Summary of the invention
The object of the present invention is to provide in a kind of radio communication network cell based on direct-push confidence machine and hypothesis testing
Disconnected detection method.Using the method in the present invention, can be obtained between the accuracy and complexity of judgement by adjusting threshold value
Coordinate.
The present invention passes through that direct-push confidence machine provides data to be tested point and normal cell or abnormal cell first
With degree, hypothesis testing is then carried out according to matching degree information, defining a kind of judgement of data point according to matching degree is normal cell
With the discriminative information amount of abnormal cell, sequential detection is then carried out according to discriminative information amount, carries out conclusive judgement.
1 direct-push confidence machine
Direct-push confidence machine is wide using a kind of use scope of the algorithmic theory of randomness theory foundation of Kolmogorov
Machine learning confidence level mechanism.It is used to measure the credibility that a sample is belonging respectively to existing several classifications.It is logical
Confidence level is estimated frequently with a kind of random detection function for meeting Kolmogorov randomness theory, it is this with machine examination
The value for surveying function is referred to as P value, and sample generally defined as to be sorted belongs to the probability of already present sample space, relative to
The value of certain class sample space is bigger, then shows that a possibility that it belongs to the sample space is bigger.
Direct-push confidence machine is substantially a kind of sorting algorithm, and in processing classification problem, it attempts for sample to be classified as
There is certain one kind in classification.It passes through the method that distance calculates, and is classified according to classified data set to observation sampling point.For
Calculating P value, first defines singular value:
Di yIndicate certain data point x to sort from small to largeiWith classification data concentrate all data points of a certain class y away from
From.Similar, Di -yIndicate the data point x to sort from small to largeiWith in classified data set all data points of-y class away from
From.- y indicates all classes that y class is removed in classified data set.K indicates the number for the neighbour's data point taken, and the value of k can
It is voluntarily determined by operator according to network condition.αi yAs certain data point xiThe singular value of y class is concentrated with classification data.
After given singular value, a kind of definition method of P value is as follows:
# indicates the data point number for meeting condition in the gesture of set, that is, set.αtest yIndicate data point to be tested with
The singular value of y class, α in classified data setj y(j=1,2 ..., n) indicate y class all data points singular value, n is y
The number of class data point.Py(αtest y) be data point in y class singular value be greater than data point to be tested relative to this class singular value
Ratio.It is not difficult to find out Py(αtest y) value interval of value is [0,1], and its value is bigger, shows sample data point to be tested
A possibility that belonging to class y is bigger.Sample to be sorted is divided into maximum P by direct-push confidence machiney(αtest y) class corresponding to value.
In calculating process, concentrates the singular value of the sample of a certain class to be less than when data to be tested and training data and be used for
It when calculating k minimum range of this sample singular value, needs to recalculate singular value for such all sample, to be to be sorted
Sample recalculates P value.For every one kind in training data, data to be tested have a corresponding P value to calculate.
2 cell outage detection methods
Cell outage detection process is divided into two parts, i.e., monitoring and test are established and interrupted to interruption detection model.In model
Establish the following steps are included:
1) different states, i.e. normal condition or interrupt status are in by manually setting cell, to acquire inhomogeneity
The training data vector of type (modeling class) is used for model foundation.Training data vector can select signal reception power (RSRP),
Signal to Interference plus Noise Ratio (SINR), cutting off rate etc., specific vector can voluntarily be determined by operator according to network operation situation.
2) trained number is calculated with the method for direct-push confidence machine to the training data vector of the different modeling types of acquisition
According to the singular value relative to each modeling class and preservation.
It interrupts in monitoring and detection process, by monitoring network state, acquires data vector to be tested, carry out interruption inspection
It surveys.Detection process is as follows:
1) P value of the data point to be tested relative to each modeling class is calculated according to direct-push confidence machine first, saved.
2) hypothesis testing is carried out according to P value.
(1) assume to establish: assuming that data to be tested vector is Xn={ x1,x2,...,xi,...,xn}.Modeling class number is
M.It interrupts detection process and follows following hypothesis:
H0:Xn~P0
H1:Xn~P1
(3)
…
HM:Xn~PM
P0~PMFor the P state of value space of different modeling classes.H0~HMRespectively assume that data point to be tested belongs to accordingly
Modeling class.
(2) determine that a decision criteria differentiates which modeling class is data point to be tested belong to.
Direct-push confidence machine gives P value of each test data point relative to training class, that is, a kind of test data
The measurement of point and the matching degree of training class.
The discriminative information amount for the y class that single test data point provides is defined as:
D is the number of class in-y.PyIt is worth bigger, indicates that attribution data to be tested is bigger in the y class a possibility that, so y class
Discriminative information amount Iy(xi) bigger.P-yIt is bigger, indicate that a possibility that attribution data is with other classes are bigger, so the judgement of y class
Information content Iy(xi) accordingly reduce.
Decision rule are as follows: Λ* y={ Xn:Iy(Xn) > Tn(λ) }, i.e., if certain a kind of accumulative sentence information content exceeds certain
One thresholding, judgement test data vector belong to the modeling class.Thresholding TnThe value of (λ) can by operator according to network condition voluntarily
It determines.
(3) adjudicate: the method for taking sequential detection carries out conclusive judgement.The discriminative information amount for initializing each class first is
Zero.For each test data, the discriminative information amount for all classes that data point provides is calculated according to formula (4), then to each
The information content of illegally occupying of class adds up, until the discriminative information amount of some class reaches a decision threshold Tn(λ), judgement test number
According to such is belonged to, stop detection.Thresholding Tn(λ) is bigger, detects that accurate provisioning request is higher, and corresponding detection complexity is higher.
The invention has the following advantages: the matching of data point and cell that the present invention is provided according to direct-push confidence machine
It spends information and carries out hypothesis testing.The discriminative information amount that a kind of judgement of data point is normal cell and abnormal cell is defined, then
Sequential detection judgement is carried out according to discriminative information amount.This method adjudicates accuracy height, and false alarm rate is low.
It is had the following advantages based on cell outage judgement of the invention:
1. training vector is selected by operator's self assemble, there is high flexibility.
2. if then the definition of information content only needs operator's selection is modeled by the parameter of user or base station measurement
Local information can be realized by distributed algorithm.
3. operator can voluntarily be adjusted between detection complexity and accuracy in detection by adjusting decision threshold
Section.
4. using Sequential Detection, there is high detection efficiency.