Microelectronic system sensor anomaly detection method based on improved kernel principal component analysis

Abstract

The invention discloses a microelectronic system sensor anomaly detection method based on improved kernel principal component analysis, and relates to the technical field of microelectronic sensing, and the method comprises the following steps: S1, obtaining steady-state normal operation data and fault data of a target sensor, and constructing an input matrix according to the steady-state normal operation data; s2, standardizing the input matrix, and training the improved KPCA model by using a standardized training sample; s3, obtaining an orthogonality enhanced feature vector matrix; s4, calculating the T2 statistic of the weighted mahalanobis distance and the SPE statistic of the local density; and S5, calculating a self-adaptive empirical threshold, and carrying out anomaly detection by using the self-adaptive empirical threshold. According to the method, the complete normalized training feature matrix is stored in the training stage, the consistency of kernel space construction in the prediction stage is ensured, and the engineering applicability and deployment stability of the model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of microelectronic sensing technology, and in particular to a microelectronic system sensor anomaly detection method based on improved kernel principal component analysis. BACKGROUND

[0002] With the continuous development of microelectronic technology and integrated circuit technology, various electronic sensors (such as temperature sensors, pressure sensors, etc.) are integrated into electronic devices, industrial equipment and intelligent terminals for real-time monitoring of system operation status and providing basis for control strategy. However, sensors are easily affected by electromagnetic noise, power fluctuations, mutual interference, environmental temperature and humidity changes, device aging, material fatigue, and sudden changes in actual working conditions of equipment during long-term operation, resulting in signal drift, mutation, and increased jitter, etc. Abnormalities, if not detected in time, will lead to system performance degradation, control misjudgment and even safety risks. The commonly used anomaly detection methods include threshold method, sliding window rule method, PCA linear dimension reduction method based on statistical characteristics, and autoencoder or neural network method based on deep learning. However, these methods have limitations in the electronic sensor scene: simple threshold method cannot handle complex nonlinear relationships and has high maintenance cost; linear PCA cannot depict the common nonlinear coupling characteristics and manifold structure in sensor signals; deep learning methods are highly dependent on the amount of training data, have poor interpretability, and are difficult to deploy in resource-constrained microelectronic systems; traditional KPCA has nonlinear modeling capability, but its kernel matrix often uses non-strict centering method, feature vector normalization is insufficient, SPE / T2 threshold depends on theoretical approximation, and kernel parameters need to be adjusted manually, resulting in high false alarm and missed alarm rates in actual detection, making it difficult to adapt to changes in different sensor types and different operating conditions. Therefore, the existing technology urgently needs an anomaly detection method that can accurately depict the nonlinear feature manifold of electronic sensors, has strict mathematical derivation and stable engineering performance, and can realize threshold and parameter adaptation, to improve the reliability and intelligence level of the system.

[0003] In summary, the present application develops an electronic sensor anomaly detection method based on improved kernel principal component analysis, which realizes anomaly detection of sensor signals in microelectronic systems through adaptive local density weighted kernel function, structure preserving centering and statistical quantity adaptive threshold discrimination. SUMMARY

[0004] The present application proposes a microelectronic system sensor anomaly detection method based on improved kernel principal component analysis to overcome the shortcomings of traditional principal component analysis methods.

[0005] The technical solution of the present application is: a microelectronic system sensor anomaly detection method based on improved kernel principal component analysis, comprising the following steps:

[0006] S1. Obtain the steady-state normal operation data and fault data of the target sensor, and construct the input matrix based on the steady-state normal operation data;

[0007] S2. Standardize the input matrix and use the standardized training samples to train the improved KPCA model;

[0008] S3. Input the fault data of the target sensor into the trained improved KPCA model to obtain the feature vector matrix with enhanced orthogonality;

[0009] S4. Based on the eigenvector matrix with enhanced orthogonality, calculate the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density.

[0010] S5. Calculate the adaptive empirical threshold based on the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density, and use the adaptive empirical threshold for anomaly detection.

[0011] Furthermore, in S1, the input matrix The expression is:

[0012] ;

[0013] in, This is the feature vector of the first sample. This is the feature vector of the second sample. For the first The feature vector of each sample For the first The feature vector of each sample For the set of real numbers, The feature dimension for each sample.

[0014] Furthermore, S3 includes the following sub-steps:

[0015] S31. Input the fault data of the target sensor into the trained improved KPCA model and calculate the Pearson correlation coefficient matrix between the target sensor and other sensors.

[0016] S32. Construct a fusion kernel based on the correlation between the target sensor and other sensors;

[0017] S33. Construct a structure preservation matrix based on the Pearson correlation coefficient matrix between the target sensor and other sensors;

[0018] S34. Construct a structure-preserving centralized kernel matrix based on the fusion kernel and the structure-preserving matrix;

[0019] S35. Perform spectral normalization and eigenvalue decomposition on the structure-preserving centered kernel matrix;

[0020] S36, using improved adaptive cumulative contribution rate for principal component selection, orthogonalizing the feature vector matrix corresponding to the screened feature value to obtain an orthogonality-enhanced feature vector matrix.

[0021] Further, in S32, the fusion kernel is expressed as:

[0022] ;

[0023] wherein, is the weight coefficient of the first base kernel, is the total number of sensors, is the first base kernel matrix.

[0024] Further, in S33, the structure-preserving matrix is expressed as:

[0025] ;

[0026] ;

[0027] wherein, is an adjacency matrix constructed based on the correlation degree of the sensors, is an identity matrix, is a node degree matrix, is an element of the adjacency matrix, is a Pearson correlation coefficient matrix of the target sensor and other sensors, is a judgment threshold.

[0028] Further, in S34, the structure-preserving centralized kernel matrix is expressed as:

[0029] ;

[0030] wherein, is a structure-preserving matrix, is a transpose, is a fusion kernel.

[0031] Further, in S36, the improved adaptive cumulative contribution rate is expressed as:

[0032] ;

[0033] ;

[0034] wherein, is the first a characteristic value, the first characteristic value of the kernel matrix, the largest characteristic value of the kernel matrix, the first characteristic value of the kernel matrix, the total number of characteristic values, an adaptive cumulative contribution rate threshold, a reference contribution rate threshold, an adjustment factor;

[0035] In S36, the expression of the characteristic vector matrix with enhanced orthogonality is as follows:

[0036] ;

[0037] wherein, the screened characteristic vector matrix, is a Gram-Schmidt orthogonal operator.

[0038] Further, S4 comprises the following sub-steps:

[0039] S41, calculating a normalized projection of the kernel space according to the characteristic vector matrix with enhanced orthogonality;

[0040] S42, calculating a T2 statistic of weighted Mahalanobis distance and an SPE statistic of local density according to the normalized projection of the kernel space.

[0041] Further, in S41, the expression of the normalized projection of the kernel space is as follows:

[0042] ;

[0043] wherein, is a diagonal matrix, is the first column of the centered kernel matrix, is the characteristic vector matrix with enhanced orthogonality;

[0044] In S42, the expression of the T2 statistic of weighted Mahalanobis distance is as follows:

[0045] ;

[0046] wherein, is the characteristic value corresponding to the first principal component;

[0047] In S42, the expression of the SPE statistic of local density The expression of S5 is:

[0048] ;

[0049] wherein, is the feature vector of the i-th sample, is the k-th nearest neighbor center of the i-th sample, is the local density of the i-th sample, is the local density of the k-th nearest neighbor center, is a small positive number to prevent the denominator from being zero. Further, in S5, the expression of the adaptive empirical threshold is: ;

[0050]

[0051] ;

[0052] ;

[0053] wherein, is the adaptive empirical threshold of the SPE statistic, is the percentile function, is the SPE statistic of the local density, is the percentile rank, is the adaptive empirical threshold of the T2 statistic, is the T2 statistic of the weighted Mahalanobis distance. 2

[0054] In S5, the expression of the adaptive empirical threshold for anomaly detection is:

[0055] ;

[0056] ;

[0057] ;

[0058] wherein, is the Boolean judgment result of whether it is an anomaly, is the Boolean judgment result of whether the SPE statistic exceeds the threshold value, is the Boolean judgment result of whether the T2 statistic exceeds the threshold value, is the SPE statistic value of the current test sample, is the T2 statistic value of the current test sample, is the adaptive empirical threshold of the T2 statistic. ​​​​​​​

[0059] The beneficial effects of the present application are:

[0060] (1) The present application introduces an adaptive local density weighted kernel function, effectively improving the description ability of nonlinear features; using a structure preserving centralization method to process the kernel matrix, the inherent geometric structure of sensor data in the feature space is better preserved, and the accuracy of anomaly detection is significantly enhanced;

[0061] (2) The present application proposes a projection method based on eigenvalue weighting and a feature vector processing mechanism enhanced by Gram-Schmidt orthogonalization, effectively avoiding the distortion problem of principal component projection, and improving the stability of feature expression;

[0062] (3) The present application innovatively uses an adaptive empirical threshold without distribution assumption to set the discriminant threshold of T2 and SPE statistics, overcoming the limitations of traditional chi-square theory approximation, and reducing error accumulation;

[0063] (4) The present application saves the complete normalized training feature matrix in the training stage, ensuring the consistency of kernel space construction in the prediction stage, and improving the engineering applicability and deployment stability of the model; These improvements work together to comprehensively improve the accuracy, robustness and engineering practicability of the present application in microelectronic system sensor anomaly detection. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is a flow chart of the microelectronic system sensor anomaly detection method based on improved kernel principal component analysis;

[0065] Figure 2 is an anomaly detection result graph based on improved kernel principal component analysis;

[0066] Figure 3 is an anomaly detection result graph based on kernel principal component analysis. DETAILED DESCRIPTION

[0067] The embodiments of the present application will be further described below in conjunction with the accompanying drawings.

[0068] As shown in Figure 1 , the present application provides a microelectronic system sensor anomaly detection method based on improved kernel principal component analysis, comprising the following steps:

[0069] S1, obtaining the steady-state normal operation data and fault data of the target sensor, and constructing an input matrix according to the steady-state normal operation data;

[0070] S2, standardizing the input matrix, and training the improved KPCA model using the standardized training samples;

[0071] S3. Input the fault data of the target sensor into the trained improved KPCA model to obtain the feature vector matrix with enhanced orthogonality;

[0072] S4. Based on the eigenvector matrix with enhanced orthogonality, calculate the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density.

[0073] S5. Calculate the adaptive empirical threshold based on the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density, and use the adaptive empirical threshold for anomaly detection.

[0074] In this embodiment of the invention, in S1, the input matrix The expression is:

[0075] ;

[0076] in, This is the feature vector of the first sample. This is the feature vector of the second sample. For the first The feature vector of each sample For the first The feature vector of each sample For the set of real numbers, The feature dimension for each sample.

[0077] The expression for standardizing the input samples is:

[0078] ;

[0079] in, It is the mean vector. This is the standard deviation vector. All features are guaranteed to be of the same magnitude, and the standardized samples are used as a unified benchmark for the kernel function and prediction. This invention stores the normalized training sample matrix in the following form for constructing the kernel matrix during subsequent prediction:

[0080] ;

[0081] in, The normalized training sample matrix, This is the normalized feature vector of the first sample. For the first The normalized feature vector of each sample.

[0082] The improved KPCA model was trained using the normalized training samples. The obtained SPE statistic threshold and T2 statistic threshold were saved for model testing.

[0083] In this embodiment of the invention, S3 includes the following sub-steps:

[0084] S31. Input the fault data of the target sensor into the trained improved KPCA model and calculate the Pearson correlation coefficient matrix between the target sensor and other sensors.

[0085] S32. Construct a fusion kernel based on the correlation between the target sensor and other sensors;

[0086] S33. Construct a structure preservation matrix based on the Pearson correlation coefficient matrix between the target sensor and other sensors;

[0087] S34. Construct a structure-preserving centralized kernel matrix based on the fusion kernel and the structure-preserving matrix;

[0088] S35. Perform spectral normalization and eigenvalue decomposition on the structure-preserving centered kernel matrix;

[0089] S36. Principal component selection is performed using the improved adaptive cumulative contribution rate. The eigenvector matrices corresponding to the selected eigenvalues ​​are orthogonalized to obtain eigenvector matrices with enhanced orthogonality.

[0090] Compared with traditional KPCA fixing Unlike the RBF kernel, this invention uses adaptive local density weighting as the kernel function, dynamically adjusting the kernel width based on the local density of the training samples. This ensures that high-density regions maintain high resolution while low-density regions maintain smoothness. Let the samples... Local density for:

[0091] ;

[0092] in, The number of neighbors selected when conducting density assessment in a local area. For the sample of Nearest neighbor set.

[0093] The kernel function formula is:

[0094] ;

[0095] in, For the sample The local density estimate, For the sample The local density estimate, To avoid tiny positive numbers with a denominator of zero, this kernel function balances local geometry and noise robustness.

[0096] For the strong coupling characteristics of multi-sensor, first calculate the Pearson correlation coefficient matrix of target sensor and other sensors :

[0097] ;

[0098] wherein, is the time series data vector of the target sensor, is the time series data vector of the first auxiliary sensor, is the time series data vector of the first auxiliary sensor, is the Pearson correlation coefficient function, is the sample mean of the target sensor data, is the sample mean of the first auxiliary sensor.

[0099] In the embodiment of the application, in S32, the fusion kernel The expression is:

[0100] ;

[0101] wherein, is the weight coefficient of the first base kernel, is the total number of sensors, is the first base kernel matrix. In the embodiment of the application, in S33, the structure preserving matrix The expression is:

[0102] ;

[0103] ;

[0104] ;

[0105] wherein, is the adjacency matrix based on the correlation degree of the sensor, is the unit matrix, is the node degree matrix, is the element of the adjacency matrix, is the Pearson correlation coefficient matrix of the target sensor and other sensors, is the determination threshold.

[0106] In the embodiment of the application, in S34, the structure preserving type centralized kernel matrix The expression is:

[0107] ;

[0108] wherein, is the structure preserving matrix, is the transpose,​​ For fusion core.

[0109] In this embodiment of the invention, in S36, the adaptive cumulative contribution rate is improved. The expression is:

[0110] ;

[0111] ;

[0112] in, The first kernel matrix 1 eigenvalue, The first kernel matrix 1 eigenvalue, The largest eigenvalue of the kernel matrix. The first kernel matrix 1 eigenvalue, The total number of eigenvalues. For adaptive cumulative contribution rate threshold, As the baseline contribution rate threshold, As a regulating factor;

[0113] In S36, the eigenvector matrix with enhanced orthogonality The expression is:

[0114] ;

[0115] in, The selected feature vector matrix, It is a Gram-Schmidt orthogonal operator.

[0116] In this embodiment of the invention, S4 includes the following sub-steps:

[0117] S41. Calculate the normalized projection of the kernel space based on the eigenvector matrix with enhanced orthogonality.

[0118] S42. Based on the normalized projection of the kernel space, calculate the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density.

[0119] In this embodiment of the invention, in S41, the normalized projection of the kernel space... The expression is:

[0120] ;

[0121] in, It is a diagonal matrix. The first of the centered kernel matrix List, The eigenvector matrix is ​​for enhancing orthogonality;

[0122] T2 statistic of weighted Mahalanobis distance in S42 The expression of is:

[0123] ;

[0124] Wherein, is the eigenvalue corresponding to the th principal component;

[0125] SPE statistic of local density in S42 The expression of is:

[0126] ;

[0127] Wherein, is the eigenvector of the th sample, is the nearest neighbor center of the th sample, is the local density of the th sample, is the local density of the nearest neighbor center, is a small positive number to prevent the denominator from being zero. In the embodiment of the application, in S5, the expression of the adaptive empirical threshold is: ;

[0128] Wherein,

[0129] is the adaptive empirical threshold of the SPE statistic, is the percentile function, is the SPE statistic of the local density,

[0130] is the percentile rank, is the adaptive empirical threshold of the T2 statistic, is the T2 statistic of the weighted Mahalanobis distance; 2

[0131] S5, the expression of the adaptive empirical threshold for anomaly detection is: ; ;

[0132]

[0133] ;

[0134] ;

[0135] ; ​​​​​​

[0136] in, The result of a Boolean judgment indicating whether it is an anomaly. This is a Boolean judgment result indicating whether the SPE statistic exceeds the threshold. This is a Boolean judgment result indicating whether the T² statistic exceeds the threshold. For the current test sample The SPE statistic value, For the current test sample The T² statistic value, This is the adaptive empirical threshold for the T² statistic.

[0137] The following description is based on specific embodiments.

[0138] First, data from temperature sensors in a microelectronic system under normal operating conditions is collected as the training set. Simultaneously, data from four typical failure modes—fixed deviation, drift, accuracy degradation, and complete failure—are collected as the test set, constructing a complete dataset. Then, the training and test set data are preprocessed, including normalization and sliding window-based feature extraction, to standardize the data and enhance the expressive power of temporal features. In the model building stage, an adaptive local density-weighted kernel function is introduced, and a fusion kernel matrix is ​​constructed by combining the correlation between the target sensor and multiple auxiliary sensors, followed by structure-preserving centering. Next, the processed kernel matrix undergoes eigenvalue decomposition, and a dual feature selection mechanism is used to determine the number of principal components. After model training, based on the SPE and T² statistic distributions of normal data, an empirical quantile method without distribution assumptions is used to determine the anomaly detection threshold. In the final application stage, real-time sensor data is input into the trained model, and its SPE and T² statistic values ​​are calculated. Real-time comparison with the threshold automatically determines whether the sensor's operating state is abnormal.

[0139] Figure 2 Based on improved kernel principal component analysis, the anomaly detection results are presented in the following figures. Subfigures (a), (b), (c), and (d) show the anomaly detection results of the temperature sensor under four fault modes (fixed deviation, drift, accuracy degradation, and complete failure), respectively. The horizontal axis represents time (in seconds), and the vertical axis represents the magnitude of the statistical values. The blue curve represents stable fluctuations in statistical values ​​under normal operating conditions, while the red curve represents the statistical value curve under abnormal conditions. The orange dashed line represents the statistical threshold. The anomaly detection results show that at the four fault occurrence times of the sensor, both the SPE statistic and the T2 statistic exceeded the set thresholds, achieving reliable anomaly detection.

[0140] Figure 3For the abnormality detection result based on the kernel principal component analysis, through the result graph, subgraphs (a), (b), (c) and (d) are respectively abnormality detection result graphs of the temperature sensor under four fault modes (fixed deviation, drift, precision drop and complete failure), the horizontal axis represents time (unit: s), the vertical axis represents the size of the statistical value, the blue curve represents the statistical value fluctuation under the normal working condition, and the red curve represents the statistical value curve of the abnormal state, and the orange dotted line represents the statistical threshold. It can be seen from the abnormality detection result that at the occurrence time of the four faults of the sensor, only the SPE statistical quantity exceeds the set threshold, the T2 statistical quantity does not exceed the set threshold, and the abnormality detection result is obviously poor.

[0141] Tables 1 and 2 are respectively the accuracy and running time of two abnormality detection models (KPCA model and improved KPCA model). From the accuracy comparison in Table 1, it can be seen that the improved KPCA model has significantly improved the detection performance of the four fault types, wherein the detection accuracy of the fixed deviation fault is greatly improved from 0.1246 to 0.9987, the recognition ability of the precision drop fault is optimized from 0.2507 to 0.9984, and the detection accuracy of the drift fault and the complete failure fault reaches excellent levels of 0.9756 and 0.9878 respectively. At the same time, the running time data in Table 2 shows that the improved model has optimized the calculation efficiency while maintaining the detection accuracy, and the processing time of the four types of faults is stably maintained in the interval of 8.80e-05 seconds to 8.88e-05 seconds, which shows better calculation performance than the running time of 9.03e-05 seconds to 9.90e-05 seconds of the original model. These results show that the improved model not only has a qualitative leap in the recognition ability of different fault types, but also maintains high calculation efficiency.

[0142] Table 1

[0143] Sample fault type KPCA model accuracy Improved KPCA model accuracy Fixed bias 0.1246 0.9987 Drift 0.7440 0.9756 Precision degradation 0.2507 0.9984 Complete failure 0.2725 0.9878

[0144] Table 2

[0145] Sample fault type KPCA model run time Improved KPCA model run time Fixed bias 9.45e-05 8.83e-05 Drift 9.90e-05 8.80e-05 Precision degradation 9.03e-05 8.86e-05 Complete failure 9.23e-05 8.88e-05

[0146] The application will be further described in detail below.

[0147] (1) Collect normal operation data and fault data containing four typical fault modes of fixed bias, drift, precision degradation and complete failure of temperature sensors in microelectronic systems; (2) Preprocess the collected data, including normalization processing and feature extraction based on sliding window, to eliminate the dimension effect and enhance the time series feature expression; (3) Construct an adaptive local density weighted kernel function, combine the correlation measure of the target sensor and auxiliary sensors to construct a fusion kernel matrix, and process it using a structure preserving centering method; (4) Perform eigenvalue decomposition on the centered kernel matrix, determine the number of principal components using a double feature screening mechanism, and perform orthogonalization enhancement processing on the feature vectors; (5) In the training stage, calculate the SPE and T² statistics based on normal data, and determine the adaptive detection threshold using the distribution-free empirical quantile method; (6) In the test stage, input the real-time collected sensor data into the trained model, and calculate the corresponding SPE and T² statistics values; (7) Automatically judge and output the abnormal detection result of the sensor working state by comparing the statistics values with the preset threshold in real time.

[0148] Those skilled in the art will appreciate that the embodiments described herein are presented for the purpose of helping the reader understand the principles of the present application, and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method for detecting sensor anomalies in microelectronic systems based on improved kernel principal component analysis, characterized in that, Includes the following steps: S1. Obtain the steady-state normal operation data and fault data of the target sensor, and construct the input matrix based on the steady-state normal operation data; S2. Standardize the input matrix and use the standardized training samples to train the improved KPCA model; S3. Input the fault data of the target sensor into the trained improved KPCA model to obtain the feature vector matrix with enhanced orthogonality; S4. Based on the eigenvector matrix with enhanced orthogonality, calculate the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density. S5. Calculate the adaptive empirical threshold based on the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density, and use the adaptive empirical threshold for anomaly detection.

2. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 1, characterized in that, In S1, the input matrix The expression is: ； in, This is the feature vector of the first sample. This is the feature vector of the second sample. For the first The feature vector of each sample For the first The feature vector of each sample For the set of real numbers, The feature dimension for each sample.

3. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 1, characterized in that, S3 includes the following sub-steps: S31. Input the fault data of the target sensor into the trained improved KPCA model and calculate the Pearson correlation coefficient matrix between the target sensor and other sensors. S32. Construct a fusion kernel based on the correlation between the target sensor and other sensors; S33. Construct a structure preservation matrix based on the Pearson correlation coefficient matrix between the target sensor and other sensors; S34. Construct a structure-preserving centralized kernel matrix based on the fusion kernel and the structure-preserving matrix; S35. Perform spectral normalization and eigenvalue decomposition on the structure-preserving centered kernel matrix; S36. Principal component selection is performed using the improved adaptive cumulative contribution rate. The eigenvector matrices corresponding to the selected eigenvalues ​​are orthogonalized to obtain eigenvector matrices with enhanced orthogonality.

4. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 3, characterized in that, In S32, the fusion core The expression is: ； in, For the first The weighting coefficients of each basic kernel The total number of sensors, For the first One basic kernel matrix.

5. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 3, characterized in that, In S33, the structure preservation matrix The expression is: ； ； in, This is an adjacency matrix constructed based on sensor correlation. As a unit array, Let the node degree matrix be... These are the elements of the adjacency matrix. The Pearson correlation matrix between the target sensor and other sensors. To determine the threshold.

6. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 3, characterized in that, In S34, the structure-preserving centralized kernel matrix The expression is: ； in, To preserve the matrix of the structure, For transpose, For fusion core.

7. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 3, characterized in that, In S36, the adaptive cumulative contribution rate is improved. The expression is: ； ； in, For the kernel matrix 1 eigenvalue, For the kernel matrix 1 eigenvalue, The largest eigenvalue of the kernel matrix. The first kernel matrix 1 eigenvalue, The total number of eigenvalues. For adaptive cumulative contribution rate threshold, As the baseline contribution rate threshold, As a regulating factor; In S36, the eigenvector matrix with enhanced orthogonality The expression is: ； in, The selected feature vector matrix, It is a Gram-Schmidt orthogonal operator.

8. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Calculate the normalized projection of the kernel space based on the eigenvector matrix with enhanced orthogonality. S42. Based on the normalized projection of the kernel space, calculate the T2 statistic of the weighted Mahalanobis distance and the SPE statistic of the local density.

9. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 8, characterized in that, In S41, the normalized projection of the kernel space The expression is: ； in, It is a diagonal matrix. The first of the centered kernel matrix List, The eigenvector matrix is ​​for enhancing orthogonality; In S42, the T2 statistic of the weighted Mahalanobis distance The expression is: ； in, For the first The eigenvalues ​​corresponding to each principal component; In S42, the SPE statistic of local density The expression is: ； in, For the first The feature vector of each sample For the first one sample Near the center, For the first Local density of a sample For the neighborhood center Local density, To prevent extremely small positive numbers with a denominator of zero.

10. The microelectronic system sensor anomaly detection method based on improved kernel principal component analysis according to claim 1, characterized in that, In step S5, the expression for the adaptive empirical threshold is: ； ； in, An adaptive empirical threshold for the SPE statistic. For percentile functions, The SPE statistic is the local density. Percentile level, The adaptive empirical threshold for the T² statistic. T is the weighted Mahalanobis distance 2 Statistic; In step S5, the expression for anomaly detection using an adaptive empirical threshold is: ； ； ； in, The result of a Boolean judgment indicating whether it is an anomaly. This is a Boolean judgment result indicating whether the SPE statistic exceeds the threshold. This is a Boolean judgment result indicating whether the T² statistic exceeds the threshold. For the current test sample The SPE statistic value, For the current test sample The T² statistic value, This is the adaptive empirical threshold for the T² statistic.

Images