An online target recognition method based on dynamic micro-cluster representation of data streams

Abstract

The present invention proposes an online target recognition method based on dynamic micro-cluster representation of data streams, aiming to improve the ability to accurately recognize rapidly changing targets, and is particularly suitable for complex and dynamic real-world scenarios. The method comprises the following steps: first, feature extraction is performed on incremental data through an encoder-decoder architecture, and target classification is performed using a target recognition network. At the same time, the extracted features are used to update the dynamic micro-cluster model online. On this basis, the emergence of new classes or the occurrence of concept drift is detected in real time by calculating the distance between the incremental data and the center of the micro-cluster. If a new class is detected, the method will add a new micro-cluster and adjust the target recognition network; if concept drift occurs, the dynamic micro-cluster model is adjusted in combination with the micro-cluster and the new data. This technology effectively overcomes the problem that the recognition performance of existing methods drops sharply when the target changes, has significant practical value, and is particularly suitable for real-time and dynamically changing application scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of online target recognition, and more specifically, relates to an online target recognition method based on data stream dynamic micro-cluster representation. Background Art

[0002] In modern target recognition tasks, the characteristics, position and morphology of targets change significantly as the environment changes. In particular, in the recognition of sea and air targets, traditional target recognition methods face great challenges due to the differences in the performance of targets at different times and in different environmental conditions. Traditional methods usually rely on the static characteristics of the target, such as position and morphology, which often leads to a significant drop in the recognition accuracy of the model when the target changes. The limitation of this method lies in its sensitivity to target characteristics. Once the shape, posture or position of the target changes, the traditional target recognition system will find it difficult to adapt, resulting in a decrease in its recognition accuracy.

[0003] In addition, as time goes by, the performance, parameter settings and working status of the sensor may also change, resulting in the phenomenon of "concept drift" in the data collected by the sensor. This drift means that the data collected by the sensor at different times and under different environmental conditions may show significant differences, which in turn affects the performance and stability of the target recognition model. The concept drift problem is another major factor affecting recognition accuracy, because traditional machine learning methods usually assume that the distribution of data is stable, but the data distribution in the actual environment will change over time, resulting in a decrease in the accuracy of the model in the later stage of deployment.

[0004] In view of this background, the present invention proposes an online target recognition method based on data stream micro-cluster representation and small sample learning. By constructing a dynamic micro-cluster model and a data stream incremental learning strategy, it not only solves the problem of target feature change and sensor drift in the target recognition task, but also can achieve relatively robust target recognition in the absence of a large amount of labeled data. This method can effectively cope with the challenges in the recognition of sea and air targets with weak features, and significantly improve the stability and accuracy of the target recognition system. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem that the model accuracy of the existing technology gradually decreases over time in real scenarios. On the one hand, traditional target recognition methods are usually very sensitive to target features such as position and shape. Once the target changes, the recognition rate will drop significantly; on the other hand, as time goes by, sensor parameters may change, causing concept drift in sensor data, affecting recognition accuracy.

[0006] To solve this problem, existing methods usually rely on a large amount of labeled data. This data includes targets in different poses and positions, so that the model can learn more robust target semantic knowledge. However, for the task of recognizing targets with weak features, there are few target supervision samples and the situation is very complicated, so traditional methods often cannot handle it effectively.

[0007] To achieve the above-mentioned invention object, the present invention provides an online target recognition method based on data stream micro-cluster representation and small sample learning to solve this problem. The method includes the following steps:

[0008] (1) Construct an unsupervised data feature extraction model. Using an encoder-decoder architecture, for each incremental data in the data stream, after embedding it into a vector representation, it is first encoded using a pre-trained encoder (such as Vision Transformer, ViT), and then decoded by a decoder (such as BigGAN). The feature extraction model is trained by calculating the reconstruction error between the features after decoding and before encoding.

[0009] (2) Construct a dynamic micro-cluster model. For each incremental data, divide it into clusters according to its distance to the center of each micro-cluster in the micro-cluster model, and update the micro-cluster model in this way.

[0010] (3) Perform downstream target recognition, input the image features extracted in the previous step into the RCNN model for target detection tasks, and train it by calculating the cross entropy loss;

[0011] (4) Detect new categories and construct new micro-clusters based on the distance from the incremental data to the micro-clusters. When there are enough data points in the constructed micro-clusters, it is considered that new categories have appeared, and the downstream target recognition network is dynamically adjusted.

[0012] (5) Detect concept drift. Detect concept drift based on the distance from incremental data to micro-clusters: If multiple data can be divided into multiple clusters at the same time, it is considered that concept drift has occurred. Create a new data set with the micro-cluster information representing the old class and the window data representing the new class, and rebuild the micro-cluster model.

[0013] The object of the present invention is achieved in this way.

[0014] The present invention is based on the online target recognition method of data stream micro-cluster representation and small sample learning. By building and maintaining a dynamic micro-cluster model, the target recognition task under weak features is achieved. It can also cope well with the occurrence of concept drift and has good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of a specific implementation method of the online target recognition method based on data stream dynamic micro-cluster characterization of the present invention;

[0016] Figure 2 Schematic diagram of detecting concept drift and new classes in the present invention, where each triangle represents a micro-cluster center and the square represents a sample x identified as a new class. i , the circle represents the sample x that is identified as having concept drift j ;

[0017] Figure 3 It is a system framework diagram of target recognition of a specific implementation method of the online target recognition method based on data stream dynamic micro-cluster characterization of the present invention. DETAILED DESCRIPTION

[0018] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0019] Figure 1 Flowchart of the online object recognition method based on data stream micro-cluster representation and small sample learning.

[0020] In this embodiment, if Figure 1 As shown, the online target recognition method for data stream micro-cluster representation and small sample learning includes a step: the present invention completes the target detection task in the data stream through a feature extractor and a dynamic micro-cluster model. The main idea is to use the dynamic micro-cluster model to identify and adjust new classes in the data stream, detect and adjust when concept drift occurs, so that the model can adapt to the small sample learning target recognition task in the data stream. The specific steps are as follows:

[0021] S1: Training feature extractor

[0022] 1.1) Use a sliding window W of size N to save the data of the data stream, embed the incremental data into representation X, and obtain the feature representation x of the data through a pre-trained encoder (such as Vision Transformer, ViT). Then obtain X through the decoder ′ .X ′ It can indicate the accuracy of the features obtained after feature extraction.

[0023] 1.2) Take the current window as the training set and calculate X ′ and the reconstruction error of X to train the feature extractor, and the loss function uses the mean square error (MSE). i , i∈(1,N) represents the representation of the i-th data in the window, X i ′ It represents the representation of the i-th data feature in the window after passing through the decoder.

[0024]

[0025] The feature extractor is trained and optimized by minimizing the loss function.

[0026] S2: Building a dynamic micro-cluster model

[0027] The present invention uses the micro-cluster structure in the data stream to detect new categories in the data and whether concept drift occurs. In streaming data, data points arrive continuously, and traditional clustering algorithms often cannot adapt to this ever-changing data stream. The micro-cluster method can efficiently handle large-scale and dynamic data streams by maintaining a smaller and dynamically updated clustering structure. Consider a data sequence x1, x2, ..., x1 that arrives over time. n ∈S k , then the microcluster of this sequence can be constructed as in In addition to these two important features, micro-clusters can also accommodate more time series features. The formula for calculating the radius of a micro-cluster is as follows:

[0028]

[0029] where x i represents the i-th data point, n is the number of points in the cluster, c k It is the micro-cluster center.

[0030] In the present invention, after the incremental data is extracted by the encoder in S1, the feature representation x is calculated. i The distance d from each cluster center is classified into clusters whose d is less than the cluster radius D, and the cluster information is updated. In streaming data clustering, a time decay function is used to weight data points. Given the timestamp of a data point, the formula for calculating the cluster center is and the formula for calculating the distance from the sample to the cluster center is as follows:

[0031]

[0032] f(t)=e -λ(T-t) ,λ>0

[0033] d = dist(x i ,c k )

[0034] where x i represents the i-th data point, c k represents the cluster center of the kth micro-cluster, f(t) is the time decay function, dist(x i ,c k ) represents x i With c kThe Euclidean distance between .

[0035] S3: Training the object recognition network

[0036] The present invention uses RCNN for target recognition and detection. RCNN (Region-based Convolutional Neural Networks) is a classic target detection method that combines the traditional candidate region generation method with a deep convolutional neural network to significantly improve the performance of target detection. The goal of RCNN is to identify the target category in the image and accurately locate the target bounding box. The specific training steps are as follows:

[0037] 3.1) The incremental data (x i ,y i ) is input into the pre-trained RCNN model. i is the data feature extracted from S1, y i is the true label (target category). Get the category probability distribution predicted by the model:

[0038]

[0039] in Represents sample x i The predicted probability of belonging to class k. C is the total number of classes.

[0040] 3.2) Calculate the cross entropy loss and use the cross entropy loss to train the target recognition network

[0041] For each candidate region i, the cross entropy loss can be expressed as:

[0042]

[0043] y ik is the true label y i One-hot encoding of is the model prediction x i The probability of belonging to class k.

[0044] In the target detection task, in addition to classifying the target, it is also necessary to accurately locate the target's bounding box. The goal of bounding box regression is to optimize the model by predicting the bounding box coordinates of each target so that the target is more accurately located. In this invention, smooth L1 loss is used as the bounding box regression loss, and the formula is as follows:

[0045]

[0046] The total loss function is:

[0047]

[0048] λ is a hyperparameter used to balance the weights of classification loss and regression loss.

[0049] S4: Detection and adaptation of new classes in the data stream

[0050] The present invention is applied to data stream scenarios. When new classes appear in incremental data, the new classes can be detected according to the dynamic micro-cluster model constructed in S2, and the micro-cluster model and the target recognition network can be adjusted.

[0051] 4.1) New Class Detection

[0052] like Figure 2 Medium i ,when The distance between the feature of the incremental data and the center of each cluster is greater than the radius of the cluster, that is, x i When it is impossible to enter the range of all clusters, it can be regarded as a new micro cluster c new , and x i ∈c new .

[0053] 4.2) Adjust the dynamic micro-cluster model and target recognition network

[0054] For the dynamic micro-cluster model, according to x i Add a new microcluster Like other microclusters, its radius is also D.

[0055] 4.3) For the target recognition network, modify the softmax layer of RCNN. The softmax function before the new class appears is as follows:

[0056]

[0057] Among them, z k is the original score for each category. When a new category appears, the modified softmax function is as follows:

[0058]

[0059] S5: Concept Drift Detection and Model Adaptation

[0060] Concept drift refers to a phenomenon in which the statistical distribution of data or the implicit pattern changes over time, resulting in a decrease in model performance. It usually occurs in problems that require processing dynamic data streams or non-stationary environments. The present invention uses a dynamic micro-cluster model to detect and adjust concept drift in the following specific ways:

[0061] 5.1) If Figure 2 Medium j , when a new data x appearsj , d(x j , c i ), c j , c j ), <D, d(x i and c j are both less than D from the cluster centers of the two micro-clusters c i and c j , that is, when simultaneously assigned to cluster c i and cluster c j , a concept drift is detected.

[0062] 5.2), For the feature extractor, reconstruct the dataset. Take the cluster centers of all clusters c1, c 2,... in the dynamic micro-cluster model as the nodes x c1 , x c2 , … representing the old state, and combine them with the current sliding window W to form a new dataset W new .

[0063]

[0064] Then, train the feature extractor according to the new dataset W new .

[0065] 5.3), For the dynamic micro-cluster model, after converting the cluster centers of the original micro-clusters into nodes and incorporating them into the sliding window as a new dataset, delete all clusters in the micro-cluster model. During the training process of the feature extractor, for each data in the new dataset W new , after extracting features, construct a new dynamic micro-cluster model in the way of S2. There are two types of data in W new : the cluster center data representing the old micro-cluster model, and the data in the new state W after concept drift. The new dynamic micro-cluster model constructed by these two types of data can well adapt to the features of the old data and the features after concept drift, and can effectively handle concept drift in the data stream.

[0066] Figure 3 is the system framework diagram for analyzing and detecting anomalies in a specific implementation of the online target recognition method based on data flow dynamic micro-cluster representation of the present invention.

[0067] In this embodiment, as Figure 3 shown, when the online target recognition system runs, it includes the following steps: (1) Extract intermediate features from the original representation through the encoder-decoder architecture; (2) Construct a small-sample dynamic micro-cluster model according to the intermediate features; (3) Detect whether new classes and concept drift occur according to the dynamic micro-cluster model, and adjust the model according to the structure; (4) Identify the target through the downstream target recognition network for the obtained intermediate representation.

[0068] Although the above describes the illustrative specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. An online target recognition method based on data stream dynamic micro-cluster characterization, characterized in that: The following steps are involved: (1) Collect MRI scan results of patients with and without Alzheimer's disease as (1) and build an unsupervised data feature extraction model. Adopt an encoder-decoder architecture. For each incremental data in the data stream, after embedding it into a vector representation, it is first encoded using a pre-trained encoder (such as Vision Transformer, ViT), and then decoded by a decoder (such as BigGAN). The feature extraction model is trained by calculating the reconstruction error between the features after decoding and before encoding. (2) Build a dynamic micro-cluster model and divide each incremental data into clusters. If a new class is detected, adjust the target recognition network. If concept drift occurs during the detection, readjust the data set and adjust the dynamic micro-cluster model. (3) Perform downstream target recognition by inputting the image features extracted in the previous step into the RCNN model for target detection tasks and performing training by calculating the cross entropy loss.

2. The online target recognition method according to claim 1, characterized in that in step (2), the problem of new classes and concept drift in incremental data is solved based on dynamic micro-clusters, wherein the method for processing the emergence of new classes is: 2.1) When the distance between the feature of the incremental data and the center of each cluster is greater than the radius of the cluster, that is, x i When it is impossible to enter the range of all clusters, it can be regarded as a new micro cluster c new , and x i ∈c new . 2.2) For the dynamic micro-cluster model, according to x i Add a new microcluster Like other microclusters, its radius is also D. 2.3) For the target recognition network, modify the softmax layer of RCNN. The softmax function before the new class appears is as follows: in, z k is the original score for each category. When a new category appears, the modified softmax function is as follows: The solution to concept drift is: 2.4) When a new data x appears j , which is related to the two microclusters c i and c j The distances between cluster centers are all less than D, that is, they can be simultaneously classified into cluster c i and cluster c j , it is detected as concept drift. 2.5) For the feature extractor, reconstruct the dataset. Set the cluster centers of all clusters c1, c2, ... in the dynamic micro-cluster model As node x representing the old state c1 ,x c2 ,…combine it with the sliding window W at this time as the new data set W new . According to the new data set W new Train the feature extractor. 2.6) For the dynamic micro-cluster model, after the cluster centers of the original micro-clusters are converted into nodes and incorporated into the sliding window as a new data set, all clusters in the micro-cluster model are deleted. During the training process of the feature extractor, for the new data set W new After extracting the features of each data, a new dynamic micro-cluster model is constructed using the S2 method. new There are two kinds of data in : the cluster center data representing the old micro-cluster model, and the data in the new state W after concept drift occurs. The new dynamic micro-cluster model constructed by these two types of data can well adapt to the characteristics of old data and the characteristics after concept drift occurs, and can well cope with concept drift in data streams.

Images