High-dimensional data stream new category identification and dynamic imbalance processing method based on deep learning

Abstract

A new category identification and dynamic imbalance processing method in an unsteady high-dimensional data stream based on deep learning comprises the following steps: preprocessing an input data stream, assuming that the data stream is a high-dimensional vector stream, and standardizing or normalizing each data sample; constructing a feature encoder based on a convolutional neural network, performing feature extraction on the high-dimensional sample of the data stream, and generating feature representation; a prototype ensemble classification framework is utilized, classification is carried out by maintaining prototypes of all categories, and category judgment is carried out according to the distance of newly input samples; the prototype set is dynamically updated through a feature correction mechanism, so that the prototype can better represent the current category distribution; in the classification process, new category detection is achieved through Mahalanobis distance calculation, if the distance between a sample and an existing category exceeds a set threshold value, the sample is judged as a new category, and a prototype set is dynamically expanded; a label distribution perception edge loss function is adopted, and the size of a loss edge is adaptively adjusted according to the number of category samples; a delay rebalance optimization strategy is utilized, a standard empirical risk minimization method is used during initial training, an LDAM loss function is switched to during later training, and the performance of the model is re-weighted and optimized for minority categories.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and big data processing technology, in particular to high

[0002] A new category recognition and dynamic imbalance processing method for high-dimensional data streams is used to process new categories in non-steady high-dimensional data streams.

[0003] Class recognition and dynamic imbalance problems, applicable to real-time monitoring, financial transaction analysis, network security and medical

[0004] Data processing and other fields. Background Art

[0005] In many modern applications, data streams are high-dimensional and non-stationary, and the statistical distribution of data varies with time.

[0006] Changes may introduce new categories or cause significant changes in the distribution of existing categories.

[0007] Learning models usually perform poorly when dealing with novel categories and class imbalance.

[0008] Most of these models rely on static training data and cannot adapt to dynamic changes in data distribution.

[0009] To address these challenges, existing technologies have proposed some methods, such as incremental learning and novelty detection.

[0010] for identifying new classes; and resampling and weighting methods for dealing with class imbalance.

[0011] Laws are often insufficient to solve both problems simultaneously when faced with a dynamic environment. Summary of the invention

[0012] The present invention aims to overcome the shortcomings of the prior art and provide a new category recognition method for high-dimensional data streams based on deep learning.

[0013] A method and device for processing dynamic imbalance.

[0014] The present invention realizes the classification of high-dimensional data streams through feature correction in a non-steady-state environment of structured data

[0015] and new category recognition, improving the recognition performance of the model in category imbalanced data. The device comprises a base

[0016] The feature representation and classification module of the convolutional neural network and a label assignment mechanism for multi-level classification are described.

[0017] The feature representation module includes a feature encoder and a multi-level prototype ensemble classifier to classify the input data stream.

[0018] The multi-level label assignment mechanism includes a label distribution-aware

[0019] Fine-grained separation mechanism to address the class imbalance problem.

[0020] The purpose of the present invention is achieved by the following technical solutions: A new category of high-dimensional data stream based on deep learning

[0021] The method for identifying and dealing with dynamic imbalance includes the following steps:

[0022] Step 1: Preprocess the input data stream. Assuming that the data stream is a high-dimensional vector stream,

[0023] The samples are standardized or normalized to reduce the impact of dimensional differences on model training.

[0024] Step 2: Construct a feature encoder based on a convolutional neural network to characterize the high-dimensional samples of the data stream.

[0025] Extract and generate feature representation.

[0026] Step 3: Use the prototype integrated classification framework to classify by maintaining the prototype of each category and

[0027] The category is determined based on the distance of the new input sample.

[0028] Step 4: Dynamically update the prototype set through the feature correction mechanism so that the prototype can better represent the current category distribution, thereby improving the generalization ability of the model.

[0029] Step 5: During the classification process, new category detection is achieved through Mahalanobis distance calculation. If the distance between a sample and an existing category exceeds the set threshold, the sample is judged as a new category and the prototype set is dynamically expanded.

[0030] Step 6: Use the label distribution-aware edge loss function to adaptively adjust the loss edge size according to the number of category samples, enhance the generalization ability of minority categories, and further optimize the category imbalance problem.

[0031] Step 7: Use the delayed rebalancing optimization strategy, use the standard empirical risk minimization method in the initial training, switch to the LDAM loss function in the later training, and re-weight the minority categories to optimize the model performance.

[0032] Furthermore, the data stream preprocessing described in step (1) includes:

[0033] Assume that the data stream is a high-dimensional vector stream, represented by X = {x1, x2, ..., x T},in is the sample at time t, and d is the dimension of the data. Each data sample is standardized or normalized so that it has the same scale in different dimensions to reduce the impact of dimensional differences on model training.

[0034] Furthermore, the convolutional neural network feature extraction in step (2) includes:

[0035] Construct a multi-layer convolutional neural network (CNN), which consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract local features in the data stream, the pooling layer is used to reduce the dimension of the feature, and the fully connected layer is used to generate the final feature representation for each data sample x t , extract its feature representation h through the CNN network t ,Right now:

[0036] h t =CNN(x t )#(1)

[0037] Among them, h t Represents the feature vector of the data sample at time t.

[0038] Furthermore, the prototype integration framework classification as described in step (3) includes:

[0039] Based on the feature representation h extracted by CNN t , using the CPE framework for classification. The CPE framework maintains a set of prototypes Each prototype P k Represents the center of a category for a new data sample h t , calculate its Euclidean distance with each prototype, and classify according to the prototype with the minimum distance, that is:

[0040]

[0041] in, For sample h t The predicted category.

[0042] Further, dynamically updating the prototype set according to the feature correction mechanism as described in step (4) includes:

[0043] When a new sample h t When correctly classified, update the prototype P of the corresponding category k , so that the prototype better represents the current category distribution. The update formula is:

[0044] p k←αp k +(1-α)h t #(3)

[0045] Among them, α is the learning rate parameter, which controls the speed of prototype update.

[0046] Furthermore, the new category detection mechanism described in step (5) includes:

[0047] Mahalanobis distance calculation: In the classification process, in addition to calculating the sample h t In addition to the Euclidean distance with each prototype, the Mahalanobis distance between the sample and the prototype is also calculated to measure the similarity between the sample and the existing category.

[0048]

[0049] Among them, S k is the covariance matrix of class k.

[0050] Determination of new categories: If the Mahalanobis distance between a sample and all existing prototypes exceeds a certain set threshold τ, it is considered that the sample may belong to a new category.

[0051]

[0052] When new categories are detected, the prototype set is dynamically expanded to add new prototypes k =h t To represent the new category, and the sample h t Mark this new category.

[0053] Furthermore, the label distribution aware marginal loss function (LDAM) as described in step (6) includes:

[0054] Loss function definition: During model training, the label distribution-aware edge loss function (LDAM) is used to adjust the edge size of each category. The specific loss function is:

[0055]

[0056] in, C is an adjustable hyperparameter, n y is the number of samples of category y, z y is the model's prediction score for category y.

[0057] Adaptive adjustment of margins: Dynamically adjust the margin size according to the distribution of category samples, and the margin of minority categories is larger, thereby improving its generalization ability. Through this adaptive adjustment, the model is better able to handle the problem of category imbalance.

[0058] Furthermore, the delayed rebalancing optimization strategy (DRW) as described in step (7) includes:

[0059] Initial training stage: In the early stage of model training, the standard empirical risk minimization (ERM) method is used for training, that is, the standard cross entropy loss function is used:

[0060]

[0061] No class weighting or resampling is performed in this stage, the purpose is to allow the model to learn basic feature representations.

[0062] Late optimization stage: In the later stage of training, switch to LDAM loss function and reweight the minority classes to further optimize the classification performance of the model. The specific optimization steps are:

[0063]

[0064] Among them, n y is the number of samples of category y.

[0065] The present invention is used to solve the problems of category imbalance and new category detection when the data stream changes dynamically. The present invention combines convolutional neural network (CNN), prototype integration framework (CPE), new category detection mechanism, label distribution aware edge loss function (LDAM) and delayed rebalancing optimization strategy (DRW). CNN is used to extract data features, and combined with the CPE framework for classification and new category detection, LDAM is used to dynamically adjust the edge size in the loss function, and DRW is used for re-weighted optimization of minority categories. This method can be widely used in financial monitoring, network security, medical data processing and other fields, and can significantly improve the performance of new category recognition and minority category detection.

[0066] Compared with the prior art, the invention has the following beneficial effects: the invention uses prototype integration and incremental learning strategies, and the model can detect and adapt to new categories appearing in the data stream in real time. Through the LDAM loss function and DRW optimization strategy, the model is more stable when processing a small number of categories, avoiding overfitting or underfitting. The model of the invention can respond to changes in data streams in real time, and is suitable for application scenarios that require efficient real-time processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. In the absence of conflict, the features of the following embodiments and implementations can be combined with each other.

[0069] Example 1

[0070] This embodiment relates to a production line product quality data stream classification method using the high-dimensional data stream new category recognition and dynamic imbalance processing method based on deep learning of the present invention.

[0071] Step (1) obtaining a labeled sample data set and an unlabeled sample data set to generate a production line product quality data set and performing data stream preprocessing;

[0072] Specifically, the step (1) includes the following sub-steps:

[0073] Step (1.1) assumes that the quality data stream generated by the production line is a high-dimensional vector stream, represented by X = {x1, x2, ..., x T},in is the sample at time t, and d is the dimension of the data. The data stream contains high-dimensional parameters such as temperature, pressure, and size, which are used to reflect the quality status of the product in the production process in real time.

[0074] Step (1.2) standardizes or normalizes each data sample so that it has the same scale in different dimensions to reduce the impact of dimensional differences on model training.

[0075] Step (2) constructs a feature encoder based on a convolutional neural network, extracts features from high-dimensional samples of the data stream, and generates feature representations. Specifically, step (2) includes the following sub-steps:

[0076] Step (2.1) constructs a multi-layer convolutional neural network (CNN), which consists of multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract local features in the data stream, the pooling layer is used to reduce the dimension of the features, and the fully connected layer is used to generate the final feature representation.

[0077] Step (2.2) For each data sample x t , extract its feature representation h through the CNN network t ,Right now:

[0078] h t =CNN(x t )#(1)

[0079] Among them, h t Represents the quality feature vector at time t.

[0080] Step (3) utilizes the prototype ensemble classification framework to perform classification by maintaining the prototype of each category and making category decisions based on the distance of the new input sample.

[0081] Specifically, the step (3) includes the following sub-steps:

[0082] Step (3.1) The feature representation h extracted based on CNN t , using the CPE framework for classification. The CPE framework maintains a set of prototypes Each prototype P k Represents a quality category (e.g., normal, slightly defective, severely defective).

[0083] Step (3.2) for the new data sample h t , calculate its Euclidean distance with each prototype, and classify according to the prototype with the minimum distance, that is:

[0084]

[0085] in, For sample h t The prediction quality category, P k is the prototype representing the kth class.

[0086] Step (4) dynamically updates the prototype set through the feature correction mechanism so that the prototype can better represent the current category distribution, thereby improving the generalization ability of the model. The details are as follows:

[0087] When a new quality sample h t When correctly classified, update the prototype P of the corresponding category k , so that the prototype better represents the current category distribution. The update formula is:

[0088] p k ←αp k +(1-α)h t #(3)

[0089] Among them, α is the learning rate parameter, which controls the speed of prototype update. Through this online update method, the model can adapt to small quality fluctuations in the production process and maintain the accuracy of real-time classification.

[0090] In step (5), during the classification process, new category detection is achieved through Mahalanobis distance calculation. If the distance between a sample and an existing category exceeds a set threshold, the sample is judged as a new category and the prototype set is dynamically expanded.

[0091] Specifically, the step (5) includes the following sub-steps:

[0092] Step (5.1) Mahalanobis distance calculation:

[0093] In the quality classification process, in addition to calculating the sample h t In addition to the Euclidean distance with each prototype, the Mahalanobis distance between the sample and the prototype is also calculated to measure the similarity between the sample and the existing category.

[0094]

[0095] Among them, S k is the covariance matrix of category k, which is used to capture the distribution information of the category.

[0096] Step (5.1) Determination of new categories: If the Mahalanobis distance between a sample and all existing prototypes exceeds a certain set threshold τ, it is considered that the sample may belong to a new category (i.e., a new type of quality defect). When new categories are detected, the prototype set is dynamically expanded to add new prototypes k =h t To represent the new category, and the sample h t The process realizes the adaptive identification of sudden quality problems in the production process.

[0097] Step (6) uses the label distribution-aware marginal loss function to adaptively adjust the loss margin size according to the number of category samples, enhance the generalization ability of minority categories, and further optimize the category imbalance problem.

[0098] Specifically, step (6) includes the following sub-steps:

[0099] Step (6.1) loss function definition:

[0100] During model training, the label distribution-aware edge loss function (LDAM) is used to adjust the edge size of each category. The specific loss function is:

[0101]

[0102] in, is the edge size adaptively adjusted according to the distribution of category samples, and N is the batch size. This loss function helps improve the model's ability to detect minority quality defect categories by increasing the edge of minority categories.

[0103] Step (6.2) Adaptive adjustment of the edge: Dynamically adjust the edge size according to the distribution of the category samples, and the edge of the minority quality defect category is larger, thereby improving its generalization ability. Through this adaptive adjustment, the model can better identify the rare but important defect types in the production process and reduce the misjudgment rate.

[0104] Step (7) utilizes the delayed rebalancing optimization strategy, uses the standard empirical risk minimization method in the initial training, switches to the LDAM loss function in the later training, and reweights the minority classes to optimize the model performance.

[0105] Specifically, the step (7) includes the following sub-steps:

[0106] Step (7.1) Initial training stage: In the early stage of model training, the standard empirical risk minimization (ERM) method is used for training, that is, the standard cross entropy loss function is used:

[0107]

[0108] No class weighting or resampling is performed at this stage, the goal is to allow the model to learn the basic feature representation of product quality.

[0109] Step (7.2) Late optimization stage: In the later stage of training, switch to the LDAM loss function and reweight the minority class to further optimize the classification performance of the model. The specific optimization steps are:

[0110]

[0111] Among them, w c is the weight of category c, and |C| is the total number of categories. This strategy reweights samples of rare defect categories to make the model more stable in detecting rare quality defects.

[0112] Step (8) Real-time quality monitoring and adaptation to new defects In the application scenario of product quality inspection, new quality defects may appear with changes in production processes or the introduction of new product specifications. To cope with this situation, the model has real-time monitoring capabilities, can dynamically identify new defect categories based on new product quality data streams and expand classifiers in a timely manner, thereby ensuring continuous monitoring of production quality and rapid response to abnormalities.

[0113] Example 2

[0114] This embodiment is based on the ISIC 2019 skin disease image dataset to verify the effectiveness of the method of the present invention. The ISIC 2019 dataset contains multiple skin disease categories, and the data is high-dimensional, complex, and unbalanced. The goal is to effectively identify newly emerging skin disease categories through the method of the present invention and improve the detection accuracy of minority categories.

[0115] Experimental setup

[0116] 1. Data preprocessing

[0117] Skin disease images were extracted from the ISIC 2019 dataset, scaled to a uniform size (224x224 pixels), and normalized so that each pixel value was between 0 and 1.

[0118] 2. Feature extraction and classification

[0119] Use the pre-trained ResNet-50 model as the basic network to extract features from the pre-processed image. The extracted feature vector h t Classification is performed through the prototype integration framework (DCPE) of the present invention.

[0120] Initially, the prototype set contains the existing skin disease categories in the ISIC 2019 dataset, and a prototype p is generated for each category. k As data flows in, the model dynamically updates these prototypes.

[0121] 3. New Category Detection

[0122] During the data stream input process, it is assumed that some samples belong to new skin disease categories that have not been seen before. By calculating the Mahalanobis distance between these samples and the existing prototypes, when the distance exceeds the threshold τ, it is determined to be a new category and added to the prototype set.

[0123] 4. Loss Function and Optimization Strategy

[0124] The LDAM loss function is used to perform adaptive margin adjustment on minority class samples to improve the model’s recognition ability for these classes.

[0125] In the initial training phase, the standard empirical risk minimization (ERM) is used for model training, and in the later stage, the DRW strategy is switched to reweight the minority categories to further optimize the performance of the model on these categories.

[0126] Evaluation indicators

[0127] In order to evaluate the performance of the method of the present invention in the detection of ISIC 2019 skin disease dataset, the following evaluation indicators are used:

[0128] 1. Precision

[0129] It measures the proportion of correct predictions made by the model when predicting a certain class. It is defined as follows:

[0130]

[0131] Among them, TP is the number of new category samples detected correctly, and FP is the number of known category samples that are incorrectly detected as new categories.

[0132] 2. Recall

[0133] It measures the proportion of samples that the model can correctly identify as belonging to a certain class. It is defined as follows:

[0134]

[0135] Among them, FN is the actual number of new category samples that are incorrectly classified as known categories.

[0136] 3. New category misclassification rate (Mnew)

[0137] Measures the proportion of new category samples that are misclassified, defined as follows:

[0138]

[0139] Among them, Nc is the total number of new category samples.

[0140] 4. Known category misclassification rate (Fnew)

[0141] Measures the proportion of known category samples that are misclassified as new categories, defined as follows:

[0142]

[0143] Among them, N is the total number of samples, and Nc is the number of new category samples.

[0144] 5. Accuracy

[0145] Measures the proportion of samples correctly classified by the model as a whole. It is defined as follows:

[0146]

[0147] Among them, Anew is the number of correctly classified new category samples, Aknown is the number of correctly classified known category samples, and m is the total number of samples.

[0148] Experimental Results

[0149] Through the test of ISIC 2019 data set, the method of the present invention shows significant advantages in multiple evaluation indicators. The specific results are as follows:

[0150] Precision and recall: The method of the present invention shows high precision and recall in the detection of new categories and minority categories, indicating its effectiveness in practical applications.

[0151] New Class Misclassification Rate (Mnew): The misclassification rate of new classes is low, proving that the model has high accuracy in detecting new classes of skin diseases.

[0152] Known category misclassification rate (Fnew): The proportion of known category samples that are misclassified as new categories is low, indicating that the model can distinguish between known categories and new categories well.

[0153] Accuracy: The overall classification accuracy of the model is high, indicating that it has excellent generalization ability in multi-category scenarios.

[0154] Summarize

[0155] Through experiments on the ISIC 2019 skin disease dataset, the method of the present invention performs well in detecting new categories and handling category imbalance. The experimental results show that this method can not only effectively identify newly emerging skin disease categories, but also significantly improve the detection accuracy of minority categories, and has high practical application value.

Claims

1. A method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning, characterized in that: The method comprises the following steps: Step 1. Preprocess the input data stream. Assuming that the data stream is a high-dimensional vector stream, standardize or normalize each data sample to reduce the impact of dimensional differences on model training. Step 2. Construct a feature encoder based on a convolutional neural network to extract features from high-dimensional samples of the data stream and generate feature representations; Step 3. Use the prototype ensemble classification framework to perform classification by maintaining the prototype of each category and make category decisions based on the distance of the new input sample; Step 4. Dynamically update the prototype set through the feature correction mechanism so that the prototype can better represent the current category distribution, thereby improving the generalization ability of the model; Step 5. During the classification process, new category detection is achieved through Mahalanobis distance calculation. If the distance between a sample and an existing category exceeds the set threshold, the sample is judged as a new category and the prototype set is dynamically expanded; Step 6. Use the label distribution-aware edge loss function to adaptively adjust the loss edge size according to the number of category samples, enhance the generalization ability of minority categories, and further optimize the category imbalance problem; Step 7. Use the delayed rebalancing optimization strategy to use the standard empirical risk minimization method in the initial training, switch to the LDAM loss function in the later training, and reweight the minority class to optimize the model performance.

2. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1 is characterized in that: The data stream preprocessing in step 1 specifically includes: For each data sample x t Standardize and normalize so that Where μ and σ are the mean and standard deviation of the features respectively; each data sample is standardized and normalized so that it has the same scale in different dimensions to reduce the impact of dimensional differences on model training.

3. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1 is characterized in that: The convolutional neural network feature extraction in step 2 includes: Construct a multi-layer convolutional neural network (CNN), which consists of multiple convolutional layers, pooling layers, and fully connected layers; the convolutional layer is used to extract local features in the data stream, the pooling layer is used to reduce the dimension of the feature, and the fully connected layer is used to generate the final feature representation for each data sample x t , extract its feature representation h through the CNN network t ,Right now: h t =CNN(x t )#(1) Among them, h t Represents the feature vector of the data sample at time t.

4. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1 is characterized in that: The classification of data samples by the prototype integrated framework CPE described in step 3 includes: based on the feature representation h extracted by CNN t , using the prototype ensemble (CPE) framework for classification; the CPE framework maintains a prototype collection Each prototype P k Represents the center of a category for a new data sample h t , calculate its Euclidean distance with each prototype, and classify according to the prototype with the minimum distance, that is: in, For sample h t The predicted category.

5. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1, characterized in that: Step 4 describes the dynamic updating of the prototype set according to the feature correction mechanism, including: when a new sample h t When correctly classified, update the prototype P of the corresponding category k , so that the prototype better represents the current category distribution; the update formula is: p k ←αp k +(1-a)h t #(3) Among them, α is the learning rate parameter, which controls the speed of prototype update.

6. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1, characterized in that: The new category detection mechanism described in step 5 includes: Mahalanobis distance calculation: In the classification process, in addition to calculating the sample h t In addition to the Euclidean distance with each prototype, the Mahalanobis distance between the sample and the prototype is also calculated to measure the similarity between the sample and the existing category; Among them, S k is the covariance matrix of category k; Determination of new categories: If the Mahalanobis distance between a sample and all existing prototypes exceeds a certain set threshold τ, then it is considered that the sample may belong to a new category; When new categories are detected, the prototype set is dynamically expanded to add new prototypes k =h t To represent the new category, and the sample h t Mark this new category.

7. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1, characterized in that: The label distribution-aware marginal loss function LDAM described in step 6 includes: Loss function definition: During the model training process, the label distribution-aware edge loss function (LDAM) is used to adjust the edge size of each category. The specific loss function is: in, C is an adjustable hyperparameter, n y is the number of samples of category y, z y is the model's prediction score for category y; Adaptive adjustment of margins: The margin size is dynamically adjusted according to the distribution of category samples, and the margin of minority categories is larger, thereby improving its generalization ability; through this adaptive adjustment, the model is able to better handle the problem of category imbalance.

8. The method for identifying new categories and processing dynamic imbalance in non-steady-state high-dimensional data streams based on deep learning according to claim 1, characterized in that: The delayed rebalancing optimization strategy DRW described in step 7 includes: Initial training stage: In the early stage of model training, the standard empirical risk minimization (ERM) method is used for training, that is, the standard cross entropy loss function is used: No class weighting or resampling is performed at this stage, so that the model can learn basic feature representations; Late optimization stage: In the later stage of training, switch to the LDAM loss function and reweight the minority categories to further optimize the classification performance of the model. The specific optimization steps are: Among them, n y is the number of samples of category y.

Images