A positive and negative slope forest-based missile anomaly point rapid detection algorithm

Abstract

The application provides a missile abnormal point rapid detection method based on a positive and negative slope forest model, and is used for solving the problems of false detection, missed detection and low detection efficiency of the existing missile abnormal detection method on missile quality data, and realizes the steps of: potential abnormal point definition; positive and negative slope tree construction; positive and negative slope forest construction; and abnormal point detection. The application compresses the data space based on the definition of "potential abnormal point" of the slope, solves the problem of low detection efficiency, realizes the accurate detection of the abnormal value of the missile quality monitoring data by constructing the positive and negative slope forest, selects and integrates the abnormal point value according to the shortest path principle, effectively improves the efficiency of the abnormal detection method, has good performance in time, and further provides decision support for the fault detection of the missile.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data mining information processing, and relates to an algorithm for detecting abnormal values, in particular to a rapid abnormal point detection method based on positive and negative slope forests. BACKGROUND

[0002] In recent years, with the increasing complexity, automation and informatization of military equipment, the requirements for reliability, maintainability and technical support capability are becoming increasingly urgent. Adopting fault detection and diagnosis technology is an important measure to improve the reliability of weapon equipment and the maintainability support capability. Abnormal detection is one of the important tasks in the field of data mining, and plays a crucial role in missile weapon state monitoring, air quality monitoring and other fields. Therefore, abnormal value detection has received more and more attention in the academic and industrial circles.

[0003] Abnormal value detection refers to identifying sample points deviating from most data from a large amount of data. Among them, time series data abnormal value detection is a branch of abnormal value detection. The existing abnormal value detection research work at present mainly includes distance-based anomaly detection, density-based anomaly value detection, statistics-based anomaly value detection, classification and clustering-based anomaly value detection, and graph-based anomaly value detection. However, these abnormal value detection methods have problems such as high time complexity and low accuracy when detecting missiles. The quality evaluation-based isolation forest solves the problems of high time complexity and low accuracy. However, the quality evaluation-based isolation forest abnormal detection method cannot effectively detect missile weapon abnormal value points in real time on a large amount of missile state time series data. Missile state time series data anomaly detection is to find objects with obvious differences in missile state data. These difference objects are usually accompanied by serious changes in missile state indicators. Through analysis of these abnormal phenomena, the changes of missile state can be obtained, thereby providing decision support for missile fault detection. SUMMARY

[0004] The purpose of the present application is to provide an improved time series missile abnormal point detection method based on positive and negative slope forest model to accurately and effectively detect abnormal values in a large amount of missile state time series data.

[0005] To achieve the above purpose, the technical solution adopted by the present application includes the following steps:

[0006] (1) Collecting time series data of missile state;

[0007] Using n sensors to sample at a sampling frequency f z Collecting n missile state time series signals of a to-be-detected missile with a length of T, wherein n≥1, T≥4687500, f s≥1500Hz;

[0008] (2) Definition of potential outliers:

[0009] Let one-dimensional time series data be (T,Y)={(t i ,y i )}, where t i ∈T,y i ∈Y and t i <t i+1 (i = 1, 2, ..., N), First, calculate the slope between data points at adjacent time points, then retain the slope where k ≠ 0, and finally store the potential outliers with positive changes and the potential outliers with negative changes into the set of potential outliers with positive changes and the set of potential outliers with negative changes, respectively.

[0010] (3) Construct a positive and negative slope tree model:

[0011] (3a) In the process of constructing the positive slope tree model, from the training set K + A sample of size ψ is randomly selected as the root node, and then a slope split point k is randomly selected from the sample. + The remaining data points are compared with the split point. Data points smaller than the split point are divided into left nodes, and data points greater than or equal to the split point are divided into right nodes. This process is repeated recursively until potential outlier points are separated or the height limit of the tree is reached.

[0012] (3b) In the process of constructing the negative slope tree model, since the smaller the value, the greater the probability of being judged as a real outlier, data points greater than or equal to the split point are divided into left nodes, and data points less than the split point are divided into right nodes.

[0013] (4) Constructing a forest model with positive and negative slopes:

[0014] When constructing the positive and negative slope forest model, the positive slope tree model and the negative slope tree model are combined to construct the positive slope forest model and the negative slope forest model respectively. During the construction of the forest model, the maximum number of trees that the forest can hold is t, and the number of trees t controls the size of the forest ensemble tree. The results show that the accuracy reaches its maximum when t = 100 (unless otherwise specified, t = 100 is used as the default value for the experiment).

[0015] (45) Perform outlier detection:

[0016] In outlier detection, the outlier proportion η of the entire dataset is first set, and then the number of outliers in the positive slope forest model is calculated separately. The number of outliers in the negative slope forest model Next, repeat the traversal of the data to be tested, traverse the left and right nodes of each tree, and record the height of each tree, sort the slope values from large to small, select the slope value according to the shortest path principle until the number of abnormal points is met, and finally, according to y i+1 = k(t i+1 -t i )+y i Return the abnormal data points.

[0017] Compared with the prior art, the present application has the following advantages:

[0018] 1. The present application combines random forest and slope characteristics to realize efficient and accurate time series division of missile anomaly point detection. According to the slope, a potential abnormal point definition is proposed to compress the data space and improve the timeliness of the algorithm; according to the positive and negative values of the slope, positive and negative slope anomaly detection tree models are constructed, and accurate detection of abnormal values is realized through classification of abnormal values; by using positive and negative slope trees, a positive and negative slope anomaly detection forest model is constructed to realize accurate detection of missile abnormal value points and analyze the parameters such as the size of the forest.

[0019] 2. The present application verifies its detection accuracy on six real data sets such as AQI, and verifies its detection efficiency on white noise simulation data sets of different sizes. Experiments show that the accuracy of the algorithm has been significantly improved, and the time efficiency has been improved several times compared with traditional machine learning and deep learning anomaly detection methods. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the implementation flowchart of the present application;

[0021] Figure 2 is the time performance diagram of the PNSF algorithm. DETAILED DESCRIPTION

[0022] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0023] Referring to Figure 1 , the positive and negative slope forest missile anomaly point rapid detection method comprises the following steps:

[0024] Step 1) Collecting time series data of missile state:

[0025] Using n sensors to sample at a sampling frequency f z Collecting n time series signals of missile state with length T of the missile to be tested, wherein n≥1, T≥4687500, f s ≥1500Hz;

[0026] Step 2) Definition of potential abnormality:

[0027] Let one-dimensional time series data (T, Y) = {(t i ,y i )}, where t i ∈T,y i ∈Y and t i <t i+1 (i = 1, 2,..., N). First, the slope between adjacent time data points (t i ,y i ) and (t i+1 ,y i+1 ) is calculated. Then the slope with k ≠ 0 is retained, and finally the positive change potential abnormal point and the negative change potential abnormal point are stored in the positive change potential abnormal point set K + and the negative change potential abnormal point set K - respectively.

[0028] For the potential abnormal points in the K + set, the greater the value, the greater the probability of being judged as a real abnormal point. For the potential abnormal points in the K - set, the smaller the value, the smaller the probability of being judged as a real abnormal point.

[0029] Step 3) Constructing positive and negative slope tree model:

[0030] Step 3a) In the process of constructing the positive slope tree model, randomly extract sample size ψ from the training set K + as the root node, then randomly select the slope split point k + from the sample, and compare the remaining data points with the split point. The data smaller than the split point is divided into the left node and the data greater than or equal to the split point is divided into the right node Recursive execution in turn until the potential abnormal value points are separated or the height limit (l = log2ψ) of the tree is reached.

[0031] Step 3b) In the process of constructing the negative slope tree model, since the smaller the value, the greater the probability of being judged as a real abnormal point. Therefore, the data greater than or equal to the split point is divided into the left node and the data smaller than the split point is divided into the right node

[0032] The construction of the above positive and negative slope tree model uses five parameters, which are training sample K + , K - , sub-sample size ψ, current tree height e and tree height limit l. The value and explanation of each parameter are as follows:

[0033] · Training sample K +and K - is the output of Algorithm 1. The data samples of the positive change potential abnormal points and the negative change potential abnormal points are stored in K + and K - respectively;

[0034] • The subsample size ψ controls the size of the sampling. Experimental results show that the influence of the subsampling quantity on the experimental results is small with the increase of the sampling size. According to the isolation forest, ψ is set to 256; (unless otherwise specified, ψ = 256 is the experimental default value)

[0035] • The current tree height e is the height of the positive slope tree model and the negative slope tree model, and is mainly used to obtain the abnormal values of the time series data;

[0036] • The limit height l of the tree is the maximum height of the positive slope tree model and the negative slope tree model, so as to prevent overfitting, and the height limit l = log2ψ;

[0037] Step 4) Constructing the positive and negative slope forest model:

[0038] In the construction of the positive and negative slope forest model, the positive slope tree model and the negative slope tree model are combined to construct the positive slope forest model and the negative slope forest model respectively. In the forest model construction process, the maximum number of trees contained by the forest is t.

[0039] The number t of trees in the construction of the positive and negative slope forest model controls the size of the forest integrated tree. The results show that when t = 100, the precision reaches the maximum (unless otherwise specified, t = 100 will be taken as the default value of the experiment);

[0040] Step 5) Performing abnormal point detection:

[0041] In the abnormal value detection, firstly, the abnormal value proportion η of the entire data set is set, and then the abnormal value number of the positive slope forest model and the abnormal value number of the negative slope forest model are calculated respectively. Next, the data to be tested is repeatedly traversed, the left and right nodes of each tree are traversed, and the height of each tree is recorded. The slope value is sorted from large to small, and the slope value is selected according to the shortest path principle until the number of abnormal points is met. Finally, according to y i+1 = k(t i+1 -t i ) + y i , the abnormal data points are returned.

[0042] The technical effects of the present application are further described in combination with the simulation experiment:

[0043] 1. Simulation conditions and contents

[0044] The experiment adopts six real data sets with similar characteristics to missiles for testing. Among them, the AQI data set is a one-dimensional time series data with a length of 8642 collected by 35 AQI monitoring stations. The time series data of 10 randomly selected AQI monitoring stations are used to test the proposed PNSF algorithm and the comparative algorithms (SVM, LOF, IF, RNN, and LSTM). The other five real data sets are ECG data set, average population data set, and climate data (daily temperature data, daily wind speed data, and daily humidity data) set. The simulation platform uses Inter(R) Core(TM) i7-7700 processor with 4 cores, 16 GB memory, Windows 10 operating system, and Python to implement the PNSF and comparative algorithms (SVM, LOF, IF, RNN, and LSTM).

[0045] 2. Simulation result analysis

[0046] Firstly, the algorithm is evaluated on the AQI data set of 10 monitoring stations and compared with SVM, LOF, IF, RNN, and LSTM algorithms. The specific results are shown in Table 1.

[0047] Table 1 Detection results of PNSF method on AQI data (average ± standard deviation)

[0048]

[0049]

[0050] From Table 1, it can be seen that the PNSF method shows the best detection results on the AQI data of 10 monitoring stations compared with SVM, LOF, IF, RNN, and LSTM algorithms. The PNSF algorithm improves the average Precision by 25.4%, 51.7%, 62.7%, 58.4%, and 54.9% compared with SVM, LOF, IF, RNN, and LSTM algorithms, respectively.

[0051] Then, the algorithm is evaluated on the other five data sets and compared with SVM, LOF, IF, RNN, and LSTM algorithms. The specific results are shown in Table 2.

[0052] Table 2 Detection results of PNSF method on other five data (average ± standard deviation)

[0053]

[0054] As shown in Table 2, the PNSF method shows the best detection results on other data sets compared with the SVM, LOF, IF, RNN and LSTM methods. The PNSF method improves the Positive Precision by 16.8%, 34.8%, 25.4%, 39.6% and 31.2% respectively compared with the SVM, LOF, IF, RNN and LSTM methods.

[0055] In addition, the time performance of the PNSF method and the comparative methods is also evaluated. The data sets are generated using a sine wave with 10,000 to 100,000 different scales, where the outliers of each data set are Gaussian noise, and the proportion of outliers of each data set is 0.01. Figure 2 The average results of the test time of six methods running 10 times on different length data sets are given, that is, the time performance chart of the PNSF method.

[0056] From Figure 2 It can be seen that the running time of the PNSF method under different scales of data has linear stability, and is less than the running time of the other five methods. This is mainly because the definition of "potential outliers" in the PNSF algorithm fundamentally compresses the data space, thereby making the method have good performance in time. Therefore, the PNSF method can quickly and accurately detect outliers in very large data sets.

[0057] The above experimental results and experimental analysis can show that the proposed missile rapid outlier detection method based on the positive and negative slope forest model has significantly improved detection accuracy, and the time performance of the proposed PNSF method is evaluated under different scales of simulation data, and the results show that the PNSF method has low time cost and stability.

Claims

1. A rapid missile anomaly detection method based on positive and negative slope forest, characterized in that, The steps include the following: (1) Collect time-series data on missile status: Using n sensors at a sampling frequency f z Collect time-series signals of n missile states of length T, where n≥1, T≥4687500, and f s ≥1500Hz; (2) Definition of potential outliers: Let one-dimensional time series data be (T,Y)={(t i ,y i )}, where t i ∈T,y i ∈Y and t i <t i+1 (i = 1, 2, ..., N), First, calculate the slope between data points at adjacent time points, then retain the slope where k ≠ 0, and finally store the potential outliers with positive changes and the potential outliers with negative changes into the set of potential outliers with positive changes and the set of potential outliers with negative changes, respectively. (3) Construct a positive and negative slope tree model: (3a) In the process of constructing the positive slope tree model, from the training set K + A sample of size ψ is randomly selected as the root node, and then a slope split point k is randomly selected from the sample. + The remaining data points are compared with the split point. Data points smaller than the split point are divided into left nodes, and data points greater than or equal to the split point are divided into right nodes. This process is repeated recursively until potential missile outlier points are separated or the tree height limit is reached. (3b) In the process of constructing the negative slope tree model, since the smaller the value, the greater the probability of being judged as a real outlier, data points greater than or equal to the split point are divided into left nodes, and data points less than the split point are divided into right nodes. (4) Constructing a forest model with positive and negative slopes: When constructing the positive and negative slope forest model, the positive slope tree model and the negative slope tree model are combined to construct the positive slope forest model and the negative slope forest model respectively. In the process of constructing the forest model, the maximum number of trees that the forest can hold is t, and the number of trees t controls the size of the forest ensemble tree. The results show that the accuracy reaches its maximum when t = 100. (5) Perform outlier detection: In outlier detection, the outlier proportion η of the entire dataset is first set, and then the number of outliers in the positive slope forest model is calculated separately. The number of outliers in the negative slope forest model Next, repeatedly traverse the data to be tested, iterating through the left and right nodes of each tree and recording the height of each tree. Sort the slope values ​​from largest to smallest, and select the slope value according to the shortest path principle until the number of outliers is satisfied. Finally, based on y... i+1 =k(t) i+1 -t i )+y i Returns the abnormal data point.

2. The method for rapid detection of missile anomalies based on positive and negative slope forests according to claim 1, characterized in that, The specific steps for proposing potential outliers in the one-dimensional time series described in step (2) are as follows: 2a) Given a one-dimensional time series (T, Y) = {(t i ,y i )}, where t i ∈T,y i ∈Y and t i <t i+1 (i = 1, 2, ..., N); 2b) Calculate t i The value of y at time t i With t i+1 The value of y at time t i+1 The slope between 2c) Judgment The relationship between the value of and 0, if Then it is stored in the set K of potential outliers with positive changes. + Conversely, it is stored in the set K of potential outliers with negative changes. - ; For K + For potential outliers in the set, the larger the value, the greater the probability that it will be identified as a real outlier. - The smaller the value of a potential outlier in the set, the lower the probability that it will be identified as a real outlier.

3. The method for rapid detection of missile anomalies based on positive and negative slope forests according to claim 1, characterized in that, The specific steps for constructing the positive and negative slope forest tree model in step (3) are as follows: 3a) From the training set K = {K} + ,K - Randomly select a sample of size ψ from the root node: K′ + ←sample(K + ,ψ) K′ - ←sample(K - ,ψ) 3b) Set the height limit of each tree to l = log2ψ, and if the height e of the current tree is less than l or |K| > 1, then execute steps 2c)-2f); otherwise, return to the outer node |K| → Size. 3c) From sample K ′+ and K′ - Randomly select slope division points k in each. + and k - ; k + ←sampleK′ + k - ←sampleK′ - 3d) If K ′+ Any remaining data point in the sequence is less than the split point k. + If the data is positive, it is divided into left nodes; otherwise, it is divided into right nodes. This process is repeated recursively until potential outliers are separated or the tree height limit is reached. 3e) If K′ - Any remaining data point in the sequence is greater than or equal to the split point k. - If the data is positive, it is divided into left nodes; otherwise, it is divided into right nodes. This process is repeated recursively until potential outliers are separated or the tree height limit is reached. 3f) Return the internal nodes of the tree The above steps involve 5 parameters, namely, the training sample K. + K - The subsample size is ψ, the current tree height is e, and the tree's height limit is l. The specific values ​​and explanations of each parameter are as follows: Training samples K + and K - This is the output of Algorithm 1. The data samples of potential outliers with positive and negative changes are stored in K respectively. + and K - middle; The subsample size ψ controls the sampling size. As the sampling size increases, the subsample size has little impact on the experimental results. Based on the isolated forest method, ψ is set to 256. The current tree height e represents the height of the positive slope tree model and the negative slope tree model, and is mainly used to obtain outliers in time series data. The height limit l of the tree is the maximum height of both the positive slope tree model and the negative slope tree model to prevent overfitting. The height limit l = log2ψ.

4. The method for rapid detection of missile anomalies based on positive and negative slope forests according to claim 1, characterized in that, The steps for constructing positive slope forests and negative slope forests described in step (4) are as follows: 4a) Given the number of trees t and the input data K = {K} + ,K - }; 4b) From K = {K + ,K - A sample of size ψ is randomly selected from}. 4c) Iterate through the pre-defined number of trees t starting from i=1 and perform the addition operation. Positive Slope Forest←Positive Slope Forest∪Positive Slope iTree(K′ + ,0,l) Negative Slope Forest←Negative Slope Forest∪Negative Slope iTree(K′ - ,0,l) The number of trees t in the above steps controls the size of the forest ensemble tree. The results show that the accuracy reaches its maximum when t = 100.

Images