between the node and the anchor point. The time decay factor is negatively correlated with the time interval; the smaller the time interval, the larger the time decay factor.

[0155] The feature significance factor is determined based on the interval type identifier and feature vector of the pattern feature cluster. The basic significance of the variable pattern feature cluster is higher than that of the stationary pattern feature cluster. The feature significance factor is obtained by superimposing the normalized value of the variance of each element of the feature vector on the basic significance.

[0156] Multiply the time decay factor by the feature saliency factor to obtain the edge weight value of the directed edge, and store the edge weight value as an additional attribute of the directed edge.

[0157] Specifically, the edge weight value is determined by two components: the time decay factor and the feature significance factor.

[0158] The calculation process of the time decay factor is as follows: Obtain the time interval value between the current directed edge connection's pattern feature cluster node and the anchor point. This time interval value has been calculated and stored as an edge attribute in Example 1. Set a time decay baseline value, which is determined based on the time length of the forward background window or backward influence window. Divide the time decay baseline value by the sum of the time decay baseline value and the time interval value to obtain the time decay factor. This calculation method ensures that the time decay factor is negatively correlated with the time interval value: when the time interval value is zero, the time decay factor reaches its maximum value of 1; as the time interval value increases, the time decay factor gradually decreases and approaches 0.

[0159] The calculation process for the feature significance factor is as follows: First, the baseline significance value is determined based on the interval type identifier of the pattern feature cluster. If the interval type identifier is variable, the baseline significance value is set to a higher value; if the interval type identifier is stationary, the baseline significance value is set to a lower value.

[0160] Then, the variance of the feature vector of the pattern feature cluster is calculated. The variance is calculated as follows: first, the arithmetic mean of all elements of the feature vector is calculated; then, the square of the difference between each element and the mean is calculated; finally, all squares are summed and divided by the number of elements to obtain the variance. The variance reflects the degree of dispersion of the values ​​of each dimension of the feature vector. The larger the variance, the more discriminative the features of the pattern feature cluster are.

[0161] Normalize the variance values. Set a normalization benchmark value, divide the variance value by the normalization benchmark value, and if the result is greater than 1, truncate it to 1 to obtain the normalized variance value.

[0162] The feature significance factor is obtained by adding the baseline significance value to the variance normalization value.

[0163] Finally, the time decay factor is multiplied by the feature saliency factor to obtain the edge weight value of the directed edge. The edge weight value is stored as an additional attribute of the directed edge, along with the time interval value, in the anchor point association graph data structure.

[0164] In the subsequent bidirectional prediction verification and fusion processing, the edge weight values ​​can be used to weight the node features, so that nodes with larger edge weight values ​​contribute more to the background representation vector and dynamic feature components.

[0165] Through the aforementioned dynamic calculation mechanism of edge weights, the anchor point association graph can more precisely express the differences in the association strength between different nodes and anchor points, enabling subsequent processing to focus on pattern feature clusters that are closer in time and more significant in features, thereby improving the accuracy of screening results.

[0166] By comprehensively considering the time decay factor and the feature saliency factor to dynamically calculate the edge weight value, it is possible to distinguish the differences in the contribution of different pattern feature clusters to the anchor point association, so that the anchor point association map has a stronger information expression ability, provides a quantitative basis for the importance of nodes for subsequent prediction verification and fusion processing, and improves the accuracy of screening results.

[0167] After generating alignment confidence labels, this embodiment adds an abnormal anchor point identification and isolation mechanism to prevent continuously low-quality anchor point association records from participating in the fusion process and affecting the reliability of the screening results.

[0168] After generating the alignment confidence labels, the following steps are included:

[0169] For each anchor point, maintain a verification counter to record the number of verifications and the number of consecutive low confidence levels experienced by that anchor point;

[0170] When the alignment confidence label of an anchor-related record is updated to a low confidence level, the number of consecutive low confidence values ​​for that record is incremented by one; when the alignment confidence label is updated to a medium or high confidence level, the number of consecutive low confidence values ​​is reset to zero.

[0171] When the number of consecutive low confidence levels exceeds the preset anomaly detection threshold, the record associated with the anchor point is marked as an anomaly and an anomaly flag is set.

[0172] Anchor point association records marked as abnormal are not included in subsequent fusion processing and screening result output, but are recorded in the abnormal log for subsequent analysis.

[0173] Specifically, a validation counter data structure is maintained for each anchor-related record. This data structure contains two fields: validation count and consecutive low-confidence count. The validation count records the total number of times the anchor-related record has undergone bidirectional prediction validation; the consecutive low-confidence count records the number of times the anchor-related record has been consecutively marked as low-confidence. When the anchor-related record is created, both fields of the validation counter are initialized to 0.

[0174] After each bidirectional prediction validation and alignment confidence label update, the validation counter is updated synchronously. The specific update rules are as follows: Increment the validation count field to record this validation. Determine the alignment confidence label level after this update: If the alignment confidence label is updated to a low confidence level, increment the consecutive low confidence count field; if the alignment confidence label is updated to a medium or high confidence level, reset the consecutive low confidence count field to 0.

[0175] Set an anomaly detection threshold, which is a positive integer representing the maximum number of consecutive low confidence levels allowed.

[0176] After each update of the verification counter, it is determined whether the number of consecutive low confidence counts exceeds the anomaly judgment threshold. If the number of consecutive low confidence counts exceeds the anomaly judgment threshold, the anchor point associated record is determined to be in an abnormal state, and the following operations are performed: an anomaly flag is set in the anchor point associated record, and the value of the flag is set to true, indicating that the record has been marked as an abnormal state; the key information of the anchor point associated record is written to the anomaly log, including the anchor point identifier, collection timestamp, number of verifications, number of consecutive low confidence counts, and the most recent bidirectional consistency score.

[0177] In the subsequent fusion processing and screening result output stages, when screening the set of records to be fused, in addition to selecting those with a high alignment confidence level, it is also necessary to exclude anchor-related records with a true anomaly flag. That is, only anchor-related records with a high alignment confidence level and a false anomaly flag can enter the set of records to be fused for subsequent processing.

[0178] The aforementioned anomaly anchor point marking and isolation mechanism automatically identifies and isolates anchor point association records that fail quality checks despite multiple verifications. This prevents low-quality data from negatively impacting screening results during fusion processing, thus improving stability and reliability. Furthermore, the anomaly log provides data support for subsequent analysis of anomaly causes and optimization of data collection and processing workflows.

[0179] By introducing a verification counter and a tracking mechanism for consecutive low-confidence counts, it is possible to identify abnormal anchor points that consistently fail the bidirectional prediction verification, mark and isolate them, prevent abnormal data from participating in the fusion processing and contaminating the screening results, improve stability and reliability of the screening results, and retain anomaly logs for easy problem tracing and system optimization.

[0180] This embodiment introduces the three optimization mechanisms mentioned above, achieving the following technical improvements based on Embodiment 1: In terms of data processing efficiency, the pattern feature cluster redundancy detection and merging mechanism effectively reduces the length of the pattern feature cluster sequence and the number of nodes in the anchor point association graph, lowering the computational complexity of bidirectional prediction verification and fusion processing; in terms of association representation capability, the dynamic calculation mechanism of edge weights enables the anchor point association graph to quantify the differences in association contribution of different nodes, providing more refined node importance information for subsequent processing; in terms of result reliability, the abnormal anchor point marking and isolation mechanism can automatically identify and exclude persistently low-quality anchor point association records, preventing abnormal data from contaminating the screening results and improving the stability of the system.

[0181] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent screening method based on multi-source data, characterized in that, include: Obtain high-frequency time series data from the first data source, calculate the difference between adjacent data points, and divide the high-frequency time series data into state change intervals and state stability intervals; Dynamic features are extracted from the state change intervals, and statistical features are extracted from the state stationary intervals to generate a pattern feature cluster sequence. Low-frequency data from a second data source is acquired, and the timestamps are used as anchor points to determine the forward background window and backward influence window of the anchor points. The predecessor and successor nodes of the pattern feature clusters are obtained by combining the pattern feature cluster sequence, and the anchor point association map is generated. Bidirectional prediction verification is performed based on the anchor point association graph: the positive prediction bias is obtained by predicting the low-frequency data features of the anchor point using the feature sequence of the predecessor node, and the negative prediction bias is obtained by predicting the pattern type of the successor node based on the low-frequency data features; the bidirectional consistency score is calculated and the alignment confidence label is generated. Monitor the incremental updates of high-frequency time series data. When the incremental data falls into the backward influence window of an existing anchor point, process the pattern feature cluster sequence and update the successor node set, re-predict and verify, update the alignment confidence label, filter the anchor point association records with high confidence, perform fusion processing on low-frequency data features and pattern feature clusters, and output the screening results.

2. The intelligent screening method based on multi-source data according to claim 1, characterized in that: The generation of pattern feature cluster sequence includes: traversing high-frequency time series data in chronological order, calculating the difference value between adjacent data points, comparing the difference value with a preset change threshold; marking the state change start point when five consecutive difference values ​​exceed the change threshold, marking the state change end point when five consecutive difference values ​​are lower than the change threshold, marking the data interval between the start point and the end point as the state change interval, and marking the remaining data interval as the state stable interval. The state change interval is divided by sliding segmentation using the first window size. Within each sub-window, the window start value, window end value, extreme point position within the window, change direction identifier, and average change rate are extracted as dynamic features. For the stable interval, a second window size larger than the first window size is used for sliding segmentation, and the window mean, window variance and window duration are extracted as statistical features in each sub-window. Dynamic features and statistical features are assembled into pattern feature clusters. Each pattern feature cluster contains a cluster identifier, a time start point, a time end point, an interval type identifier, and a feature vector. All pattern feature clusters are arranged in chronological order to form a pattern feature cluster sequence.

3. The intelligent screening method based on multi-source data according to claim 1, characterized in that: Determine the foreground and background windows and the background influence window of the anchor point, including: Extract the acquisition timestamp of each data record from the low-frequency data, and use the acquisition timestamp as the time reference of the anchor point; determine the time length parameters of the forward background window and the backward influence window based on the acquisition cycle of the low-frequency data and the time coverage of the high-frequency time series data; extend the time length of the forward background window in the negative direction of the time axis based on the acquisition timestamp of the anchor point, and determine the start time boundary and end time boundary of the forward background window; extend the time length of the backward influence window in the positive direction of the time axis, and determine the start time boundary and end time boundary of the backward influence window.

4. The intelligent screening method based on multi-source data according to claim 1, characterized in that: Generating an anchor point association graph includes: creating an anchor point record for each anchor point, the anchor point record containing an anchor point identifier, a collection timestamp, and the original feature vector of low-frequency data; retrieving pattern feature clusters whose time range falls within the forward background window from the pattern feature cluster sequence based on the time boundary of the anchor point's forward background window, and using these as the set of predecessor nodes for the anchor point; retrieving pattern feature clusters whose time range falls within the backward influence window from the pattern feature cluster sequence based on the time boundary of the anchor point's backward influence window, and using these as the set of successor nodes for the anchor point; using the anchor point as the central node of the graph, establishing directed edges of historical background type between the central node and each predecessor node, and establishing directed edges of subsequent response type between the central node and each successor node, the attributes of each directed edge including the time interval value between the node and the anchor point; and associating and storing the original feature vector of the anchor point with the feature vector sequences of the predecessor nodes and the feature vector sequences of the successor nodes to form the anchor point association graph.

5. The intelligent screening method based on multi-source data according to claim 1, characterized in that: The positive prediction bias acquisition process includes: acquiring the feature vectors of all predecessor nodes within the forward background window of the anchor point, arranging them in chronological order to form a predecessor feature vector sequence; inputting the predecessor feature vector sequence into a sequence encoding module to encode the variable-length feature vector sequence, generating a fixed-dimensional background representation vector; inputting the background representation vector into a feature mapping module to map the background representation vector to the same dimensional space as the low-frequency data features, outputting a predicted value vector for the low-frequency data features of the anchor point; comparing the predicted value vector with the actual low-frequency data feature vector of the anchor point, calculating the vector distance between the predicted value vector and the actual feature vector, and using the vector distance as the positive prediction bias.

6. The intelligent screening method based on multi-source data according to claim 1, characterized in that: The reverse prediction bias acquisition process includes: acquiring the original feature vector and forward background representation vector of the low-frequency data of the anchor point, and concatenating them to form a joint input vector; dividing the backward influence window of the anchor point into multiple time slots according to a preset time interval, with each time slot corresponding to a pattern type label to be predicted, including rising change type, falling change type, fluctuating change type, and stationary type; inputting the joint input vector into the pattern type prediction module, and outputting the predicted probability distribution of the pattern type that should appear in each time slot; acquiring the interval type identifier of the actual pattern feature cluster corresponding to each time slot in the backward influence window as the actual pattern type distribution; comparing the predicted probability distribution with the actual pattern type distribution, calculating the distribution difference metric between the predicted distribution and the actual distribution, and using the distribution difference metric as the reverse prediction bias.

7. The intelligent screening method based on multi-source data according to claim 1, characterized in that: Calculate the bidirectional consistency score and generate alignment confidence labels, including: normalizing the positive prediction bias to obtain the positive normalized bias value; normalizing the negative prediction bias to obtain the negative normalized bias value; and calculating the bidirectional consistency score based on the positive and negative normalized bias values. Set a high confidence threshold and a low confidence threshold. Anchor point association records with bidirectional consistency scores higher than the high confidence threshold are marked as high confidence level, those with scores between the high and low confidence thresholds are marked as medium confidence level, and those with scores lower than the low confidence threshold are marked as low confidence level. The confidence level is then attached as an alignment confidence label to the corresponding anchor point association record.

8. The intelligent screening method based on multi-source data according to claim 1, characterized in that: The steps for obtaining screening results include: selecting anchor-related records with a high confidence level alignment from all anchor-related records to form a set of records to be merged; For each anchor-related record in the record set to be fused, extract the original low-frequency data feature vector of the anchor, the pattern feature cluster feature vector sequence of the predecessor node, and the pattern feature cluster feature vector sequence of the successor node; The original feature vector of low-frequency data is used as the static feature component, and the feature vector sequences of predecessor and successor nodes are concatenated in time order to be used as the dynamic feature component. The static and dynamic feature components are jointly processed at the feature level to generate a fused feature representation vector. The fused feature representation vector is input into the screening and classification module, which outputs the screening conclusion category and the corresponding confidence value. The screening conclusion category and the confidence value are then organized into a screening result data structure and output.

Images