Online incremental learning method and system based on adaptive B + tree index

Abstract

The invention discloses an online incremental learning method and system based on a self-adaptive B + tree index, and relates to the technical field of artificial intelligence, and the method comprises the steps: constructing an extended self-adaptive B + tree index structure, inserting a new sample, and calculating a trajectory disturbance value; generating a disturbance score by combining the disturbance value and a node state, and adding a new sample and an associated node sample into an incremental learning scheduling queue when a condition is met; when a node splitting, merging or reconstructing event occurs, acquiring an affected node and updating a structural code of the affected node; monitoring sample label track change, adjusting a label confidence mark when a threshold value is exceeded, and limiting participation frequency according to a screening mechanism; and performing forgetting compression operation on the leaf nodes with low access frequency and small label fluctuation degree. By constructing an adaptive B + tree index and combining disturbance-driven scheduling, structure change response and forgetting compression strategies, sample management refinement, learning update high efficiency and structure maintenance controllability are realized, and the stability and resource utilization efficiency of online incremental learning are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an online incremental learning method and system based on adaptive B+ tree indexing. Background Art

[0002] With the widespread application of artificial intelligence in real-time data processing, intelligent decision-making, and edge computing, online incremental learning has become a key research direction for adapting to dynamic environmental changes and continuously improving model performance. Online incremental learning systems typically require the following capabilities: accepting new samples in real time, dynamically updating models, managing changes in sample distribution, and evaluating and controlling sample quality and structural stability.

[0003] Most existing online learning methods rely on global model optimization and fail to respond promptly to micro-local perturbations and structural changes in sample distribution. This can lead to delayed model updates, the accumulation of redundant samples, or low-quality samples interfering with the training process. Furthermore, traditional index structures suffer from the following deficiencies in sample insertion and dynamic structure maintenance: They lack explicit modeling mechanisms for changing sample trajectories, making it impossible to characterize the degree of deviation between the insertion path of new samples and the historical path; they lack support for real-time evaluation and dynamic intervention of local node states; and they lack a controllable cleanup strategy for redundant samples within nodes, resulting in reduced storage efficiency and model performance over long periods of time. Summary of the Invention

[0004] In view of the shortcomings of the prior art described above, the purpose of the present invention is to provide an online incremental learning method and system based on adaptive B+ tree indexing to solve the above technical problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an online incremental learning method based on adaptive B+ tree index, comprising:

[0006] S1: Construct an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes include feature vectors, timestamps, labels, label confidence marks, and structure codes. The state parameters of the leaf nodes include the number of visits, label fluctuation degree, and structure evolution rate.

[0007] S2: Obtain the feature vector of the new sample, insert it into the B+ tree index structure, and calculate the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample;

[0008] S3: When a new sample is inserted into the B+ tree index structure, a disturbance score is generated based on the trajectory disturbance value and the state parameters of the leaf node. If the disturbance score is greater than the first threshold and the label confidence is marked as reliable or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue;

[0009] S4: When a node split, merge, or reconstruction operation triggers a structure change event, obtain the affected node subset and update the structure code of the nodes in the subset;

[0010] S5: Obtain the label trajectory of the sample, update the label confidence mark of the sample whose label trajectory change is greater than the second threshold, update the label confidence mark to unstable or suspicious, and limit the participation frequency of the sample according to the preset semi-supervised screening mechanism;

[0011] S6: Perform a forgetting compression operation on the leaf nodes whose access frequency is less than the third threshold and whose label fluctuation degree is less than the fourth threshold, and delete the samples in the nodes.

[0012] The present invention is further configured such that, in the sample field, the feature vector is used to record the multidimensional representation of the inserted sample after being mapped by the feature extraction function; the timestamp is used to record the time when the sample is inserted into the index structure; the label is used to record the supervision category information of the sample; the label confidence mark is used to record the credibility status of the label source, including reusable, credible, unstable and suspicious, and is initially set to credible; the structural code is used to represent the local structural distribution status of the leaf node where the sample is located, and is generated based on the feature vectors of all samples in the node through an aggregate mapping function;

[0013] Among the state parameters, the number of visits is updated cumulatively each time a node is inserted or queried; the degree of label fluctuation is obtained by statistically analyzing the category distribution entropy or variation amplitude of the sample labels within the node within a set time window; the structural evolution rate is obtained by performing differential calculations on the structural coding sequence generated by the node in a continuous time period.

[0014] The present invention is further configured such that step S2 includes:

[0015] Obtain the feature vector of the new sample and insert it into the adaptive B+ tree index structure according to the preset indexing rules. During the insertion process, record the node sequence passed through and generate the insertion path.

[0016] Retrieving a preset number of historical sample insertion paths from the index structure, where the historical sample insertion paths are composed of node access sequences recorded during the historical sample insertion process;

[0017] The insertion path is compared with the insertion path of historical samples for sequence differences, and the trajectory perturbation value is calculated based on the edit distance of the node sequence, the jump depth and the distance between the structural embedding vectors. When there is no historical sample in the index structure, the trajectory perturbation value is set to the maximum value.

[0018] The present invention is further configured such that step S3 includes:

[0019] When a new sample is inserted into the B+ tree index structure, the trajectory disturbance value and the label fluctuation degree and structure evolution rate of the leaf node where the new sample is located are obtained;

[0020] A disturbance score is generated by performing a joint analysis based on the trajectory disturbance value, label fluctuation degree and structure evolution rate;

[0021] When the disturbance score is greater than a preset first threshold and the label confidence mark of the sample in the leaf node is credible or unstable, the new sample and the sample set in the leaf node are added to the incremental learning scheduling queue.

[0022] The present invention is further configured to generate a disturbance score by performing a joint analysis based on the trajectory disturbance value, the label fluctuation degree, and the structure evolution rate, including:

[0023] When the trajectory disturbance value is greater than the first disturbance threshold, and the label fluctuation degree is greater than the second disturbance threshold or the structure evolution rate is greater than the third disturbance threshold, the disturbance score is judged to be high;

[0024] When the trajectory disturbance value is greater than the first disturbance threshold, the label fluctuation degree is less than or equal to the second disturbance threshold, and the structure evolution rate is less than or equal to the third disturbance threshold, the disturbance score is determined to be medium;

[0025] When the trajectory disturbance value is less than or equal to the first disturbance threshold, the disturbance score is determined to be low.

[0026] The present invention is further configured such that, in step S4, when the number of samples after the target node is inserted is greater than a preset upper threshold, a node splitting event is triggered; when a node splitting event occurs, the original node before the split, the new node generated by the split, and the parent node are set as the affected node subsets;

[0027] When the number of samples in the target node is less than the preset lower limit after the samples are deleted, and it can be merged with the adjacent sibling node, the node merge event is triggered; when the node merge event occurs, the nodes and parent nodes involved in the merge are set as the affected node subset;

[0028] When split or merge events occur continuously in a subtree area until the height difference between the left and right subtrees exceeds the preset height difference threshold, a node reconstruction event is triggered; when a node reconstruction event occurs, all nodes involved in the reconstruction are set to the affected node subset;

[0029] For each node in the affected node subset, the node's hierarchical position, index number, type mark of whether it is a leaf node, and the current number of child nodes are collected, and a structure coding value is generated through the structure coding function. The structure coding value is updated and written into the structure coding field of the corresponding node.

[0030] The present invention is further configured such that step S5 includes:

[0031] establishing a label trajectory record table with the feature vector as the primary key in the leaf node, when the label of the sample changes, obtaining the current timestamp and label value, and generating a label trajectory sequence;

[0032] According to the label trajectory sequence, the number of occurrences of a pair of continuous different labels is counted, and is normalized to the ratio of the total trajectory length, which is defined as the label change rate;

[0033] According to the label trajectory sequence, the total number of label categories is counted, and the label value range is calculated in the regression task;

[0034] When the label change rate and / or the label value range is greater than the second threshold, the label confidence marker of the sample is updated to an unstable state, and when it is greater than the second threshold for a continuous preset time window, it is updated to a suspicious state;

[0035] Based on the updated label confidence marker, when performing incremental learning scheduling, a preset semi-supervised screening strategy is applied, wherein the unstable state sample is set to have an upper limit of sampling probability, and the suspicious state sample is rejected to enter the scheduling task queue.

[0036] The application further provides that step S6 comprises:

[0037] A forgetting compression strategy triggered at a fixed period is set, the leaf nodes are traversed in the period, the access frequency and the label fluctuation degree of the nodes in a preset sliding time window are calculated, wherein the access frequency is the number of node accesses divided by the window length; and the label fluctuation degree is the average of the label fluctuation degree of the nodes;

[0038] When the access frequency of the leaf node is less than a third threshold and the label fluctuation degree is less than a fourth threshold, a forgetting compression operation is performed, all sample fields stored in the leaf node are deleted, the node state is reset to reusable, and the structure code is retained.

[0039] The application further provides that it further comprises:

[0040] A periodic explainability output mechanism is set, in a preset sample processing period, the running state information of the adaptive B+ tree index structure is collected, and an explainability data file including a node access heat map and a structure change map is generated;

[0041] The node access heat map calculates the standardized heat value based on the access frequency of the leaf node in the sliding window, and generates a structured index map with heat annotation;

[0042] The structure change map records the event type, occurrence time and node evolution relationship by monitoring the node splitting, merging or reconstruction events in the index structure, and generates a time series structure graph.

[0043] The present invention also provides an online incremental learning system based on adaptive B+ tree indexing, which is used to implement the above-mentioned online incremental learning method based on adaptive B+ tree indexing, including:

[0044] Structure construction module: Constructs an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes include feature vectors, timestamps, labels, label confidence marks, and structure codes. The state parameters of the leaf nodes include the number of visits, label fluctuation degree, and structure evolution rate.

[0045] Sample insertion module: obtains the feature vector of the new sample, inserts it into the B+ tree index structure, and calculates the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample;

[0046] Score generation module: When a new sample is inserted into the B+ tree index structure, a perturbation score is generated based on the trajectory perturbation value and the state parameters of the leaf node. If the score is greater than the first threshold and the label confidence is marked as credible or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue;

[0047] Event update module: When a node split, merge, or reconstruction operation triggers a structure change event, it obtains the affected node subset and updates the structure code of the nodes in the subset;

[0048] Label update module: obtains the label trajectory of the sample, updates the label confidence mark of the sample whose label trajectory changes are greater than the second threshold, updates the label confidence mark to unstable or suspicious, and limits the participation frequency of the sample according to the preset semi-supervised screening mechanism;

[0049] Forgetting compression module: Performing forgetting compression operation on leaf nodes whose access frequency is less than the third threshold and whose label fluctuation is less than the fourth threshold, and deleting samples in the nodes.

[0050] The application provides an online incremental learning method and system based on adaptive B+ tree index, the method constructs an extended adaptive B+ tree index structure, wherein the sample field stored in the leaf node includes a feature vector, a timestamp, a label, a label confidence mark and a structure code, and the state parameters of the leaf node include the access frequency, the label fluctuation degree and the structure evolution rate; the feature vector of the new sample is obtained, the B+ tree index structure is inserted, and the trajectory disturbance value is calculated according to the sequence difference between the insertion path of the new sample and the insertion path of the historical sample; when the new sample is inserted into the B+ tree index structure, the disturbance score is generated according to the trajectory disturbance value and the state parameters of the leaf node, when the disturbance score is greater than the first threshold value and the label confidence mark is credible or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue; when the node splitting, merging or reconstruction operation triggers the structure change event, the affected node subset is obtained, and the structure code of the node in the subset is updated; the label trajectory of the sample is obtained, the label confidence mark of the sample with a label trajectory change greater than the second threshold value is updated, the label confidence mark is updated to unstable or suspicious, and the participation frequency of the sample is limited according to the preset semi-supervised screening mechanism; the forgetful compression operation is performed on the leaf node with an access frequency less than the third threshold value and a label fluctuation degree less than the fourth threshold value, the sample in the node is deleted, and the beneficial effects include:

[0051] 1. Realize sample-level structured management and state awareness: by defining the sample field and state parameters of the leaf node in the extended adaptive B+ tree index structure, the dynamic perception and continuous management of the structure position, evolution trajectory and label credibility of each sample are realized, and the adaptability to the change of streaming data is enhanced;

[0052] 2. Introduce a disturbance-driven incremental learning scheduling mechanism: based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample, the trajectory disturbance value is calculated, and the disturbance score is generated in combination with the state parameters of the leaf node, so that the model is only updated incrementally at the node with representative changes in structure disturbance or label state, and the risk of redundant training and overfitting is effectively inhibited;

[0053] 3. Support adaptive response and code update of structure change event: for node splitting, merging and reconstruction events, the structure change detection mechanism is triggered, the affected node subset is dynamically identified, and the structure code update is performed, so as to maintain the logical consistency and interpretability of the index structure, and provide support for subsequent visualization and structure analysis;

[0054] 4. Provide a controllable forgetful compression strategy: set the joint judgment mechanism of the access frequency and the label fluctuation degree in the sliding time window, perform the compression and deletion operation on the leaf node with long-time low frequency and stable fluctuation, avoid the expansion of the index structure and the accumulation of samples, and improve the resource utilization rate.

[0055] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the following specific embodiments of the present application can be implemented according to the content of the description, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative labor. In the drawings:

[0057] Figure 1 A flow chart of an online incremental learning method based on adaptive B+ tree index is shown for an exemplary embodiment of the present application;

[0058] Figure 2 A structural schematic diagram of an online incremental learning system based on adaptive B+ tree index is shown for an exemplary embodiment of the present application. DETAILED DESCRIPTION

[0059] The embodiments of the present application will be described below with reference to the drawings and preferred embodiments, and those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in the specification. The present application can also be implemented or applied by different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, but not for limiting the protection scope of the present application.

[0060] It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and only the components related to the present application are shown in the diagrams, but not the number, shape and size of the components when actually implemented. The actual implementation of each component may be randomly changed in type, number and proportion, and the layout pattern of the components may also be more complex.

[0061] In the following description, a large number of details are discussed to provide a more thorough explanation of the embodiments of the present application, however, it is obvious to those skilled in the art that the embodiments of the present application can be implemented without these specific details, and in other embodiments, the known structures and devices are shown in the form of block diagrams rather than in the form of details, so as not to make the embodiments of the present application difficult to understand.

[0062] Embodiment one:

[0063] An online incremental learning method based on adaptive B+ tree index, such as Figure 1 Shown, including:

[0064] S1: Construct an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes include feature vectors, timestamps, labels, label confidence marks, and structure codes. The state parameters of the leaf nodes include the number of visits, label fluctuation degree, and structure evolution rate.

[0065] S2: Obtain the feature vector of the new sample, insert it into the B+ tree index structure, and calculate the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample;

[0066] S3: When a new sample is inserted into the B+ tree index structure, a disturbance score is generated based on the trajectory disturbance value and the state parameters of the leaf node. If the disturbance score is greater than the first threshold and the label confidence is marked as reliable or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue;

[0067] S4: When a node split, merge, or reconstruction operation triggers a structure change event, obtain the affected node subset and update the structure code of the nodes in the subset;

[0068] S5: Obtain the label trajectory of the sample, update the label confidence mark of the sample whose label trajectory change is greater than the second threshold, update the label confidence mark to unstable or suspicious, and limit the participation frequency of the sample according to the preset semi-supervised screening mechanism;

[0069] S6: Perform a forgetting compression operation on the leaf nodes whose access frequency is less than the third threshold and whose label fluctuation degree is less than the fourth threshold, and delete the samples in the nodes.

[0070] B+ tree is a self-balancing tree structure widely used in databases and file systems, mainly used for storing and managing large amounts of data, its characteristics are that all records are stored in leaf nodes, and leaf nodes are linked by linked lists, which is convenient for range query and sequential traversal, it is a multi-way search tree with efficient insertion, deletion and lookup operations, the extended B+ tree is based on the traditional B+ tree, and supports adaptive management of samples, by introducing new fields and state parameters, the B+ tree can not only perform efficient index lookup, but also adapt to dynamic changes of data and self-adjustment. For example, it can dynamically optimize the structure, manage samples according to the access frequency, label fluctuation and node structure change, in step S1, the application is further provided, in the sample field, the feature vector is used to record the multi-dimensional representation of the inserted sample after mapping by the feature extraction function; the timestamp is used to record the time when the sample is inserted into the index structure; the label is used to record the supervised class information of the sample; the label confidence mark is used to record the confidence state of the label source, including reusable, reliable, unstable and suspicious, and is initially set to reliable; the structure code is used to represent the local structure distribution state of the leaf node where the sample is located, which is generated based on the feature vectors of all samples in the node through an aggregation mapping function; specifically, the feature vector is a vector representation in a multi-dimensional space obtained by mapping the original data of the sample through a feature extraction function (such as PCA, CNN, traditional statistical method in machine learning), which is a high-dimensional representation of the sample and is used as input for classification, regression and other tasks, wherein the original sample data includes image, text or sensor data; the original data is mapped into a vector in a multi-dimensional feature space through a series of feature extraction methods (such as convolutional neural network (CNN) for image, TF-IDF for text), which is not described here, the obtained vector is stored in the sample field of the leaf node, so as to facilitate subsequent calculation and index lookup; the timestamp represents the time when the sample is inserted into the B+ tree index structure, which is used to record the historical state of the sample and provide a basis for the time series analysis of the sample, whenever a new sample is inserted into the B+ tree index, the current time (such as Unix timestamp or UTC time) is obtained and stored in the timestamp field of the sample, the timestamp is used for time series analysis, which can track the historical state of the sample and support time-dependent update in incremental learning; the label is the supervised class information of the sample (class label in classification task, numerical value in regression task), which is the target variable of the sample and the basis for model learning, the label is provided by the labeler and stored in the leaf node, so as to be indexed and managed together with the sample feature vector and learned; the label confidence mark is used to represent the confidence state of the label. There are four states of label confidence: reusable, reliable, unstable and suspicious.Initially, the label confidence mark is set to be trusted, which means that the sample label comes from a trusted source and can be used multiple times; trusted means that the label is relatively accurate, and the model or labeling process has been verified; unstable means that the label has uncertainty or large fluctuation, and may need to be recalibrated due to data fluctuation; suspicious means that the label is highly uncertain and cannot be trusted, and needs to be further checked; the label confidence mark is dynamically updated according to the label trajectory, classification result and label fluctuation of the sample. In incremental learning, if the sample label fluctuates frequently or is unstable, the confidence mark of the updated label is unstable or suspicious; the structure code is used to represent the local structure state of the leaf node, which is calculated based on the feature vectors of all samples in the node, and is a summary value of the node, reflecting the overall information of the data distribution in the node; the structure code is obtained by aggregating the feature vectors of all samples in the node (such as taking the average, weighted average, etc.); the structure code can be generated by an aggregation mapping function, such as PCA dimensionality reduction or vector clustering of the feature vector, and then calculating the aggregation result (such as mean or center point) of the feature of all samples in the node, which is a prior art, and will not be described here; the structure code is stored in the leaf node as one of the index values of the node.

[0071] In the state parameters, the access times are updated by accumulation every time the node is inserted or queried; the label fluctuation degree is obtained by calculating the class distribution entropy or variation amplitude of the labels of the samples in the node within a set time window; the structure evolution rate is obtained by performing difference calculation on the structure code sequence generated by the node in a continuous time period; specifically, the access times represent the number of times the node is accessed, which is used to measure the activity of the node, and the access times of the node are updated by one every time a sample is inserted into the node; each leaf node maintains an "access times" field, which is accumulated every time the node is accessed, facilitating subsequent analysis of the activity of the node; the label fluctuation degree reflects the change amplitude of the labels of the samples in the node, which is measured by calculating the entropy or variation amplitude of the label class distribution in the node; the frequency distribution of all sample labels in the node is calculated, and the entropy of the label distribution is calculated based on the frequency distribution: H = -∑ i p i log(p i ), wherein p i is the probability of each label category; the structure evolution rate measures the speed of the change of the structure of the node, which represents the structure change frequency of the node in a period of time, and the difference calculation is performed on the structure code generated by the node in a continuous time period to measure the amplitude or frequency of the change. If the structure code changes frequently in a short time, it indicates that the node is in a high evolution rate state, and the structure evolution rate S r The difference between the current code and the code of the previous time period is calculated as follows: wherein C current and C previousare the structural codes for the current and previous time periods respectively, and Δt is the time difference.

[0072] The present invention is further configured such that step S2 includes:

[0073] Obtain the feature vector of the new sample, insert the feature vector into the adaptive B+ tree index structure according to the preset indexing rules, record the node sequence passed during the insertion process, and generate an insertion path; specifically, input the feature vector of the new sample into the constructed adaptive B+ tree index structure; during the indexing process, each time an intermediate node or leaf node is passed, its node number or unique structure identifier is written into the "insertion path sequence" to form a complete insertion trajectory of the sample; furthermore, the preset indexing rule is the minimum distance matching rule based on the cluster center distance: each child node maintains the center point of the sample feature vector within it, and the intermediate node saves the center point set of all child nodes. For the feature vector of the new sample, its Euclidean distance with the center points of all child nodes is calculated in turn, and the path selection follows the minimum distance matching principle: the child node with the smallest distance is selected as the next hop, and the sample is inserted into the corresponding leaf node. If the number of samples exceeds the upper limit, the leaf node is split;

[0074] Retrieve a preset number of historical sample insertion paths from the index structure. The historical sample insertion path consists of the node access sequence recorded during the historical sample insertion process. Specifically, retrieve three historical sample insertion paths from the same leaf node in the B+ tree as the reference trajectory set. Each historical path is the node access sequence recorded when these historical samples were inserted. The historical path must meet a certain time window or sample selection strategy (such as the most recently inserted or samples with a specific label).

[0075] The insertion path is compared with the historical sample insertion path for sequence differences, and the trajectory perturbation value is calculated based on the edit distance, jump depth and distance between the structural embedding vectors of the node sequence. When there is no historical sample in the index structure, the trajectory perturbation value is set to the maximum value; specifically, the new sample insertion path is compared with each historical sample path for sequence differences: the number of modification steps of the two path sequences at the node level is evaluated, recorded as the edit distance; the number of non-continuous jump levels appearing in the path is recorded, recorded as the jump depth; the structural embedding vector (structural encoding) of each node is extracted, and the average distance between the new sample path and the historical path in the embedding space is measured, recorded as the structural embedding vector difference. The above three indicators are combined to form a trajectory perturbation value as a quantitative result of the degree of structural perturbation; further, the trajectory perturbation value = 0.5*edit distance + 0.2*jump depth + 0.3*embedding distance mean, where the edit distance, jump depth and embedding distance mean are all normalized values. When there is no historical sample in the index structure, the trajectory perturbation value is set to 1.

[0076] The present invention is further configured such that step S3 includes:

[0077] When a new sample is inserted into the B+ tree index structure, the trajectory disturbance value and the label fluctuation degree and structural evolution rate of the leaf node where the new sample is located are obtained. Specifically, when a new sample is inserted into the B+ tree index structure, the logic for obtaining the trajectory disturbance value, label fluctuation degree, and structural evolution rate has been explained in the previous article and will not be repeated here.

[0078] A disturbance score is generated by performing a joint analysis based on the trajectory disturbance value, the label fluctuation degree, and the structural evolution rate. The present invention is further configured to perform a joint analysis based on the trajectory disturbance value, the label fluctuation degree, and the structural evolution rate to generate a disturbance score, including: when the trajectory disturbance value is greater than a first disturbance threshold, and the label fluctuation degree is greater than a second disturbance threshold, or the structural evolution rate is greater than a third disturbance threshold, the disturbance score is determined to be high; when the trajectory disturbance value is greater than the first disturbance threshold, and the label fluctuation degree is less than or equal to the second disturbance threshold, and the structural evolution rate is less than or equal to the third disturbance threshold, the disturbance score is determined to be medium; when the trajectory disturbance value is less than or equal to the first disturbance threshold, the disturbance score is determined to be low. Specifically, after extracting the above three types of disturbance-related parameters, the joint analysis process of the disturbance score is entered. Three threshold parameters are set: the first perturbation threshold is used to determine whether the trajectory perturbation value is significant, the second perturbation threshold is used to determine whether the label fluctuation degree is severe; the third perturbation threshold is used to determine whether the structural evolution rate is fast; it is set through historical statistical data; if the trajectory perturbation value is greater than the first perturbation threshold, and any value of the label fluctuation degree or the structural evolution rate is greater than the corresponding threshold, the perturbation score is "high"; if the trajectory perturbation value is greater than the first perturbation threshold, and neither the label fluctuation degree nor the structural evolution rate exceeds the corresponding threshold, the perturbation score is "medium"; if the trajectory perturbation value does not exceed the first perturbation threshold, the perturbation score is "low". The scoring results are graded ("high, medium, low"). Compared with the continuous numerical form, the grade identification can execute scheduling decisions through simple conditional branches when making judgments, without the need for refined numerical comparisons, reducing the logic complexity of the model and improving the system's operating efficiency and robustness;

[0079] When the perturbation score is greater than the preset first threshold and the label confidence mark of the sample in the leaf node is credible or unstable, the new sample and the sample set in the leaf node are added to the incremental learning scheduling queue; the first threshold (middle) for triggering the scheduling of the perturbation score is set. When the perturbation score caused by the current sample insertion exceeds the threshold, the label confidence mark status of the sample set in the leaf node to which the sample belongs is further checked. If the label confidence mark of the sample set in the leaf node is "credible" or "unstable", it means that although not all samples in the area are manually labeled, they have certain reference value. The current sample and all samples in the node are packaged into a learning scheduling unit and pushed to the incremental learning scheduling queue as candidate data for model update; if the perturbation score does not exceed the threshold, or the confidence mark is in the "suspicious" state, the scheduling is not triggered to avoid redundant training or the introduction of low-quality samples.

[0080] The present invention is further configured such that, in step S4, when the number of samples after the target node is inserted is greater than a preset upper threshold, a node splitting event is triggered; when a node splitting event occurs, the original node before the split, the new node generated by the split, and the parent node are set as the affected node subset; specifically, after a new sample is inserted into a target leaf node, the number of samples N_node within the node exceeds a preset upper threshold T_upper, triggering a node splitting event; performing sample partitioning on the node, dividing the samples into two groups, left and right, in order according to keywords; creating a new leaf node, migrating a group of samples to the new node; registering the new node with its parent node; and adding the node before the split, the new node after the split, and the parent node to the "affected node subset";

[0081] When the number of samples in the target node is less than the preset lower limit threshold after the sample is deleted, and it can be merged with the adjacent sibling node, a node merge event is triggered; when a node merge event occurs, the nodes involved in the merge and the parent node are set as the affected node subset; specifically, after a target node sample is deleted, its sample number N_node is less than the preset lower limit threshold T_lower, and there is a mergeable sibling node, a node merge event is triggered; the samples of the node and its adjacent sibling node are merged into one node; the indexes of the original two nodes are deleted in the parent node; the merged node and its parent node are added to the "affected node subset";

[0082] When split or merge events occur continuously in a subtree area until the height difference between the left and right subtrees exceeds a preset height difference threshold, a node reconstruction event is triggered. When a node reconstruction event occurs, all nodes involved in the reconstruction are set as the affected node subset. Specifically, a subtree undergoes continuous node mergers / splits, resulting in the height difference ΔH between the left and right subtrees exceeding the preset height difference threshold T_height. A node reconstruction event is triggered. Structural rebalancing (such as AVL-like rotation or B-tree rearrangement) is performed on the entire subtree structure. All involved structural adjustment nodes are recorded. All affected nodes are added to the "affected node subset".

[0083] For each node in the affected node subset, the node's hierarchical position, index number, type mark of whether it is a leaf node, and current number of child nodes are collected, a structure coding value is generated through the structure coding function, and the structure coding value is updated and written into the structure coding field of the corresponding node; specifically, the following operations are performed on each node in the affected node subset: node structure state parameters are collected, including hierarchical position: recording the depth L of the node in the B+ tree structure; index number: recording the ordinal position Idx of the node in its parent node; node type mark: Boolean variable isLeaf marks whether it is a leaf node; current number of child nodes: recording the number C of direct child nodes it has (if it is a leaf node, it can be set to 0); input the above structure state parameter set S = {L, Idx, isLeaf, C}; generate a structure coding value E_struct through the structure coding function F_enc(S); further, the structure coding function F_enc(S) is an ordered numerical vector encoding, which directly maps the structure state to a standardized structure vector.

[0084] The present invention is further configured such that step S5 includes:

[0085] In the leaf node, a label trajectory record table is established with the feature vector as the primary key. When the label of a sample changes, the current timestamp and label value are obtained to generate a label trajectory sequence. Specifically, the label changes experienced by each sample in its life cycle are continuously tracked to form a label trajectory. A label trajectory record table is established in each leaf node with the sample feature vector v as the primary key. Whenever the sample is accessed again, if the label value is different from the last record, the current timestamp t is recorded. i , current label value y i ; Sequence form: T v ={(t1,y1),(t2,y2),…,(t n ,y n )}, if the sample feature vector appears for the first time or there is no record in the table, initialize the label track;

[0086] According to the label trajectory sequence, the number of occurrences of consecutive different label pairs is counted and normalized to the ratio of the total trajectory length, which is defined as the label change rate. Specifically, from the sequence T v Extract all label change events, i.e. y i ≠y i-1 , let the total number of change events be Δy, and define the label change rate: R change =Δy / (n-1), the label change rate reflects the instability of the label state;

[0087] According to the label trajectory sequence, the total number of label categories is counted, and the label value range is calculated in the regression task; when the label change rate and / or the label value range is greater than the second threshold, the label confidence mark of the sample is updated to an unstable state. When it is continuously greater than the second threshold within a continuous preset time window, it is updated to a suspicious state; specifically, the number of different categories in the label trajectory is counted C count , calculate the range R of the labels in the trajectory range =max(y i )-min(y i ), the second threshold θ2 is the judgment threshold set for label confidence update, and corresponding thresholds are set for label change rate and label value range, w t is the system rolling time window, k is the continuous window threshold counter, when R range >θ2 and C count >θ2, or when R range >θ2 or C count >θ2, set the current sample confidence mark to "unstable". If the above conditions are met in k consecutive time windows: update the confidence mark to "suspicious" and update the label confidence mark written into the sample field;

[0088] Based on the updated label confidence mark, a preset semi-supervised screening strategy is applied when executing incremental learning scheduling, in which an upper limit of the sampling probability is set for unstable state samples, and suspicious state samples are rejected from entering the scheduling task queue; specifically, the impact of unreliable samples on model training is controlled, the risk of label error propagation is reduced, and a preset sampling probability upper limit P1(0.2)∈(0,1) is set; when the label confidence mark is reusable, it enters the scheduling queue directly; when the label confidence mark is credible, it enters the scheduling queue directly; when the label confidence mark is unstable, random sampling is performed with probability P1 to determine whether to schedule; when the label confidence mark is suspicious, it is rejected from entering the scheduling queue.

[0089] The present invention is further configured such that step S6 includes:

[0090] A forgetfulness compression strategy is set up that triggers a fixed period. During this period, leaf nodes are traversed and the node access frequency and label fluctuation within a preset sliding time window are calculated. The access frequency is the number of node accesses divided by the window length; the label fluctuation is the average node label fluctuation. Specifically, the system presets a compression period parameter T_gc. A scheduler is set up to trigger the traversal compression process once every T_gc period. During the execution, all leaf nodes are evaluated sequentially.

[0091] When the access frequency of a leaf node is less than the third threshold and the label fluctuation is less than the fourth threshold, a forget compression operation is performed to delete all sample fields stored in the leaf node, reset the node state to be reusable, and retain the structural encoding; specifically, the state fields maintained by the node are: cumulative access count V_total; current system clock t_now; last compression check time t_last_gc; sliding window length W. If t_now-t_last_gc≥T_gc, the access frequency within the window is defined as: f access =W window / V,W window Represents the number of visits to the node within W time; when the access frequency of a leaf node is less than the third threshold and the label fluctuation degree is less than the fourth threshold, the node is marked as a "low activity + stable" node, and forgetting compression is performed, including: deleting all sample fields in the node (including feature vectors, timestamps, labels, label confidence marks); retaining the structure encoding field (for structure visualization and subsequent reconstruction); setting the status mark to "reusable node" for priority filling when subsequent samples are inserted; resetting the number of visits to 0, and clearing the label fluctuation buffer.

[0092] The present invention is further configured to include:

[0093] Set up a periodic interpretability output mechanism to collect the operating status information of the adaptive B+ tree index structure within the preset sample processing cycle, and generate interpretable data files including node access heat maps and structure change maps;

[0094] The node access heat map calculates the normalized heat value based on the access frequency of leaf nodes in the sliding window and generates a structured index map with heat annotations. Specifically, for all leaf nodes in the adaptive B+ tree, the access records within the sliding time window W are collected. Each leaf node stores an access timestamp buffer or access counter. The number of accesses within the sliding window V is counted. i , calculate the access frequency f i =V i / W, all access frequencies f i Normalized to heat value h i ∈[0,1], using Min-Max normalization: h i =(fi -f min ) / (f max -f min +∈), where f max 、f min are the minimum and maximum access frequencies of all leaf nodes in the sliding window, and ∈ is a small constant to avoid division by zero; traverse the B+ tree structure and construct a graph structure (node-edge topology) based on the current index topology; the heat value h of each leaf node is i Map to color / brightness level; mark to the corresponding node primitive to generate heat map;

[0095] The structural change graph generates a time series structure graph by monitoring node splitting, merging or reconstruction events in the index structure, recording the event type, occurrence time and node evolution relationship. Specifically, the structural change records are aggregated and a time series graph is constructed in chronological order. The evolution path of the node is expressed in a graph structure, and a visual graph (PDF / PNG) or a structural graph (GraphML / JSON) can be generated, preserving the timestamp, structural change chain, and the relationship graph of the involved nodes.

[0096] Example 2:

[0097] See also Figure 2 The exemplary online incremental learning system based on adaptive B+ tree index is used to implement the above-mentioned online incremental learning method based on adaptive B+ tree index, including:

[0098] Structure construction module: Constructs an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes include feature vectors, timestamps, labels, label confidence marks, and structure codes. The state parameters of the leaf nodes include the number of visits, label fluctuation degree, and structure evolution rate.

[0099] Sample insertion module: obtains the feature vector of the new sample, inserts it into the B+ tree index structure, and calculates the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample;

[0100] Score generation module: When a new sample is inserted into the B+ tree index structure, a perturbation score is generated based on the trajectory perturbation value and the state parameters of the leaf node. If the score is greater than the first threshold and the label confidence is marked as credible or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue;

[0101] Event update module: When a node split, merge, or reconstruction operation triggers a structure change event, it obtains the affected node subset and updates the structure code of the nodes in the subset;

[0102] Label update module: obtains the label trajectory of the sample, updates the label confidence mark of the sample whose label trajectory changes are greater than the second threshold, updates the label confidence mark to unstable or suspicious, and limits the participation frequency of the sample according to the preset semi-supervised screening mechanism;

[0103] Forgetting compression module: Performing forgetting compression operation on leaf nodes whose access frequency is less than the third threshold and whose label fluctuation is less than the fourth threshold, and deleting samples in the nodes.

[0104] It should be noted that the online incremental learning system based on adaptive B+ tree indexing provided in the above embodiment and the online incremental learning method based on adaptive B+ tree indexing provided in the above embodiment are of the same concept, wherein the specific manner in which each module and unit performs operations has been described in detail in the method embodiment and will not be repeated here. In actual applications, the online incremental learning system based on adaptive B+ tree indexing provided in the above embodiment can allocate the above functions to different functional modules as needed, that is, divide the internal structure of the system into different functional modules to complete all or part of the functions described above, and this is not limited here.

[0105] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0106] It should be understood that the term "and / or" as used herein simply describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. A and B can be singular or plural. Furthermore, the character " / " as used herein generally indicates an "or" relationship between the associated objects, but it may also indicate an "and / or" relationship. For specific understanding, please refer to the context.

[0107] In this application, "at least one" means one or more, and "plurality" means two or more. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or plural.

[0108] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0109] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0110] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0111] In the several embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0112] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0113] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0114] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0115] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. An online incremental learning method based on adaptive B+ tree index, characterized in that: include: S1: Construct an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes are It includes feature vectors, timestamps, labels, label confidence marks and structural encodings. The state parameters of leaf nodes include the number of visits, label fluctuation degree and structural evolution rate. S2: Obtain the feature vector of the new sample, insert it into the B+ tree index structure, and calculate the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample; S3: When a new sample is inserted into the B+ tree index structure, a disturbance score is generated based on the trajectory disturbance value and the state parameters of the leaf node. If the disturbance score is greater than the first threshold and the label confidence is marked as reliable or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue; S4: When a node split, merge, or reconstruction operation triggers a structure change event, obtain the affected node subset and update the structure code of the nodes in the subset; S5: Obtain the label trajectory of the sample, update the label confidence mark of the sample whose label trajectory change is greater than the second threshold, update the label confidence mark to unstable or suspicious, and limit the participation frequency of the sample according to the preset semi-supervised screening mechanism; S6: Perform a forgetting compression operation on the leaf nodes whose access frequency is less than the third threshold and whose label fluctuation degree is less than the fourth threshold, and delete the samples in the nodes.

2. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: In the sample field, the feature vector is used to record the multidimensional representation of the inserted sample after mapping through the feature extraction function; the timestamp is used to record the time when the sample is inserted into the index structure; the label is used to record the supervision category information of the sample; the label confidence mark is used to record the credibility status of the label source, including reusable, trustworthy, unstable and suspicious, and is initially set to trustworthy; the structural code is used to represent the local structural distribution status of the leaf node where the sample is located, and is generated based on the feature vectors of all samples in the node through the aggregation mapping function; Among the state parameters, the number of visits is updated cumulatively each time a node is inserted or queried; the degree of label fluctuation is obtained by statistically analyzing the category distribution entropy or variation amplitude of the sample labels within the node within a set time window; the structural evolution rate is obtained by performing differential calculations on the structural coding sequence generated by the node in a continuous time period.

3. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: Step S2 includes: Obtain the feature vector of the new sample and insert it into the adaptive B+ tree index structure according to the preset indexing rules. During the insertion process, record the node sequence passed through and generate the insertion path. Retrieving a preset number of historical sample insertion paths from the index structure, where the historical sample insertion paths are composed of node access sequences recorded during the historical sample insertion process; The insertion path is compared with the insertion path of historical samples for sequence differences, and the trajectory perturbation value is calculated based on the edit distance of the node sequence, the jump depth and the distance between the structural embedding vectors. When there is no historical sample in the index structure, the trajectory perturbation value is set to the maximum value.

4. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: Step S3 includes: When a new sample is inserted into the B+ tree index structure, the trajectory disturbance value and the label fluctuation degree and structure evolution rate of the leaf node where the new sample is located are obtained; A disturbance score is generated by performing a joint analysis based on the trajectory disturbance value, label fluctuation degree and structure evolution rate; When the disturbance score is greater than a preset first threshold and the label confidence mark of the sample in the leaf node is credible or unstable, the new sample and the sample set in the leaf node are added to the incremental learning scheduling queue.

5. The online incremental learning method based on adaptive B+ tree index according to claim 4, characterized in that: A joint analysis is performed based on the trajectory perturbation value, label fluctuation degree, and structure evolution rate to generate a perturbation score, including: When the trajectory disturbance value is greater than the first disturbance threshold, and the label fluctuation degree is greater than the second disturbance threshold or the structure evolution rate is greater than the third disturbance threshold, the disturbance score is judged to be high; When the trajectory disturbance value is greater than the first disturbance threshold, the label fluctuation degree is less than or equal to the second disturbance threshold, and the structure evolution rate is less than or equal to the third disturbance threshold, the disturbance score is determined to be medium; When the trajectory disturbance value is less than or equal to the first disturbance threshold, the disturbance score is determined to be low.

6. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: In step S4, when the number of samples after the target node is inserted exceeds the preset upper threshold, a node split event is triggered; when a node split event occurs, the original node before the split, the new node generated by the split, and the parent node are set as the affected node subsets; When the number of samples in the target node is less than the preset lower limit after the samples are deleted, and it can be merged with the adjacent sibling node, the node merge event is triggered; When a node merge event occurs, the nodes and parent nodes involved in the merge are set as the affected node subsets; When split or merge events occur continuously in a subtree area until the height difference between the left and right subtrees exceeds the preset height difference threshold, a node reconstruction event is triggered; when a node reconstruction event occurs, all nodes involved in the reconstruction are set to the affected node subset; For each node in the affected node subset, the node's hierarchical position, index number, type mark of whether it is a leaf node, and the current number of child nodes are collected, and a structure coding value is generated through the structure coding function. The structure coding value is updated and written into the structure coding field of the corresponding node.

7. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: Step S5 includes: In the leaf node, a label trajectory record table is established with the feature vector as the primary key. When the label of a sample changes, the current timestamp and label value are obtained to generate a label trajectory sequence. The number of occurrences of consecutive different label pairs is counted according to the label trajectory sequence, and normalized to the ratio of the total trajectory length, which is defined as the label change rate; Count the total number of label categories based on the label trajectory sequence and calculate the range of label values ​​in the regression task; When the label change rate and / or the label value range is greater than the second threshold, the label confidence mark of the sample is updated to an unstable state. When it is continuously greater than the second threshold within a continuous preset time window, it is updated to a suspicious state. Based on the updated label confidence mark, a preset semi-supervised screening strategy is applied when performing incremental learning scheduling. In this strategy, an upper limit on the sampling probability of unstable state samples is set, and suspicious state samples are rejected from entering the scheduling task queue.

8. The online incremental learning method based on adaptive B+ tree index according to claim 1, characterized in that: Step S6 includes: Set a forget compression strategy triggered by a fixed period, traverse the leaf nodes within the period, and calculate the node access frequency and label fluctuation within the preset sliding time window. The access frequency is the number of node visits divided by the window length; the label fluctuation is the mean node label fluctuation. When the access frequency of a leaf node is less than the third threshold and the label fluctuation is less than the fourth threshold, a forget compression operation is performed to delete all sample fields stored in the leaf node, reset the node state to be reusable, and retain the structural coding.

9. The online incremental learning method based on adaptive B+ tree index according to claim 8, characterized in that: Also includes: Set up a periodic interpretability output mechanism to collect the operating status information of the adaptive B+ tree index structure within the preset sample processing cycle, and generate interpretable data files including node access heat maps and structure change maps; The node access heat map calculates the normalized heat value based on the access frequency of leaf nodes in the sliding window and generates a structured index map with heat annotations. The structural change graph generates a time series structure graph by monitoring the node splitting, merging or reconstruction events in the index structure, recording the event type, occurrence time and node evolution relationship.

10. An online incremental learning system based on adaptive B+ tree index, used to implement the online incremental learning method based on adaptive B+ tree index according to any one of claims 1 to 9, characterized in that: include: Structure construction module: Constructs an extended adaptive B+ tree index structure, where the sample fields stored in the leaf nodes include feature vectors, timestamps, labels, label confidence marks, and structure codes. The state parameters of the leaf nodes include the number of visits, label fluctuation degree, and structure evolution rate. Sample insertion module: obtains the feature vector of the new sample, inserts it into the B+ tree index structure, and calculates the trajectory disturbance value based on the sequence difference between the insertion path of the new sample and the insertion path of the historical sample; Score generation module: When a new sample is inserted into the B+ tree index structure, a perturbation score is generated based on the trajectory perturbation value and the state parameters of the leaf node. If the score is greater than the first threshold and the label confidence is marked as credible or unstable, the new sample and the sample set of the leaf node are added to the incremental learning scheduling queue; Event update module: When a node split, merge, or reconstruction operation triggers a structure change event, it obtains the affected node subset and updates the structure code of the nodes in the subset; Label update module: obtains the label trajectory of the sample, updates the label confidence mark of the sample whose label trajectory changes are greater than the second threshold, updates the label confidence mark to unstable or suspicious, and limits the participation frequency of the sample according to the preset semi-supervised screening mechanism; Forgetting compression module: Performing forgetting compression operation on leaf nodes whose access frequency is less than the third threshold and whose label fluctuation is less than the fourth threshold, and deleting samples in the nodes.

Images