Nuclide identification method based on KNN classification

Abstract

The invention discloses a nuclide identification method based on KNN classification. The nuclide identification method comprises the following steps: 1) performing dimension reduction on original high-dimensional gamma energy spectrum data of a training sample by using a principal component analysis (PCA); 2) performing attribute reduction on gamma energy spectrum features of the training sample after PCA dimension reduction through a neighborhood rough set, and rejecting redundant features and retaining a feature set with the most discriminant ability through heuristic search based on attribute dependency; and 3) utilizing the feature set obtained in the step 2) to realize rapid nuclide identification of the to-be-detected sample through a K-nearest neighbor classification method. According to the method, PCA dimension reduction, attribute reduction and KNN classification methods are fused, high recognition precision can be kept, meanwhile, better calculation efficiency is achieved, and the method is suitable for portable equipment with limited resources to achieve real-time and accurate nuclide recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of gamma spectrum analysis of nuclear detection technology, and in particular to a nuclide identification method based on KNN classification. BACKGROUND

[0002] Gamma nuclide identification technology plays a key role in the fields of radiation monitoring, nuclear emergency response and nuclear non-proliferation, and its core goal is to quickly and accurately identify the nuclide type from a complex gamma spectrum. The traditional method represented by the characteristic peak matching method relies on the comparison of characteristic peak energy and nuclide database, although the principle is simple and the calculation complexity is low, but in the low count rate, high background or mixed nuclide scene, the characteristic peak is easily submerged by the Compton continuous region or disturbed by statistical fluctuations, causing missed and false judgments.

[0003] To improve the identification performance under complex spectrum, intelligent methods such as machine learning and deep learning have gradually become mainstream. They automatically extract deep features through full-spectrum data analysis, and show better recognition accuracy and robustness to overlapping peaks and complex spectra. However, its model parameters are large and the calculation structure is complex, resulting in high computational overhead, which is difficult to deploy on portable devices with limited computing power.

[0004] Another method based on time series analysis uses prior information such as pulse arrival time and energy distribution to build an identification model, which has made some progress in recent years and is suitable for fast nuclide identification. However, it is difficult to effectively analyze superimposed signals with similar energies, and the classification accuracy is obviously insufficient in the mixed nuclide scene.

[0005] Therefore, how to maintain high recognition accuracy and computational efficiency under limited computing power has become a key problem that needs to be broken through. SUMMARY

[0006] In view of the above-mentioned deficiencies in the prior art, the purpose of the present application is to solve the problems existing in the prior art, and to provide a nuclide identification method based on KNN classification, which can maintain high recognition accuracy while having better computational efficiency by fusing PCA dimension reduction, attribute reduction and KNN classification method, and is suitable for resource-limited portable devices to realize real-time and accurate nuclide identification.

[0007] To solve the above technical problems, the technical scheme adopted by the present application is as follows: a nuclide identification method based on KNN classification, characterized in that it comprises the following steps:

[0008] 1) Use principal component analysis (PCA) to reduce the dimension of the original high-dimensional gamma spectrum data of the training sample to eliminate linear redundancy;

[0009] 2) Attribute reduction of the gamma spectrum features of the training samples after PCA dimension reduction by neighborhood rough set, and through heuristic search based on attribute dependency, redundant features are removed and the most discriminative feature set is retained;

[0010] 3) Using the feature set obtained in step 2), through K-neighbor classification method, the nuclide identification of the test sample is realized.

[0011] Further, in step 2), the specific process is as follows:

[0012] 21) According to the gamma spectrum training sample set U, The mean μ and the standard deviation σ of the sum and the sum of the squares of the training sample set are calculated, and the sample is standardized: U norm =(U-μ) / σ;

[0013] Where x i is the original gamma spectrum vector, y i ∈{ω1,ω2,…,ω c} is the nuclide category label, and n is the total number of samples;

[0014] 22) Train the PCA model on U norm , sort by feature contribution rate, select the first P principal components with cumulative contribution rate ≥ 99%, and get the projection matrix W ∈ R P×1024 ; Project the standardized data to the PCA space: Get the reduced condition attribute set A = {1,2,…,P};

[0015] 23) Construct a neighborhood decision system <U p , A, D>, D is the decision attribute, and initialize the reduction set and

[0016] 24) Calculate

[0017] 25) Select a k such that SIG(a k ,reduct,D) = max{SIG(a i ,reduct,D)};

[0018] 26) If SIG(a k , reduct, D) > 0, then reduct = reduct ∪ a k , return to step 24); otherwise, go to step 27);

[0019] 27) According to reduct, the condition attributes in U P are screened to obtain the reduced training sample set U'.​

[0020] Further, the specific process of step 3) is as follows:

[0021] 31) Standardize the sample to be tested:

[0022] 32) Project the standardized data to the PCA space:

[0023] 33) Based on the reduct, select the low-dimensional feature vector u' from U P

[0024] 34) Using the K-Nearest Neighbor classification method, calculate the Euclidean distance of u' and all samples in U', select the K samples with the smallest distance as the neighbors; then for each neighbor u i assign a weight where ud i is the Euclidean distance, and ε is a very small positive number to prevent division by zero error; then for each nuclide category ω', calculate the weighted voting score: Finally, calculate the confidence:

[0025] 35) Output the nuclide category ω whose confidence exceeds the threshold.

[0026] Compared with the prior art, the present application has the following advantages:

[0027] The present application is based on the nuclide identification method of neighborhood rough set and K-Nearest Neighbor classification, which fuses PCA dimension reduction, attribute reduction and KNN classification method. KNN does not need complex training, only relies on distance measurement and neighbor voting mechanism, and significantly reduces the computational complexity. While maintaining high recognition accuracy, it has better computational efficiency, and is suitable for resource-limited portable devices to realize real-time and accurate nuclide identification. DETAILED DESCRIPTION

[0028] The present application will be further described below in conjunction with examples.

[0029] Embodiment: A nuclide identification method based on KNN classification, comprising the following steps:

[0030] 1) Use the principal component analysis method PCA to reduce the dimension of the original high-dimensional gamma spectrum data of the training sample, so as to eliminate linear redundancy; wherein the principal component analysis method PCA (Principal Component Analysis) is a mature multivariate statistical analysis method, which is widely used in high-dimensional data dimension reduction in the fields of signal processing, pattern recognition, data mining, etc. Using this method to reduce the dimension of the gamma spectrum can eliminate the linear correlation and redundant information between channels. ​

[0031] 2) Attribute reduction of the gamma spectrum features of the training samples after PCA dimensionality reduction by neighborhood rough set, through heuristic search based on attribute dependency, eliminate redundant features and retain the most discriminative feature set.

[0032] Wherein, the neighborhood rough set theory is specifically:

[0033] A, given a non-empty finite set U = {u1, u2, …, u n} on the real number space, for any object u i on U, define δ(u i ) = {u | u ∈ U ∧ Δ(u, u i ) ≤ δ};

[0034] Where δ is the neighborhood radius; Δ is the metric function on the N-dimensional real number space R N , select the Euclidean distance as the metric function, the core role is to quantify the difference between samples.

[0035] B, given a three tuple structure <U, A, D>, where U represents the object set (called the domain), denoted as U = {u1, u2, …, u n}; A is a continuous numerical condition attribute set; D is the decision attribute, which divides U into mutually exclusive decision classes {U1, U2, …, U N}, each element in the set can be regarded as a sample set with the same nuclide class label, called <U, A, D> a neighborhood decision system (NDS); A in the system can induce the neighborhood relation on the domain U.

[0036] C, in the neighborhood decision system NDS = <U, A, D>, the decision attribute D divides the domain U into N mutually exclusive equivalence classes {U1, U2, …, U N}; For any condition attribute subset Define the upper and lower approximation sets and boundary domain of decision attribute D about B:

[0037] Upper approximation set: Where

[0038] Lower approximation set: Where

[0039] D, the dependency degree of decision attribute D on condition attribute subset B is: γ B (D) quantifies the explanatory ability of the condition attribute subset B to the decision classification:

[0040] When γ B (D)→1, it means that most samples in the domain fall into the lower approximation set of decision attribute D with respect to B, indicating that B can distinguish most decision classes.

[0041] When γ B (D)→0, it means that most samples in the domain are located in the boundary domain, indicating that B lacks classification discrimination ability.

[0042] E, the attribute importance of conditional attribute a with respect to subset is defined as:

[0043] SIG(a, B, D) = γ B∪a (D)-γ B (D).

[0044] In the implementation of the scheme, the specific process is as follows:

[0045] 21) According to the γ spectrum training sample set U, the mean μ and the standard deviation σ of the sum and the sum of squares of the training sample set are calculated, and the samples are standardized: U norm =(U-μ) / σ.

[0046] Where, x i is the original γ spectrum vector, y i ∈{ω1, ω2, …, ω c} is the nuclide class label, and n is the total number of samples;

[0047] 22) Train the PCA model on U norm , sort by feature contribution rate, select the first P principal components with cumulative contribution rate ≥99%, and get the projection matrix W ∈ R P×1024 ; Project the standardized data to the PCA space: Get the reduced continuous numerical conditional attribute set A = {1, 2, …, P};

[0048] 23) Construct a neighborhood decision system <U p , A, D>, D is the decision attribute, and initialize the reduction set and

[0049] 24) Calculate

[0050] 25) Select a k such that SIG(a k ,reduct,D) = max{SIG(a i ,reduct,D)}.

[0051] 26) If SIG(a k ​If reduct, D) > 0, then let reduct = reduct U a k Go to step 27); otherwise, go to step 24).

[0052] 27) Filter the condition attributes in U according to reduct, and obtain the reduced training sample set U'. P

[0053] 3) Use the feature set obtained in step 2) to realize fast nuclide identification of the test sample by K-Nearest Neighbor (KNN) classification method.

[0054] The KNN classification method is an instance-based lazy learning algorithm. It assumes that similar samples are close to each other in the feature space, and the classification decision does not need to be trained in advance, but by calculating the distance between the sample to be identified and all labeled samples in the feature space, the K nearest neighbors (i.e. neighbors) are selected, and the distance weighted voting mechanism is used to determine the class to which the sample to be identified belongs. The formula for measuring the distance between samples in the KNN classification algorithm used in this paper is the Euclidean distance, which is consistent with the measurement function of the difference between samples in the neighborhood rough set.

[0055] The specific process is as follows:

[0056] 31) Standardize the test sample:

[0057] 32) Project the standardized data into the PCA space:

[0058] 33) Filter the features in U based on reduct, and obtain the low-dimensional feature vector u'; P

[0059] 34) Use the K-Nearest Neighbor (KNN) classification method to calculate the Euclidean distance between u' and all samples in U', select the K samples with the smallest distance as the neighbors; then assign a weight to each neighbor u i where ud i is the Euclidean distance, and ε is a small positive number to prevent division by zero error; then for each nuclide class ω', calculate the weighted voting score: Finally, calculate the confidence:

[0060] 35) Output the nuclide class ω whose confidence exceeds the threshold.

[0061] By using the scheme, the total time complexity T total ​​​O(chl) + O(chl x P) + O(n x d + n x log k + k x c); wherein, O is the time complexity of a certain calculation process, chl is the total number of spectral channel addresses, P is the number of principal components reserved after PCA dimension reduction, n is the number of training samples, k is the number of neighbors, and c is the number of nuclide categories. By comparing the magnitudes of each single complexity (usually chl > n > P and n >> k, n >> c), it can be found that the PCA projection operation O(chl x P) is the dominant term of the entire time complexity, and the subsequent classification algorithm has little effect on the calculation efficiency. It can be seen that the present scheme is very suitable for deployment on an embedded hardware platform.

[0062] As a specific embodiment,

[0063] Given a PCA-reduced neighborhood decision system NDS = <U, A, D>. Where the domain U = {e1, e2, …, e9} contains 9 samples; the condition attribute set A = {a1, a2, a3, a4, a5} is a 5-dimensional feature vector; the decision attribute D divides the samples into three categories of nuclides: 241 Am, 133 Ba, 137 Cs. The specific attribute values and category labels of the samples are shown in Table 1; the sample vector to be identified is u = {0.0, 0.00112397, -0.005, 0.255, 1}. The calculation parameters are set as neighborhood radius δ = 0.03 and number of neighbors K = 3.

[0064] Table 1

[0065]

[0066] First, the decision attribute D divides the domain U into three equivalent classes:

[0067] U / D = {U1, U2, U3} = {{e1, e2, e3}, {e4, e5, e6}, {e7, e8, e9}};

[0068] Then, step 2) is performed to initialize the reduction set Take the importance of attribute a4 as an example.

[0069] The neighborhood relation induced by {a4} is:

[0070] δ(e1): {e1, e2};

[0071] δ(e2): {e1, e2};

[0072] δ(e3): {e3};

[0073] δ(e4) to δ(e9): {e4, e5, e6, e7, e8, e9}.

[0074] Calculate the approximate set: (because its neighborhood is all contained in U1), Therefore

[0075] Dependency

[0076] Attribute importance

[0077] Similarly, calculate the importance of other attributes: SIG(a1,reduct,D)=0, SIG(a2,reduct,D)=0, SIG(a3,reduct,D)=0, SIG(a5,reduct,D)= 1 / 3. Select the attribute a4 with the highest importance and add it to the reduction set, i.e., reduct = {a4}. Repeat the above process to obtain the optimal reduction set reduct = {a4, a2, a1}. Based on this result, the features are filtered to obtain the reduced neighborhood decision system, as shown in Table 2.

[0078] Table 2

[0079]

[0080] Next, step 3) is used to predict the category of sample u.

[0081] First, filter u by reduct to get u′={0.0,0.00112397,0.255}.

[0082] Then, the Euclidean distance between u′ and each sample in Table 2 was calculated and sorted in ascending order of distance. The results are shown in Table 3.

[0083] Select the first K = 3 nearest neighbor samples (e6, e4, e5), all of which belong to the same category. 133 Ba. Using distance-weighted voting, the result is calculated as follows: 133 The total weight of Ba is 1 / (d6+ε)+1 / (d4+ε)+1 / (d5+ε), and the confidence level is conf=1.0.

[0084] Therefore, the sample to be tested was identified as a nuclide. 133 Ba, with a confidence level of 100%.

[0085] Table 3

[0086]

[0087]

[0088] Finally, it needs to be explained that the above examples are only used to illustrate the technical solutions of the present application but not to limit the technical solutions, and those of ordinary skill in the art should understand that the technical solutions of the present application are modified or equivalently replaced without departing from the purpose and scope of the technical solutions, which should be covered in the scope of claims of the present application.

Claims

1. A nuclide identification method based on KNN classification, characterized in that, Comprising the following steps: 1) Dimensionality reduction is performed on the original high-dimensional gamma spectrum data of the training samples by using principal component analysis (PCA) to eliminate linear redundancy; 2) Attribute reduction is performed on the gamma spectrum features of the training samples after dimensionality reduction by using neighborhood rough set, and through heuristic search based on attribute dependency, redundant features are removed and the most discriminative feature set is retained; 3) Through K-nearest neighbor classification method, the feature set obtained in step 2) is used to realize fast nuclide identification of the test sample.

2. The nuclide identification method based on KNN classification according to claim 1, characterized in that, In step 2), the specific process is as follows: 21) According to the gamma spectrum training sample set U, Calculate the mean μ and standard deviation σ of the training sample set and and standardize the samples: U nor m = (U - μ) / σ; Wherein, x i is the original gamma spectrum vector, y i ∈ {ω1, ω2, …, ω c} is the nuclide category label, and n is the total number of samples. 22) In U norm train the PCA model, sort by feature contribution rate, select the first P principal components with cumulative contribution rate ≥ 99%, and get the projection matrix W ∈ R P×1024 ; Project the normalized data to the PCA space: Get the reduced condition attribute set A = {1, 2, ω, P}. 23) Constructing neighborhood decision system <U p , A, D> where D is the decision attribute, initialize the reduced set and 24) to computing 25) Select a k such that SIG(a k , reduct, D) = max{SIG(a i , reduct, D)}. 26) If SIG(a k , reduct, D) > 0, then let reduct = reduct U a k , go to step 24); else go to step 27); 27) According to reduct, filter the condition attributes in U P to get the reduced training sample set U'.

3. The nuclide identification method based on KNN classification according to claim 1, characterized in that, The specific process of step 3) is as follows: 31) Standardization of the sample to be tested: 32) Project the normalized data into the PCA space: 33) based on reducts for U P Feature selection is performed to obtain a low-dimensional feature vector u'. 34) Using the K-Nearest Neighbor classification method, compute the Euclidean distance of u' to all samples in U', and select the K samples with the smallest distance as the neighbors; then for each neighbor u i assign a weight where ud i is the Euclidean distance, and ε is a small positive number to prevent division by zero error; then for each isotope class ω', compute the weighted vote score: Finally compute the confidence: 35) Output the nuclide category ω whose confidence exceeds the threshold value.

Images