A heterogeneous target association retrieval method based on multi-dimensional feature data

Abstract

The application relates to the field of big data analysis and discloses a heterogeneous target correlation retrieval method based on multi-dimensional feature data, which comprises the following steps: extracting a time axis component of an initial high-dimensional feature tensor to establish an inverted hash table structure, writing morphological feature vectors and correlation feature vectors as logical pointers into time bucket nodes to complete shunting; calculating an information entropy gating threshold according to feature component information entropy and bus congestion degree data and matching a candidate set; filtering interference nodes by using the threshold and assembling a discrete local adjacency matrix in a local memory space; introducing a random discrete time sequence delay jitter amount between 5 mu s and 12 mu s into a diagonal line calibration operator; calculating the local graph topology connectivity of the discrete local adjacency matrix and outputting a result; the application converts the traversal of a retrieval space into local convergence under the jurisdiction of a time axis, eliminates the dependence of calculation overhead on data scale, and stabilizes bus delay disturbance.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and more specifically, to a heterogeneous target association retrieval method based on multidimensional feature data. Background Technology

[0002] Currently, cross-modal data association and heterogeneous target dependency analysis serve as the underlying mechanism supporting the retrieval of complex indicators in large-scale clusters. The current approach typically employs a global spatial coordinate alignment method, which uses feature transformation rules to map spatiotemporal components, materialized morphological features, and entity dependencies into a unified high-dimensional dense continuous vector space. A global spatial similarity metric matrix is ​​then established, and the geometric distance between all feature vectors is calculated using a nearest neighbor traversal scanning algorithm to determine target similarity.

[0003] The methods described above exhibit stable mathematical convergence when processing uniformly distributed static datasets. However, streaming data exhibits a non-uniform topological distribution with locally high density and globally sparse distribution during actual flow. Using a unified feature alignment matrix for forced spatial alignment can easily lead to geometric distortions in high-dimensional vector spaces, causing substantial breaks in the original local adjacency topological dependencies between targets. This introduces structural projection biases that cannot be eliminated in situ during the feature transformation process. Meanwhile, in the context of high-frequency concurrent input of massive streaming features, the computational cost of similarity comparison of the full distance matrix is ​​strongly dependent on the total data size. Frequent computational loads can easily accumulate bus latency, causing overall response performance degradation and hindering data engine updates. To cope with retrieval overhead, clusters often adopt the approach of adding computational nodes to distribute the data or configuring static hashing mechanisms to achieve coarse-grained spatial partitioning. However, adding nodes results in expensive physical layer deployment costs, while static hashing cannot be aware of external bus transmission conditions and cannot absorb the 5μs to 12μs random timing delay disturbances introduced by clock slippage.

[0004] Therefore, the technical problem to be solved by this invention is how to get rid of the mapping limitation of global space brute-force alignment, construct a local topology inverted graph convergence mechanism that can be adaptively adjusted according to external operating conditions through feature orthogonal decoupling, reduce the latency of streaming feature retrieval and improve the fault tolerance of the system. Summary of the Invention

[0005] To address the problems in the background art, this invention provides a heterogeneous target association retrieval method based on multidimensional feature data, comprising: Step S1: Extract the one-dimensional time axis component from the initial high-dimensional feature tensor as the main constraint benchmark, and establish an inverted hash table structure in memory space; use the morphological feature vector and the associated feature vector as mounting attributes, and seamlessly write them into the time bucket nodes of the inverted hash table structure in the form of logical pointers to complete the data storage and distribution of the large-scale dataset in the time axis dimension, and construct the premise for locally convergent data processing. Step S2: In response to the received query request, extract the baseline time features of the query target and match the local data candidate set in the inverted hash table structure; obtain the static stability balance coefficient between 0.6 and 1.4 based on the system initialization, and collect the information entropy value of the target feature component carried by the query request and the bus congestion data fed back by the network bus. Calculate the information entropy gating threshold by dividing the product of the static stability balance coefficient and the information entropy value by the bus congestion data. Step S3: Based on the information entropy gate threshold, filter the interfering nodes in the local data candidate set, use the entity association dependency relationship between the remaining nodes to assemble a discrete local adjacency matrix in the local memory space, and introduce the random discrete disturbance value between 5μs and 12μs, which represents the data synchronization delay fluctuation, into the diagonal calibration operator of the discrete local adjacency matrix in situ to smooth out the data bus throughput delay and timing delay disturbance, and realize boundary adaptive adjustment. Step S4: Call the underlying discrete Boolean graph operator to calculate the local graph topological connectivity of the discrete local adjacency matrix. When the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into the final heterogeneous target association retrieval result and output it.

[0006] Preferably, in step S2, the static stability balance coefficient, the information entropy value of the target feature component carried by the query request, and the bus congestion data fed back by the network bus jointly constrain the dynamic fluctuation range of the information entropy gating threshold; wherein, the information entropy gating threshold maintains a positive linear dependence on the static stability balance coefficient, a positive linear dependence on the information entropy value, and an inverse proportional dependence on the bus congestion data.

[0007] Preferably, step S1 includes the following steps: Step S11, extracting multi-source feature streams from a large-scale heterogeneous dataset and extracting nanosecond-level timestamps for each data item, combining them into one-dimensional time axis components; Step S12, allocating contiguous memory blocks in the local cache and constructing an inverted hash table structure with one-dimensional time axis components as index keys; Step S13, calculating the joint hash code of morphological feature vectors and associated feature vectors, generating logical pointers containing absolute memory addresses, and writing the logical pointers into the time bucket nodes of the inverted hash table structure.

[0008] Preferably, step S3 includes the following steps: Step S31, calculate the multidimensional semantic similarity and topological hop count between each pair of nodes in the local data candidate set. When the multidimensional semantic similarity is greater than the similarity threshold and the topological hop count is less than the distance threshold, establish the entity association dependency relationship between each pair of nodes; Step S32, assemble a discrete local adjacency matrix in the local memory space based on the entity association dependency relationship; Step S33, obtain the time synchronization delay variation value of the data flow during the distributed synchronization process, convert the time synchronization delay variation value into a random discrete disturbance value between 5μs and 12μs, and directly accumulate the random discrete disturbance value into the diagonal calibration operator of the discrete local adjacency matrix to adaptively adjust the data boundary and suppress the data flow divergence under high-frequency concurrency.

[0009] Preferably, after filtering interfering nodes in the local data candidate set based on the information entropy gating threshold in step S3, the cross-temporal nonlinear deviation adaptive calibration path is run under abnormal data transmission conditions, including the following steps: Step S34, when the packet loss rate of the morphological feature vector exceeds the packet loss threshold or an asymmetric data interruption occurs, the cross-temporal nonlinear deviation adaptive calibration path is triggered; Step S35, the distributed system log is called to retrieve the high-frequency dependency topology map under the historical steady-state section as a substitute benchmark; Step S36, the weight coefficient of the associated feature vector in the discrete local adjacency matrix is ​​increased in a discrete stepwise manner, skipping the matching process of the morphological feature vector, and driving the subsequent connectivity convergence through the entity association topology dependency chain of the high-frequency dependency topology map, thereby maintaining the continuous running stability of the data analysis engine.

[0010] Preferably, in step S34, the triggering of the cross-temporal nonlinear deviation adaptive calibration path is completed through a two-stage qualification removal test. The two-stage qualification removal test includes the following steps: Step S341, during the first-stage test, the duration of continuous interruption of the incoming characteristic data packets within the sliding window is counted. When the duration of continuous interruption exceeds the delay limit, the interruption flag of the morphological feature vector is triggered; Step S342, during the second-stage test, the symmetry of the data packet flow direction of the multi-source channel is calculated. When the asymmetric packet loss coefficient exceeds the safety threshold, the cross-temporal nonlinear deviation adaptive calibration path is activated.

[0011] Preferably, step S4 includes the following steps: Step S41, obtain the discrete local adjacency matrix, call the underlying discrete Boolean graph operator, and convert the high-dimensional calculation process into a local graph topological state transition deduction; Step S42, calculate the algebraic connectivity of the discrete local adjacency matrix in the graph topological space to obtain the local graph topological connectivity representing the data dependency relationship; Step S43, when the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into a heterogeneous target association retrieval result and output it.

[0012] Preferably, before extracting the one-dimensional time axis component from the initial high-dimensional feature tensor as a constraint benchmark in step S1, an orthogonal dimensionality reduction and topology reconstruction method for multi-dimensional feature data is included, which includes the following steps: Step S01, acquiring massive multi-dimensional source data across modalities including text, images, and audio / video, and extracting the original high-dimensional feature vector through a heterogeneous feature extraction network; Step S02, performing orthogonal transformation on the original high-dimensional feature vector using an orthogonal matrix to eliminate nonlinear redundant features, outputting morphological feature vectors and associated feature vectors, and combining them to construct the initial high-dimensional feature tensor.

[0013] Preferably, step S02 includes the following steps: Step S021, calling the singular value decomposition operator to perform orthogonal projection on the original high-dimensional feature vector, extracting the feature vector corresponding to the largest singular value of the first term as an orthogonal basis, and completing the orthogonal dimensionality reduction of the multidimensional feature data; Step S022, based on the dimensionality-reduced basis vectors, converting the semantic space similarity into topological adjacency relations, and after removing cross-modal mapping noise, outputting morphological feature vectors and associated feature vectors.

[0014] The embodiments of the present invention have at least the following beneficial effects: 1. In heterogeneous target association retrieval of multidimensional feature data, the original data features are split along orthogonal dimensions to separate mutually independent spatiotemporal feature vectors, morphological feature vectors, and association feature vectors, eliminating coupling interference between different feature dimensions; a one-dimensional discrete time bucket index structure is established using spatiotemporal feature vectors, and morphological feature vectors and association feature vectors are compressed into mounted pointers to achieve physical splitting of high-dimensional datasets on the time axis; during retrieval, the data candidate set is limited by the target spatiotemporal interval, and the output target is determined by morphological feature filtering and local topological adjacency matrix connectivity, transforming the global traversal scan of high-dimensional space into local convergence under the time axis, eliminating the strong dependence of retrieval computation on the total data size of the database table.

[0015] 2. The retrieval unit dynamically reconstructs the judgment boundary of the retrieval space based on the ratio of the information entropy of the feature components of the query target to the real-time concurrent flow congestion density fed back by the network bus. When the threshold is crossed, the search radius is adaptively truncated. After removing interference nodes whose morphological feature differences are greater than the gate threshold, the discrete local adjacency matrix is ​​assembled in the local memory space using the entity association dependency relationship between the remaining nodes. The random discrete disturbance value between 5μs and 12μs generated by the clock slip is introduced in situ into the diagonal calibration operator of the discrete local adjacency matrix. This transforms the bus congestion and timing delay disturbance under high-frequency concurrency into the driving source of adaptive boundary adjustment, eliminating the logical divergence in the traditional calculation process.

[0016] 3. The system is configured with a backup defense mechanism for non-ideal streaming conditions. When a large-scale interruption or asymmetric packet loss is detected in the morphological feature data vector, the adaptive calibration path for cross-time nonlinear deviation is automatically activated through a two-stage qualification removal check. This path uses the high-frequency dependency topology graph under the historical steady-state section in the distributed log as a substitute benchmark, and forcibly increases the weight coefficient of the associated feature vector in the local adjacency matrix in a discrete step-by-step manner. It skips the traditional morphological feature vector matching process and drives the subsequent connectivity convergence by the pure entity association graph topology dependency chain. This maintains the operational stability of the data analysis engine under non-ideal edge conditions of sudden data chain breakage and avoids logical deadlock in the system. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein: Figure 1 This is a flowchart of the heterogeneous target association retrieval method using inverted hash structure and discrete Boolean graph operators of the present invention; Figure 2 This is a parallel processing diagram of offline feature weight solidification and dynamic maintenance of local adjacency matrix in this invention. Detailed Implementation

[0018] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0019] A heterogeneous target association retrieval method based on multidimensional feature data includes: Step S1: Extract the one-dimensional time axis component from the initial high-dimensional feature tensor as the main constraint benchmark, and establish an inverted hash table structure in memory space; use the morphological feature vector and the associated feature vector as mounting attributes, and seamlessly write them into the time bucket nodes of the inverted hash table structure in the form of logical pointers to complete the data storage and distribution of the large-scale dataset in the time axis dimension, and construct the premise for locally convergent data processing. Step S2: In response to the received query request, extract the baseline time features of the query target and match the local data candidate set in the inverted hash table structure; obtain the static stability balance coefficient between 0.6 and 1.4 based on the system initialization, and collect the information entropy value of the target feature component carried by the query request and the bus congestion data fed back by the network bus. Calculate the information entropy gating threshold by dividing the product of the static stability balance coefficient and the information entropy value by the bus congestion data. Step S3: Based on the information entropy gate threshold, filter the interfering nodes in the local data candidate set, use the entity association dependency relationship between the remaining nodes to assemble a discrete local adjacency matrix in the local memory space, and introduce the random discrete disturbance value between 5μs and 12μs, which represents the data synchronization delay fluctuation, into the diagonal calibration operator of the discrete local adjacency matrix in situ to smooth out the data bus throughput delay and timing delay disturbance, and realize boundary adaptive adjustment. Step S4: Call the underlying discrete Boolean graph operator to calculate the local graph topological connectivity of the discrete local adjacency matrix. When the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into the final heterogeneous target association retrieval result and output it.

[0020] Preferably, in step S2, the static stability balance coefficient, the information entropy value of the target feature component carried by the query request, and the bus congestion data fed back by the network bus jointly constrain the dynamic fluctuation range of the information entropy gating threshold; wherein, the information entropy gating threshold maintains a positive linear dependence on the static stability balance coefficient, a positive linear dependence on the information entropy value, and an inverse proportional dependence on the bus congestion data.

[0021] Preferably, step S1 includes the following steps: Step S11, extracting multi-source feature streams from a large-scale heterogeneous dataset and extracting nanosecond-level timestamps for each data item, combining them into one-dimensional time axis components; Step S12, allocating contiguous memory blocks in the local cache and constructing an inverted hash table structure with one-dimensional time axis components as index keys; Step S13, calculating the joint hash code of morphological feature vectors and associated feature vectors, generating logical pointers containing absolute memory addresses, and writing the logical pointers into the time bucket nodes of the inverted hash table structure.

[0022] Preferably, step S3 includes the following steps: Step S31, calculate the multidimensional semantic similarity and topological hop count between each pair of nodes in the local data candidate set. When the multidimensional semantic similarity is greater than the similarity threshold and the topological hop count is less than the distance threshold, establish the entity association dependency relationship between each pair of nodes; Step S32, assemble a discrete local adjacency matrix in the local memory space based on the entity association dependency relationship; Step S33, obtain the time synchronization delay variation value of the data flow during the distributed synchronization process, convert the time synchronization delay variation value into a random discrete disturbance value between 5μs and 12μs, and directly accumulate the random discrete disturbance value into the diagonal calibration operator of the discrete local adjacency matrix to adaptively adjust the data boundary and suppress the data flow divergence under high-frequency concurrency.

[0023] Preferably, after filtering interfering nodes in the local data candidate set based on the information entropy gating threshold in step S3, the cross-temporal nonlinear deviation adaptive calibration path is run under abnormal data transmission conditions, including the following steps: Step S34, when the packet loss rate of the morphological feature vector exceeds the packet loss threshold or an asymmetric data interruption occurs, the cross-temporal nonlinear deviation adaptive calibration path is triggered; Step S35, the distributed system log is called to retrieve the high-frequency dependency topology map under the historical steady-state section as a substitute benchmark; Step S36, the weight coefficient of the associated feature vector in the discrete local adjacency matrix is ​​increased in a discrete stepwise manner, skipping the matching process of the morphological feature vector, and driving the subsequent connectivity convergence through the entity association topology dependency chain of the high-frequency dependency topology map, thereby maintaining the continuous running stability of the data analysis engine.

[0024] Preferably, in step S34, the triggering of the cross-temporal nonlinear deviation adaptive calibration path is completed through a two-stage qualification removal test. The two-stage qualification removal test includes the following steps: Step S341, during the first-stage test, the duration of continuous interruption of the incoming characteristic data packets within the sliding window is counted. When the duration of continuous interruption exceeds the delay limit, the interruption flag of the morphological feature vector is triggered; Step S342, during the second-stage test, the symmetry of the data packet flow direction of the multi-source channel is calculated. When the asymmetric packet loss coefficient exceeds the safety threshold, the cross-temporal nonlinear deviation adaptive calibration path is activated.

[0025] Preferably, step S4 includes the following steps: Step S41, obtain the discrete local adjacency matrix, call the underlying discrete Boolean graph operator, and convert the high-dimensional calculation process into a local graph topological state transition deduction; Step S42, calculate the algebraic connectivity of the discrete local adjacency matrix in the graph topological space to obtain the local graph topological connectivity representing the data dependency relationship; Step S43, when the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into a heterogeneous target association retrieval result and output it.

[0026] Preferably, before extracting the one-dimensional time axis component from the initial high-dimensional feature tensor as a constraint benchmark in step S1, an orthogonal dimensionality reduction and topology reconstruction method for multi-dimensional feature data is included, which includes the following steps: Step S01, acquiring massive multi-dimensional source data across modalities including text, images, and audio / video, and extracting the original high-dimensional feature vector through a heterogeneous feature extraction network; Step S02, performing orthogonal transformation on the original high-dimensional feature vector using an orthogonal matrix to eliminate nonlinear redundant features, outputting morphological feature vectors and associated feature vectors, and combining them to construct the initial high-dimensional feature tensor.

[0027] Preferably, step S02 includes the following steps: Step S021, calling the singular value decomposition operator to perform orthogonal projection on the original high-dimensional feature vector, extracting the feature vector corresponding to the largest singular value of the first term as an orthogonal basis, and completing the orthogonal dimensionality reduction of the multidimensional feature data; Step S022, based on the dimensionality-reduced basis vectors, converting the semantic space similarity into topological adjacency relations, and after removing cross-modal mapping noise, outputting morphological feature vectors and associated feature vectors.

[0028] Example 1: In a big data analysis system, the heterogeneous targets to be retrieved include spatiotemporal features, morphological features, and relational features. The system acquires the original feature dataset of the heterogeneous targets to be retrieved and splits it along the orthogonal dimension into independent spatiotemporal feature vectors, morphological feature vectors, and relational feature vectors. Further, the original feature dataset is pre-processed and generated by a cross-modal heterogeneous feature extraction network. This network performs high-dimensional vector mapping on text data using the embedding layer of a pre-trained language model, and extracts multi-scale visual representations from image and audio / video data using a deep convolutional network and a temporal attention mechanism module. A globally fully connected layer projects the heterogeneous original features from different modalities into a basic feature space of the same dimension, thus completing the initial alignment of cross-modal features and providing standardized feature inputs for subsequent orthogonal dimension decoupling and splitting. The system extracts the time axis component from the spatiotemporal feature vector, using nanosecond-level timestamps as the primary constraint benchmark, and caches it locally. A contiguous memory block is allocated to construct an inverted hash table structure. This inverted hash table structure uses time axis components as index keys. By calculating the joint hash code of the morphological dimension feature vector and the relational dimension feature vector, a logical pointer containing the absolute memory address is generated. The morphological dimension feature vector and the relational dimension feature vector are used as mounting attributes and written into the corresponding time bucket node. This completes the data storage and distribution of large-scale datasets along the time axis, constructing a prerequisite for locally convergent data processing. Specifically, each time bucket node corresponds to a pre-allocated fixed-step linked list structure in the local cache, and the logical pointer stores the absolute address of the head pointer of this linked list. When large-scale heterogeneous data streams are written concurrently at high frequency, the system uses a mutex lock mechanism to lock the address of the time bucket corresponding to the current nanosecond stamp without conflict. The absolute address generated by the joint hash code is seamlessly mounted as the tail insertion node, thereby realizing discrete block isolation of data based on the time axis at the physical memory level and eliminating bus computation accumulation caused by global scanning.

[0029] When the retrieval subject receives a retrieval request, it extracts the target spatiotemporal interval, and the system initializes the static stability balance coefficient. The value is 1.0, and the information entropy value of the target feature components carried in the current query request is collected. And bus congestion data fed back from the network bus. The system uses formulas Calculate the information entropy gating threshold Regarding the collection and calculation of search parameters, the search engine obtains bus congestion data fed back from the network bus by reading the network statistics interface of the operating system kernel. Bus congestion data The packet loss count corresponds to the weighted sum of the ring buffer packet loss count and the real-time queuing depth in the bus transmission queue, and the values ​​are distributed as follows: to Between; the system statistically analyzes the normalized probability density distribution of the target feature components carried in the query request within the current sliding time window and calculates the information entropy value of the generated target feature components. Information entropy gating threshold The calculation formula is ,in, The information entropy gating threshold. The static stability equilibrium coefficient, The information entropy value of the target feature component. The system uses bus congestion data and calculates the information entropy threshold to control the congestion level. As a hard cutoff boundary, when the difference between feature components is greater than the information entropy gating threshold. Interference nodes in the corresponding local data candidate set are directly removed. The cross-domain parameter coordination mechanism here lies in information entropy. The degree of uncertainty in the data inherent in the query target itself, and the degree of bus congestion. This reflects the real-time carrying capacity limit of the underlying hardware bus; when network throughput deteriorates and bus congestion increases... When the information entropy rises sharply, the calculated information entropy gating threshold is... The system adaptively reduces the matching radius of the retrieval space, forcing the data analysis engine to converge within a highly localized, minimal feature subset. This reduces reliance on data transmission from the distributed bus, thus mitigating congestion risks at the physical network layer by proactively shrinking the data layer's boundaries. The system selects target spatiotemporal intervals from an inverted hash table structure and extracts morphological feature vectors within these intervals. It calculates the multidimensional semantic similarity and topological hop count between each pair of nodes in the candidate set. When the multidimensional semantic similarity exceeds a preset similarity threshold and the topological hop count is less than a distance threshold, the system establishes entity association dependencies between the pairs of nodes. The system uses information entropy-gated thresholds. Filter out interfering nodes in the candidate set, and assemble a discrete local adjacency matrix in the local memory space using the entity association dependencies between the remaining nodes.

[0030] During distributed data synchronization, the system acquires the time synchronization delay variation value, converts it into an 8μs random discrete perturbation value, and directly accumulates it into the diagonal calibration operator of the discrete local adjacency matrix. Regarding the data processing of the diagonal calibration operator, the system acquires the time synchronization delay variation value during the distributed synchronization process. Time synchronization delay variation value The perturbation coefficients were determined by calculating the difference between the network card's hardware clock stamp counter, with values ​​ranging from 5μs to 12μs. To eliminate the time dimension and match the topological weights of the discrete local adjacency matrix, the system invoked a normalized scaling operator to calculate the dimensionless perturbation coefficients. perturbation coefficient The calculation formula is ,in, For perturbation coefficients, The static stability equilibrium coefficient, This represents the time synchronization delay variation value. The maximum tolerable delay threshold is preset and fixed at 20μs; the calculated perturbation coefficients are... Multiplying by the identity matrix and directly accumulating to the diagonal elements of the discrete local adjacency matrix corrects the feature divergence caused by asynchronous updates of distributed nodes at the algebraic feature spectrum level. In practice, the dimensionless perturbation coefficients are... The dot product with the identity matrix and the summation to the diagonal essentially intervenes in the spectral radius of the graph matrix by dynamically adjusting the virtual edge weights of node self-loops without changing the original topological structure of the network. This ensures that graph state jitter caused by physical latency is completely absorbed by the feature shifts of diagonal elements, guaranteeing the algebraic convergence of the entire data analysis engine under distributed multi-node asynchronous update conditions. The system calls the underlying discrete Boolean graph operator to calculate the algebraic connectivity of the discrete local adjacency matrix in the graph topological space, i.e., the local graph topological connectivity. Specifically, the underlying discrete Boolean graph operator performs parallel bitwise operations on the reachability matrix of the Boolean adjacency matrix to quickly filter out closed-loop topological connected branches that satisfy the connectivity basis. Based on this, the non-zero Laplace matrix structure of the local adjacency matrix is ​​extracted. Then, the system's built-in iterative power method is used to approximate and calculate its second minimum eigenvalue within the local topological connectivity defined by Boolean operators. This eigenvalue is used as an algebraic connectivity metric to characterize the global topological convergence state, thus achieving a spatial transformation from discrete topology to algebraic continuous measurement while ensuring computational efficiency. When the topological connectivity of the local graph reaches the convergence criterion, the system converts the corresponding topological node identifier into heterogeneous target association retrieval results and outputs them. In this retrieval process, the combination of spatiotemporal dimension feature vectors and inverted hash table structure realizes the physical splitting of the full data along the time axis, avoiding high-dimensional mapping of the full matrix and information entropy gating threshold. The introduction of this feature enables the system to adaptively adjust the search radius based on real-time bus congestion, eliminate noisy nodes, and introduce random discrete disturbance values ​​into the diagonal calibration operator. This provides in-situ hedging against timing jitter in high-frequency concurrent scenarios, ensuring the convergence stability of the retrieval logic under extreme dynamic environments. Regarding the weight adjustment in the adaptive calibration path for cross-temporal nonlinear deviation, when the packet loss rate of the morphological feature vector exceeds the 5% packet loss threshold, the system initiates incremental iterative control and sets a fixed iteration step size. The value is 0.05; the system accumulates the weight coefficient of the current iteration step plus the iteration step size in each iteration loop. This is used to increase the computational weight of associated feature vectors in the discrete local adjacency matrix, and after each accumulation, the graph algebraic connectivity value of the current discrete local adjacency matrix is ​​calculated through the underlying discrete Boolean graph operator. The weight coefficient for the current loop step is also included. The calculation formula is ,in, The weight coefficient for the current loop step. The weight coefficients from the previous iteration. The iteration step size is set when the absolute value of the first derivative of the graph algebraic connectivity value is less than the preset convergence tolerance threshold of 0.0001 or the current iteration step weight coefficient. When the preset safety upper limit threshold of 1.8 is reached, the system terminates the step iteration and locks the current weight coefficients. Instead, it uses the high-frequency dependent topology graph under the historical steady-state section to drive the subsequent connectivity convergence determination. During this process, the 5% packet loss threshold and the 0.05 iteration step size are set according to the empirical value of the typical retransmission cycle of the network hardware to maintain a stable and gradual weighted compensation. The convergence tolerance threshold of 0.0001 is a micro-truncation criterion to ensure the algorithm's order of magnitude accuracy, while the safety upper limit threshold of 1.8 is used to limit the maximum scaling ratio of the associated weights to prevent excessively large weight values ​​from obscuring the original edge node connection relationships in the graph structure.

[0031] Example 2: In a big data analytics scenario, to verify the robustness and technical performance optimization of the proposed heterogeneous target association retrieval method, a test dataset containing 5,000,000 heterogeneous entity records was constructed based on a distributed cluster platform. This dataset covers heterogeneous sensor feature data with different sampling frequencies and original accuracies to simulate the diversity and complexity of data sources in real-world applications. The experimental design includes one experimental sample group and three control sample groups at different levels. The experimental sample group adopts the technical solution of this invention, and its key control parameter is set as follows: static stability balance coefficient. The value is 1.0, which is the information entropy gating threshold. Dynamic calculation was performed, with the random discrete perturbation value set to 8 μs. The parameter configurations for the three control groups were as follows: The first control group removed the associated feature vector mounting step and directly performed morphological feature matching; the second control group adjusted the static stability balance coefficient. The value was set to 0.3, which is below the effective working range; the third control group had a static stability equilibrium coefficient. Set to 2.2, which is higher than the effective working range.

[0032] During the experiment, the system initialized each sample group and input simulated data with 15% random noise and 8% packet loss rate into the database. By monitoring the heterogeneous target association retrieval index output by each sample group, the following key data were statistically recorded. Under the 15% noise level, the experimental sample group successfully identified and associated 96.4% of the target heterogeneous entities, with an average retrieval response time of 12.8ms. The local graph topological connectivity convergence curve reached a stable plateau region in the 14th iteration. The first control sample group, lacking the support of association feature vectors, saw its association success rate drop to 62.1%, and the retrieval response time increased to 45.2ms. The second control sample group, due to the lack of static stability balance coefficient... The threshold is too low, resulting in a low information entropy threshold. The calculation showed a significant bias, with a large number of noisy nodes incorrectly included in the local adjacency matrix, resulting in an association success rate of only 58.7%, exhibiting association oscillation. The third control group also showed a problem due to its static stability equilibrium coefficient. The index was too high, excessively excluding marginal and heterogeneous entities, causing the association success rate to drop to 65.3%, and multiple retrieval result interruptions occurred when processing complex association links.

[0033] To verify the necessity of introducing random discrete perturbation values, the system's performance was additionally measured under a 50Hz power frequency interference environment. Without the introduction of random discrete perturbation values, the diagonal calibration operator of the discrete local adjacency matrix exhibited continuous temporal drift, resulting in a 22.5% confidence bias in the association retrieval results. After introducing an 8μs random discrete perturbation value for in-situ compensation, this confidence bias decreased to 2.1%. The data evolution trend indicates that when the static stability balance coefficient... When the value is within the range of 0.6 to 1.4, the association success rate and response time show an optimized performance window. Below 0.6, the system correlation accuracy shows a significant decrease, indicating that the sensitivity to noise exceeds the system suppression boundary; when Above 1.4, the retrieval recall decreases, indicating that over-filtering leads to the loss of effective related information. This performance inflection point demonstrates the crucial role of this numerical range in balancing the system's retrieval accuracy and adaptability to complex environments. Experimental data supports the invention's achievement of the expected technical effect through adaptive thresholding and perturbation compensation in the multi-dimensional feature data processing flow. Engineering principles show that the static stability balance coefficient... The physical lower bound of 0.6 is determined by the minimum search radius that guarantees the basic retrieval recall rate. A value below 0.6 will cause the information entropy gating threshold to shrink excessively, resulting in the wrong elimination of a large number of edge valid nodes. The physical upper bound of 1.4 is determined by the maximum allowable feature divergence boundary under bus congestion conditions. A value above 1.4 will introduce excessive environmental burst noise, causing retrieval oscillations in the local adjacency matrix due to the accumulation of noisy nodes.

[0034] Example 3: This example combines Figures 1 to 2 This section describes a heterogeneous target association retrieval method based on multidimensional feature data, such as... Figure 1 As shown, step S1 extracts the one-dimensional time axis component from the initial high-dimensional feature tensor as the primary constraint benchmark, and establishes an inverted hash table structure in memory space. The morphological feature vector and associated feature vector are used as mounting attributes and written into the time bucket nodes of this inverted hash table structure in the form of logical pointers, completing the data storage and distribution of the large-scale dataset in the time axis dimension. Next, step S2 extracts the benchmark time features of the query target and matches local data candidate sets in the inverted hash table structure to obtain the static stability balance coefficient between 0.6 and 1.4 during system initialization. It also collects the information entropy value of the query request carrying the target feature component and the bus congestion data fed back by the network bus, using the static stability balance coefficient... The information entropy threshold is calculated by multiplying the information entropy value by the bus congestion data. Step S3 filters the interfering nodes in the local data candidate set using the information entropy threshold. The entity association dependency relationship between the remaining nodes is used to assemble a discrete local adjacency matrix in the local memory space. The random discrete disturbance value between 5μs and 12μs, which represents the data synchronization delay fluctuation, is introduced in situ into the diagonal calibration operator of the discrete local adjacency matrix to achieve boundary adaptive adjustment. Finally, step S4 calls the underlying discrete Boolean graph operator to calculate the local graph topological connectivity of the discrete local adjacency matrix. When the local graph topological connectivity reaches the convergence criterion, the corresponding topological node identifier is converted into heterogeneous target association retrieval results and output.

[0035] like Figure 2As shown, the system architecture includes parallel feature dimension extraction logic and time-triggered calibration logic. The left branch starts by extracting training samples with clearly defined association labels from the offline state of the training sample set input node. It then proceeds downwards through the multi-dimensional feature vector extraction node, which reads and separates morphological feature vectors from associated feature vectors. Next, the feature component information gain value calculation node calculates the information gain of the morphological feature components and the associated feature components respectively. Finally, it enters the dimension weight matrix construction node, which performs normalization operations and dimension concatenation operators to solidify the dimension weight matrix. This node then connects to the right-side edge weight distribution comprehensive scan node via a unidirectional dotted line. The bottom of the left branch connects to the real-time weighted dot product and strength calibration node, which performs a weighted dot product operation on the combined vectors and outputs a scalar product as the association strength between nodes. The right branch starts from the top low-load state... The system automatically starts the calibration process at 02:00 every day during the low-load period, triggered by the segmented time-series trigger node. It then proceeds downwards to the edge weight distribution full scan node, which performs a full scan of the edge weight distribution of all connected components in the adjacency matrix. The process flows to the time decay reduction processing node, which performs time decay factor reduction processing on outdated association paths. The process then proceeds to the path cleaning and spatial measurement node, which deletes association paths below the threshold and calculates the current remaining memory space. Finally, the process reaches the segmented data archiving execution node, which triggers the segmented data archiving logic based on the available memory ratio. The real-time weighted dot product and intensity calibration node at the bottom left and the segmented data archiving execution node at the bottom right converge to point to the backend persistent database at the bottom of the global system to migrate association information that has been running for more than 30 days and dynamically maintain the structural sparsity of the local adjacency matrix.

[0036] Example 4: In a big data correlation analysis system, due to inherent network jitter and multipath interference in the physical transmission of heterogeneous data streams generated by terminal devices, the system experiences a decrease in retrieval hit rate during heterogeneous target correlation retrieval due to a shift in the reference timing. This example demonstrates a technical solution to address the aforementioned timing disturbances and environmental noise. The system constructs the following physical monitoring architecture: a synchronous clock module physically coupled to the distributed data bus is set up. This module generates a reference pulse signal and records the arrival time of all heterogeneous data streams to the data processing node through a signal acquisition interface. To characterize the degree of timing instability during transmission, the system establishes a dynamic delay observation window with a window length of 50ms, continuously collecting the arrival time deviation of the data streams on the bus at a sampling frequency of 2kHz to obtain a series of raw delay deviation sequences. .

[0037] To effectively suppress timing disturbances, the system constructs a nonlinear deviation adaptive calibration path. This calibration path is triggered when the packet loss rate of the data stream exceeds a 5% packet loss threshold or when an asymmetric interruption of the data link is detected. Specifically, the symmetry of the data packet flow is determined by real-time statistical analysis of the ratio of the number of characteristic data packets flowing into and out of the distributed bidirectional channel within the same observation period. The asymmetric packet loss coefficient is defined as the absolute deviation of this ratio from the standard equilibrium state of 1.0. When a sudden congestion occurs in the unidirectional network, causing the asymmetric packet loss coefficient to exceed the set safety threshold, the nonlinear deviation adaptive calibration path is triggered. The system calls the distributed system log to retrieve the high-frequency dependency topology map recorded by the system under a preset steady-state cross-section of the past 10 seconds, using it as a substitute benchmark for the current retrieval operation. Within the calibration logic, the system introduces a discrete step-by-step weight update mechanism to correct the distribution weights of heterogeneous data features in the association adjacency matrix. The weight coefficient update formula is as follows: ,in, These are the updated weight coefficients of the associated feature vector. These are the basic weighting coefficients under the historical steady-state cross-section. This is the step adjustment factor, with a value range of 0.1 to 0.3, and the unit is a dimensionless pure number; This is the normalized time delay deviation value.

[0038] System basis By adjusting the computational contribution of the associated feature vector in the discrete local adjacency matrix and increasing the weight allocation of steady-state features in the high-frequency dependent topology map, the system skips the direct matching process of high-noise morphological features under abnormal data transmission conditions. Instead, it infers the correlation of heterogeneous targets through the structural correlation information of the historical steady-state topology map. This process eliminates local invalid connections that are greatly affected by disturbances by performing Boolean logic judgment on each correlation path in the discrete local adjacency matrix, thereby maintaining the target correlation retrieval logic within the preset accuracy range. Performance tests in this embodiment show that under non-steady-state conditions with a network packet loss rate of 8%, after introducing the aforementioned cross-temporal nonlinear deviation adaptive calibration path, the proportion of erroneous correlation paths caused by noise in the system's output correlation retrieval results decreased from the original 15.6% to 2.4%. This data indicates that by calling the historical steady-state topology benchmark and combining it with step-by-step weight adjustment, the system effectively offsets the second-order temporal error caused by uncontrollable factors in the transmission environment, achieving stable output of retrieval performance under complex big data analysis conditions.

[0039] Example 5: In a big data correlation analysis system, the system needs to perform initial parameter matching to address the differences in read / write performance of different storage media. The system introduces a discrete storage response calibration procedure to ensure the uniformity of retrieval response latency during heterogeneous data retrieval. The system calls the system configuration parameter interface to read the baseline throughput of the current storage unit. Compared with instantaneous average response delay The system establishes a storage response calibration mapping table, which is accessed through... The storage throughput scheduling factor is calculated. ,in, This is a preset media correction factor, ranging from 0.85 to 1.15, and is a dimensionless parameter. It is used when the storage unit detects that the read / write latency deviation of five consecutive data packets exceeds the system-set latency jitter limit. At that time, the system immediately adjusts the storage throughput scheduling factor. Dynamically expand or shrink the current concurrent read / write task thread pool, adjusting the number of threads accordingly. ,in The default number of threads configured for system initialization.

[0040] In the high-dimensional feature matching stage of heterogeneous target association retrieval, the system sets up a multi-dimensional feature semantic difference measurement procedure, and the system extracts the morphological feature vectors corresponding to the heterogeneous targets. Associated feature vectors Construct combined vectors To quantify the contribution weight of different dimensional features to the final association result, the system introduces a dimensional contribution calibration procedure: 2000 training samples with clear association labels are extracted offline, and the information gain value of each dimension of the morphological feature vector is calculated. Information gain values ​​of each dimension of the associated feature vector The system constructs a dimension weight matrix. Defined as ,in, This indicates the dimension concatenation operator, and the weight matrix... The vectors are stored in the runtime memory of the retrieval unit. When the system performs an associative retrieval, it processes the combined vectors input in real time. The system performs a weighted dot product operation, using the output scalar product as the final measure of the strength of the association between nodes. This is achieved through a dimensionality-weight matrix obtained during offline training. This achieves feature enhancement of key dimensions in a multi-dimensional feature space, avoiding the problem of decreased target retrieval hit rate caused by sparse distribution of associated features.

[0041] Example 6: When the data processing system faces the problem of high-dimensional topology graph structure aging and data storage performance degradation caused by long-term operation of the associated network, the system activates an offline topology benchmark calibration and knowledge base update mechanism. The system starts this calibration process during the low-load period at 02:00 every day, by scanning the edge weight distribution of all connected components in the associated adjacency matrix and based on the time decay factor. Outdated related paths are subject to weight reduction, and the weight update formula is as follows: ,in, The updated associated path weight coefficients, The base weights currently stored in the local cache are dimensionless. This is the time decay factor, with a value ranging from 0.05 to 0.15. The system recalculates the weight of all associated paths daily based on the formula described above, taking into account the number of days the path has been running since the last update, and then deletes them. For association paths below the 0.01 threshold, memory space is freed up. After the system performs path cleanup, it calculates the current remaining memory space. If the available memory ratio is less than 20%, it triggers segmented data archiving logic, which migrates association information that has been running for more than 30 days to the backend persistent database. Through this closed-loop management of dynamic decay and segmented archiving, the system maintains the structural sparsity of the local adjacency matrix, ensuring that the high-dimensional retrieval logic maintains a fixed algorithm execution complexity in long-term operation.

[0042] In the high-dimensional feature matching stage of heterogeneous target association retrieval, the system executes the feature dimension contribution evaluation procedure and reads the morphological feature vectors from the training sample set. Information gain value Associated feature vectors Information gain value The weights of each dimension are determined through normalization operations, resulting in a dimension weight matrix. The definition is as follows: ,in, For the dimension weight matrix, For morphological feature component information gain, For the information gain of associated feature components, This represents the dimension concatenation operator, which the system uses to perform real-time association retrieval on the input combined vector. By performing a weighted dot product operation and using the output scalar product as a measure of the strength of association between nodes, this weight matrix enhances the processing of key features in high-dimensional space and ensures the stability of the retrieval logic under heterogeneous data input conditions.

[0043] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with the technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A heterogeneous target association retrieval method based on multidimensional feature data, characterized in that, include: Step S1: Extract the one-dimensional time axis component from the initial high-dimensional feature tensor as the main constraint benchmark, and establish an inverted hash table structure in memory space. Use the morphological feature vector and associated feature vector as mounting attributes and seamlessly write them into the time bucket nodes of the inverted hash table structure in the form of logical pointers to complete the data storage and distribution of the large-scale dataset in the time axis dimension; and construct the premise for locally convergent data processing. Step S2: In response to the received query request, extract the baseline time features of the query target and match the local data candidate set in the inverted hash table structure; obtain the static stability balance coefficient between 0.6 and 1.4 based on the system initialization, and collect the information entropy value of the target feature component carried by the query request and the bus congestion data fed back by the network bus. Calculate the information entropy gating threshold by dividing the product of the static stability balance coefficient and the information entropy value by the bus congestion data. Step S3: Based on the information entropy gate threshold, filter the interfering nodes in the local data candidate set, use the entity association dependency relationship between the remaining nodes to assemble a discrete local adjacency matrix in the local memory space, and introduce the random discrete disturbance value between 5μs and 12μs, which represents the data synchronization delay fluctuation, into the diagonal calibration operator of the discrete local adjacency matrix in situ to smooth out the data bus throughput delay and timing delay disturbance, and realize boundary adaptive adjustment. Step S4: Call the underlying discrete Boolean graph operator to calculate the local graph topological connectivity of the discrete local adjacency matrix. When the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into the final heterogeneous target association retrieval result and output it.

2. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, In step S2, the static stability balance coefficient, the information entropy value of the target feature component carried by the query request, and the bus congestion data fed back by the network bus jointly constrain the dynamic fluctuation range of the information entropy gating threshold; wherein, the information entropy gating threshold maintains a positive linear dependence on the static stability balance coefficient, a positive linear dependence on the information entropy value, and an inverse proportional dependence on the bus congestion data.

3. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, Step S1 includes the following steps: Step S11, extract multi-source feature streams from the large-scale heterogeneous dataset, extract the nanosecond-level timestamp of each data item, and combine them into a one-dimensional time axis component; Step S12, allocate contiguous memory blocks in the local cache, and construct an inverted hash table structure with the one-dimensional time axis component as the index key; Step S13, calculate the joint hash code of the morphological feature vector and the associated feature vector, generate a logical pointer containing the absolute memory address, and write the logical pointer into the time bucket node of the inverted hash table structure.

4. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, Step S3 includes the following steps: Step S31, calculate the multidimensional semantic similarity and topological hop count between each pair of nodes in the local data candidate set. When the multidimensional semantic similarity is greater than the similarity threshold and the topological hop count is less than the distance threshold, establish the entity association dependency between each pair of nodes; Step S32, assemble a discrete local adjacency matrix in the local memory space based on the entity association dependency; Step S33, obtain the time synchronization delay variation value of the data flow during the distributed synchronization process, convert the time synchronization delay variation value into a random discrete disturbance value between 5μs and 12μs, and directly accumulate the random discrete disturbance value into the diagonal calibration operator of the discrete local adjacency matrix to adaptively adjust the data boundary and suppress the data flow divergence under high-frequency concurrency.

5. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, After filtering interfering nodes in the local data candidate set based on the information entropy gating threshold in step S3, the cross-temporal nonlinear deviation adaptive calibration path is run under abnormal data transmission conditions, including the following steps: Step S34, when the packet loss rate of the morphological feature vector exceeds the packet loss threshold or an asymmetric data interruption occurs, the cross-temporal nonlinear deviation adaptive calibration path is triggered; Step S35, the distributed system log is called to retrieve the high-frequency dependency topology map under the historical steady-state section as a substitute benchmark; Step S36, the weight coefficient of the associated feature vector in the discrete local adjacency matrix is ​​increased in a discrete stepwise manner, skipping the matching process of the morphological feature vector, and driving the subsequent connectivity convergence through the entity association topology dependency chain of the high-frequency dependency topology map, thus maintaining the continuous running stability of the data analysis engine.

6. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 5, characterized in that, In step S34, the cross-temporal nonlinear deviation adaptive calibration path is triggered through a two-stage qualification removal test. The two-stage qualification removal test includes the following steps: Step S341, during the first-stage test, the duration of continuous interruption of the incoming characteristic data packets within the sliding window is counted. When the duration of continuous interruption exceeds the delay limit, the interruption flag of the morphological feature vector is triggered. Step S342, during the second-stage test, the symmetry of the data packet flow direction of the multi-source channel is calculated. When the asymmetric packet loss coefficient exceeds the safety threshold, the cross-temporal nonlinear deviation adaptive calibration path is activated.

7. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, Step S4 includes the following steps: Step S41, obtain the discrete local adjacency matrix, call the underlying discrete Boolean graph operator, and convert the high-dimensional calculation process into a local graph topological state transition deduction; Step S42, calculate the algebraic connectivity of the discrete local adjacency matrix in the graph topological space to obtain the local graph topological connectivity representing the data dependency relationship; Step S43, when the local graph topological connectivity reaches the convergence criterion, convert the corresponding topological node identifier into a heterogeneous target association retrieval result and output it.

8. The heterogeneous target association retrieval method based on multidimensional feature data according to claim 1, characterized in that, Before extracting the one-dimensional time axis component from the initial high-dimensional feature tensor as a constraint benchmark in step S1, an orthogonal dimensionality reduction and topology reconstruction method for multi-dimensional feature data is also included, which includes the following steps: Step S01, acquiring massive multi-dimensional source data across modalities including text, images, and audio / video, and extracting the original high-dimensional feature vector through a heterogeneous feature extraction network; Step S02, performing orthogonal transformation on the original high-dimensional feature vector using an orthogonal matrix to eliminate nonlinear redundant features, outputting morphological feature vectors and associated feature vectors, and combining them to construct the initial high-dimensional feature tensor.

9. A heterogeneous target association retrieval method based on multidimensional feature data according to claim 8, characterized in that, Step S02 includes the following steps: Step S021, calling the singular value decomposition operator to perform orthogonal projection on the original high-dimensional feature vector, extracting the feature vector corresponding to the largest singular value of the first term as an orthogonal basis, and completing the orthogonal dimensionality reduction of the multidimensional feature data; Step S022, based on the dimensionality-reduced basis vectors, converting the semantic space similarity into topological adjacency relations, and after removing cross-modal mapping noise, outputting morphological feature vectors and associated feature vectors.

Images