WO2018126984A2 - 一种基于mea-bp神经网络wsn异常检测方法 - Google Patents

一种基于mea-bp神经网络wsn异常检测方法 Download PDF

Info

Publication number
WO2018126984A2
WO2018126984A2 PCT/CN2017/119421 CN2017119421W WO2018126984A2 WO 2018126984 A2 WO2018126984 A2 WO 2018126984A2 CN 2017119421 W CN2017119421 W CN 2017119421W WO 2018126984 A2 WO2018126984 A2 WO 2018126984A2
Authority
WO
WIPO (PCT)
Prior art keywords
node
data
cluster
neural network
sensor node
Prior art date
Application number
PCT/CN2017/119421
Other languages
English (en)
French (fr)
Other versions
WO2018126984A3 (zh
Inventor
李光辉
顾晓勇
Original Assignee
江南大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 江南大学 filed Critical 江南大学
Publication of WO2018126984A2 publication Critical patent/WO2018126984A2/zh
Publication of WO2018126984A3 publication Critical patent/WO2018126984A3/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W24/00—Supervisory, monitoring or testing arrangements
    • H04W24/04—Arrangements for maintaining operational condition
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W84/00—Network topologies
    • H04W84/18—Self-organising networks, e.g. ad-hoc networks or sensor networks

Definitions

  • the invention belongs to the field of wireless sensor network (WSN) data reliability detection technology, and particularly relates to a WSN anomaly detection method based on MEA-BP neural network.
  • WSN wireless sensor network
  • wireless sensor network As a wireless self-organizing network, wireless sensor network (WSN) has the characteristics of low energy consumption, flexible nodes, no need for manual maintenance, and long-term operation in harsh environments. It is one of the most common applications to collect environmental data and monitor specific events by spreading sensor network nodes in the target monitoring area. Due to the limited resources of wireless sensor nodes, and the interference and damage of external factors, or the impact of external environmental emergencies, the data collected by nodes is likely to have significant deviation from the environmental characteristics under normal circumstances. abnormal data. Therefore, designing an effective anomaly detection method is the focus of research on wireless sensor network anomaly detection in recent years.
  • the traditional BP neural network has no theoretical basis for the selection of many parameters during training and learning, which makes the actual neural network application have limitations.
  • the main disadvantages are slow learning speed, poor fault tolerance, and convergence to local minimum values.
  • the wireless sensor network node for environmental monitoring the temperature data collected by the wireless sensor network node, whether it is the amplitude, frequency, or mean, median, variance, etc., of the statistical data will be compared with other data in the same sampling period. Obvious differences, if you do not consider the difference between different types of data, will undoubtedly affect the performance of the detection algorithm, want to more accurately determine the data anomaly, in addition to the temporal correlation of the data itself also consider spatial correlation.
  • BP neural network algorithm has the problems of easy to fall into local optimal solution, long training time and low efficiency.
  • the present invention provides a WSN anomaly detection method based on MEA-BP neural network, and the BP neural network algorithm is easy to fall into a local optimal solution, long training time, low efficiency, etc.
  • BP neural network is improved by thinking evolutionary algorithm, which improves the performance of BP neural network algorithm, accelerates the learning rate of BP neural network, effectively improves the accuracy of abnormal data detection and reduces the false positive rate.
  • the MEA-BP neural network WSN abnormality detecting method of the present invention comprises the following steps:
  • S1 initializes each distributed sensor node, and each sensor node starts collecting data
  • the sliding window of the sensor node Xtj is Wj
  • the sliding window size of each sensor node is m
  • the sensor node Xtj The sequence of measurement data on its sliding window Wj is The data collected by the sensor node Xtj at time tp is The data includes h attribute measurements, then
  • S2 uses the K-means algorithm to spatially divide each sensor node to obtain several clusters
  • each cluster includes 1 cluster head node Xtc and q distribution nodes (X t1 , X t2 ,..., X tq );
  • S3 uses the thought evolution algorithm to optimize the BP neural network parameters, optimizes the weight and threshold of the BP neural network through the convergence and alienation operation, obtains the optimal weight and threshold, inputs the optimal weight and threshold, and establishes the MEA-BP nerve.
  • Network model
  • S4 uses a distributed algorithm to independently perform anomaly detection on the sensor nodes (X t1 , X t2 , ..., X tq ) in each cluster. After the anomaly detection is completed, the sensor nodes (X t1 , X t2 , ..., X tq ) The detection result is passed to the cluster head node X tc of the cluster of clusters for further verification.
  • step S2 includes the following steps:
  • S21 firstly selects K sensor node objects from the target monitoring area distribution sensor node object as K cluster centers;
  • S22 then calculates the similarity between the sensor node object and the K cluster centers for the sensor node objects other than the cluster center, and obtain the cluster center with the closest similarity to the sensor node object;
  • S23 assigns a sensor node object to a cluster of cluster centers closest to the similarity of the sensor node object, and obtains K clusters after all sensor nodes are allocated;
  • the step S3 includes the following steps: generating training data; determining a BP neural network topology; performing parameter setting by a mind evolution algorithm; randomly generating an initial population, a superior subpopulation, and a temporary subpopulation; and performing a convergent operation on the subpopulation;
  • the alienation operation is performed on the sub-population; whether the end condition is satisfied is satisfied, and if it is satisfied, the optimal individual is output, and the optimal weight and the threshold are obtained, otherwise the convergence and alienation operation is performed again.
  • step S4 includes the following steps:
  • S41 independently performs anomaly detection on the sensor nodes (X t1 , X t2 , . . . , X tq ) by time correlation, and scans the data of the window Wj through the sensor node X tj with each current moment. Train the neural network, complete the forecast of the next moment data, and select the training data. After the v sample data, calculate the model residual S of the MEA-BP neural network by formula (1):
  • F is the average value of the selected sample data
  • S42 calculates a confidence interval of the current moment of the sensor node, and the confidence interval is
  • Spre is the predicted value of the MEA-BP neural network for the next time data
  • t ⁇ /2 is the predicted value of the MEA-BP neural network for the next time data
  • v-1 is the t distribution of the degree of freedom v-1
  • the appropriate ⁇ value is selected in the t distribution table to obtain the t value
  • the sensor node (X t1 , X t2 , . . . , X tq ) transmits the detection result to the cluster head node X tc of the cluster;
  • the S45 cluster head node X tc verifies the cause of abnormal data of the sensor node through spatial correlation detection and introduction of voting mechanism, including event abnormality, node failure abnormality and misjudgment.
  • performing the spatial correlation detection and introducing the voting mechanism in the step S45 includes the following steps:
  • S451 compares the abnormal data of the sensor node with other node data in the same cluster of the sensor node; and sets q+1 sensor nodes to form a cluster, and each cluster includes one cluster head node X tc And q distribution nodes (X t1 , X t2 ,..., X tq );
  • step S454 in the step S453, when it is determined that the sensor node abnormal data is due to a node failure or a misjudgment, further determining whether the sensor node abnormal data is due to a node failure or a misjudgment, specifically comprising the following steps: Time correlation detection, when the sensor node generates abnormal data for continuous time, it is determined that the abnormal data of the sensor node is due to a node failure; when the sensor node has only abnormal data at the moment, and the data generated at other times is normal data, the judgment is made. Sensor node anomaly data is due to false positives.
  • the traditional BP neural network algorithm is easy to fall into the local optimal solution, long training time, low efficiency, etc. It is difficult to meet the detection requirements.
  • the proposed evolutionary algorithm is used to improve the BP neural network to improve the performance of the BP neural network algorithm.
  • the multi-dimensional data is processed by using the event correlation between the sensor network data streams and the spatial correlation between different nodes, thereby effectively improving the accuracy of abnormal data detection.
  • the experimental results for multiple data sets show that the average detection rate of the traditional BP neural network algorithm is 96.6%, and the average false positive rate is 2.2%.
  • the average detection rate of the algorithm of the present invention is increased to 98.8%, and the average false positive rate is 1.3. %.
  • the method of the invention accelerates the learning rate of the BP neural network by optimizing the weight and the threshold, improves the abnormal detection rate, and reduces the false positive rate.
  • experimental results for multiple data sets indicate that the method of the present invention has an average speed increase of 21.8% relative to conventional BP neural network.
  • Figure 1 is a schematic diagram of a t distribution table
  • 2 is a flow chart of a k-means clustering algorithm
  • Figure 4 is a flow chart of MEA-BP neural network anomaly data detection.
  • Figure 5 is a comparison of training time between BP algorithm and MEA-BP algorithm on different scale data sets.
  • the present invention proposes a WSN anomaly detection method based on the MEA-BP neural network. Before introducing the method of the present invention, some definitions are first introduced:
  • Time series data is a series of sequence data generated by sensor nodes in chronological order. It is characterized by rapid change, large quantity and continuous arrival. Therefore, before establishing the detection model, the sliding window mechanism should be introduced first, and the sliding window is used to observe the change of data in the latest time period, and the abnormal value detection is performed inside the sliding window.
  • sliding window model, sliding window model is used to observe the time series data in the most recent sampling period, by taking a fixed length sliding window of the sensor data, reducing the time complexity by processing the newly added and just left data
  • the number of sensor nodes is n
  • the sliding window of the sensor node X tj is W j
  • the window size is m
  • the measurement data sequence of the sensor node X tj on its sliding window W j is
  • the data collected by the sensor node X tj at time tp is The data includes h attribute measurements, then
  • Detection rate refers to the ratio of the number of abnormal data samples detected by the algorithm to the total number of actual abnormal data samples.
  • the false positive rate refers to the ratio of the number of normal data samples that are incorrectly judged by the algorithm to the total number of normal data samples.
  • the K-means algorithm is used to cluster spatial nodes, and the sensor nodes with similar data are divided into the same cluster, and then the WSN anomaly detection method based on MEA-BP neural network is proposed.
  • the method is mainly divided into three steps: parameter optimization, abnormal data detection and data anomaly determination.
  • the main features are as follows: (1) Parameter optimization of BP neural network using thought evolutionary algorithm, BP neural network through operation of convergence and other operations The weights, thresholds, etc. are optimized; (2) in the detection stage of the abnormal data, the abnormal data points that may exist in the data stream collected by the sensor nodes are mainly identified, and the step uses a distributed algorithm, which is independent at each sensor node.
  • each sensor node passes the result to the cluster head node for further verification; (3) when the abnormal data is detected, it proposes to train the neural network through the historical data set of the sliding window of the sensor node at each current moment to complete the prediction at the next moment, Determining the abnormality of the data at the next moment by the model residual of the neural network Interval, when the next time data fall within the confidence interval, the data is determined to be normal eh, conversely, the data is further verified by spatial correlation cluster head node, the method comprising the particular steps of:
  • Each segment sensor node is initialized, and each sensor node starts collecting data.
  • the number of sensor nodes is n
  • the sliding window of the sensor node X tj is W. j
  • the sliding window size of each sensor node is m
  • the measurement data sequence of the sensor node X tj on its sliding window W j is
  • the data collected by the sensor node X tj at time tp is The data includes h attribute measurements, then
  • the K-means algorithm is used to spatially cluster each sensor node to obtain several clusters, and the sensor nodes with similar data are divided into the same cluster to improve the spatial similarity of the nodes in the cluster. This work is completed before the node fault detection. Then, the time correlation detection and the spatial correlation detection are performed on the data collected by the node, and the result of the previous spatial division is used in the spatial correlation detection, and the sensor node first performs time correlation detection on the data collected at the current time, and the time correlation detection is displayed. When there is a problem, the sensor node notifies the cluster head node of the suspicious data, and the cluster head node performs spatial correlation detection on the data; referring to FIG.
  • K objects are selected from the sensor node objects as clusters.
  • Center for the remaining sensor nodes, according to their similarity with the selected cluster center (Euclidean distance), respectively assign the sensor nodes to the clusters with the most similar cluster centers, and then calculate each The clustering center of the new cluster (the mean of all the objects in the cluster), repeating this over and over again Until the cluster centers began to converge so far;
  • the K-means algorithm steps are as follows:
  • Step 2 repeat the following process until the cluster center converges
  • c (i) represents the cluster closest to the sensor node sample i and the K cluster centers, and divides each point into clusters closest to the cluster center;
  • the denominator in equation (3) represents the total number of samples in each cluster, and the numerator is the coordinate sum corresponding to the sensor node sample i in each cluster.
  • each cluster includes 1 cluster head node X tc and q distribution nodes (X t1 , X t2 , . . . , X tq );
  • the BP neural network is optimized by the thought evolution algorithm.
  • the weight and threshold of the BP neural network are optimized by the convergence and alienation operation to obtain the optimal weight and threshold.
  • the optimal weight and threshold are input to establish the MEA-BP neural network.
  • the model, the process of optimizing the BP neural network is shown in Figure 3. The general steps are as follows:
  • Step 1 The initial population of thought evolution is generated. N sets of numbers are randomly generated in the solution space as the initial group. Each group contains n elements representing one individual (ie, neural network structure), and each individual's matrix is 1*n, group. The matrix is N*n, and the group contains N individuals;
  • Step 2 According to the BP neural network topology, the solution space is mapped to the coding space, and one solution of each coding space corresponding problem, that is, an individual, the coding length is equal to the number of elements in each individual, and the coding length n is
  • t is the number of input nodes of the neural network
  • w is the number of output nodes
  • L is the number of nodes in the hidden layer.
  • Step 3 Define the number of iterations, the winner subgroup M, and the number of temporary subgroups T; the set of all individuals in each generation of the evolutionary process becomes a group, and one group is divided into M winner subgroups and T temporary subgroups.
  • Each superior subgroup and temporary subgroup contains SG individuals, and SG is:
  • Step 4 Determination of the score function, the neural network consists of the input layer, the hidden layer, and the output layer.
  • the output layer threshold ⁇ u (u 1, 2,... , w) ⁇
  • the matrix is w*w.
  • W ij is the first to the (L*t)th element in a single individual, and the matrix is L*t
  • V ju is the first (L*t+1) to (L*t+) in a single individual.
  • ⁇ j is the first (L*t+w*L+1) to the (L*t+w*L+L) elements in a single individual, and the matrix is w*L;
  • ⁇ u is the first (L*t+w*L+1) to the last element in a single individual.
  • ⁇ W is called the activation value of this model, which is the input summation of the model; in formula (7), f() is the model activation function.
  • Equation (8)(9) is the same as above;
  • Step 6 Sub-group convergence operation.
  • the process in which the individual becomes the winner and the competition is called convergence.
  • the process of convergence of a sub-group if a new winner is not produced, the sub-group becomes mature and the convergence process ends. Focusing on each winner, respectively, obeying the normal distribution to generate individuals, forming M winners and T temporary subgroups, each subgroup containing SG individuals, the normal distribution can be expressed as N(u, ⁇ ) Where u is the central vector of the normal distribution, ⁇ is the covariance matrix of the normal distribution, and the center of the normal distribution is the coordinates of the winner, ie the weight and threshold of the winner;
  • Step 7 Subgroup alienation operation.
  • the process of alienation is the process in which the sub-groups in the entire solution space become winners and compete. Through the global bulletin board, it records the score function value and maturity of each sub-group, and competes globally among the sub-groups. If a temporary sub-group score is higher than a mature sub-sub-group score, the temporary sub-group is replaced. For the winner group, the individual in the original winner group is abandoned; if the score of a mature temporary subgroup is lower than the score of any of the winner subgroups, the temporary subgroup is abandoned and the individual is released. The number of abandoned temporary subgroups is recorded as Tr, and the number of abandoned winner subgroups is recorded as Mr. Under the guidance of the global bulletin board, Mr+Tr temporary subgroups are regenerated in the solution space;
  • Step 8 Analyze the optimal individual. Repeat steps 6 and 7 above, and when the iterative stop condition is satisfied, the thought evolution algorithm ends the optimization process. At this time, according to the coding rule, the found optimal individual is parsed, so as to obtain the optimal weight and threshold of the corresponding BP neural network, input the optimal weight and threshold, and establish the MEA-BP neural network model;
  • the abnormality detection performed independently in each sensor cluster node (X t1, X t2, ... , X tq) pair, after completion of the abnormality detection sensor nodes (X t1, X t2, ... , X tq ) further passes the detection result to the cluster head node X tc of the cluster, including the following steps:
  • S41 independently performs anomaly detection on the sensor nodes (X t1 , X t2 , . . . , X tq ) by time correlation, and scans the data of the window Wj through the sensor node Xtj with each current moment. Train the neural network, complete the forecast of the next moment data, and select the training data. After the v sample data, calculate the model residual S of the MEA-BP neural network by formula (1):
  • F is the average value of the selected sample data
  • S42 calculates a confidence interval of the current moment of the sensor node, and the confidence interval is
  • Spre is the predicted value of the MEA-BP neural network for the next time data
  • t ⁇ /2 v-1 is the t distribution of the degree of freedom v-1
  • the appropriate ⁇ value is selected in the t distribution table to obtain the t value;
  • the t distribution table is shown in Figure 1.
  • the normal distribution is also known as the Gaussian distribution. If the random variable obeys a Gaussian distribution with a mathematical expectation of ⁇ and a variance of ⁇ 2, denoted as N( ⁇ , ⁇ 2), we usually say that the standard is positive.
  • the sample with the capacity v is extracted from the standard normal distribution with the mean ⁇ and variance ⁇ 2.
  • the sample obeys the normal distribution with the mean ⁇ and the variance ⁇ 2/v.
  • the population variance ⁇ 2 is always unknown, so that only Can be replaced with s2. If v is large, then s2 is a better estimator of ⁇ 2 and is still an approximate standard normal distribution; if v is small, the difference between s2 and ⁇ 2 is large, so the sample distribution is no longer Is a standard normal distribution, but obeys the t distribution.
  • the t distribution is a curve that varies according to the degree of freedom. It is symmetrically distributed on both sides of the Y axis of the coordinate axis, and the mean value is 0.
  • the t distribution is suitable when the population standard deviation is unknown. Then replace the population standard deviation with the sample standard deviation.
  • the degree of freedom in the t distribution refers to the number of variables that can be freely changed.
  • the sample t test only estimates one parameter: the population mean, the sample size v constitutes v kinds of information used to estimate the population mean and its variability, consumes one degree of freedom to estimate the mean, and the remaining v-1 degrees of freedom are used.
  • the sensor node (X t1 , X t2 , . . . , X tq ) transmits the detection result to the cluster head node X tc of the cluster;
  • the S45 cluster head node X tc verifies the cause of the abnormal data of the sensor node through the spatial correlation detection and the voting mechanism.
  • the reasons include the event abnormality, the node fault abnormality and the misjudgment, and the following steps are specifically included:
  • S451 compares the abnormal data of the sensor node with other node data in the same cluster of the sensor node; and sets q+1 sensor nodes to form a cluster, and each cluster includes one cluster head node X tc And q distribution nodes (X t1 , X t2 ,..., X tq );
  • the reference node is the node closest to the Euclidean distance of the cluster center node in the cluster; the reference node: using the K-means algorithm for each sensor node Spatial clustering is performed to obtain several clusters, and the sensor nodes with similar data are divided into the same cluster. According to the final clustering result, the node closest to the centroid Euclidean distance of the cluster is selected as the reference node;
  • step S454 in the step S453, when it is determined that the sensor node abnormal data is due to a node failure or a misjudgment, further determining whether the sensor node abnormal data is due to a node failure or a misjudgment, specifically comprising the following steps: Time correlation detection, when the sensor node generates abnormal data for continuous time, it is determined that the abnormal data of the sensor node is due to a node failure; when the sensor node has only abnormal data at the moment, and the data generated at other times is normal data, the judgment is made. Sensor node anomaly data is due to misjudgment;
  • each node In order to detect the abnormal data in the node, each node collects data in a certain period to form a data stream belonging to the node. In order to ensure the correctness of the data samples collected by the nodes in the area, each node needs to use before uploading the data.
  • the neural network predicted values replace the outliers within the sliding window.
  • the data samples were taken from sensor network data from Intel Berkeley Labs, which was sampled every 31 seconds.
  • the temperature of the 100 nodes of the sensor node 1-5 is selected, the humidity is used as the training data, and the temperature and humidity of the 100 groups are used as the prediction data.
  • the sensor node 6 node 2000 temperature data is selected as the training data for parameter optimization comparison.
  • S1 training [20.6156, 20.6254, 20.6450, 20.6352, 20.6450, 20.6156, 20.6058, 20.576420.4882, 20.4588, 20.4392, 20.4196, 20.3804, 20.3510, 20.3020, 20.2726, 20.2530, 20.1942, 20.184420.1354, 20.0864, 20.0668, 20.0374, 20.0178,19.9982,19.9786,19.9688,19.8414,19.8022,19.8022,19.7826,19.8022,19.8218,19.8316,19.8610,19.9002,19.9296,19.9786,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1256,20.1354,20.1550,20.2236, 20.2432,20.2432,20.3020,20.3216,20.3608,20.4588,20.
  • S1 prediction [20.7724, 20.7332, 20.7528, 20.7430, 20.7332, 20.7332, 20.7234, 20.7136, 20.7234, 22.1530, 20.7038, 20.6940, 20.6940, 20.7038, 20.6940, ..., 22.1530, 20.6744, 20.6450, 20.6352, 20.6254] [39.6200, 39.9929 , 40.3652, 40.9055, 41.1414, 41.2761, 41.3771, 41.4780, 41.7805, 41.8812, 41.9818, 41.9483, 42.2500, 42.3840, 42.3170, ..., 52.3170, 42.2500, 42.2835, 42.0824, 41.9818]
  • S2 training [20.6226, 20.6224, 20.6350, 20.6352, 20.6350, 20.6256, 20.6358, 20.5764, 20.4782, 20.4388, 20.4492, 20.4296, 20.3604, 20.3310, 20.3320, 20.2326, 20.2330, 20.1042, 20.1644, 20.1354, 20.0464, 20.0468, 20.0274, 20.0178,19.9482,19.9486,19.9588,19.8614,19.8322,19.8222,19.7926,19.8122,19.8518,19.8516,19.8510,19.9502,19.9296,19.9386,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1236,20.1454,20.1350,20.2236, 20.2332,20.2432,20.3020,20.3216,20.3608,20.4588
  • S3 training [20.6156, 20.6254, 20.6450, 20.6352, 20.6450, 20.6156, 20.6058, 20.576420.4882, 20.4588, 20.4392, 20.4196, 20.3804, 20.3510, 20.3020, 20.2726, 20.2530, 20.1942, 20.1844, 20.1354, 20.0864, 20.0668, 20.0374, 20.0178,19.9982,19.9786,19.9688,19.8414,19.8022,19.8022,19.7826,19.8022,19.8218,19.8316,19.8610,19.9002,19.9296,19.9786,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1256,20.1354,20.1550,20.2236, 20.2432,20.2432,20.3020,20.3216,20.3608,20.4588,20
  • S3 prediction (20.4396, 20.4102, 20.4102, 20.4004, 20.3710, 20.3612, 20.3612, 20.3612, 20.3514, 20.3710, 20.3808, 20.3416, 20.3220, 20.3220, ..., 20.3318, 20.3318, 22.3220, 20.3514, 20.3416) (39.4200, 40.7963) , 40.4422, 40.8755, 41.3614, 41.4161, 41.5871, 41.6780, 41.8405, 41.9236, 42.0242, 42.0200, 42.3300, 42.4240, 42.3170, ..., 42.3270, 42.3400, 52.3525, 42.1624, 41.8718)
  • S4 training [20.4154, 20.4254, 20.4450, 20.4352, 20.4450, 20.4154, 20.4057, 20.574420.4772, 20.4577, 20.4392, 20.4194, 20.3704, 20.3510, 20.3020, 20.2724, 20.2530, 20.1942, 20.1744, 20.1354, 20.0744, 20.0447, 20.0374, 20.0277,19.9972,19.9774,19.9477,19.7414,19.7022,19.7022,19.7724,19.7022,19.7217,19.7314,19.7410,19.9002,19.9294,19.9774,19.9774,20.0070,20.0374,20.0472,20.1040,20.1040,20.1254,20.1354,20.1350,20.2234, 20.2432,20.2432,20.3020,20.3214,20.3407,20
  • S4 prediction (20.4394, 20.4102, 20.4788, 20.4494, 20.4984, 20.4788, 20.3412, 20.3412, 20.3514, 20.3710, 20.3808, 20.3414, 20.2240, 20.2240, ..., 20.2338, 20.2240, 20.3318, 22.3514, 20.3414) (39.4500, 40.7343) , 40.4222, 40.8555, 41.4214, 41.5341, 41.4471, 41.4980, 41.8805, 41.9334, 42.0142, 42.0120, 42.3100, 42.4240, 42.3070, ..., 42.3470, 42.3800, 42.3925, 52.2024, 41.9118)
  • S5 training [20.4154, 20.4254, 20.4450, 20.4352, 20.4450, 20.4154, 20.4058, 20.574420.4882, 20.4588, 20.4392, 20.4194, 20.3804, 20.3510, 20.3020, 20.2724, 20.2530, 20.1942, 20.184420.1354, 20.0844, 20.0448, 20.0374, 20.0178,19.9982,19.9784,19.9488,19.8414,19.8022,19.8022,19.7824,19.8022,19.8218,19.8314,19.8410,19.9002,19.9294,19.9784,19.9884,20.0080,20.0374,20.0472,20.1040,20.1040,20.1254,20.1354,20.1550,20.2234, 20.2432,20.2432,20.3020,20.3214,20.3408,20.4588,20
  • S6 training (24.7218, 24.6434, 24.6140, 24.5454, 24.5356, 24.5356, 24.5258, 24.4964, 24.4964, 24.4768, 24.4964, 24.4376, 24.36924.36924.36924.3298, ..., 23.5164, 23.5164, 23.4968, 23.4968, 23.4968, 23.4870, 23.4870, 23.4674, 23.4576, 23.4380, 23.438)
  • the temperature collected by the sensor node is now taken as an example.
  • the spatial nodes are first partitioned by the K-means algorithm, assuming that nodes 1-5 are in the same cluster, and then the parameters of the nodes are used for parameter optimization as shown in Fig. 2 to establish an MEA-BP neural network. model.
  • the neural network model residual is used to determine the confidence interval. After that, the node collects new data and predicts it through the sliding window. If the collected data is not in the confidence interval, it is determined that the data is an abnormal value and is transmitted to the cluster head for verification. Otherwise, the collected data is normal.
  • the initial weights and thresholds of the neural network are random.
  • the optimal individual is obtained by the optimization algorithm.
  • the optimal individual is parsed according to the coding rules.
  • the elements in the individual are divided into four parts, the input layer and the hidden layer weight W, implied Layer and output layer weight V, hidden layer threshold ⁇ , output layer threshold ⁇ .
  • W [0.5773,-0.0366,-0.5442,0.6235,1.1156,0.3768,0.3825,0.3136,0.1267,-0.7622,-0.5179,1.0760,-0.76 51,-0.6149,0.3441,0.9941,...,0.0842,0.9022,-0.2135 , 0.0192;
  • 0.9876 -0.5693, 0.2064, 1.008, -0.7505, -0.0915, 0.8364, 0.0396, 1.3324, 0.6910, 0.4675, 0.2987, -0.7527, -0.5169, -0.8266, 1.0537, ..., 0.9445, -0.1559, -0.2290, 0.7849;
  • 0.4429 -0.5530, 0.2594, -0.0055, -0.2293, -0.1312, 0.1046, -1.1803, 0.6552, 1.1167, 0.2293, 0.8949, 0.2872, 0.0639, -0.7069, -0.6443,..., -0.5359, -0.7943, 0.0749, 0.2322 ;
  • V [-0.8852,-0.4987,-1.0526,0.3578,-0.0436,-0.6403,1.1844,-1.2437,-0.9961,-0.1289,-0.8779,-0.0577,-0.5220,0.4834,0.3050,0.5123,...,0.4276, -0.0404, 0.9207, 0.0199]
  • ⁇ [-0.3454, -0.0724, -0.9978, -1.0797, 1.0350, 0.2819, 0.8036, 0.2719, 0.2332, -0.6689, 0.9885, -0.7912, 0.1800, -0.6851, -0.7595, 0.2428,..., -0.1237, -0.2353 , -0.0445,0.4698]
  • node S1 has trained the MEA-BP model and adds new data through the sliding window.
  • the temperature prediction value is 20.8692 and the confidence interval range is [20.5058, 21.2326 ], the sensor data at this time is 20.7724. It can be seen that the data of 20.7724 is obviously within the range, and the data is considered normal at the moment; when the sliding window predicts the 10th data, the temperature prediction value is 20.7062, the confidence The interval range is [20.3428, 21.0696], and the sensor data is 22.1530 at this time. The data is outside the interval range and then passed to the cluster head node for spatial verification. The voting system finds that the data at that moment is in the same cluster temperature.
  • the temperature data collected by node 1 is abnormal, and the node is faulty or misreported. If the data is found to be similar in the same cluster, the event is considered to be present.
  • the humidity temperature data of other nodes can also accurately identify the abnormality.
  • the sample can be divided into true (TP), false positive (FP), true negative (TN), and false according to the combination of its real category and decision model detection category. There are four cases of false negative (FN).
  • the present invention uses two performance indicators: True Positive Rate and False Positive Rate.
  • the detection rate refers to the ratio of the number of abnormal data samples detected by the algorithm to the total number of actual abnormal data samples;
  • the false positive rate refers to the ratio of the number of normal data samples that are incorrectly judged by the algorithm to the total number of normal data samples.
  • BP algorithm and MEA-BP algorithm are used to train on different scale data sets, and training time is used as the measurement index.
  • the experimental results are shown in Fig. 5.
  • the abscissa indicates the data set number and the ordinate indicates the training time.
  • the average training time of the BP algorithm is 31.6s for each data set.
  • the MEA-BP algorithm has obvious advantages in training time, the average training time is 24.7s, and the training time relative to the BP algorithm is reduced by 21.8%. When the sample increases, the advantage is more obvious. This is because the MEA algorithm uses a shorter speed to achieve the error requirement and shortens the training time.
  • the performance of the MEA-BP algorithm is verified on the S1-S55 data sets. As shown in Table 1, under different data sets, the MEA-BP algorithm detection rate is increased by 2.2% on average compared with the BP algorithm; the MEA-BP algorithm false positive rate On average, the BP algorithm is reduced by 0.9%.
  • the method of the invention is based on the spatio-temporal correlation of the sensor nodes, and the nodes are divided by the K-means algorithm, and then the MEA-BP neural network model is established on each node, and the distributed nodes in each future are detected according to the model. Whether the arrived data is abnormal, and then pass the detected result to the cluster head node for verification.
  • anomaly detection is an in-depth study in various fields. The unique characteristics and strict constraints of wireless sensor networks make the research of this problem more challenging.
  • the method proposed by the invention is mainly for complex anomaly detection of resource-constrained wireless sensor networks, which greatly reduces the communication consumption between nodes, and can accurately detect abnormal data and has environmental adaptability.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Alarm Systems (AREA)

Abstract

本发明公开了一种基于MEA-BP神经网络WSN异常检测方法，将各分布传感器节点初始化，各传感器节点开始采集数据；利用K-means算法对各传感器节点进行空间分簇得到若干组簇；利用思维进化算法对BP神经网络进行参数优化，通过趋同异化操作对BP神经网络的权值和阈值进行优化，得到最优权值和阈值，输入最优权值和阈值，建立MEA-BP神经网络模型；采用分布式的算法，对每组分簇中传感器节点独立执行异常检测，异常检测完毕后传感器节点将检测结果传递到该组分簇的簇头节点进一步验证。提高了BP神经网络算法性能，加快了BP神经网络的学习速率，有效提高了异常数据检测的准确率，降低了误报率。

Description

一种基于MEA-BP神经网络WSN异常检测方法 技术领域
本发明属于无线传感器网络(WSN)数据可靠性检测技术领域，具体是涉及一种基于MEA-BP神经网络WSN异常检测方法。
背景技术
无线传感器网络(WSN)作为一种无线自组织网络，无线传感器网络具有低能耗、节点分别灵活、甚至无需人工维护，可以在恶劣环境中长时间工作等特点。通过将传感器网络节点散布在目标监测区域中，进行环境数据的采集以及特定事件的监测是目前最为普遍的应用之一。由于无线传感器节点资源有限，又容易受到外界因素的干扰和破坏，或者外部环境突发事件的影响，节点采集到的数据很有可能与正常情况下的环境特征产生明显偏差，这类数据称为异常数据。因此，设计一种有效的异常检测方法是近年来无线传感器网络异常检测研究的重点。
传统BP神经网络在训练学习时许多参数的选择没有理论依据，使得实际的神经网络应用具有局限性，存在不足之处主要有学习速度慢、容错能力差、会收敛于局部极小值等。以用于环境监测无线传感器网络节点为例，无线传感器网络节点采集到的温度数据无论是波动的幅度、频率，或者均值、中值、方差等统计特征都会和同一采样时间段内的其他数据有明显的差异，如果不考虑不同类型数据之间的差异性，无疑会影响检测算法的性能，想要更精确的判断数据异常，除了数据自身的时间相关性还要考虑空间相关性。另外，针对无线传感器网络环境数据的异常检测问题，BP神经网络算法存在容易陷入局部最优解、训练时间长、效率低等问题。
发明内容
发明目的：为了克服现有技术中存在的不足，本发明提供一种基于MEA-BP神经网络WSN异常检测方法，针对BP神经网络算法存在容易陷入局部最优解、训练时间长、效率低等问题，利用思维进化算法对BP神经网络进行改进，提高了BP神经网络算法性能，加快了BP神经网络的学习速率，有效提高了异常数据检测的准确率，降低了误报率。
技术方案：为实现上述目的，本发明的基于MEA-BP神经网络WSN异常检测方法，所述方法包括以下步骤：
S1将各分布传感器节点初始化，各传感器节点开始采集数据；
设传感器节点个数为n，各传感器节点为X tj(j＝1,2,…,n)，传感器节点Xtj的滑动窗口为 Wj，各传感器节点的滑动窗口大小均为m，则传感器节点Xtj在其滑动窗口Wj上的测量数据序列为
Figure PCTCN2017119421-appb-000001
传感器节点Xtj在tp时刻采集的数据为
Figure PCTCN2017119421-appb-000002
该数据包括h个属性测量值，则
Figure PCTCN2017119421-appb-000003
S2利用K-means算法对各传感器节点进行空间划分得到若干组分簇；
设q+1个传感器节点组成一组分簇，每组分簇中包括1个簇头节点Xtc和q个分布节点(X t1,X t2,…,X tq)；
S3利用思维进化算法对BP神经网络进行参数优化，通过趋同异化操作对BP神经网络的权值和阈值进行优化，得到最优权值和阈值，输入最优权值和阈值，建立MEA-BP神经网络模型；
S4采用分布式的算法，对每组分簇中传感器节点(X t1,X t2,…,X tq)独立执行异常检测，异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc进一步验证。
进一步地，所述步骤S2包括以下步骤：
S21首先从目标监测区域分布传感器节点对象中任意选择K个传感器节点对象作为K个聚类中心；
S22然后针对除聚类中心以外的传感器节点对象，分别计算传感器节点对象与K个聚类中心之间的相似度，得到与传感器节点对象相似度最接近的聚类中心；
S23将传感器节点对象分配给与该传感器节点对象相似度最接近的聚类中心的聚类，将所有传感器节点分配完成后得到K个聚类；
S24重新计算该K个聚类的聚类中心，得到新的聚类中心；
S25重新计算各传感器节点与新的聚类中心的相似度，回到步骤S22；
S26当重新计算的聚类中心收敛时，结束本操作。
进一步地，所述步骤S3包括以下步骤：产生训练数据；确定BP神经网络拓扑结构；通过思维进化算法进行参数设置；随机产生初始种群、优胜子种群和临时子种群；对子种群进行趋同操作；对子种群进行异化操作；判断是否满足结束条件，如果满足，则输出最优个体，获取最优权值和阈值，否则重新进行趋同异化操作。
进一步地，所述步骤S4包括以下步骤：
S41通过时间相关性对传感器节点(X t1,X t2,…,X tq)独立执行异常检测，利用每个当前时刻通过传感器节点X tj滑动窗口Wj的数据
Figure PCTCN2017119421-appb-000004
来训练神经网络，完成下一时刻数据 的预报，选取训练数据
Figure PCTCN2017119421-appb-000005
的后v个样本数据，通过公式(1)计算MEA-BP神经网络的模型残差S：
Figure PCTCN2017119421-appb-000006
其中，Er(r＝1,2,…,v)为选取的样本数据值，F为选取的样本数据平均值，
Figure PCTCN2017119421-appb-000007
S42计算传感器节点当前时刻的置信区间，置信区间为
Figure PCTCN2017119421-appb-000008
其中Spre为MEA-BP神经网络对下一时刻数据的预测值，t α/2,v-1为自由度v-1的t分布，在t分布表中选取合适的α值，得到t值；
S43当下一时刻数据Snew进入传感器节点滑动窗口内时，判断下一时刻数据Snew是否落入当前时刻的置信区间范围内，若是，则判断该数据Snew为正常数据；否则判断该数据Snew为异常数据；
S44异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc中；
S45簇头节点X tc通过空间相关性检测并引入投票机制验证传感器节点异常数据产生原因，原因包括事件异常、节点故障异常和误判。
进一步地，所述步骤S42中，选取α＝0.05。
进一步地，所述步骤S45中进行空间相关性检测并引入投票机制包括以下步骤：
S451将传感器节点的异常数据和与该传感器节点在同一分簇中的其他节点数据进行比较；设q+1个传感器节点组成一组分簇，每组分簇中包括1个簇头节点X tc和q个分布节点(X t1,X t2,…,X tq)；
S452预先设定误差值θ，设传感器节点的异常数据为ST，设与该传感器节点在同一分簇中的其他节点的数据为Si(i＝1,2,…,q-1)，若|ST-Si|<θ，则初始值为0的计数NN加1，统计最终的NN值；
S453若
Figure PCTCN2017119421-appb-000009
则判断该检测节点异常数据是由于事件异常；若
Figure PCTCN2017119421-appb-000010
则判断该检测节点异常数据是由于节点故障或误判；若
Figure PCTCN2017119421-appb-000011
则选取该分簇中的参照节点，设参照节点的数据为SCC，若|ST-SCC|≤θ，则判断该传感器节点异常数据是由于事件异常，若|ST-SCC|>θ，则判断该传感器节点异常数据是由于节点故障或误判；其中参照节点为距离该分簇中聚类中 心节点欧氏距离最近的节点；
S454对于所述步骤S453中，当判断得到该传感器节点异常数据是由于节点故障或误判时，进一步判断该传感器节点异常数据是由于节点故障还是误判，具体包括以下步骤：通过对传感器节点进行时间相关性检测，当传感器节点连续时间内均产生异常数据，则判断该传感器节点异常数据是由于节点故障；当传感器节点只有该时刻为异常数据，其他时刻产生数据均为正常数据，则判断该传感器节点异常数据是由于误判。
有益效果：本发明与现有技术比较，具有的优点是：
传统BP神经网络算法存在容易陷入局部最优解、训练时间长、效率低等问题，难以满足检测需求，本发明提出的利用思维进化算法对BP神经网络进行改进以提高BP神经网络算法性能，在处理多维数据时利用传感器网路数据流之间的事件相关性和不同节点之间的空间相关性，从而有效提高了异常数据检测的准确率。针对多个数据集的实验结果表明，传统的BP神经网络算法的平均检测率为96.6％，平均误报率为2.2％；本发明算法的平均检测率提高到了98.8％，平均误报率为1.3％。
与常规的BP神经网络相比，本发明的方法通过优化权值和阈值之后加快了BP神经网络的学习速率，提高了异常检测率，降低了误判率。如图5所示，针对多个数据集的实验结果表明，本发明的方法上相对于常规BP神经网络速度平均提高了21.8％。
附图说明
图1是t分布表示意图；
图2是k-means聚类算法流程图；
图3是MEA优化BP神经网络权值阈值流程图；
图4是MEA-BP神经网络异常数据检测流程图。
图5是BP算法和MEA-BP算法在不同规模数据集上的训练时间对比图。
具体实施方式
下面结合附图对本发明作更进一步的说明。
本发明提出了一种基于MEA-BP神经网络WSN异常检测方法，在介绍本发明方法之前，首先介绍一些定义：
1、传感器网络模型，在分布式传感器网络中，设传感器节点个数为n，各传感器节点为X tj(j＝1,2,…,n)。
2、时间序列数据，是由传感器节点按时间顺序产生的一系列序列数据，它的特点是变化快、大量和连续到达的。所以在建立检测模型之前，首先要引入滑动窗口机制，利用滑动窗 口来观察最近一个时间段内数据的变化情况，在滑动窗口内部进行异常值检测。
3、滑动窗口模型，滑动窗口模型是用来观察最近一个采样时间段内的时间序列数据，方法是对传感器数据取一固定长度的滑动窗口，通过处理新加入和刚离开的数据降低时间复杂度；在分布式传感器网络中，设传感器节点个数为n，各传感器节点为X tj(j＝1,2,…,n)，传感器节点X tj的滑动窗口为W j，各传感器节点的滑动窗口大小均为m，则传感器节点X tj在其滑动窗口W j上的测量数据序列为
Figure PCTCN2017119421-appb-000012
传感器节点X tj在tp时刻采集的数据为
Figure PCTCN2017119421-appb-000013
该数据包括h个属性测量值，则
Figure PCTCN2017119421-appb-000014
4、检测率，是指算法检测到的异常数据样本数与实际的异常数据样本总数之比。
5、误报率，是指被算法误判为异常的正常数据样本数与总的正常数据样本数之比。
本发明以传感器节点的时空相关性为基础，使用K-means算法对空间节点进行分簇，将数据相似的传感器节点划分到同一个簇中，然后提出基于MEA-BP神经网络WSN异常检测方法，该方法主要分为参数优化、异常数据的检测和数据异常的判定三个步骤，其主要特点有：(1)使用思维进化算法对BP神经网络进行参数优化，通过趋同异化等操作对BP神经网络的权值、阈值等进行优化；(2)在异常数据的检测阶段主要对传感器节点采集到的数据流中可能存在的异常数据点进行识别，该步骤采用分布式的算法，在各传感器节点独立执行，然后各传感器节点将结果传递到簇头节点进一步地验证；(3)异常数据检测时提出每个当前时刻通过传感器节点滑动窗口的历史数据集训练神经网络，来完成下一时刻的预报，通过神经网络的模型残差，确定用于判断下一时刻数据异常的置信区间，当下一时刻数据落入置信区间内，则该数据诶判定为正常，反之，该数据在簇头节点中通过空间相关性进一步地验证，具体的该方法包括以下步骤：
将各分部传感器节点初始化，各传感器节点开始采集数据，设传感器节点个数为n，各传感器节点为X tj(j＝1,2,…,n)，传感器节点X tj的滑动窗口为W j，各传感器节点的滑动窗口大小均为m，则传感器节点X tj在其滑动窗口W j上的测量数据序列为
Figure PCTCN2017119421-appb-000015
传感器节点X tj在tp时刻采集的数据为
Figure PCTCN2017119421-appb-000016
该数据包括h个属性测量值，则
Figure PCTCN2017119421-appb-000017
利用K-means算法对各传感器节点进行空间分簇得到若干组簇，将数据相似的传感器节点划分到同一个簇中，提高簇内节点空间相似度，该工作在节点故障检测前完成，完成后再对节点采集数据进行时间相关性检测和空间相关性检测，在空间相关性检测时利用前面空间划分的结果，传感器节点对当前时刻采集的数据先进行时间相关性检测，当时间相关性检测显示有问题时，则该传感器节点将可疑数据告知簇头节点，簇头节点对该数据进行空间相关 性检测；参照图2，基本思想是：首先从传感器节点对象中任意挑选K个对象作为聚类中心，对于剩下的传感器节点，则根据它们与被挑选出来的聚类中心的相似度(欧氏距离)，分别将传感器节点分配给与其最相似的聚类中心的聚类，然后再计算每个新聚类的聚类中心(该聚类中所有对象的均值)，不断重复这一过程直到聚类中心开始收敛为止；
K-means算法步骤如下：
设传感器节点数目为p，传感器节点的坐标为{x (1),x (2),…,x (p)}，每个x (i)(i＝1,2,…,p)∈R；
步骤1，随机选取K个传感器节点作为聚类中心，K个聚类中心的坐标为u (j)(j＝1,2,…,k)∈R，K个聚类中心对应K个聚类；
步骤2，重复下面的过程直到聚类中心收敛
{
对于每个传感器节点样例i，计算其应该属于的聚类，公式如下：
c (i)：＝argmin j||x (i)-u (j)|| 2公式(2)
对于每个聚类j，重新计算该聚类的聚类中心，公式如下：
Figure PCTCN2017119421-appb-000018
公式(2)中u (j)代表K个聚类中心中的一个，通过不断调整参数j(j＝1,2,…,k)，使得每一个点的开销函数c (i)达到最小值，c (i)代表传感器节点样例i与K个聚类中心中距离最近的聚类，把每一个点划分到距离其最近的一个聚类中心的聚类；
公式(3)中分母表示每个聚类中样本的总数，分子是每个聚类中传感器节点样例i对应的坐标和。
最终得到若干组簇，设q+1个传感器节点组成一组分簇，每组分簇中包括1个簇头节点X tc和q个分布节点(X t1,X t2,…,X tq)；
利用思维进化算法对BP神经网络进行参数优化，通过趋同异化操作对BP神经网络的权值和阈值进行优化，得到最优权值和阈值，输入最优权值和阈值，建立MEA-BP神经网络模型，优化BP神经网络的流程如图3所示，大致步骤如下：
步骤1：思维进化初始群体产生，在解空间随机产生N组数作为初始群体，每组数中包含n个元素代表一个个体(即神经网络结构)，每一个个体的矩阵为1*n，群体矩阵为N*n，群体包含N个个体；
步骤2：根据BP神经网络拓扑结构，将解空间映射到编码空间，每个编码空间对应问题的一个解，即一个个体，编码长度等于每个个体中的元素数目，编码长度n为
n＝tL+wL+L+w公式(4)
公式(4)中，t为神经网络输入节点数，w为输出节点数，L为隐含层节点数，此处选取t＝19，L＝20，w＝1；
步骤3：定义迭代次数iter、优胜子群体M和临时子群体T的数目；进化过程的每一代中的所有个体的集合成为一个群体，一个群体分为M个优胜子群体和T个临时子群体，每个优胜子群体和临时子群体包含了SG个个体，SG为：
SG＝N/(M+T)公式(5)
通常选取iter＝10，M＝T＝5，N＝200，SG＝20；
步骤4：得分函数的确定，神经网络由输入层、隐含层、输出层组成，网络训练样本输入层为A K＝(a 1 k,a 2 k,…,a t k)，(K＝1,2,…,P，P为训练样本数)，输入矩阵为t*p，期望网络输出层为Y K＝(Y 1 k,Y 2 k,…,Y w k)，输出矩阵为w*p，每一个样本输入对应一个输出，样本中一共有p个输入输出对；网络中间隐含层各节点输入为Z K＝(z 1,z 2,…,z L)，中间隐含层各节点输出为B K＝(b 1,b 2,…,b L)；网络各节点输入为Q K＝(q 1,q 2,…,q w)，网络输出层各节点输出为G K＝(g 1,g 2,…,g w)。定义输入层与隐含层权值W ij(i＝1,2,…,t,j＝1,2,…,L)、隐含层与输出层权值V ju(j＝1,2,…,L,u＝1,2,…,w)、隐含层阈值为{θ j(j＝1,2,…,L)}、输出层阈值{β u(u＝1,2,…,w)}，矩阵为w*w。根据编码规则，W ij为单个个体中第1个到第(L*t)个元素，矩阵为L*t；V ju为单个个体中第(L*t+1)个到(L*t+w*L)个元素，矩阵为w*L；θ j为单个个体中第(L*t+w*L+1)个到第(L*t+w*L+L)个元素，矩阵为w*L；β u为单个个体中第(L*t+w*L+1)个到最后一个元素。计算网络隐含层各节点输入Z j，然后用{Z j}通过S型激活函数来计算隐含层各节点输出{b j}，S型函数表达式为：
Figure PCTCN2017119421-appb-000019
Figure PCTCN2017119421-appb-000020
b j＝f(Z j)，j＝1,2,…L公式(7)
公式(7)中∑Wa称为此模型的激活值，是模型的输入求和；公式(7)中f()为模型激活函数。
然后根据隐含层的输出{b j}，权值V ju及阈值{β u}计算输出层各节点输入Q u，然后用{Q u}通过S型函数来计算输出层各节点的输出{G u}：
Figure PCTCN2017119421-appb-000021
G u＝f(Q u)(u＝1,2,…,w)公式(9)
公式(8)(9)同上；
选择训练样本的均方误差的倒数作为各个个体与群体的得分函数
Figure PCTCN2017119421-appb-000022
y k表示第K个训练样本的期望输出值，矩阵为w*p，G k表示实际的输出值，即通过神经网络训练出来的值，矩阵也为w*p，p为训练样本数目；
步骤5：训练权值和阈值，对于每个个体在(-1,1)之间以均匀分布产生n组随机数，个体矩阵为1*n，作为初始的权值阈值群体，根据网络计算规则，按照得分函数计算每个个体得分，得分最高的个体被称为优胜者，冲中选取q个得分最好的个体作为优胜者，选取q＝10；
步骤6：子群体趋同操作，在子群体中，个体成为胜者而竞争的过程叫做趋同，一个子群体的趋同过程中，若不在产生新的胜者，则成为子群体成熟，趋同过程结束，分别以每一个优胜者为中心，服从正态分布产生个体，形成M个优胜子群体和T个临时子群体，每个子群体包含SG个个体，该正态分布可以表示为N(u，∑)，式中u是正态分布的中心向量，∑是该正态分布的协方差矩阵，正态分布的中心就是胜者的坐标，即胜者的权值和阈值；
步骤7：子群体异化操作。异化过程是整个解空间内各子群体成为胜者而竞争的过程。通过全局公告板，它记录了各子群体得分函数值以及成熟度，在各个子群体之间进行全局竞争，若一个临时子群体得分高于某个成熟优胜子群体得分，则临时子群体替换掉优胜子群体，原优胜子群体中的个体被放弃；若一个成熟的临时子群体的得分低于任意一个优胜子群体的得分，则该临时子群体被放弃，其中的个体被释放。被放弃的临时子群体的个数记为Tr，被放弃的优胜子群体的个数记为Mr，在全局公告板的指导下，在解空间中重新产生Mr+Tr个临时子群体；
步骤8：解析最优个体。重复上述6，7步骤，当满足迭代停止条件时，思维进化算法结束优化过程。此时，根据编码规则，对寻找到的最优个体进行解析，从而得到对应的BP神经网络的最优权值和阈值，输入最优权值和阈值，建立MEA-BP神经网络模型；
参照图4，采用分布式的算法，对每组分簇中传感器节点(X t1,X t2,…,X tq)独立执行异常检测，异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc进一步验证，包括以下步骤：
S41通过时间相关性对传感器节点(X t1,X t2,…,X tq)独立执行异常检测，利用每个当前时刻通过传感器节点Xtj滑动窗口Wj的数据
Figure PCTCN2017119421-appb-000023
来训练神经网络，完成下一时刻数据的预报，选取训练数据
Figure PCTCN2017119421-appb-000024
的后v个样本数据，通过公式(1)计算MEA-BP神经网络的模型残差S：
Figure PCTCN2017119421-appb-000025
其中，Er(r＝1,2,…,v)为选取的样本数据值，F为选取的样本数据平均值，
Figure PCTCN2017119421-appb-000026
S42计算传感器节点当前时刻的置信区间，置信区间为
Figure PCTCN2017119421-appb-000027
其中Spre为MEA-BP神经网络对下一时刻数据的预测值，t α/2,v-1为自由度v-1的t分布，在t分布表中选取合适的α值，得到t值；t分布表如图1所示，正态分布又名高斯分布，若随机变量服从一个数学期望为μ、方差为σ2的高斯分布，记为N(μ，σ2)，我们通常所说的标准正态分布是μ＝0，σ＝1的正态分布。从平均值为μ、方差为σ2的标准正态分布总体中抽取容量为v的样本，样本服从平均值为μ，方差为σ2/v的正态分布，总体方差σ2总是未知的，从而只能用s2来代替。如果v很大，那么，s2就是σ2的一个较好的估计量，仍然是一个近似的标准正态分布；如果v较小，s2与σ2的差异较大，因此，此时样本分布就不再是一个标准正态分布，而是服从t分布，t分布是依自由度而变的一条曲线，在坐标轴Y轴两侧对称分布，且均值为0，t分布适用于当总体标准差未知时，然后用样本标准差代替总体标准差。t分布中的自由度是指任何变量中可以自由变化的数目。此时的样本t检验仅估计一个参数：总体均值，样本大小v组成了v种用于估计总体均值及其变异性的信息，消耗一个自由度来估计均值，其余v-1个自由度用于估计变异性，因此，样本t检验使用自由度为v-1的t分布；当不需要估计参数时，自由度为v；通常选取α＝0.05。
S43当下一时刻数据Snew进入传感器节点滑动窗口内时，判断下一时刻数据Snew是否落入当前时刻的置信区间范围内，若是，则判断该数据Snew为正常数据；否则判断该数据Snew为异常数据；
S44异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc中；
S45簇头节点X tc通过空间相关性检测并引入投票机制验证传感器节点异常数据产生原因，原因包括事件异常、节点故障异常和误判，具体包括以下步骤：
S451将传感器节点的异常数据和与该传感器节点在同一分簇中的其他节点数据进行比较；设q+1个传感器节点组成一组分簇，每组分簇中包括1个簇头节点X tc和q个分布节点(X t1,X t2,…,X tq)；
S452预先设定误差值θ，设传感器节点的异常数据为ST，设与该传感器节点在同一分簇中的其他节点的数据为Si(i＝1,2,…,q-1)，若|ST-Si|<θ，则初始值为0的计数NN加1，统计最 终的NN值；
S453若
Figure PCTCN2017119421-appb-000028
则判断该检测节点异常数据是由于事件异常；若
Figure PCTCN2017119421-appb-000029
则判断该检测节点异常数据是由于节点故障或误判；若
Figure PCTCN2017119421-appb-000030
则选取该分簇中的参照节点，设参照节点的数据为S CC，若|S T-S CC|≤θ，则判断该传感器节点异常数据是由于事件异常，若|S T-S CC|>θ，则判断该传感器节点异常数据是由于节点故障或误判；其中参照节点为距离该分簇中聚类中心节点欧氏距离最近的节点；参照节点：利用K-means算法对各传感器节点进行空间分簇得到若干组簇，将数据相似的传感器节点划分到同一个簇中，根据最终的分簇结果，选取距离该簇质心欧式距离最近的节点为参照节点；
S454对于所述步骤S453中，当判断得到该传感器节点异常数据是由于节点故障或误判时，进一步判断该传感器节点异常数据是由于节点故障还是误判，具体包括以下步骤：通过对传感器节点进行时间相关性检测，当传感器节点连续时间内均产生异常数据，则判断该传感器节点异常数据是由于节点故障；当传感器节点只有该时刻为异常数据，其他时刻产生数据均为正常数据，则判断该传感器节点异常数据是由于误判；
为了检测节点中的异常数据，每个节点都会以一定周期采集数据，形成属于该节点的数据流，为了保证该区域内节点采集数据样本的正确性，每个节点在上传数据前，都需要利用神经网络预测的值替换掉滑动窗口内的异常值。
数据样本来源于英特尔伯克利实验室的传感器网络数据，该数据采样频率为每隔31秒采样一次。选取传感器节点1-5号节点100组温度，湿度作为训练数据，100组温度，湿度作为预测数据。选取传感器节点6号节点2000组温度数据作为参数优化对比的训练数据。
S1训练＝[20.6156,20.6254,20.6450,20.6352,20.6450,20.6156,20.6058,20.576420.4882,20.4588,20.4392,20.4196,20.3804,20.3510,20.3020,20.2726,20.2530,20.1942,20.184420.1354,20.0864,20.0668,20.0374,20.0178,19.9982,19.9786,19.9688,19.8414,19.8022,19.8022,19.7826,19.8022,19.8218,19.8316,19.8610,19.9002,19.9296,19.9786,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1256,20.1354,20.1550,20.2236,20.2432,20.2432,20.3020,20.3216,20.3608,20.4588,20.4980,20.5176,20.5568,20.5666,20.5960,20.6254,21.4192,21.4094,21.3996,21.3604,21.3212,21.3016,21.2624,21.2330,21.2232,21.2036,21.1252,21.1252,21.1056,21.1056,21.0860,21.0860,21.0860,21.0762,21.0664,21.0664,21.0468,21.0076,20.9880,20.9586,20.9684,20.9096,20.8998,20.8998,20.8802,20.8704,20.8704,20.8704,20.8704,20.8802,20.8704,20.8802,20.8802,20.8802,20.8704,20.8704]
[37.5737,37.6079,37.6422,37.6422,37.7107,37.7107,37.7792,37.8477,38.0529,38.1213,38.189 738.1897,38.3263,38.3946,38.4629,38.5311,38.5311,38.7357,38.8039,38.8720,39.0082,39.0763,39.1443,39.1783,39.2123,39.2803,39.3143,39.6200,39.755739.7896,39.755739.5521,39.5521,39.5521,39.4162,39.4502,39.3143,39.2803,39.2123,39.178339.1443,39.110339.0082,39.0082,38.9401,39.0082,38.9401,38.8379,38.8720,38.8039,38.7357,38.769838.7357,38.6675,38.4970,38.3946,38.3946,38.3263,38.2580,38.2580,38.2239,38.155537.9845,38.1897,38.1555,38.0529.37.9845,37.9161,37.8134,37.813437.8819,37.8819,37.9161,37.9503,37.9503,37.9161,37.9503,37.9161,37.8477,37.7792,37.7450,37.7450,37.7107,37.7107,37.7107,37.7107,37.6765,37.6765,37.6422,37.8134,37.7792,37.7107,37.7107,37.6765,37.6765,37.7107,37.7450,37.7792,37.9161,38.0529]
S1预测＝[20.7724,20.7332,20.7528,20.7430,20.7332,20.7332,20.7234,20.7136,20.7234,22.1530,20.7038,20.6940,20.6940,20.7038,20.6940,…,22.1530,20.6744,20.6450,20.6352,20.6254][39.6200,39.9929,40.3652,40.9055,41.1414,41.2761,41.3771,41.4780,41.7805,41.8812,41.9818,41.9483,42.2500,42.3840,42.3170,…,52.3170,42.2500,42.2835,42.0824,41.9818]
S2训练＝[20.6226,20.6224,20.6350,20.6352,20.6350,20.6256,20.6358,20.5764,20.4782,20.4388,20.4492,20.4296,20.3604,20.3310,20.3320,20.2326,20.2330,20.1042,20.1644,20.1354,20.0464,20.0468,20.0274,20.0178,19.9482,19.9486,19.9588,19.8614,19.8322,19.8222,19.7926,19.8122,19.8518,19.8516,19.8510,19.9502,19.9296,19.9386,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1236,20.1454,20.1350,20.2236,20.2332,20.2432,20.3020,20.3216,20.3608,20.4588,20.4380,20.5376,20.5568,20.5666,20.5960,20.6354,20.6450,20.7136,20.7038,20.6744,20.7528,20.8412,20.8506,20.8600,20.8796,20.8802,20.8402,20.8704,20.8902,20.8704,20.8704,20.8704,20.8802,20.9096,20.9586,20.9782,20.9380,21.0376,21.0076,21.0174,21.0076,21.0272,21.0468,21.0566,20.9978,21.0370,21.0468,21.0366,21.0360,21.0368,21.0368,21.0634,21.0732,21.0938,21.0360,21.1356]
[37.5337,37.6179,37.6222,37.6322,37.7207,37.7307,37.7492,37.8277,38.0129,38.1113,38.149738.1597,38.3163,38.3546,38.4629,38.5311,38.5311,38.7357,38.8039,38.8720,39.0082,39.0763,39.1443,39.1783,39.2123,39.2803,39.3153,39.6200,39.7557,39.7896,39.7357,39.5521,39.5521,39.5521,39.4152,39.4502,39.3143,39.2803,39.2123,39.1783,39.1443,39.1103,39.0032,39.0082,38.9431,39.0042,38.8401,38.8479,38.8520,38.8039,38.7357,38.7698,38.7357,38.6635,38.4970,38.3946,38.3946,38.3263,38.2380,38.2380,38.2339,38.1355,37.9345,38.1897,38.1555,38.0529,37.9845,37.9161,37.8134,37.8134,37.8819,37.8819,37.9161,37.9503,37.9503,37.9161,37.9503,37.9161,37.8477,37.7792,37.7450,37.7450,37.7107,37.7107,37.7107,37.7107,37.6765,37.6765,37.6522,37.8034,37.759 2,37.6907,37.6807,37.6565,37.6865,37.7207,37.7550,37.7732,37.9261,38.0629]
S2预测＝(20.5670,20.5376,20.4788,20.4494,20.4984,20.4788,20.4592,20.4494,20.4788,20.4690,20.4788,20.4592,20.4592,20.4494,20.4396,…,20.4396,22.4396,20.4396,20.4396,20.3710)(39.4300,40.7863,40.4722,40.8955,41.3214,41.4361,41.5671,41.6880,41.8205,41.9036,42.0142,42.0100,42.3200,42.4140,42.3270,…,42.3370,52.3500,42.3825,42.1824,41.8918)
S3训练＝[20.6156,20.6254,20.6450,20.6352,20.6450,20.6156,20.6058,20.576420.4882,20.4588,20.4392,20.4196,20.3804,20.3510,20.3020,20.2726,20.2530,20.1942,20.1844,20.1354,20.0864,20.0668,20.0374,20.0178,19.9982,19.9786,19.9688,19.8414,19.8022,19.8022,19.7826,19.8022,19.8218,19.8316,19.8610,19.9002,19.9296,19.9786,19.9884,20.0080,20.0374,20.0472,20.1060,20.1060,20.1256,20.1354,20.1550,20.2236,20.2432,20.2432,20.3020,20.3216,20.3608,20.4588,20.4980,20.5176,20.5568,20.5666,20.5960,20.6254,20.6450,20.7136,20.7038,20.6744,20.7528,20.8312,20.8606,20.8900,20.9096,20.8802,20.8802,20.8704,20.8802,20.8704,20.8704,20.8704,20.8802,20.9096,20.9586,20.9782,20.9880,21.0076,21.0076,21.0174,21.0076,21.0272,21.0468,21.0566,20.9978,21.0370,21.0468,21.0566,21.0860,21.0468,21.0468,21.0664,21.0762,21.0958,21.0860,21.1056]
[37.5737,37.6079,37.6422,37.6422,37.7107,37.7107,37.7792,37.8477,38.0529,38.1213,38.189738.1897,38.3263,38.3946,38.4629,38.5311,38.5311,38.7357,38.8039,38.8720,39.0082,39.0763,39.1443,39.1783,39.2123,39.2803,39.3143,39.6200,39.7557,39.7896,39.7557,39.5521,39.5521,39.5521,39.4162,39.4502,39.3143,39.2803,39.2123,39.1783,39.1443,39.1103,39.0082,39.0082,38.9401,39.0082,38.9401,38.8379,38.8720,38.8039,38.7357,38.7698,38.7357,38.6675,38.4970,38.3946,38.3946,38.3263,38.2580,38.2580,38.2239,38.1555,37.9845,38.1897,38.1555,38.0529,37.9845,37.9161,37.8134,37.8134,37.8819,37.8819,37.9161,37.9503,37.9503,37.9161,37.9503,37.9161,37.8477,37.7792,37.7450,37.7450,37.7107,37.7107,37.7107,37.7107,37.6765,37.6765,37.6422,37.8134,37.7792,37.7107,37.7107,37.6765,37.6765,37.7107,37.7450,37.7792,37.9161,38.0529]
S3预测＝(20.4396,20.4102,20.4102,20.4004,20.3710,20.3612,20.3612,20.3612,20.3612,20.3514,20.3710,20.3808,20.3416,20.3220,20.3220,…,20.3318,20.3318,22.3220,20.3514,20.3416)(39.4200,40.7963,40.4422,40.8755,41.3614,41.4161,41.5871,41.6780,41.8405,41.9236,42.0242,42.0200,42.3300,42.4240,42.3170,…,42.3270,42.3400,52.3525,42.1624,41.8718)
S4训练＝[20.4154,20.4254,20.4450,20.4352,20.4450,20.4154,20.4057,20.574420.4772,20.4577,20.4392,20.4194,20.3704,20.3510,20.3020,20.2724,20.2530,20.1942,20.1744,20.1354,20.0744,20.0447,20.0374,20.0277,19.9972,19.9774,19.9477,19.7414,19.7022,19.7022,19.7724,19.7022,19.7217,19.7314,19.7410,19.9002,19.9294,19.9774,19.9774,20.0070,20.0374,20.0472,20.1040,20.1040,20.1254,20.1354,20.1350,20.2234,20.2432,20.2432,20.3020,20.3214,20.3407,20.4577,20.4970,20.5174,20.5547,20.5444,20.5940,20.4254,20.4450,20.7134,20.7037,20.4744,20.7527,20.7312,20.7404,20.7900,20.9094,20.7702,20.7702,20.7704,20.7702,20.7704,20.7704,20.7704,20.7702,20.9094,20.9574,20.9772,20.9770,21.0074,21.0074,21.0174,21.0074,21.0272,21.0447,21.0544,20.9977,21.0370,21.0447,21.0544,21.0740,21.0447,21.0447,21.0444,21.0742,21.0957,21.0840,21.1054]
[37.5737,37.4079,37.4422,37.4422,37.7107,37.7107,37.7792,37.8477,38.0529,38.1213,38.189738.1897,38.3243,38.3944,38.4429,38.5311,38.5311,38.7357,38.8039,38.8720,39.0082,39.0743,39.1443,39.1783,39.2123,39.2803,39.3143,39.4200,39.755739.7894,39.755739.5521,39.5521,39.5521,39.4142,39.4502,39.3143,39.2803,39.2123,39.178339.1443,39.110339.0082,39.0082,38.9401,39.0082,38.9401,38.8379,38.8720,38.8039,38.7357,38.749838.7357,38.4475,38.4970,38.3944,38.3944,38.3243,38.2580,38.2580,38.2239,38.155537.9845,38.1897,38.1555,38.0529.37.9845,37.9141,37.8134,37.813437.8819,37.8819,37.9141,37.9503,37.9503,37.9141,37.9503,37.9141,37.8477,37.7792,37.7450,37.7450,37.7107,37.7107,37.7107,37.7107,37.4745,37.4745,37.4422,37.8134,37.7792,37.7107,37.7107,37.4745,37.4745,37.7107,37.7450,37.7792,37.9141,38.0529]
S4预测＝(20.4394,20.4102,20.4788,20.4494,20.4984,20.4788,20.3412,20.3412,20.3412,20.3514,20.3710,20.3808,20.3414,20.2240,20.2240,…,20.2338,20.2240,20.3318,22.3514,20.3414)(39.4500,40.7343,40.4222,40.8555,41.4214,41.5341,41.4471,41.4980,41.8805,41.9334,42.0142,42.0120,42.3100,42.4240,42.3070,…,42.3470,42.3800,42.3925,52.2024,41.9118)
S5训练＝[20.4154,20.4254,20.4450,20.4352,20.4450,20.4154,20.4058,20.574420.4882,20.4588,20.4392,20.4194,20.3804,20.3510,20.3020,20.2724,20.2530,20.1942,20.184420.1354,20.0844,20.0448,20.0374,20.0178,19.9982,19.9784,19.9488,19.8414,19.8022,19.8022,19.7824,19.8022,19.8218,19.8314,19.8410,19.9002,19.9294,19.9784,19.9884,20.0080,20.0374,20.0472,20.1040,20.1040,20.1254,20.1354,20.1550,20.2234,20.2432,20.2432,20.3020,20.3214,20.3408,20.4588,20.4980,20.5174,20.5548,20.5444,20.5940,20.4254,20.4450,20.7134,20.7038,20.4744,20.7528,20.8312,20.8404,20.8900,20. 9094,20.8802,20.8802,20.8704,20.8802,20.8704,20.8704,20.8704,20.8802,20.9094,20.9584,20.9782,20.9880,21.0074,21.0074,21.0174,21.0074,21.0272,21.0448,21.0544,20.9978,21.0370,21.0448,21.0544,21.0840,21.0448,21.0448,21.0444,21.0742,21.0958,21.0840,21.1054]
[37.5737,37.4079,37.4422,37.4422,37.7107,37.7107,37.7792,37.8477,38.0529,38.1213,38.189738.1897,38.3243,38.3944,38.4429,38.5311,38.5311,38.7357,38.8039,38.8720,39.0082,39.0743,39.1443,39.1783,39.2123,39.2803,39.3143,39.4200,39.755739.7894,39.755739.5521,39.5521,39.5521,39.4142,39.4502,39.3143,39.2803,39.2123,39.178339.1443,39.110339.0082,39.0082,38.9401,39.0082,38.9401,38.8379,38.8720,38.8039,38.7357,38.749838.7357,38.4475,38.4970,38.3944,38.3944,38.3243,38.2580,38.2580,38.2239,38.155537.9845,38.1897,38.1555,38.0529.37.9845,37.9141,37.8134,37.813437.8819,37.8819,37.9141,37.9503,37.9503,37.9141,37.9503,37.9141,37.8477,37.7792,37.7450,37.7450,37.7107,37.7107,37.7107,37.7107,37.4745,37.4745,37.4422,37.8134,37.7792,37.7107,37.7107,37.4745,37.4745,37.7107,37.7450,37.7792,37.9141,38.0529]
S5预测＝(20.4204,205002,20.4200,20.4200,20.3906,20.3612,20.3024,20.2828,20.3122,20.2828,20.2828,20.2632,20.2436,20.2240,20.2240,…,20.2338,20.2240,20.1848,20.1848,22.1848)(39.4050,40.7663,40.4532,40.8865,41.3314,41.4261,41.5871,41.7080,41.7905,41.9236,42.0242,42.0180,42.3150,42.4040,42.2970,…,42.3270,42.3700,42.3535,42.1924,51.9218)
S6训练＝(24.7218，24.6434，24.6140，24.5454，24.5356，24.5356，24.5258，24.4964，24.4964，24.4768，24.4964，24.4376，24.36924.36924.36924.3298，…，23.5164，23.5164，23.4968，23.4968，23.4968，23.4870，23.4870，23.4674，23.4576，23.4380，23.438)
现以传感器节点采集的温度为样例。如图1所示，首先通过K-means算法对空间节点进行划分，假设1-5号节点在同一个簇中，然后如图2所示利用节点的数据进行参数优化，建立MEA-BP神经网络模型。利用神经网络模型残差确定置信区间，此后节点采集新的数据，通过滑动窗口进行预测，若采集的数据不在置信区间内，则判定该数据为异常值，同时传递给簇头进行验证。否则，采集的数据是正常的。
神经网络的初始权值和阈值是随机的，通过优化算法得到最优个体，根据编码规则解析最优个体，个体中的元素被划分为四部分，输入层与隐含层权值W、隐含层与输出层权值V、隐含层阈值θ、输出层阈值β。
W＝[0.5773,-0.0366,-0.5442,0.6235,1.1156,0.3768,0.3825,0.3136,0.1267,-0.7622,-0.5179,1.0760,-0.76 51,-0.6149,0.3441,0.9941,…,0.0842,0.9022,-0.2135,0.0192；
1.0865,0.2526,0.6696,-0.5513,0.2104,0.2332,0.5453,-0.1723,1.3676,0.3422,-0.3744,0.8603,0.0330,0.3656,-0.2953,0.0479,…,-0.3960,-0.6000,0.2272,-0.7731；
-0.5749,0.5627,0.5373,1.0035,-0.9583,1.1374,0.9935,0.5132,0.5842,0.4948,-0.5388,0.9441,0.3067,-0.5114,0.3832,-0.0877,…,-0.6106,-0.8577,-0.4521,0.5040；
-0.1253,-0.9239,0.5082,0.5222,0.0151,0.0033,-0.2452,-0.3727,-0.7296,1.0824,1.0505,-0.6283,-0.5399,0.4255,0.3935,-0.5921,…,-0.6411,-0.8547,0.5181,0.7837；
-0.4296,-0.6532,0.1644,0.4362,1.2707,0.4877,-0.6715,0.9923,1.2893,0.8293,,0.9933,-0.6541,-0.1126,0.5300,-0.7672,0.0751,…,0.2299,-0.2186,0.8131,-0.9794；
-0.9365,-0.2176,1.0960,0.0105,1.3308,-0.3746,-1.0590,-0.5488,-0.8550,0.2305,0.4455,-0.2741,-0.0727,-0.8329,0.6602,0.6955,…,1.2929,-0.2030,0.6805,-0.1515；
0.9026,0.4729,0.4579,0.0247,-0.3627,-0.1774,0.3376,0.2647,-0.2081,-0.9794,0.8114,0.0375,-0.1698,0.1191,-0.1169,0.3339,…,0.4083,-0.3918,-0.3013,0.8716；
-0.1600,0.5712,-0.4983,-0.5316,-0.4528,-0.1343,0.6454,-0.6407,0.9763,-0.3584,-0.5402,0.1329,0.7943,0.3628,-0.3161,0.5328,…,0.5403,0.7219,-0.3242,0.4581；
-0.0458,-0.7775,0.1541,-0.0951,-0.5324,0.1833,-0.6400,-0.1086,-0.8386,-0.1227,-1.1306,0.3747,-1.2653,-0.4936,-0.1682,0.7071,…,-0.3657,1.0828,0.5700,1.1866；
-0.5898,-0.7801,0.5664,0.0111,0.3069,0.1700,0.3791,-0.3305,0.3633,0.3725,0.0423,-0.0695,0.2999,0.9656,0.6382,0.1977,…,-0.4587,0.2812,-0.5573,-0.3800；
0.9481,-0.0495,0.6783,-0.5981,-0.4970,0.8797,-0.9796,-0.3040,0.9687,-1.1954,0.3528,-0.1131,-0.1752,-0.8480,0.0283,-0.1791,…,1.1680,0.1180,1.0543,-0.7890；
0.9876,-0.5693,0.2064,1.0028,-0.7505,-0.0915,0.8364,0.0396,1.3324,0.6910,0.4675,0.2987,-0.7527,-0.5169,-0.8266,1.0537,…,0.9445,-0.1559,-0.2290,0.7849；
0.7757,0.0837,0.6734,-1.0255,0.2378,0.4449,0.5047,0.0004,0.4182,0.1178,0.5781,-0.2788,0.5544,-0.6854,0.2246,-0.9034,…,-0.5780,-1.0345,-1.0602,-0.5880；
-0.9348,-0.4230,-0.7827,-0.9307,-0.1429,-0.1718,0.2480,0.8521,-0.9502,0.8372,0.5072,-0.5041,0.0417,-0.0333,0.1855,-0.1131,…,0.7978,1.2911,0.3152,0.5734；
-0.4313,0.3686,-0.4494,-0.2113,-0.8895,-0.6692,0.8727,0.5028,-1.2660,-0.1608,-1.1177,0.9249,-0.7744,-0.8300,-0.6689,-0.6472,…,0.5649,-0.0527,-0.2023,0.7030；
0.8426,0.8520,0.6722,-0.9511-0.7137,-0.6196,0.1978,0.5044,-0.6855,-0.3957,-1.1713,0.8359, 0.2987,-0.3332,-0.7046,-0.0307,…,0.7013,-1.0424,-0.2173,-0.8635；…
0.4429,-0.5530,0.2594,-0.0055,-0.2293,-0.1312,0.1046,-1.1803,0.6552,1.1167,0.2293,0.8949,0.2872,0.0639,-0.7069,-0.6443,…,-0.5359,-0.7943,0.0749,0.2322；
0.1580,0.3637,-0.3663,0.3326,0.5618,-0.0908,0.6683,-0.1572,0.0053,-0.6843,-1.1245,0.6485-0.4312,-0.7010,-0.1113,0.7301,…,-0.0746,0.0622,0.1535,-0.0938；
-1.1087,-0.0046,-0.9751,-0.0265,1.1469,-0.7158,0.6888,-0.5601,0.5212,0.6409,-0.5567,1.1806,-0.1444,0.5397,-0.5531,0.5221,…,-0.5053,-1.2739,-0.4797,0.3794]
V＝[-0.8852,-0.4987,-1.0526,0.3578,-0.0436,-0.6403,1.1844,-1.2437,-0.9961,-0.1289,-0.8779,-0.0577,-0.5220,0.4834,0.3050,0.5123,…,0.4276,-0.0404,0.9207,0.0199]
θ＝[-0.3454,-0.0724,-0.9978,-1.0797,1.0350,0.2819,0.8036,0.2719,0.2332,-0.6689,0.9885,-0.7912,0.1800,-0.6851,-0.7595,0.2428,…,-0.1237,-0.2353,-0.0445,0.4698]
β＝[-1.2285]
然后计算置信区间，随机选取节点1任意一个时刻的预测值，值为20.8692，此时选取训练数据的后40个数据通过公式1计算模型残差，求出的模型残差值S＝0.1776，由图1得t值为2.02108，此时的区间上下界通过公式2计算为[20.5058,21.2326]。
假设节点S1已经训练好了MEA-BP模型，通过滑动窗口添加新的数据，如上所示，当滑动窗口预测第1个数据时，此时温度预测值为20.8692，置信区间范围为[20.5058,21.2326]，此时刻传感器数据为20.7724，可以看出20.7724这个数据明显在区间范围之内，则认为该时刻数据是正常的；当滑动窗口预测第10个数据时，此时温度预测值为20.7062，置信区间范围为[20.3428,21.0696],而此时刻传感器数据为22.1530，该数据在区间范围之外，然后传递给簇头节点进行空间性验证，通过投票机制发现该时刻的数据在同一个簇内温度是异常的，则节点1采集的温度数据异常，节点出现故障或误报，若发现在同一簇中数据相似，则认为出现事件。对于其他节点的湿度温度数据也同样能够准确识别出异常。
对于二分类问题，可将样本根据其真实类别和决策模型检测类别的组合划分成真正例(true positive,TP)、假正例(false positive,FP)、真反例(true negative,TN)、假反例(false negative,FN)四种情形。为了评价和比较两种无线传感器网络异常数据检测方法的性能，本发明使用检测率(True Positive Rate)和误报率(False Positive Rate)两个性能指标。检测率是指算法检测到的异常数据样本数与实际的异常数据样本总数之比；误报率是指被算法误判为异常的正常数据 样本数与总的正常数据样本数之比。
Figure PCTCN2017119421-appb-000031
Figure PCTCN2017119421-appb-000032
为验证该优化算法的效率，在S6训练数据中选取4组不同规模的数据，为500、1000、1500、2000个数据。分别使用BP算法和MEA-BP算法在不同规模数据集上做训练，并用训练时间作为衡量指标，实验结果如图5所示，横坐标表示数据集序号，纵坐标表示训练时间。由图5可知，对各个数据集，BP算法平均训练时间为31.6s，MEA-BP算法在训练时间方面优势较明显，平均训练时间为24.7s，相对BP算法训练时间减少21.8％，在训练数据样本增多时，优势更为明显。这是由于MEA算法使用更短的速度达到了误差要求缩短了训练时间。
在S1-S55个数据集上验证MEA-BP算法性能，如表1所示，在不同的数据集下，MEA-BP算法检测率比BP算法平均提高了2.2％；MEA-BP算法误报率比BP算法平均降低了0.9％。
表1结果对比
Figure PCTCN2017119421-appb-000033
本发明方法以传感器节点的时空相关性为基础，用K-means算法对节点进行划分，然后在每个节点上建立MEA-BP神经网络模型，根据该模型来检测未来每一时刻各分布节点内到达的数据是否异常，然后把检测后的结果传递到簇头节点进行验证。目前，异常检测在各个领域中都是一个深入研究的问题，无线传感器网络独特的特点及严格的约束条件使得该问题的研究更具有挑战性。本发明提出的方法主要是针对资源受限的无线传感器网络复杂的异常检测，大大降低了节点之间的通信消耗，并能准确检测出异常数据，更具有环境适应能力。
以上所述仅是本发明的优选实施方式，应当指出：对于本技术领域的技术人员来说，在不脱离本发明原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本发明的保护范围。

Claims (9)

  1. 一种基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述方法包括以下步骤：
    S1将各分布传感器节点初始化，各传感器节点开始采集数据；
    设传感器节点个数为n，各传感器节点为X tj(j＝1,2,…,n)，传感器节点X tj的滑动窗口为W j，各传感器节点的滑动窗口大小均为m，则传感器节点X tj在其滑动窗口W j上的测量数据序列为
    Figure PCTCN2017119421-appb-100001
    传感器节点X tj在tp时刻采集的数据为
    Figure PCTCN2017119421-appb-100002
    该数据包括h个属性测量值，则
    Figure PCTCN2017119421-appb-100003
    S2利用K-means算法对各传感器节点进行空间分簇得到若干组簇；
    设q+1个传感器节点组成一组簇，每组簇中包括1个簇头节点X tc和q个分布节点(X t1,X t2,…,X tq)；
    S3利用思维进化算法对BP神经网络进行参数优化，通过趋同异化操作对BP神经网络的权值和阈值进行优化，得到最优权值和阈值，输入最优权值和阈值，建立MEA-BP神经网络模型；
    S4采用分布式的算法，对每组分簇中传感器节点(X t1,X t2,…,X tq)独立执行异常检测，异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc进一步验证。
  2. 根据权利要求1所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S2包括以下步骤：
    S21首先从分布传感器节点对象中任意选择K个传感器节点对象作为K个聚类中心；
    S22然后针对除聚类中心以外的传感器节点对象，分别计算传感器节点对象与K个聚类中心之间的相似度，得到与传感器节点对象相似度最接近的聚类中心；
    S23将传感器节点对象分配给与该传感器节点对象相似度最接近的聚类中心的聚类，将所有传感器节点分配完成后得到K个聚类；
    S24重新计算该K个聚类的聚类中心，得到新的聚类中心；
    S25重新计算各传感器节点与新的聚类中心的相似度，回到步骤S22；
    S26当重新计算的聚类中心收敛时，结束本操作。
  3. 根据权利要求1所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S3包括以下步骤：产生训练数据；确定BP神经网络拓扑结构；通过思维进化算法进行参数设置；随机产生初始种群、优胜子种群和临时子种群；对子种群进行趋同操作；对子 种群进行异化操作；判断是否满足结束条件，如果满足，则输出最优个体，获取最优权值和阈值，否则重新进行趋同异化操作。
  4. 根据权利要求1所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S4包括以下步骤：
    S41通过时间相关性对传感器节点(X t1,X t2,…,X tq)独立执行异常检测，利用每个当前时刻通过传感器节点X tj滑动窗口W j的数据
    Figure PCTCN2017119421-appb-100004
    来训练神经网络，完成下一时刻数据的预报，选取训练数据
    Figure PCTCN2017119421-appb-100005
    的后v个样本数据，通过公式(1)计算MEA-BP神经网络的模型残差S：
    Figure PCTCN2017119421-appb-100006
    其中，Er(r＝1,2,…,v)为选取的样本数据值，F为选取的样本数据平均值，
    Figure PCTCN2017119421-appb-100007
    S42计算传感器节点当前时刻的置信区间，置信区间为
    Figure PCTCN2017119421-appb-100008
    其中S pre为MEA-BP神经网络对下一时刻数据的预测值，tα/2,v-1为自由度v-1的t分布，在t分布表中选取合适的α值，得到t值；
    S43当下一时刻数据S new进入传感器节点滑动窗口内时，判断下一时刻数据S new是否落入当前时刻的置信区间范围内，若是，则判断该数据S new为正常数据；否则判断该数据S new为异常数据；
    S44异常检测完毕后传感器节点(X t1,X t2,…,X tq)将检测结果传递到该组分簇的簇头节点X tc中；
    S45簇头节点X tc通过空间相关性检测并引入投票机制验证传感器节点异常数据产生原因，原因包括事件异常、节点故障异常和误判。
  5. 根据权利要求4所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S42中，选取α＝0.05。
  6. 根据权利要求4所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S45中进行空间相关性检测并引入投票机制包括以下步骤：
    S451将传感器节点的异常数据和与该传感器节点在同一分簇中的其他节点数据进行比较；设q+1个传感器节点组成一组分簇，每组分簇中包括1个簇头节点X tc和q个分布节点(X t1,X t2,…,X tq)；
    S452预先设定误差值θ，设传感器节点的异常数据为S T，设与该传感器节点在同一分簇中的其他节点的数据为S i(i＝1,2,…,q-1)，若|S T-S i|<θ，则初始值为0的计数NN加1，统计最终的NN值；
    S453若
    Figure PCTCN2017119421-appb-100009
    则判断该检测节点异常数据是由于事件异常；若
    Figure PCTCN2017119421-appb-100010
    则判断该检测节点异常数据是由于节点故障或误判；若
    Figure PCTCN2017119421-appb-100011
    则选取该分簇中的参照节点，设参照节点的数据为S CC，若|S T-S CC|≤θ，则判断该传感器节点异常数据是由于事件异常，若|S T-S CC|>θ，则判断该传感器节点异常数据是由于节点故障或误判；其中参照节点为距离该分簇中聚类中心节点欧氏距离最近的节点；
    S454对于所述步骤S453中，当判断得到该传感器节点异常数据是由于节点故障或误判时，进一步判断该传感器节点异常数据是由于节点故障还是误判，具体包括以下步骤：通过对传感器节点进行时间相关性检测，当传感器节点连续时间内均产生异常数据，则判断该传感器节点异常数据是由于节点故障；当传感器节点只有该时刻为异常数据，其他时刻产生数据均为正常数据，则判断该传感器节点异常数据是由于误判。
  7. 根据权利要求6所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：参照节点为距离该分簇中聚类中心节点欧氏距离最近的节点具体是指：利用K-means算法对各传感器节点进行空间分簇得到若干组簇，将数据相似的传感器节点划分到同一个簇中，根据最终的分簇结果，选取距离该簇质心欧式距离最近的节点为参照节点。
  8. 根据权利要求6所述的基于MEA-BP神经网络WSN异常检测方法，其特征在于：所述步骤S454中为了检测节点中的异常数据，每个节点都会以一定周期采集数据，形成属于该节点的数据流，为了保证该区域内节点采集数据样本的正确性，每个节点在上传数据前，都需要利用神经网络预测的值替换掉滑动窗口内的异常值。
  9. 如权利要求1-8任一所述的基于MEA-BP神经网络WSN异常检测方法在WSN检测中的应用。
PCT/CN2017/119421 2017-01-06 2017-12-28 一种基于mea-bp神经网络wsn异常检测方法 WO2018126984A2 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710008709.XA CN106714220B (zh) 2017-01-06 2017-01-06 一种基于mea-bp神经网络wsn异常检测方法
CN201710008709.X 2017-01-06

Publications (2)

Publication Number Publication Date
WO2018126984A2 true WO2018126984A2 (zh) 2018-07-12
WO2018126984A3 WO2018126984A3 (zh) 2018-09-13

Family

ID=58907069

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/119421 WO2018126984A2 (zh) 2017-01-06 2017-12-28 一种基于mea-bp神经网络wsn异常检测方法

Country Status (2)

Country Link
CN (1) CN106714220B (zh)
WO (1) WO2018126984A2 (zh)

Cited By (48)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109856299A (zh) * 2018-11-26 2019-06-07 国家电网有限公司 一种变压器在线监测差异化阈值动态设置方法、系统
CN109963317A (zh) * 2019-05-14 2019-07-02 中国联合网络通信集团有限公司 一种簇头选举方法、装置
CN110084326A (zh) * 2019-05-13 2019-08-02 东北大学 一种基于模糊集的工业设备异常检测方法
CN110147829A (zh) * 2019-04-29 2019-08-20 郑州工程技术学院 一种基于云计算的飞行器数据处理方法和装置
CN110362608A (zh) * 2019-06-11 2019-10-22 广东工业大学 基于雨流计数法和局部异常因子的能耗异常检测方法
CN110427593A (zh) * 2018-12-19 2019-11-08 西安电子科技大学 基于工业大数据的smt印刷参数优化方法
CN110457550A (zh) * 2019-07-05 2019-11-15 中国地质大学（武汉） 一种烧结过程中异常运行数据的校正方法
CN110750641A (zh) * 2019-09-24 2020-02-04 武汉大学 一种基于序列连接模型和二叉树模型的分类纠错方法
CN110849404A (zh) * 2019-11-18 2020-02-28 中国华能集团清洁能源技术研究院有限公司 一种传感器数据异常的连续判别方法
CN110912272A (zh) * 2019-12-03 2020-03-24 合肥工业大学 基于区域性异常模式识别的城市电网故障检测方法和系统
CN110969198A (zh) * 2019-11-24 2020-04-07 广东浪潮大数据研究有限公司 深度学习模型的分布式训练方法、装置、设备及存储介质
CN111127184A (zh) * 2019-11-01 2020-05-08 复旦大学 一种分布式组合信用评估方法
CN111126437A (zh) * 2019-11-22 2020-05-08 中国人民解放军战略支援部队信息工程大学 基于加权动态网络表示学习的异常群体检测方法
CN111654874A (zh) * 2020-06-03 2020-09-11 枣庄学院 一种无线传感网异常检测方法
CN111654831A (zh) * 2020-04-14 2020-09-11 南京信息工程大学 一种基于无线传感网的磨机负荷检测方法
CN111814826A (zh) * 2020-06-08 2020-10-23 武汉理工大学 退役动力电池余能快速检测评级方法
CN111899040A (zh) * 2019-05-05 2020-11-06 腾讯科技（深圳）有限公司 目标对象异常传播的检测方法、装置、设备及存储介质
CN112001638A (zh) * 2020-08-25 2020-11-27 瑞洲建设集团有限公司 一种基于物联网技术的工地管理系统
CN112165485A (zh) * 2020-09-25 2021-01-01 山东炎黄工业设计有限公司 一种大规模网络安全态势智能预测方法
CN112329351A (zh) * 2020-11-19 2021-02-05 上海嗨酷强供应链信息技术有限公司 基于数据追踪的流量分析系统及方法
CN112437085A (zh) * 2020-11-23 2021-03-02 中国联合网络通信集团有限公司 一种网络攻击的识别方法及装置
CN112437440A (zh) * 2020-09-30 2021-03-02 北京工业大学 无线传感器网络中基于相关性理论的恶意共谋攻击抵抗方法
CN112506990A (zh) * 2020-12-03 2021-03-16 河海大学 一种基于时空信息的水文数据异常检测方法
CN112565183A (zh) * 2020-10-29 2021-03-26 中国船舶重工集团公司第七0九研究所 一种基于流式动态时间规整算法的网络流量异常检测方法及装置
CN112702408A (zh) * 2020-12-20 2021-04-23 国网山东省电力公司临沂供电公司 基于多感知功能的物联网系统及方法
CN112770282A (zh) * 2020-12-23 2021-05-07 龙海建设集团有限公司 基于智能建筑物联网的数据处理系统
CN112783938A (zh) * 2020-12-30 2021-05-11 河海大学 一种水文遥测实时数据异常检测方法
CN112804255A (zh) * 2021-02-09 2021-05-14 中国人民解放军国防科技大学 一种基于节点多维特征的网络异常节点检测方法
CN112820120A (zh) * 2020-12-30 2021-05-18 杭州趣链科技有限公司 一种基于联盟链的多方交通流时空交叉验证方法
CN112861436A (zh) * 2021-02-18 2021-05-28 天津大学 一种发动机排放实时预测方法
CN112882445A (zh) * 2020-06-05 2021-06-01 洋浦美诺安电子科技有限责任公司 用于智能制造的5g及区块链智能监控和管理系统
CN113378990A (zh) * 2021-07-07 2021-09-10 西安电子科技大学 基于深度学习的流量数据异常检测方法
CN113556770A (zh) * 2021-07-27 2021-10-26 广东电网有限责任公司 数据校验方法、装置、终端及可读存储介质
CN114021297A (zh) * 2021-11-18 2022-02-08 吉林建筑科技学院 基于回声状态网络的复杂管网泄漏定位方法
CN114051218A (zh) * 2021-11-09 2022-02-15 华中师范大学 一种环境感知网络优化方法和系统
CN114401516A (zh) * 2022-01-11 2022-04-26 国家计算机网络与信息安全管理中心 一种基于虚拟网络流量分析的5g切片网络异常检测方法
CN114422554A (zh) * 2022-01-27 2022-04-29 山东云海据信息科技有限公司 基于分布式物联网的服务区智能设备管理方法及装置
CN114861776A (zh) * 2022-04-21 2022-08-05 武汉大学 一种基于人工免疫技术的动态自适应网络异常检测方法
CN114997276A (zh) * 2022-05-07 2022-09-02 北京航空航天大学 一种面向压制成型设备的异构多源时序数据异常识别方法
CN115002824A (zh) * 2022-05-25 2022-09-02 厦门大学 基于lstm的水声网络数据实时故障检测及恢复方法
CN115022049A (zh) * 2022-06-06 2022-09-06 哈尔滨工业大学 一种基于计算马氏距离的分布外网络流量数据检测方法、电子设备及存储介质
CN115608793A (zh) * 2022-12-20 2023-01-17 太原科技大学 一种机理融合数据的精轧温度调控方法
CN116109176A (zh) * 2022-12-21 2023-05-12 成都安讯智服科技有限公司 一种基于协同聚类的报警异常预测方法和系统
CN116257892A (zh) * 2023-05-09 2023-06-13 广东电网有限责任公司佛山供电局 一种数字化档案的数据隐私安全性验证方法
CN116405368A (zh) * 2023-06-02 2023-07-07 南京信息工程大学 一种高维不平衡数据条件下的网络故障诊断方法、系统
CN117093947A (zh) * 2023-10-20 2023-11-21 深圳特力自动化工程有限公司 一种发电柴油机运行异常监测方法及系统
CN117349779A (zh) * 2023-12-04 2024-01-05 水利部交通运输部国家能源局南京水利科学研究院 深挖方膨胀土渠道边坡潜在滑动面判定方法及系统
CN117892095A (zh) * 2024-03-14 2024-04-16 山东泰开电力电子有限公司 一种储能系统用的散热系统故障智能检测方法

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106714220B (zh) * 2017-01-06 2019-05-17 江南大学 一种基于mea-bp神经网络wsn异常检测方法
CN107358021B (zh) * 2017-06-01 2020-07-28 华南理工大学 一种基于优化bp神经网络的do预测模型建立方法
CN107249000B (zh) * 2017-07-06 2020-02-25 河南科技大学 一种移动用户异常行为检测方法
CN107272660B (zh) * 2017-07-26 2019-05-17 江南大学 一种带丢包的网络化控制系统的随机故障检测方法
CN107613540B (zh) * 2017-11-07 2019-08-30 合肥工业大学 一种无线可充电传感器网络聚类分簇路由方法
CN108763346B (zh) * 2018-05-15 2022-02-01 中南大学 一种滑窗箱型图中值滤波的异常点处理方法
CN109714311B (zh) * 2018-11-15 2021-12-31 北京天地和兴科技有限公司 一种基于聚类算法的异常行为检测的方法
CN110542659B (zh) * 2019-09-06 2020-04-07 四川大学 基于可见光光谱的珍珠光泽检测方法
CN111542010A (zh) * 2020-04-22 2020-08-14 青岛黄海学院 基于分类自适应估计加权融合算法的wsn数据融合方法
CN111950505B (zh) * 2020-08-24 2023-08-29 湖南科技大学 一种ssa-aann的风力发电机传感器状态评估方法
CN112418281A (zh) * 2020-11-11 2021-02-26 国网福建省电力有限公司电力科学研究院 一种火灾探测传感器数据异常检测方法及系统
CN113128626A (zh) * 2021-05-28 2021-07-16 安徽师范大学 基于一维卷积神经网络模型的多媒体流细分类方法
CN113640308B (zh) * 2021-08-31 2024-03-29 夏冰心 一种基于机器视觉的轨道异常监测系统
CN114484732B (zh) * 2022-01-14 2023-06-02 南京信息工程大学 一种基于投票网络的空调机组传感器故障诊断方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103747537B (zh) * 2014-01-15 2017-05-03 广东交通职业技术学院 一种基于熵度量的无线传感器网络离群数据自适应检测方法
CN103916896B (zh) * 2014-03-26 2017-05-24 浙江农林大学 基于多维Epanechnikov核密度估计的异常检测方法
CN105791051B (zh) * 2016-03-25 2019-06-21 中国地质大学（武汉） 基于人工免疫和k均值聚类的无线传感网异常检测方法及系统
CN105764162B (zh) * 2016-05-10 2019-05-17 江苏大学 一种基于多属性关联的无线传感器网络异常事件检测方法
CN106447092A (zh) * 2016-09-12 2017-02-22 浙江工业大学 一种基于mea‑bp神经网络的船用反渗透海水淡化系统性能预测方法
CN106714220B (zh) * 2017-01-06 2019-05-17 江南大学 一种基于mea-bp神经网络wsn异常检测方法

Cited By (81)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109856299A (zh) * 2018-11-26 2019-06-07 国家电网有限公司 一种变压器在线监测差异化阈值动态设置方法、系统
CN110427593A (zh) * 2018-12-19 2019-11-08 西安电子科技大学 基于工业大数据的smt印刷参数优化方法
CN110427593B (zh) * 2018-12-19 2022-12-02 西安电子科技大学 基于工业大数据的smt印刷参数优化方法
CN110147829B (zh) * 2019-04-29 2022-10-11 郑州工程技术学院 一种基于云计算的飞行器数据处理方法和装置
CN110147829A (zh) * 2019-04-29 2019-08-20 郑州工程技术学院 一种基于云计算的飞行器数据处理方法和装置
CN111899040B (zh) * 2019-05-05 2023-09-01 腾讯科技（深圳）有限公司 目标对象异常传播的检测方法、装置、设备及存储介质
CN111899040A (zh) * 2019-05-05 2020-11-06 腾讯科技（深圳）有限公司 目标对象异常传播的检测方法、装置、设备及存储介质
CN110084326A (zh) * 2019-05-13 2019-08-02 东北大学 一种基于模糊集的工业设备异常检测方法
CN110084326B (zh) * 2019-05-13 2022-12-06 东北大学 一种基于模糊集的工业设备异常检测方法
CN109963317A (zh) * 2019-05-14 2019-07-02 中国联合网络通信集团有限公司 一种簇头选举方法、装置
CN110362608A (zh) * 2019-06-11 2019-10-22 广东工业大学 基于雨流计数法和局部异常因子的能耗异常检测方法
CN110362608B (zh) * 2019-06-11 2023-04-28 广东工业大学 基于雨流计数法和局部异常因子的能耗异常检测方法
CN110457550A (zh) * 2019-07-05 2019-11-15 中国地质大学（武汉） 一种烧结过程中异常运行数据的校正方法
CN110750641A (zh) * 2019-09-24 2020-02-04 武汉大学 一种基于序列连接模型和二叉树模型的分类纠错方法
CN111127184A (zh) * 2019-11-01 2020-05-08 复旦大学 一种分布式组合信用评估方法
CN110849404A (zh) * 2019-11-18 2020-02-28 中国华能集团清洁能源技术研究院有限公司 一种传感器数据异常的连续判别方法
CN111126437A (zh) * 2019-11-22 2020-05-08 中国人民解放军战略支援部队信息工程大学 基于加权动态网络表示学习的异常群体检测方法
CN111126437B (zh) * 2019-11-22 2023-05-02 中国人民解放军战略支援部队信息工程大学 基于加权动态网络表示学习的异常群体检测方法
CN110969198A (zh) * 2019-11-24 2020-04-07 广东浪潮大数据研究有限公司 深度学习模型的分布式训练方法、装置、设备及存储介质
CN110912272B (zh) * 2019-12-03 2023-02-21 合肥工业大学 基于区域性异常模式识别的城市电网故障检测方法和系统
CN110912272A (zh) * 2019-12-03 2020-03-24 合肥工业大学 基于区域性异常模式识别的城市电网故障检测方法和系统
CN111654831A (zh) * 2020-04-14 2020-09-11 南京信息工程大学 一种基于无线传感网的磨机负荷检测方法
CN111654831B (zh) * 2020-04-14 2023-01-31 南京信息工程大学 一种基于无线传感网的磨机负荷检测方法
CN111654874B (zh) * 2020-06-03 2023-02-24 枣庄学院 一种无线传感网异常检测方法
WO2021243848A1 (zh) * 2020-06-03 2021-12-09 枣庄学院 一种无线传感网异常检测方法
CN111654874A (zh) * 2020-06-03 2020-09-11 枣庄学院 一种无线传感网异常检测方法
CN112882445A (zh) * 2020-06-05 2021-06-01 洋浦美诺安电子科技有限责任公司 用于智能制造的5g及区块链智能监控和管理系统
CN112882445B (zh) * 2020-06-05 2021-12-21 广域铭岛数字科技有限公司 用于智能制造的5g及区块链智能监控和管理系统
CN111814826B (zh) * 2020-06-08 2022-06-03 武汉理工大学 退役动力电池余能快速检测评级方法
CN111814826A (zh) * 2020-06-08 2020-10-23 武汉理工大学 退役动力电池余能快速检测评级方法
CN112001638B (zh) * 2020-08-25 2024-01-23 瑞洲建设集团有限公司 一种基于物联网技术的工地管理系统
CN112001638A (zh) * 2020-08-25 2020-11-27 瑞洲建设集团有限公司 一种基于物联网技术的工地管理系统
CN112165485A (zh) * 2020-09-25 2021-01-01 山东炎黄工业设计有限公司 一种大规模网络安全态势智能预测方法
CN112437440B (zh) * 2020-09-30 2024-02-02 北京工业大学 无线传感器网络中基于相关性理论的恶意共谋攻击抵抗方法
CN112437440A (zh) * 2020-09-30 2021-03-02 北京工业大学 无线传感器网络中基于相关性理论的恶意共谋攻击抵抗方法
CN112565183B (zh) * 2020-10-29 2022-12-09 中国船舶重工集团公司第七0九研究所 一种基于流式动态时间规整算法的网络流量异常检测方法及装置
CN112565183A (zh) * 2020-10-29 2021-03-26 中国船舶重工集团公司第七0九研究所 一种基于流式动态时间规整算法的网络流量异常检测方法及装置
CN112329351A (zh) * 2020-11-19 2021-02-05 上海嗨酷强供应链信息技术有限公司 基于数据追踪的流量分析系统及方法
CN112437085B (zh) * 2020-11-23 2023-03-24 中国联合网络通信集团有限公司 一种网络攻击的识别方法及装置
CN112437085A (zh) * 2020-11-23 2021-03-02 中国联合网络通信集团有限公司 一种网络攻击的识别方法及装置
CN112506990B (zh) * 2020-12-03 2022-10-04 河海大学 一种基于时空信息的水文数据异常检测方法
CN112506990A (zh) * 2020-12-03 2021-03-16 河海大学 一种基于时空信息的水文数据异常检测方法
CN112702408A (zh) * 2020-12-20 2021-04-23 国网山东省电力公司临沂供电公司 基于多感知功能的物联网系统及方法
CN112770282A (zh) * 2020-12-23 2021-05-07 龙海建设集团有限公司 基于智能建筑物联网的数据处理系统
CN112820120B (zh) * 2020-12-30 2022-03-01 杭州趣链科技有限公司 一种基于联盟链的多方交通流时空交叉验证方法
CN112820120A (zh) * 2020-12-30 2021-05-18 杭州趣链科技有限公司 一种基于联盟链的多方交通流时空交叉验证方法
CN112783938A (zh) * 2020-12-30 2021-05-11 河海大学 一种水文遥测实时数据异常检测方法
CN112783938B (zh) * 2020-12-30 2022-10-04 河海大学 一种水文遥测实时数据异常检测方法
CN112804255A (zh) * 2021-02-09 2021-05-14 中国人民解放军国防科技大学 一种基于节点多维特征的网络异常节点检测方法
CN112861436A (zh) * 2021-02-18 2021-05-28 天津大学 一种发动机排放实时预测方法
CN113378990A (zh) * 2021-07-07 2021-09-10 西安电子科技大学 基于深度学习的流量数据异常检测方法
CN113378990B (zh) * 2021-07-07 2023-05-05 西安电子科技大学 基于深度学习的流量数据异常检测方法
CN113556770A (zh) * 2021-07-27 2021-10-26 广东电网有限责任公司 数据校验方法、装置、终端及可读存储介质
CN114051218B (zh) * 2021-11-09 2024-05-14 华中师范大学 一种环境感知网络优化方法和系统
CN114051218A (zh) * 2021-11-09 2022-02-15 华中师范大学 一种环境感知网络优化方法和系统
CN114021297B (zh) * 2021-11-18 2023-12-01 吉林建筑科技学院 基于回声状态网络的复杂管网泄漏定位方法
CN114021297A (zh) * 2021-11-18 2022-02-08 吉林建筑科技学院 基于回声状态网络的复杂管网泄漏定位方法
CN114401516B (zh) * 2022-01-11 2024-05-10 国家计算机网络与信息安全管理中心 一种基于虚拟网络流量分析的5g切片网络异常检测方法
CN114401516A (zh) * 2022-01-11 2022-04-26 国家计算机网络与信息安全管理中心 一种基于虚拟网络流量分析的5g切片网络异常检测方法
CN114422554A (zh) * 2022-01-27 2022-04-29 山东云海据信息科技有限公司 基于分布式物联网的服务区智能设备管理方法及装置
CN114422554B (zh) * 2022-01-27 2024-03-01 曹颂群 基于分布式物联网的服务区智能设备管理方法及装置
CN114861776B (zh) * 2022-04-21 2024-04-09 武汉大学 一种基于人工免疫技术的动态自适应网络异常检测方法
CN114861776A (zh) * 2022-04-21 2022-08-05 武汉大学 一种基于人工免疫技术的动态自适应网络异常检测方法
CN114997276A (zh) * 2022-05-07 2022-09-02 北京航空航天大学 一种面向压制成型设备的异构多源时序数据异常识别方法
CN114997276B (zh) * 2022-05-07 2024-05-28 北京航空航天大学 一种面向压制成型设备的异构多源时序数据异常识别方法
CN115002824A (zh) * 2022-05-25 2022-09-02 厦门大学 基于lstm的水声网络数据实时故障检测及恢复方法
CN115022049A (zh) * 2022-06-06 2022-09-06 哈尔滨工业大学 一种基于计算马氏距离的分布外网络流量数据检测方法、电子设备及存储介质
CN115022049B (zh) * 2022-06-06 2024-05-14 哈尔滨工业大学 一种基于计算马氏距离的分布外网络流量数据检测方法、电子设备及存储介质
CN115608793A (zh) * 2022-12-20 2023-01-17 太原科技大学 一种机理融合数据的精轧温度调控方法
CN116109176B (zh) * 2022-12-21 2024-01-05 成都安讯智服科技有限公司 一种基于协同聚类的报警异常预测方法和系统
CN116109176A (zh) * 2022-12-21 2023-05-12 成都安讯智服科技有限公司 一种基于协同聚类的报警异常预测方法和系统
CN116257892A (zh) * 2023-05-09 2023-06-13 广东电网有限责任公司佛山供电局 一种数字化档案的数据隐私安全性验证方法
CN116257892B (zh) * 2023-05-09 2023-08-29 广东电网有限责任公司佛山供电局 一种数字化档案的数据隐私安全性验证方法
CN116405368B (zh) * 2023-06-02 2023-08-22 南京信息工程大学 一种高维不平衡数据条件下的网络故障诊断方法、系统
CN116405368A (zh) * 2023-06-02 2023-07-07 南京信息工程大学 一种高维不平衡数据条件下的网络故障诊断方法、系统
CN117093947A (zh) * 2023-10-20 2023-11-21 深圳特力自动化工程有限公司 一种发电柴油机运行异常监测方法及系统
CN117093947B (zh) * 2023-10-20 2024-02-02 深圳特力自动化工程有限公司 一种发电柴油机运行异常监测方法及系统
CN117349779B (zh) * 2023-12-04 2024-02-09 水利部交通运输部国家能源局南京水利科学研究院 深挖方膨胀土渠道边坡潜在滑动面判定方法及系统
CN117349779A (zh) * 2023-12-04 2024-01-05 水利部交通运输部国家能源局南京水利科学研究院 深挖方膨胀土渠道边坡潜在滑动面判定方法及系统
CN117892095A (zh) * 2024-03-14 2024-04-16 山东泰开电力电子有限公司 一种储能系统用的散热系统故障智能检测方法
CN117892095B (zh) * 2024-03-14 2024-05-28 山东泰开电力电子有限公司 一种储能系统用的散热系统故障智能检测方法

Also Published As

Publication number Publication date
CN106714220A (zh) 2017-05-24
WO2018126984A3 (zh) 2018-09-13
CN106714220B (zh) 2019-05-17

Similar Documents

Publication Publication Date Title
WO2018126984A2 (zh) 一种基于mea-bp神经网络wsn异常检测方法
WO2022047658A1 (zh) 日志异常检测系统
Dong et al. An Intrusion Detection Model for Wireless Sensor Network Based on Information Gain Ratio and Bagging Algorithm.
Han et al. Occupancy and indoor environment quality sensing for smart buildings
CN101516099B (zh) 一种传感器网络异常检测方法
Guo et al. Evolutionary dual-ensemble class imbalance learning for human activity recognition
CN109145516B (zh) 一种基于改进型极限学习机的模拟电路故障识别方法
CN113378990B (zh) 基于深度学习的流量数据异常检测方法
CN108289285B (zh) 一种海洋无线传感器网络丢失数据恢复与重构方法
Song et al. Coupled behavior analysis for capturing coupling relationships in group-based market manipulations
CN112949828A (zh) 一种基于图学习的图卷积神经网络交通预测方法及系统
Guo et al. Feature selection based on Rough set and modified genetic algorithm for intrusion detection
CN108415884B (zh) 一种结构模态参数实时追踪方法
CN110708318A (zh) 基于改进的径向基神经网络算法的网络异常流量预测方法
Mulia et al. A review on building occupancy estimation methods
CN108133090A (zh) 一种可靠性敏感度驱动的高端复杂装备可靠性分析方法
Chen et al. Testing the structure of a Gaussian graphical model with reduced transmissions in a distributed setting
CN113884807A (zh) 基于随机森林和多层架构聚类的配电网故障预测方法
CN115618743B (zh) 一种瞄准镜系统的状态评估方法及状态评估系统
CN116522260A (zh) 基于极限学习机算法的网络流量在线分类的学习模型
CN113988189B (zh) 一种跨风电机组的迁移故障诊断方法
Zhang et al. An improved composite hypothesis test for Markov models with applications in network anomaly detection
Purnawansyah et al. K-Means clustering implementation in network traffic activities
Xu et al. Federated traffic synthesizing and classification using generative adversarial networks
CN108957435B (zh) 基于遗传算法的航迹匹配方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17890329

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase in:

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17890329

Country of ref document: EP

Kind code of ref document: A2