Matrix sketch based device dynamic feature selection method and system

Abstract

The application discloses a device dynamic feature selection method and system based on a matrix Sketch, and relates to the technical field of device feature data processing. The method comprises the following steps: inserting data into an initialized matrix Sketch when device state feature data is updated; performing a space release operation and updating the matrix Sketch; performing sparse learning feature selection according to a sparse feature selection algorithm; and adjusting the sparse learning feature selection frequency according to a dynamic feature selection frequency analysis evaluation result. The application performs sparse learning on the matrix Sketch. Compared with sparse learning on original data, the calculation overhead of sparse learning can be greatly reduced due to the small space of the matrix Sketch, the calculation efficiency of dynamic feature selection is improved, and the applicability is improved, so that the problem of low calculation efficiency of dynamic feature selection according to network dynamic environment changes in the prior art and the resulting insufficient applicability are solved.

Description

Technical Field

[0001] This invention relates to the field of equipment feature data processing technology, and in particular to a method and system for selecting dynamic features of equipment based on matrix Sketch. Background Technology

[0002] With the widespread adoption of technologies such as the Internet of Things (IoT), 5G communication, and the Industrial Internet, the number of devices is growing exponentially, generating massive amounts of monitoring data per second. Traditional methods require storing all data and performing offline feature selection calculations, resulting in high storage and computation costs. Industry 4.0 and smart city scenarios demand systems that can respond to changes in device status in real time. Traditional static feature selection methods cannot adapt to the dynamic shifts in data distribution, leading to the invalidation of feature subsets. Edge devices have limited computing power, making it difficult to support complex model operations. Therefore, efficient feature selection at the data source is necessary to reduce transmission and computational burdens.

[0003] Existing methods for dynamic feature selection in devices employ the following techniques: filtering, which selects features based on statistical characteristics but ignores interactions between features and cannot adapt to dynamic environments; wrapping, which relies on downstream task performance evaluation, incurs high computational overhead, and is unsuitable for streaming data; embedding, which embeds feature selection during model training but requires frequent model retraining for streaming data; sliding window, which retains only data within the most recent window but still requires storing the entire window data, resulting in high memory consumption; and online learning algorithms, which update the model sample by sample but are inefficient for high-dimensional data.

[0004] For example, the invention patent with publication number CN115952397A discloses a method, apparatus, device, storage medium, and program product for determining risk points, which relates to the field of big data processing technology. The method includes: acquiring alarm information generated within the current period; classifying each alarm information according to attribute information to obtain multiple alarm information sets; calculating the hash value of each feature value in the matched alarm information set through each sketch processor, and statistically analyzing the acquired hash values ​​to obtain a processor matrix; comparing each processor matrix with its corresponding threshold matrix to obtain the spatial location of the abnormal target, and determining the risk point based on the spatial location of the abnormal target.

[0005] For example, the invention patent with announcement number CN114330504B discloses a method for detecting malicious network traffic based on Sketch. Its implementation steps are as follows: generating a sample set and a test set for each device; generating a key feature matrix and a non-key feature matrix for each sample matrix; processing each set of non-key feature matrices for each device using Sketch; constructing three models; training the three models using the three key feature matrices for each device respectively; selecting the best model as the malicious traffic detection model; and using the malicious traffic detection model to detect network traffic.

[0006] However, in the process of implementing the inventive technical solution in the embodiments of this application, it was found that the above-mentioned technology has at least the following technical problems:

[0007] In existing technologies, traditional methods for streaming data processing (such as sliding windows and offline modeling) require storing the entire original data or the complete feature matrix, resulting in storage overhead increasing linearly with the amount and dimensionality of data. For example, in the Industrial Internet of Things (IIoT), monitoring 100,000 devices generates 100-dimensional data per second, with storage costs reaching 10GB per second. Offline feature selection requires multiple traversals of the entire dataset, consuming significant computational resources and failing to provide real-time responses. Furthermore, it suffers from low computational efficiency in dynamically selecting features based on changes in the network environment, leading to insufficient applicability. Summary of the Invention

[0008] This application provides a device dynamic feature selection method and system based on matrix Sketch, which solves the problem of insufficient applicability caused by low computational efficiency in dynamic feature selection based on changes in the network dynamic environment in the prior art, and achieves the effect of improving the computational efficiency of dynamic feature selection while improving applicability.

[0009] This application provides a device dynamic feature selection method based on a matrix sketch, including the following steps: initializing the matrix sketch; inserting data into the initialized matrix sketch when device status feature data is updated; if the data insertion in the matrix sketch is full, performing a space release operation and updating the matrix sketch; setting a sparse feature selection algorithm based on the historical device status feature data area and the current device status feature data area of ​​the matrix sketch; performing sparse learning feature selection based on the sparse feature selection algorithm; performing dynamic feature selection frequency analysis and evaluation; and adjusting the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results.

[0010] Furthermore, when the device status feature data is updated, data is inserted into the initialization matrix Sketch. Specifically, this includes: setting the feature dimension, original data size, and compression ratio according to a predefined matrix initialization scheme; forming a large matrix from the original device performance feature data based on the feature dimension and original data size; obtaining the number of matrix rows in the initialization matrix Sketch based on the compression ratio; setting the initialization matrix Sketch using the feature dimension and the number of matrix rows; and directly inserting and storing the performance feature data into the initialization matrix Sketch in a first-in-first-out manner when the performance feature data is updated sequentially.

[0011] Furthermore, a space release operation is performed and the matrix Sketch is updated. Specifically, this includes: when device status feature data is updated, checking whether the number of rows currently stored in the matrix Sketch has reached the maximum number of rows; if all rows in the matrix Sketch are occupied, a space release operation is triggered; performing singular value decomposition on the matrix Sketch to obtain a singular value diagonal matrix and a right singular matrix, and then reducing the size according to preset parameters. The singular value diagonal matrix is ​​truncated and the matrix Sketch is updated; the matrix Sketch is divided into the historical device status feature data area and the current device status feature data area according to the matrix segmentation weight; the matrix segmentation weight is dynamically adjusted according to the feature fluctuation amplitude of the matrix Sketch.

[0012] Furthermore, the specific process of dynamically adjusting the matrix segmentation weights based on the characteristic fluctuation amplitude of the matrix Sketch is as follows: The device status characteristic data recorded in the sliding window of the matrix Sketch is processed according to the sliding window size. If a space release operation is performed and the matrix Sketch is updated, the device status characteristic data in the sliding window is re-recorded. For each feature dimension of the device status characteristic data in the sliding window, the standard deviation of the device status characteristic data in the sliding window under each feature dimension is calculated, and then the overall volatility is obtained by summing and averaging. The local volatility is obtained by analyzing the overall volatility based on preset local salient values. If the local volatility is greater than the upper limit of the local volatility range threshold, the matrix segmentation weights are reduced during the next predefined number of space release operations and matrix Sketch updates. If the local volatility is within the local volatility range threshold, the matrix segmentation weights are maintained during the next predefined number of space release operations and matrix Sketch updates. If the local volatility is less than the lower limit of the local volatility range threshold, the matrix segmentation weights are increased during the next predefined number of space release operations and matrix Sketch updates.

[0013] Furthermore, a sparse feature selection algorithm is set based on the historical device status feature data area and the current device status feature data area of ​​the Sketch matrix. Specifically, this includes: performing matrix transformations on the historical and current device status feature data areas to obtain the orthogonal basis matrices corresponding to the historical and current device status feature data areas; and setting a sparse feature selection algorithm based on these matrices. The objective function of the sparse feature selection algorithm is: ; ;in, This is the historical device status feature data area of ​​the Matrix Sketch. This is the current device status feature data area of ​​the Matrix Sketch. It is the orthogonal basis matrix corresponding to the historical device status feature data area of ​​the matrix Sketch. It is the orthogonal basis matrix of the current device state feature data area of ​​the matrix Sketch; It is a feature indicator matrix that merges the current device status feature data area of ​​the MatrixSketch and the historical device status feature data area of ​​the MatrixSketch. These are the weight parameters of the manifold regularization term; It is the weight parameter of the sparse regularization term, which is used for feature selection based on the sparse feature selection algorithm.

[0014] Furthermore, feature selection is performed using a sparse feature selection algorithm, specifically including: calculating the corresponding feature score for each row of the feature indicator matrix, as shown in the following formula: ; Indicates the first The feature score of the i-th feature is used to correspond to the i-th feature in the feature indicator matrix. The L2 norm of the row; the obtained feature scores are arranged in non-increasing order, and the device features corresponding to the feature scores are selected and continuously measured according to the predefined number of feature selections.

[0015] Furthermore, a dynamic feature selection frequency analysis and evaluation is performed, specifically including: collecting and obtaining the feature score of the current sparse learning feature selection and the feature score of the previous sparse learning feature selection; directly extracting the predefined fluctuation threshold, the first weighting factor of feature change, and the second weighting factor of feature change from the device dynamic feature selection database; performing a difference analysis between the feature score of the current sparse learning feature selection and the corresponding feature score of the previous sparse learning feature selection, performing a proportion analysis with the predefined fluctuation threshold of the corresponding feature score, performing a summation and mean analysis, and then adjusting the weight with the first weighting factor of feature change to obtain the first component of dynamic feature selection frequency analysis; performing a difference analysis between the feature score of the current sparse learning feature selection and the corresponding average feature score of the full data, performing a proportion analysis with the corresponding average feature score of the full data, performing a summation and mean analysis, and then adjusting the weight with the second weighting factor of feature change to obtain the second component of dynamic feature selection frequency analysis; and obtaining the dynamic feature selection frequency analysis evaluation mutation value based on the first and second components of dynamic feature selection frequency analysis.

[0016] Furthermore, the sparse learning feature selection frequency is adjusted based on the dynamic feature selection frequency analysis evaluation results. Specifically, this includes: if the dynamic feature selection frequency analysis evaluation mutation value is less than or equal to the dynamic feature selection frequency analysis evaluation mutation threshold, then the weight parameter of the manifold regularization term is gradually reduced to the predefined lower limit of the weight parameter of the manifold regularization term in the following predefined adjustment counts; if the dynamic feature selection frequency analysis evaluation mutation value is greater than the dynamic feature selection frequency analysis evaluation mutation threshold, then the weight parameter of the manifold regularization term is gradually reduced to the predefined lower limit of the weight parameter of the manifold regularization term in the following predefined adjustment counts, the dynamic feature selection frequency analysis evaluation mutation value is subtracted from the dynamic feature selection frequency analysis evaluation mutation threshold to obtain the dynamic feature selection frequency analysis evaluation mutation difference, and the sparse learning feature selection time interval is reduced based on the dynamic feature selection frequency analysis evaluation mutation difference.

[0017] Furthermore, adjusting the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results also includes: if the dynamic feature selection frequency analysis and evaluation mutation value is less than or equal to the dynamic feature selection frequency analysis and evaluation mutation threshold, then the number of consecutive cumulative judgments is recorded; if the dynamic feature selection frequency analysis and evaluation mutation value is greater than the dynamic feature selection frequency analysis and evaluation mutation threshold, then the number of consecutive cumulative judgments is reset to zero; if the number of consecutive cumulative judgments is less than the number of consecutive cumulative judgments threshold, then no adjustment is made; if the number of consecutive cumulative judgments is equal to or greater than the number of consecutive cumulative judgments threshold, then the sparse learning feature selection time interval is gradually increased to the upper limit of the sparse learning feature selection time interval in the following predefined adjustment number of times.

[0018] This application provides a device dynamic feature selection system based on matrix Sketch, including a matrix Sketch initialization module, a matrix Sketch update module, a matrix Sketch dynamic feature selection module, and a matrix Sketch dynamic feature selection adjustment module: The matrix Sketch initialization module initializes the matrix Sketch and inserts data into the initialized matrix Sketch when device status feature data is updated; the matrix Sketch update module releases space and updates the matrix Sketch if the data insertion in the matrix Sketch is full; the matrix Sketch dynamic feature selection module sets a sparse feature selection algorithm based on the historical device status feature data area and the current device status feature data area of ​​the matrix Sketch, and performs sparse learning feature selection based on the sparse feature selection algorithm; the matrix Sketch dynamic feature selection adjustment module performs dynamic feature selection frequency analysis and evaluation, and adjusts the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results.

[0019] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0020] 1. This invention performs sparse learning on a matrix Sketch. Compared to sparse learning on the original data, the computational overhead of sparse learning can be significantly reduced because the matrix Sketch space is small. This achieves the effect of improving the computational efficiency of dynamic feature selection and improving its applicability. It solves the problem of insufficient applicability caused by the low computational efficiency of dynamic feature selection based on changes in the network dynamic environment in the prior art.

[0021] 2. By dividing the matrix Sketch into historical and current data regions, orthogonal basis matrices are extracted for each region. Combined with manifold regularization and sparse regularization, an objective function is constructed for feature selection. The historical data region preserves long-term feature patterns through compression, and manifold regularization constrains the local data structure, preventing feature failure due to data distribution drift. The current data region focuses on real-time data, using sparse regularization to filter key features. This significantly reduces the computational cost of matrix decomposition through the compact structure of the matrix Sketch, resulting in significantly reduced resource consumption compared to full-data sparse learning. Simultaneously, manifold constraints preserve the geometric relationships in the data space, reducing noise interference. Furthermore, dynamic adjustment of regularization parameters balances the impact of historical and real-time data, adapting to different scenario requirements. Specifically, in industrial IoT device monitoring, indicators such as CPU utilization and disk I / O latency of server clusters dynamically change with load. Real-time capture of abnormal fluctuations (such as sudden CPU spikes) is possible, while historical data constraints avoid misjudging short-term noise, reducing false alarm rates and improving anomaly detection efficiency.

[0022] 3. By analyzing the variation range of feature scores, fluctuation thresholds, and comparison with the full dataset, the frequency of sparse learning execution is dynamically adjusted. When feature fluctuations are drastic (e.g., sudden changes in indicators due to network attacks), the detection frequency is increased for rapid response; during periods of stable data, the frequency is reduced to minimize redundant computation. Simultaneously, by combining error-driven mechanisms (e.g., feature score difference ratio analysis) and manifold consistency verification, the failure of feature subsets due to low-frequency updates is avoided, ensuring the accuracy of downstream tasks (e.g., anomaly detection). Furthermore, the adjustment rhythm of parameters is controlled by the threshold of consecutive judgment counts, preventing instability caused by frequent changes. In the performance monitoring of local computing server clusters, the performance data of these clusters exhibits periodic fluctuations due to external business influences. The dynamic adjustment mechanism can increase the feature selection frequency during periods of high business activity, incorporating multiple key indicators in real time; and reduce the computational load and extend equipment lifespan during stable business periods, while maintaining monitoring accuracy. Attached Figure Description

[0023] Figure 1 A flowchart of a device dynamic feature selection method based on matrix Sketch provided in this application embodiment;

[0024] Figure 2 This is a structural diagram illustrating how the MatrixSketch is divided into a historical device status feature data area and a current device status feature data area.

[0025] Figure 3 This is a structural diagram of a device dynamic feature selection system based on matrix Sketch provided in an embodiment of this application. Detailed Implementation

[0026] This application provides a device dynamic feature selection method and system based on matrix Sketch, which solves the problem of insufficient applicability caused by low computational efficiency in dynamic feature selection based on changes in the network dynamic environment in the prior art, and achieves the effect of improving the computational efficiency of dynamic feature selection while improving applicability.

[0027] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0028] like Figure 1 The diagram shows a flowchart of a device dynamic feature selection method based on a matrix sketch provided in this application. This method is applied to a device dynamic feature selection system based on a matrix sketch and includes the following steps: initializing the matrix sketch; inserting data into the initialized matrix sketch when device status feature data is updated; if the data insertion in the matrix sketch is full, performing a space release operation and updating the matrix sketch; setting a sparse feature selection algorithm based on the historical device status feature data area and the current device status feature data area of ​​the matrix sketch; performing sparse learning feature selection based on the sparse feature selection algorithm; performing dynamic feature selection frequency analysis and evaluation; and adjusting the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results.

[0029] In this embodiment, a medium-sized data center generates terabytes of device monitoring data daily. Using traditional methods (such as full storage + offline modeling) would be prohibitively expensive in terms of storage and computing resources. The importance of features during peak and off-peak periods of server load may differ significantly, requiring dynamic adjustment of monitoring metrics. IoT sensor nodes typically have less than 1MB of memory, insufficient to store all data, necessitating lightweight algorithms. Network device measurements involve massive performance metrics across multiple dimensions and aspects, including network, system, environment, and applications, resulting in high measurement costs. Feature selection is the process of eliminating irrelevant, erroneous, noisy, and redundant features. It selects a subset of features with greater information content and fewer dimensions from all available features, ensuring the subset contains the original features of all features for better interpretability. Feature selection is widely used in clustering, classification, and anomaly detection. Existing feature selection methods mainly include wrapping and filtering methods.

[0030] The wrapper method treats feature selection as a search problem, determining the optimal feature combination by evaluating the performance of different feature subsets. For a given feature selection problem, such as selecting a subset of features, common search algorithms include forward-selection greedy search and backward-selection greedy search. These algorithms continuously add or remove features, or combine and exchange feature subsets, by calculating the correlation between feature subsets and the objective function, to explore the feature subset space and find the optimal feature combination. However, this search process typically requires significant computational resources and time, making it unsuitable for measurement scenarios where feature distributions change over time. Because the entire search process requires calculating the performance of subsets, it often needs to be trained in conjunction with downstream tasks, resulting in a large computational burden and a tendency to overfit. Furthermore, the feature selection process of these methods is usually tightly coupled with downstream tasks, making it difficult to generalize to other tasks.

[0031] Filtering methods offer advantages in efficiency and versatility compared to wrapping methods. They link the characteristics of the feature metrics themselves with the target loss function for selection, providing a fast way to screen features. Statistical filtering is a common feature selection method that evaluates feature importance based on the statistical properties of the data, such as mutual information, correlation coefficient, and variance. It is generally believed that components with higher variance are more dispersed and contain more information; algorithms are proposed to retain the component with the highest variance by calculating the covariance matrix between features. Another approach identifies high-performing features as candidate features, specifically by minimizing intra-class distances and maximizing inter-class distances, using these as evaluation criteria. However, this method often relies too heavily on a single statistical property for feature selection, leading to the selection of inaccurate or discriminative features that are not well-suited for downstream tasks.

[0032] However, most existing feature selection algorithms are offline algorithms, relying on historical device state data for feature selection. This makes them unsuitable for scenarios where key features change due to concept drift in dynamic network environments. To address this challenge, some feature selection algorithms employ a sliding window approach to capture new data features. These algorithms select features by storing the latest sliding window data. However, sliding window-based feature selection algorithms require storing at least one window's worth of original data, resulting in significant storage overhead. Furthermore, solving sparse learning problems and performing matrix factorization on the original data also incurs substantial computational costs.

[0033] Examples of application scenarios for this invention are as follows: 1. Industrial IoT equipment monitoring: In factories, numerous sensors collect real-time equipment operation data (such as temperature, vibration, pressure, etc.), requiring dynamic identification of key features to predict faults or optimize performance. The technology is applied to compress historical data (e.g., 100,000 vibration data points from the past hour) using matrix sketching, preserving long-term patterns; dynamic anomaly detection, combined with manifold regularization to constrain the local structure of equipment operating states, avoids misjudging short-term noise (such as sudden current fluctuations) due to data drift. 2. Server cluster load monitoring: Data centers need to analyze CPU, memory, disk I / O, and other indicators in real time to identify sudden load peaks or resource bottlenecks. The technology is applied to dynamically select core indicators such as CPU utilization and network throughput during peak traffic periods (e.g., e-commerce promotions), ignoring low-relevance features (e.g., idle memory pages). During off-peak periods, the feature selection frequency is reduced (e.g., once every 5 minutes), increasing to once per second during peak periods to balance computing resource consumption. 3. Smart city infrastructure management: Traffic lights, environmental monitoring stations, and other equipment generate massive amounts of spatiotemporal data, requiring dynamic optimization of monitoring indicators. Technology is applied to multi-dimensional feature fusion, combining traffic flow (current data area) within a sliding window with historical congestion patterns (historical data area) to identify abnormal congestion characteristics (such as sudden traffic accidents). Lightweight matrix Sketch is deployed in roadside units (RSUs) to process camera data in real time, reducing cloud transmission bandwidth. 4. Smart grid equipment health management: Power transformers, circuit breakers, and other equipment require long-term monitoring of parameters such as current and temperature to prevent potential faults. Technology is applied to preserve the local geometric relationships of equipment operating status (such as temperature gradients between adjacent monitoring points) to avoid misjudgments due to local noise. For example, during extreme weather (such as thunderstorms), the manifold regularization weight is increased to enhance feature robustness; during stable periods, the weight is reduced to capture subtle changes.

[0034] Furthermore, when the device status feature data is updated, data is inserted into the initialization matrix Sketch. Specifically, this includes: setting the feature dimension, original data size, and compression ratio according to a predefined matrix initialization scheme; forming a large matrix from the original device performance feature data based on the feature dimension and original data size; obtaining the number of matrix rows in the initialization matrix Sketch based on the compression ratio; setting the initialization matrix Sketch using the feature dimension and the number of matrix rows; and directly inserting and storing the performance feature data into the initialization matrix Sketch in a first-in-first-out manner when the performance feature data is updated sequentially.

[0035] In this embodiment, the feature dimension, original data size, and compression ratio are obtained according to a predefined matrix initialization scheme.

[0036] Feature dimension M refers to the number of predefined features in the original data (e.g., the number of server metrics: CPU utilization, memory utilization, etc.).

[0037] The original data size N represents the total number of rows of historical device status characteristic data collected (e.g., 100,000 server indicator data points collected in the past hour). The original data size varies depending on the sampling frequency of different external devices.

[0038] The compression ratio r is a predefined compression ratio (e.g., r=0.01 means compressing the data to 1% of the number of rows).

[0039] Further calculations yielded the number of rows in the matrix: For example: if N=100000 and r=0.01, then r=10000 rows; The feature dimensions representing device status characteristic data, This also represents the number of samples of equipment status feature data, so the original equipment status feature data forms a large matrix. Device status characteristic data is continuously arriving, forming a large matrix. The scale gradually increases. Using a fixed-dimensional matrix (Sketch), Approximate original matrix ,in Much smaller than First, initialize a matrix Sketch. It can only store Device status characteristic data A matrix (Sketch) is used to represent streaming data; the insertion operation represents the current... Device status characteristic data When they arrive in sequence, they are directly inserted and stored in the matrix Sketch according to the first-in-first-out principle. At this time, the first element in the matrix Sketch is... Line refers to device status characteristic data. .

[0040] Furthermore, a space release operation is performed and the matrix Sketch is updated. Specifically, this includes: when device status feature data is updated, checking whether the number of rows currently stored in the matrix Sketch has reached the maximum number of rows; if all rows in the matrix Sketch are occupied, a space release operation is triggered; performing singular value decomposition on the matrix Sketch to obtain a singular value diagonal matrix and a right singular matrix, and then reducing the size according to preset parameters. The singular value diagonal matrix is ​​truncated and the matrix Sketch is updated; the matrix Sketch is divided into the historical device status feature data area and the current device status feature data area according to the matrix segmentation weight; the matrix segmentation weight is dynamically adjusted according to the feature fluctuation amplitude of the matrix Sketch.

[0041] In this embodiment, when device status feature data Since the matrix Sketch no longer has rows with zero values, it cannot be stored. For storage When a matrix sketch performs a "free space" operation, it allows the matrix sketch to reappear rows with zero values, replacing the original rows. Stored in compressed form In the process, newly arrived device status characteristic data is stored in middle.

[0042] For the matrix Sketch Perform singular value decomposition, and reduce the singular values ​​obtained from the decomposition so that the singular values ​​of small values ​​are reduced to 0. The parameter Y is a matrix Sketch. The reduction parameters, and satisfy .

[0043] Example: Sketch of matrix Perform singular value decomposition to obtain the singular value matrix. And right singular matrix The right singular matrix represents the column-space characteristics of the data. The Y-th singular value is selected. Reduce singular values , such that it is not greater than The singular values ​​are all reduced to zero. Finally, the Sketch matrix is ​​updated. ,implement ,at this time From row Y to All rows are zero values, allowing us to continue storing new stream data.

[0044] According to the storage procedure of Matrix Sketch, such as Figure 2The diagram shows the structure of the Sketch matrix divided into a historical device status feature data area and a current device status feature data area. Matrix B is divided into... and Two parts, of which It is a sketch of historical device status feature data, which can be regarded as a compression of historical device status feature data. Stores current device status characteristic data.

[0045] Example: When updating device status feature data and reconstructing the matrix for the first time, (Matrix Sketch Historical Device Status Feature Data Area) Retain the first Y rows and store the compressed historical device status feature data. (Matrix Sketch current device status feature data area) stores the new device status feature data in the first position. The original zero-value row is filled with new data in subsequent rows. At this time, the parameter Y can also be regarded as a matrix. segmentation and The matrix partitioning weights are used. In the space release operation, the parameter Y affects the amount of historical information retained after the release operation. Obviously, the larger the parameter Y, the more historical information is retained, but this leads to more frequent release operations, resulting in high computational overhead. The smaller the parameter Y, the less historical information is retained, and the lower the overhead of the release operation for streaming data storage.

[0046] Furthermore, the specific process of dynamically adjusting the matrix segmentation weights based on the characteristic fluctuation amplitude of the matrix Sketch is as follows: The device status characteristic data recorded in the sliding window of the matrix Sketch is processed according to the sliding window size. If a space release operation is performed and the matrix Sketch is updated, the device status characteristic data in the sliding window is re-recorded. For each feature dimension of the device status characteristic data in the sliding window, the standard deviation of the device status characteristic data in the sliding window under each feature dimension is calculated, and then the overall volatility is obtained by summing and averaging. The local volatility is obtained by analyzing the overall volatility based on preset local salient values. If the local volatility is greater than the upper limit of the local volatility range threshold, the matrix segmentation weights are reduced during the next predefined number of space release operations and matrix Sketch updates. If the local volatility is within the local volatility range threshold, the matrix segmentation weights are maintained during the next predefined number of space release operations and matrix Sketch updates. If the local volatility is less than the lower limit of the local volatility range threshold, the matrix segmentation weights are increased during the next predefined number of space release operations and matrix Sketch updates.

[0047] In this embodiment, the sliding window size is defined as w (such as the most recent 100 data points or data within 5 minutes). The sliding window is dynamically updated according to the arrival time of the data. For example, the most recent 100 data points are the latest 100 data points calculated from the device status feature data of the matrix Sketch, and the 100 data points within this range are used as the data in the sliding window.

[0048] The data within the sliding window comes from the matrix Sketch. and .because This is compressed data; using it directly may introduce errors, so it should be used preferentially. The original data in. If If the capacity is insufficient (e.g., after initialization), then combine... Approximate statistics.

[0049] If the device state feature data within the sliding window is re-recorded when performing a space release operation and updating the matrix Sketch, the following are some example steps: First, initialize a double-ended queue window to store the feature vectors of the most recent w data points.

[0050] Then maintain the cumulative sum sum[m] and the sum of squares sum_sq[m] for each feature (m∈[1,M]).

[0051] When device status characteristic data Upon arrival, append the device status characteristic data x ∈ R^M to the end of the window.

[0052] Update the cumulative sum sum[m] and the sum of squares sum_sq[m] using the new features; remove expired data; if the current window size exceeds the preset window capacity, remove the earliest data set x_old[m] from the window; delete x_old[m] from sum[m]; subtract x_old[m]^2 from sum_sq[m]; calculate the feature fluctuation amplitude: calculate the mean for each feature dimension m. ; Calculate the variance: (If the window is not full, replace the denominator with w-1 to calculate the sample standard deviation), and further calculate the standard deviation: Overall volatility can be represented by the average of the standard deviations of all characteristics, i.e. Local volatility can be selected by taking the average of the Q features with the largest standard deviation to highlight key changes. The preset local highlight value is Q.

[0053] If the local volatility exceeds the upper limit of the local volatility range threshold, the matrix segmentation weight is reduced during the subsequent predefined number of space release operations and matrix Sketch updates. For example, with a sliding window size w of 100, a local volatility of 0.3, and a matrix segmentation weight of 0.1, the calculated local volatility after the sliding window update is 0.6 (higher than the upper threshold of 0.5). The matrix segmentation weight is gradually reduced to 0.05 during the three sliding window updates, allowing more new data to influence the matrix Sketch. This compresses historical data, increases the proportion of new data, and makes feature selection more focused on recent changes.

[0054] It should be noted that if, during the process of gradually reducing the matrix segmentation weights in three sliding window updates, the local volatility does not continue to exceed the upper limit of the local volatility range threshold, but instead the local volatility is less than the lower limit of the local volatility range threshold, then the operation of gradually reducing the matrix segmentation weights is interrupted, and the operation of gradually increasing the matrix segmentation weights is switched to the previous operation.

[0055] The above method can adaptively adjust the parameter λ to balance the weight of historical information and new data, thereby optimizing the compression efficiency and feature selection accuracy of the matrix Sketch.

[0056] Furthermore, a sparse feature selection algorithm is set based on the historical device status feature data area and the current device status feature data area of ​​the Sketch matrix. Specifically, this includes: performing matrix transformations on the historical and current device status feature data areas to obtain the orthogonal basis matrices corresponding to the historical and current device status feature data areas; and setting a sparse feature selection algorithm based on these matrices. The objective function of the sparse feature selection algorithm is: ; ;in, This is the historical device status feature data area of ​​the Matrix Sketch. This is the current device status feature data area of ​​the Matrix Sketch. It is the orthogonal basis matrix corresponding to the historical device status feature data area of ​​the matrix Sketch. It is the orthogonal basis matrix of the current device state feature data area of ​​the matrix Sketch; It is a feature indicator matrix that merges the current device status feature data area of ​​the MatrixSketch and the historical device status feature data area of ​​the MatrixSketch. These are the weight parameters of the manifold regularization term; It is the weight parameter of the sparse regularization term, which is used for feature selection based on the sparse feature selection algorithm.

[0057] In this embodiment, a sparse feature selection algorithm is designed on the matrix Sketch, and the sparse learning problem to be solved is as follows:

[0058] (1);

[0059] ;

[0060] (2);

[0061] As shown in the formula, and These are the orthogonal basis matrices of the new data and the historical data, respectively. It is a feature indicator matrix that integrates new and historical data; this project performs L... 2，1 Norm regularization, constraining the eigenvalue indicator matrix It features sparse expression, and key features can be selected using V.

[0062] It is worth noting that, This is a matrix factorization process. To make the factorization process more accurate, and thus ensure the accuracy of feature selection, [the following steps are taken]. Constructing the Laplacian matrix ,use To learn the manifold information of the data space, so that data that are similar in the original space are also similar in the decomposed data space, while controlling parameters. Balance the impact of historical performance data and current performance data on feature importance.

[0063] Because the historical data is reduced, manifold constraints cannot be applied. Only contains Not included ; It is the weight parameter of the manifold regularization term, which controls the influence of manifold constraints on feature selection.

[0064] Manifold structure preservation is achieved by constraining the local geometry of the data space through the Laplace matrix P, ensuring that adjacent data points in the original space remain similar in the lower-rank space.

[0065] Balance weights and adjust historical data For new data The impact. For example, larger It will impose stricter constraints. It conforms to the manifold structure, thus preserving local patterns of historical data.

[0066] In practical applications, if the data has a distinct manifold structure (such as the topology of a sensor network), increasing... This can improve the robustness of features. If the data distribution is complex or there is a lot of noise, reducing... This can avoid overfitting.

[0067] It is a sparse regularization term ( The weighting parameter of the norm controls the strength of feature sparsity.

[0068] Function: Forced feature indicator matrix The rows are sparse, thus allowing for the selection of key features.

[0069] Noise reduction capability, relatively large It will generate sparser However, it may lose information about secondary features; smaller Retaining more features may introduce noise. If the goal is to extract a small number of core metrics (such as CPU and memory in server monitoring) from high-dimensional data, increasing... Key features can be located directly.

[0070] Furthermore, feature selection is performed using a sparse feature selection algorithm, specifically including: calculating the corresponding feature score for each row of the feature indicator matrix, as shown in the following formula: ; Indicates the first The feature score of the i-th feature is used to correspond to the i-th feature in the feature indicator matrix. The L2 norm of the row; the obtained feature scores are arranged in non-increasing order, and the device features corresponding to the feature scores are selected and continuously measured according to the predefined number of feature selections.

[0071] In this embodiment, the feature indicator matrix obtained by formula (1) , No. Feature scores of each feature, corresponding matrix No. The L2 norm of the row. The higher the feature score, the more important the corresponding feature. Therefore, feature selection is calculated according to formula (2). Sort in non-increasing order and select the first... Each feature is considered an important sub-feature. For the selection... Continuous measurement of an important characteristic, such as on a server device. Take 2, calculate If the first two features in the non-increasing order are obtained as the device's CPU utilization and IO wait time, then only the important features are measured, transmitted, and reported, thereby significantly reducing overhead.

[0072] Sparse learning on a matrix sketch can significantly reduce computational overhead compared to sparse learning on the original data because the matrix sketch space is small.

[0073] Furthermore, a dynamic feature selection frequency analysis and evaluation is performed, specifically including: collecting and obtaining the feature score of the current sparse learning feature selection and the feature score of the previous sparse learning feature selection; directly extracting the predefined fluctuation threshold, the first weighting factor of feature change, and the second weighting factor of feature change from the device dynamic feature selection database; performing a difference analysis between the feature score of the current sparse learning feature selection and the corresponding feature score of the previous sparse learning feature selection, performing a proportion analysis with the predefined fluctuation threshold of the corresponding feature score, performing a summation and mean analysis, and then adjusting the weight with the first weighting factor of feature change to obtain the first component of dynamic feature selection frequency analysis; performing a difference analysis between the feature score of the current sparse learning feature selection and the corresponding average feature score of the full data, performing a proportion analysis with the corresponding average feature score of the full data, performing a summation and mean analysis, and then adjusting the weight with the second weighting factor of feature change to obtain the second component of dynamic feature selection frequency analysis; and obtaining the dynamic feature selection frequency analysis evaluation mutation value based on the first and second components of dynamic feature selection frequency analysis.

[0074] In this embodiment, the current sparse learning feature selection is numbered. This indicates the selection number of the sparse learning feature in the current iteration. , This indicates the feature selection number from the previous sparse learning iteration.

[0075] This represents the mutation value of the dynamic feature selection frequency analysis evaluation for the current sparse learning feature selection. The mutation value of the dynamic feature selection frequency analysis evaluation evaluation is used to quantify the degree of mutation of dynamic feature changes.

[0076] ; ; This represents the first sparse learning feature selection. Feature scores of each feature; This represents the first sparse learning feature selection in the previous iteration. Feature scores of each feature; Indicates the first The predefined fluctuation threshold of the feature score of each feature is directly extracted from the device dynamic feature selection database; This represents the first sparse learning feature selection. The average feature score of the full dataset, where full dataset refers to the original, unsampled, complete set of performance data. Here, it refers to the _th feature of matrix Sketch B. The average of the feature scores of all data for each feature; The first weighting factor representing feature changes is directly extracted from the device dynamic feature selection database; The second weighting factor representing feature changes is extracted directly from the device dynamic feature selection database.

[0077] Based on the corresponding historical average performance fluctuation amplitude of the equipment (the historical average performance fluctuation amplitude represents the average of the ratios of the performance data at the corresponding moment in history to the historical average performance data), a mapping relationship is constructed with the first weighting factor and the second weighting factor of the characteristic change. The smaller the historical average performance fluctuation amplitude of the equipment, the smaller the first weighting factor of the characteristic change and the larger the second weighting factor of the characteristic change.

[0078] Furthermore, the sparse learning feature selection frequency is adjusted based on the dynamic feature selection frequency analysis evaluation results. Specifically, this includes: if the dynamic feature selection frequency analysis evaluation mutation value is less than or equal to the dynamic feature selection frequency analysis evaluation mutation threshold, then the weight parameter of the manifold regularization term is gradually reduced to the predefined lower limit of the weight parameter of the manifold regularization term in the following predefined adjustment counts; if the dynamic feature selection frequency analysis evaluation mutation value is greater than the dynamic feature selection frequency analysis evaluation mutation threshold, then the weight parameter of the manifold regularization term is gradually reduced to the predefined lower limit of the weight parameter of the manifold regularization term in the following predefined adjustment counts, the dynamic feature selection frequency analysis evaluation mutation value is subtracted from the dynamic feature selection frequency analysis evaluation mutation threshold to obtain the dynamic feature selection frequency analysis evaluation mutation difference, and the sparse learning feature selection time interval is reduced based on the dynamic feature selection frequency analysis evaluation mutation difference.

[0079] In this embodiment, each feature selection requires a sparse learning process.

[0080] When drastic data fluctuations are detected, increase To enhance sparsity and quickly filter redundant features.

[0081] When the data tends to stabilize, reduce This reduces the weight of manifold constraints, avoiding overfitting to historical structures. It also improves the algorithm's adaptability to dynamic environments, indirectly optimizing the decision-making basis for feature selection frequency.

[0082] The method of improving the sparse learning feature selection time interval by evaluating the mutation difference based on dynamic feature selection frequency analysis is specifically based on historical data. A mapping relationship is constructed between the evaluation mutation difference of dynamic feature selection frequency analysis and the adjustment of the sparse learning feature selection time interval. The real-time evaluation mutation difference of dynamic feature selection frequency analysis is input to obtain the reduced sparse learning feature selection time interval. For example, if the evaluation mutation difference of dynamic feature selection frequency analysis is 1, the sparse learning feature selection time interval is shortened by 10%.

[0083] Furthermore, adjusting the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results also includes: if the dynamic feature selection frequency analysis and evaluation mutation value is less than or equal to the dynamic feature selection frequency analysis and evaluation mutation threshold, then the number of consecutive cumulative judgments is recorded; if the dynamic feature selection frequency analysis and evaluation mutation value is greater than the dynamic feature selection frequency analysis and evaluation mutation threshold, then the number of consecutive cumulative judgments is reset to zero; if the number of consecutive cumulative judgments is less than the number of consecutive cumulative judgments threshold, then no adjustment is made; if the number of consecutive cumulative judgments is equal to or greater than the number of consecutive cumulative judgments threshold, then the sparse learning feature selection time interval is gradually increased to the upper limit of the sparse learning feature selection time interval in the following predefined adjustment number of times.

[0084] In this embodiment, if the number of consecutive cumulative judgments is equal to or greater than the threshold number of consecutive cumulative judgments, it indicates that the corresponding device feature data is in a relatively stable state. Increasing the sparse learning feature selection time interval and decreasing the sparse learning feature selection frequency will reduce the consumption of real-time computing resources.

[0085] If the cumulative number of consecutive judgments is equal to or greater than the threshold for the cumulative number of consecutive judgments, the sparse learning feature selection time interval will be gradually increased to the upper limit of the sparse learning feature selection time interval in the subsequent predefined adjustment steps. A specific example is as follows: The system gradually extends the time interval between two feature selection operations (e.g., from the initial 1 minute to the upper limit of 10 minutes) to reduce computational resource consumption while ensuring data stability. If the initial time interval is 1 minute, each adjustment increases by 10% (i.e., 1.1 minutes) until the preset upper limit of 10 minutes is reached. If the data is stable after 5 consecutive judgments, it indicates that the device feature data has entered a stable period, at which point the adjustment mechanism is activated. Step size: The increment of the time interval for each adjustment (e.g., a fixed increase of 1 minute or a proportional increase). Upper limit: The maximum time interval preset based on hardware computing power and business needs (e.g., 10 minutes) to prevent excessive extension from causing response delays.

[0086] like Figure 3The diagram shown is a structural diagram of a device dynamic feature selection system based on matrix Sketch provided in this application embodiment. The system includes a matrix Sketch initialization module, a matrix Sketch update module, a matrix Sketch dynamic feature selection module, and a matrix Sketch dynamic feature selection adjustment module: The matrix Sketch initialization module initializes the matrix Sketch and inserts data into the initialized matrix Sketch when device status feature data is updated; the matrix Sketch update module releases space and updates the matrix Sketch if the data insertion in the matrix Sketch is full; the matrix Sketch dynamic feature selection module sets a sparse feature selection algorithm based on the historical device status feature data area and the current device status feature data area of ​​the matrix Sketch, and performs sparse learning feature selection based on the sparse feature selection algorithm; the matrix Sketch dynamic feature selection adjustment module performs dynamic feature selection frequency analysis and evaluation, and adjusts the sparse learning feature selection frequency based on the dynamic feature selection frequency analysis and evaluation results.

[0087] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0088] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0092] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for device dynamic feature selection based on matrix sketch, characterized in that, The method comprises the following steps: initializing a matrix Sketch, inserting data into the initialized matrix Sketch when device state feature data is updated; if the data insertion of the matrix Sketch is full, performing a space releasing operation and updating the matrix Sketch; setting a sparse feature selection algorithm according to a matrix Sketch historical device state feature data area and a matrix Sketch current device state feature data area, and performing sparse learning feature selection according to the sparse feature selection algorithm; performing dynamic feature selection frequency analysis and evaluation, and adjusting the sparse learning feature selection frequency according to the dynamic feature selection frequency analysis and evaluation result; the dynamic feature selection frequency analysis and evaluation specifically comprises: collecting and obtaining feature scores of the current sparse learning feature selection and feature scores of the last sparse learning feature selection; directly extracting a pre-defined fluctuation threshold of the feature scores, a feature change first weight factor and a feature change second weight factor from a device dynamic feature selection database; performing difference analysis on the feature scores of the current sparse learning feature selection and the corresponding feature scores of the last sparse learning feature selection, performing proportion analysis on the corresponding pre-defined fluctuation threshold of the feature scores, performing summation mean analysis, and then performing weight correction on the feature change first weight factor to obtain a dynamic feature selection frequency analysis first component; performing difference analysis on the feature scores of the current sparse learning feature selection and the corresponding average of the full-amount data feature scores, performing proportion analysis on the corresponding average of the full-amount data feature scores, performing summation mean analysis, and then performing weight correction on the feature change second weight factor to obtain a dynamic feature selection frequency analysis second component; obtaining a dynamic feature selection frequency analysis and evaluation mutation value according to the dynamic feature selection frequency analysis first component and the dynamic feature selection frequency analysis second component; the adjustment of the sparse learning feature selection frequency according to the dynamic feature selection frequency analysis and evaluation result specifically comprises: if the dynamic feature selection frequency analysis and evaluation mutation value is less than or equal to a dynamic feature selection frequency analysis and evaluation mutation threshold, gradually reducing the weight parameter of the manifold regularization term to a pre-defined lower limit of the weight parameter of the manifold regularization term in the next pre-defined adjustment times; if the dynamic feature selection frequency analysis and evaluation mutation value is greater than the dynamic feature selection frequency analysis and evaluation mutation threshold, gradually reducing the weight parameter of the manifold regularization term to the pre-defined lower limit of the weight parameter of the manifold regularization term in the next pre-defined adjustment times, subtracting the dynamic feature selection frequency analysis and evaluation mutation threshold from the dynamic feature selection frequency analysis and evaluation mutation value to obtain a dynamic feature selection frequency analysis and evaluation mutation difference value, and reducing the sparse learning feature selection time interval according to the dynamic feature selection frequency analysis and evaluation mutation difference value.

2. The device dynamic feature selection method based on matrix Sketch of claim 1, wherein, the data insertion into the initialized matrix Sketch when the device state feature data is updated specifically comprises: setting a feature dimension, an original data size and a compression ratio according to a pre-defined matrix initialization scheme; device original performance feature data is composed of a large matrix according to the feature dimension and the original data size; According to the compression ratio, the number of matrix rows of the initialization matrix Sketch is obtained; The initialization matrix Sketch is set according to the feature dimension and the number of matrix rows; When the performance feature data is updated in sequence, the first-in first-out direct insertion is performed in the initialization matrix Sketch.

3. The device dynamic feature selection method based on matrix sketch of claim 1, wherein, The space releasing operation is performed and the matrix Sketch is updated, and the specific operation includes: When the device state feature data is updated, it is checked whether the number of rows currently stored in the matrix Sketch has reached the maximum number of rows, if all the rows of the matrix Sketch are occupied, the space releasing operation is triggered: The singular value diagonal matrix and the right singular matrix are obtained by performing singular value decomposition on the matrix Sketch, and the preset reduction parameter is determined according to the singular value diagonal matrix The singular value diagonal matrix is truncated, and the matrix Sketch is updated. The matrix Sketch is divided into a matrix Sketch historical device state feature data area and a matrix Sketch current device state feature data area according to the matrix partition weight. The dynamic adjustment of the matrix partition weight is adjusted according to the feature fluctuation amplitude of the matrix Sketch.

4. The device dynamic feature selection method based on matrix sketch of claim 3, wherein, The specific process of adjusting the dynamic adjustment of the matrix partition weight according to the feature fluctuation amplitude of the matrix Sketch includes: According to the sliding window size, the device state feature data recorded in the sliding window by the matrix Sketch is obtained; If the space releasing operation is performed and the matrix Sketch is updated, the device state feature data in the sliding window is recorded again; According to each feature dimension of the device state feature data in the sliding window, the standard deviation of the device state feature data in the sliding window under each feature dimension is calculated, and then the overall volatility is obtained by summing and averaging, and the local volatility is obtained by analyzing the overall volatility according to the preset local highlight value; If the local volatility is greater than the upper limit of the local volatility range threshold, the matrix partition weight is reduced when the space releasing operation is performed and the matrix Sketch is updated for the next predefined number of times; If the local volatility is within the local volatility range threshold, the matrix partition weight is maintained when the space releasing operation is performed and the matrix Sketch is updated for the next predefined number of times; If the local volatility is less than the lower limit of the local volatility range threshold, the matrix partition weight is increased when the space releasing operation is performed and the matrix Sketch is updated for the next predefined number of times.

5. The device dynamic feature selection method based on matrix Sketch of claim 1, wherein, The sparse feature selection algorithm is set according to the matrix Sketch historical device state feature data area and the matrix Sketch current device state feature data area, and the specific operation includes: The matrix transformation is performed according to the matrix Sketch historical device state feature data area and the matrix Sketch current device state feature data area, respectively, to obtain the orthogonal basis matrix corresponding to the matrix Sketch historical device state feature data area and the orthogonal basis matrix of the matrix Sketch current device state feature data area; According to the orthogonal basis matrix corresponding to the matrix Sketch historical equipment state feature data area and the orthogonal basis matrix of the matrix Sketch current equipment state feature data area, a sparse feature selection algorithm is set, and a target function of the sparse feature selection algorithm is: ; wherein, is a matrix of Sketch historical device state feature data areas, is a matrix of Sketch current device state feature data areas; is an orthogonal basis matrix corresponding to the matrix Sketch historical device state feature data area, is an orthogonal basis matrix of the matrix Sketch current device state feature data area; is a feature indication matrix fusing the matrix Sketch current device state feature data area and the matrix Sketch historical device state feature data area; is a weight parameter for the manifold regularization term; is a weight parameter for the sparse regularization term; The feature selection is performed according to the sparse feature selection algorithm.

6. The device dynamic feature selection method based on matrix sketch of claim 5, wherein, The feature selection is performed according to the sparse feature selection algorithm, and the specific operation includes: The feature score corresponding to each row of the feature indication matrix is calculated, and the specific formula is as follows: ； a feature score representing a th feature, for a corresponding feature indicating matrix th row of the L2 norm; The obtained feature scores are arranged in a non-decreasing order, and the device features corresponding to the feature scores are selected according to the predefined feature selection quantity and continuously measured.

7. The device dynamic feature selection method based on matrix Sketch of claim 1, wherein, The adjusting the sparse learning feature selection frequency according to the dynamic feature selection frequency analysis evaluation result further comprises: If the dynamic feature selection frequency analysis evaluation mutation value is less than or equal to the dynamic feature selection frequency analysis evaluation mutation threshold value, a continuous cumulative determination number is recorded, and if the dynamic feature selection frequency analysis evaluation mutation value is greater than the dynamic feature selection frequency analysis evaluation mutation threshold value, the continuous cumulative determination number is cleared; If the continuous cumulative determination number is less than a continuous cumulative determination number threshold value, no adjustment is performed; If the continuous cumulative determination number is equal to or greater than the continuous cumulative determination number threshold value, the sparse learning feature selection time interval is gradually increased to an upper limit of the sparse learning feature selection time interval in a next predefined adjustment number.

8. A system for device dynamic feature selection based on matrix Sketch, applying the method for device dynamic feature selection based on matrix Sketch according to any one of claims 1-7, characterized in that, The matrix Sketch initialization module, the matrix Sketch updating module, the matrix Sketch dynamic feature selection module and the matrix Sketch dynamic feature selection adjustment module are included: The matrix Sketch initialization module is used for initializing the matrix Sketch, and inserting data into the initialized matrix Sketch when the device state feature data is updated; The matrix Sketch updating module is used for performing a space releasing operation and updating the matrix Sketch if the data insertion of the matrix Sketch is full; The matrix Sketch dynamic feature selection module is used for setting a sparse feature selection algorithm according to a matrix Sketch historical device state feature data area and a matrix Sketch current device state feature data area, and performing sparse learning feature selection according to the sparse feature selection algorithm; The matrix Sketch dynamic feature selection adjustment module is used for performing dynamic feature selection frequency analysis evaluation, and adjusting the sparse learning feature selection frequency according to a result of the dynamic feature selection frequency analysis evaluation.

Images