Application of Neural Network in Fault Location of Optical Transport Network

Tianyang Liu,Haoyuan Mei,＊,Qiang Sun,＊,Huachun Zhou

1 School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China

Abstract: Due to the increasing variety of information and services carried by optical networks,the survivability of network becomes an important problem in current research.The fault location of OTN is of great significance for studying the survivability of optical networks.Firstly,a three-channel network model is established and analyzing common alarm data,the fault monitoring points and common fault points are carried out.The artificial neural network is introduced into the fault location field of OTN and it is used to judge whether the possible fault point exists or not.But one of the obvious limitations of general neural networks is that they receive a fixedsize vector as input and produce a fixed-size vector as the output.Not only that,these models is even fixed for mapping operations (for example,the number of layers in the model).The difference between the recurrent neural network and general neural networks is that it can operate on the sequence.In spite of the fact that the gradient disappears and the gradient explodes still exist in the neural network,the method of gradient shearing or weight regularization is adopted to solve this problem,and choose the LSTM (long-short term memory networks)to locate the fault.The output uses the concept of membership degree of fuzzy theory to express the possible fault point with the probability from 0 to 1.Priority is given to the treatment of fault points with high probability.The concept of F-Measure is also introduced,and the positioning effect is measured by using location time,MSE and F-Measure.The experiment shows that both LSTM and BP neural network can locate the fault of optical transport network well,but the overall effect of LSTM is better.The localization time of LSTM is shorter than that of BP neural network,and the F1-score of LSTM can reach 0.961566888396156 after 45 iterations,which meets the accuracy and real-time requirements of fault location.Therefore,it has good application prospect and practical value to introduce neural network into the fault location field of optical transport network.

Keywords: optical transport networks; failure localization; artificial neural network; longshort term memory network; BP neural network; F1-Measure

I.INTRODUCTION

With the continuous development of optical network technology and the increase of carrying information and service types,the access mode and networking mode of service are also constantly improving.As the next generation backbone transport network,Optical transport network (OTN)is carrying more and more information.At the same time,there are a variety of service signals,including SDH,Ethernet IP and so on,which can be used in OTN to transmit efficiently,quickly,reliably and transparently.At present,the network rate interface of OTN can reach 400G.For such a complex,ultra-high-speed,super-capacity optical transport network,any network fault will result in the degradation of QoS of the oversized optical transport network,and even a large amount of information can be lost.Therefore,the survivability of optical networks has become an important issue in the research of optical networks and the fault location of OTN is of great significance to the study of the survivability of optical network.

In recent years,with the rapid growth of demand for large-scale data processing and in-depth analysis in various industries,data mining has become a hot and difficult point in the field of computer research.Data mining is a cross research field in computer science.Its research methods are closely related to many subjects,such as statistics,machine learning,expert system,information retrieval,social network,natural language processing,pattern recognition and so on[1].Although data processing and data mining have made great achievements in many fields,the current application of OTN is far less extensive than other fields,and its application is not mature enough.As a large capacity backbone transport network,OTN network will put forward the demand of intelligent scheduling and reduce manual operation in the future.At present,the network management of OTN is usually completed by its own independent network element management system[2].The alarm information shows separately,which makes the network fault diagnosis very complicated,and usually it can only be done manually.Experienced network managers are required to analyze and respond quickly to fault features and correlations[3].This paper is mainly based on neural network and fuzzy set theory to analyze,and use indicators such as F1-Measure to evaluate.

Artificial neural networks (ANN)have the following characteristics: non-linearity,strong robustness and fault tolerance,parallel distributed processing methods,self-learning and adaptive ability,and fast processing of quantitative and qualitative knowledge[4].Therefore,the artificial neural network can be introduced into the fault location field of OTN.The neural network model used in this paper is BPN (Back Propagation Network)and RNN (Recurrent Neural Network).BP (back propagation)neural network is a concept proposed by scientists led by Rumelhart and McClelland in 1986[5].It is a multi-layer feedforward neural network trained according to the error reverse propagation algorithm.It is the most widely used neural network at present.The main purpose of recurrent neural network (RNN)is to process and predict sequence data[6].BPN and RNN are used to analyze the alarm signal and fault point in the OTN,so as to locate the fault.The fuzzy set theory in fuzzy mathematics is introduced,and the probability value of fault location is defined as a [0,1].The probability of each possible fault point is expressed by the size of the probability,and the fault location is carried out.However,these two models also have their own problems: BPN converges slowly,and it is a fully connected network,which cannot deal with the unequal input,and output dimension data sets,that is,its application examples and network size have some contradictions; On the other hand,RNN has the problem of gradient vanishing and overfitting.

The main work of this paper is to apply the neural network to the data and network analysis method in the optical transport network,and explore the existing neural network model more suitable for this field.Select two different neural network models BPN and RNN to train the data generated by the alarm data set.The fault location effect of neural network in optical transport network model is verified by testing the actual alarm data and fault data.In the course of the experiment,not only the evaluation index MSE and location time commonly used in the field of optical network fault location are selected,but also the F1 score in statistics is introduced to measure the effect of neural network fault location in optical transport network more comprehensively.The experimental results show that the alarm data generated by the constructed optical transport network channel model have time correlation and are related to the location of the fault.Compared with the traditional BP neural network,the recurrent neural network,especially the LSTM (Long Short-Term Memory)neural network,can solve the problem of overfitting and gradient vanishing and has a better effect in the field of optical transport network fault location.

II.RELATED WORK

Many related studies have investigated ANN[7,8].Neural network is a parallel distributed processor composed of a large number of neurons (data processing units).It has the characteristics of self-learning and self-adaptation.It acquires data from the external environment (data sets)and continuously adjusts the salient weights between the interconnecting neurons using these data to minimize the difference between the desired response and the actual output signal[9].The neural network is trained with data sets until the correction of the salient weights reaches a stable state.The application of artificial neural network in data mining in Ref.[10] is analyzed in detail.Because of its strong adaptability,robustness and distributed parallel processing,the artificial neural network has a wide application prospect in the field of data mining.The author mainly studies the model structure of the convolution neural network in Ref.[11],and summarizes the application of the convolutional neural network with the examples of image classification,face recognition,audio retrieval,and target detection and so on.The local connection,weight sharing and pool operation of the convolutional neural network can effectively reduce the complexity of the network,reduce the number of training parameters,and make the model invariant to a certain extent in terms of translation,distortion and scaling.It has strong robustness and fault-tolerant ability and is easy to train and optimize.But convolution neural network is strict in dimension of input data,and it cannot achieve the desired training effect for its lack of regularity and easily disturbed data.The author in Ref.[12] used BP neural network to realize image restoration,a three-layer standard BP neural network is established and LM algorithm is used for learning and training[13,14].A neural network scene matching algorithm based on fuzzy set is proposed in Ref.[15].The use of fuzzy set has more anti-jamming ability than traditional neural network algorithm.Gradient regularization and Dropout techniques in Ref.[16] are introduced in image description to avoid model overfitting and GRU is introduced to reduce the training parameters of the model.

On the other hand,there are many related researches in the field of optical transport network fault location.A technique combining OCDR and OTDR is proposed to monitor all optical fibers and components in passive optical network and to locate faults quickly and accurately in Ref.[17].The fuzzy mathematical theory in Ref.[18] is used to analyze the problem of multi-fault location in optical network,and the fault location problem is transferred from the analysis method of random events to the analysis method of fuzzy events.But the fault membership function model needs to be verified by a large number of actual data,and the network topology of its selected lattice optical network is relatively simple at the same time.In order to solve the new problem of fault location in optical switching systems caused by multi-granularity optical channels,a network model based on channel is proposed and the binary tree algorithm is applied to fault location of multi-granularity optical networks in Ref.[19].Of course,when the binary tree algorithm is faced with a large number of alarm sets,it must need a lot of storage space.The author in Ref.[20] introduces a method of locating a fast fault link called monitoring trail (m-trail)in a WDM network,which eliminates the cyclic constraints of the M-cycle and provides a monitoring mechanism that has a smaller monitoring cost than the M-cycle.The neural network is introduced into the fault location of WDM optical network in Ref.[21].The standard BP neural network model is used and the alarm equipment set and fault equipment set are set up and trained.However,in the optical transport network,the alarm signal and the fault equipment are not always fixed dimension.BP neural network cannot deal well with a longer sequence data dimension.At the same time,in the alarm equipment set,the alarm signal has a mapping relationship that cannot be ignored.Therefore,BP neural network is not the best choice to analyze the alarm signal.Therefore,in this paper the current work aims to address the aforementioned gaps extending the work of Ref.[21].Channel-based network model is used to locate the fault of optical transport network.

III.OPTICAL NETWORK MODEL ARCHITECTURE

3.1 Fault and alarm classification of optical transport network

As shown in the figure1,failure can be divided into soft fault and hard fault according to the different forms of fault in optical transport networks.Among them,soft fault refers to those that lead to network performance decline.Generally speaking,soft fault has little effect on network performance,but it is not easy to detect.The hard fault is the fault that the transmission channel is interrupted and the transmission service is completely interrupted by the sudden event.Once there is a hard fault in the network,this should be processed immediately,otherwise it will lead to a lot of data loss[24].

The fault can also be divided into node fault and link fault according to the location of fault[25].Node failure is mainly caused by the failure of the node equipment,the power failure of the device,the unplug of the single board,the fault of the optical transmitter and some other factors.

The alarm of optical transport network can be divided into communication class,quality of service class,device alarm class,processing failure class and environment class[26].In this paper,we mainly consider: optical power anomaly class,optical signal interrupt class and bit error class.

The implementation of survivability in optical transport networks can be divided into four stages: fault detection,fault location,fault notification and fault recovery[22].However,fault detection and location are prerequisites for fault recovery in optical transport networks.Therefore,it is necessary to adopt an effective method to locate accurately and quickly and adopt the corresponding fault recovery strategy.

The fault detection mechanism of optical transport network is the premise of network fault location and network recovery.Whether the location of network fault can be detected quickly will directly affect the stability of network performance.Therefore,the transmission performance of the network is monitored in real time,and the alarm information can be transmitted to the network management system in the event of network failure[23].According to the fault alarm information collected,the management system can locate the location of the fault,and then take the corresponding fault recovery measures to ensure the effective transmission of services in the network.

Fig.1.Fault classification of optical transport network.

3.2 Optical network model construction

In OTN,as in DWDM optical network,the network is generally composed of ring network and mesh network.The ring network protection technology is mature,simple management,is the most commonly used topology at this stage.The ring structure has two main advantages: one is that OTN provides a ring protection mechanism similar to SDH and a multiplexing mapping mechanism,which can be facilitated by the rich networking experience of SDH[27].Second,the OTN technology provides a variety of protection methods under the ring structure,and the security performance of the network is guaranteed.But it also has obvious shortcomings: in order to improve the utilization rate of the system,it is necessary to carry out the loop network design according to the flow direction of the service during the early planning,and to avoid the traffic crossing the loop as far as possible,so it is difficult to design the network.

Fig.2.The channel-based network model.

In order to achieve strong network connectivity requirements,the nodes of a mesh network should be at least an optical crossover device with a crossover function.The optical crossover device should be a WDM device supporting the OTN interface in the ROAM/ PXCU OTN terminal multiplexing device[28].In other words,WDM is equivalent to a subset of OTN.The main advantages are that business scheduling is more flexible,the direct circuit can be set according to the flow direction of traffic,and the network reliability is high.The main disadvantage of the network topology is that the control and management of the structure and response are complicated,because the path arrangement of OTN will involve the wave in the optical layer.In the circuit organization of channel organization and electric layer,before the intelligent function is turned on,the path arrangement needs to be carried out manually,which not only increases the complexity of path arrangement,but also is not convenient for the later maintenance of the network.Therefore,mesh network topology is usually only used in backbone networks with high reliability.

In this paper,the channel-based network model is used to analyze.Figure 2 shows a three-channel optical transport network model.

In figure2,there are three channels,one of which is passed by:A11,R,F7,A12,F9,S1;Channel 2 passes by:A21,M1,F1,P1,F2,R,F6,A22;Channel three passes by:A31,M1,F1,P1,F2,R,F3,P2,F4,M2,F5,A32,F8,S2.

In this paper,the fault classification and alarm classification of optical transport network are analyzed.The equipment of classA,classM,classP,classSandRare all capable of providing alarm signals.ClassAequipment can provide an alarm for the entire external channel device,such as error code alarm,optical signal interrupt alarm or optical power abnormal alarm.ClassM,ClassPand ClassSequipment can provide self-alarm (selfalarm that is,provide alarm when the current equipment fails),for example,optical signal interrupt alarm or optical power abnormal alarm.ClassFdevice cannot provide alarm (as in optical fiber).The device set of the above three-channel model is sequenced to form an alarm vector sequence: {A11,A12,A21,A22,A31,A32,R,M1,M2,P1,P2,S1,S2}.If the set of devices is alerted,the vector value of the corresponding position in the sequence is changed from 0 to 1.In particular,when there is no alarm in the equipment,the corresponding vector value of the alarm sequence is set to 0.For example,whenR,A12of channel one andR,A22of channel two simultaneously send out the alarm,the sequence of alarm vectors is {0,1,0,1,0,0,1,0,0,0,0,0,0}.

At the same time,the device set of the above three-channel model generates the sequence of fault vectors: {A11,A12,A21,A22,A31,A32,R,M1,M2,P1,P2,S1,S2,F1,F2,F3,F4,F5,F6,F7,F8,F9},If the device set fails,the vector of the corresponding position in the sequence is changed from 0 to 1.In particular,when the equipment has no fault,the corresponding vector value of the fault sequence is set to 0.For example: ifF5andM2has a fault,the sequence of faults is {0,0,0,0,0,0,0,0,1,0,0,0,0,0,0,0,0,1,0,0,0,0}.

In the process of fault location,the alarm sequence caused by the fault contains meaningless alarm,error alarm,redundant alarm,or error in the course of transmission to the network.As a result,the network management may receive warnings that are missing and false,and may be repeated[29].Using the theory of fuzzy mathematics to deal with this situation,fuzzy mathematics is a quantitative processing method to study fuzzy phenomena.The concept of fuzzy mathematics is given below.

Fuzzy subset: Set a qualification function for the given universeU,combining each elementxinUwith a number in the interval [0,1].μA(x)indicates the level of qualification inA.Abecomes a fuzzy subset ofU.Here μA(x)is equivalent toCA(x),but its value is not only 0 and 1,but extends to any value in [0,1].Generally,fuzzy subsets are also called fuzzy sets,while classical sets are special cases of fuzzy sets.

Membership function: For any givenUand mapping μAofUin the closed interval [0,1],

A fuzzy subsetAofUcan be determined:

In the formula,μA(x)is called the membership function ofA,and μA(xi)is called the degree of membership of elementxi.At the time,when μA(xi)=1,xicompletely belongs to fuzzy setA; when μA(xi)=0,xidoes not belong to fuzzy setAat all; the closer μA(xi)is to 1 and the greater the degree to whichxibelongs toA[30].

The fault is represented by the membership function.The degree of the possibility of the fault occurrence position can be determined by the degree of membership of each vector of the fault sequence,and the fault position with a higher probability is preferentially processed.Degree of membership in the range of [0,0.1] is unlikely to be faulty; range in (0.1,0.5)may be faulty; range in (0.5,0.8)extremely likely to be faulty; range in (0.8,1)must be faulty.

IV.NEURAL NETWORK MODEL

The neural network models in this paper are BP neural network model and LSTM (longshort term memory networks)model),respectively.

4.1 BP neural network model

In the BP neural network model,the alarm sequence of the three-channel network model in the last text is used as input,which is composed of 13 [0,1] sequences,and the fault sequence is the output,which is composed of 22 [0,1] sequences.The BP neural network model has four layers,13 neurons in the input layer,32 and 64 neurons in the hidden layer and 22 neurons in the output layer.A BP neural network of 13×32×64×22 is constructed.The number of neurons in the hidden layer of neural networks can be obtained by empirical formula:

m: the number of neurons in the hidden layer

n: the number of neurons in the input layer

l: the number of neurons in the output layer

The number of hidden neurons is continuously changed to train the network,and the number of hidden layer neurons with the smallest network error is chosen.In this model,the number of hidden layer neurons is 32 and 64 respectively.Figure 3 shows a four-layer BP neural network model.

Given the n-dimensional input layer sequence:x1,x2,...xn,the first hidden layer sequence ish11,h12,…h1m,the second hidden layer sequence ish21,h22,…h2j,L dimensional output layer sequence isy1,y2,…yl.

Each connection weight is assigned a random number in one interval (-1,1),and the error function E is set.The learning rate is 0.01 and the number of iterations is 150 times.The first three layers use the ReLUfunction in the activation function of the four-layer neural network,and the fourth layer selects the sigmoid function.

Fig.3.Four-layer BP neural network model.

Fig.4.Single neuron structure.

In the forward pass,the output of therthneuron of the hidden layer is:

In the reverse pass,the output error of the sample is:

The update formula for any weight parameter is:

The update formula for any threshold parameter is:

Among them,Wiris the weight matrix,Sris the output of the hidden layer,bris the base,Eis the error function,W+is the updated weight,andb+is the updated threshold.

However,the BP neural network has the following drawbacks: it is easy to form local minimum values and cannot get the global optimal value; many training times make the learning efficiency low,the convergence speed is slow; the selection of hidden layers lacks theoretical guidance; learning new during training the sample has a tendency to forget the old sample.Therefore,we also use the LSTM model to analyze the data set again.

4.2 LSTM model

LSTM (Long Short Term)is a special type of RNN.Due to the fact that in the RNN,the relationship between the neurons that are far away from each other is calculated,it involves the multiplication of the Jacobian matrix so that the gradient disappears or the gradient explodes.LSTM changes the weight of the selfloop by increasing the input threshold,forgetting threshold,and output threshold.In this way,when the model parameters are fixed,the integral scale at different times can be changed dynamically,thus avoiding the gradient disappearing or the gradient expanding problem.The LSTM single neuron structure is shown in figure4.

According to its structure,ftrepresents the forgotten threshold,itrepresents the input threshold,represents the state of the neuron at the previous moment,Ctrepresents the state of the neuron at the moment,otrepresents the output threshold,htrepresents the output of the current neuron,andht-1represents the output of the previous moment.The forgetting threshold determines how much of the unit state at the last moment is retained until the current moment; the input threshold determines how much of the network input is saved to the cell state at the current moment; the output threshold controls how much of the unit state is output to the current output value of the LSTM.

Therefore:

Amongft,otandht,the activation function is sigmoid function.Set the number of LSTM network layers to three.There are 13 neurons in the input layer,the number of neurons in the hidden layer is 64,and the number of neurons in the output layer is 22.A 13×64×22 three-layer LSTM neural network is constructed.Each connection weight is assigned a random number within the interval (-1,1),and the error function E is set.The learning rate is 0.01 and the number of iterations is 150.The first two layers of the LSTM activation function use theReLUfunction,and the third layer uses the sigmoid function.

4.3 Feasibility analysis

LSTM neural network is a kind of recurrent neural network,which is a special cyclic neural network.It is suitable for processing and predicting important events with relatively long interval and delay in sequence.Sequence prediction analysis is to make use of the characteristics of the event time in the past period to predict the characteristics of the event in the future.This is a relatively complex prediction modeling problem.Different from the prediction of regression analysis model,it depends on the sequence of events,and the result of input model is different after changing the sequence of values of the same size.

The fault alarm data set of OTN can be used in LSTM neural network.The main reasons are as follows: OTN fault alarm data set is composed of a large number of alarm vectors and is time-related.The alarm vector dimension is constructed from the channel model in the network.When there is a link or node failure in the network at a certain time,the network management background sends out an alarm signal,from which the alarm vector can be determined.Fault of a single node device on a channel in a network,such as a laser power fault.In this case,not only the element value corresponding to the laser position in the alarm vector is changed from 0 to 1,but also the OLP device which detects the optical signal,for example,sends out the alarm signal.At the same time,the signal transmission direction is different; the alarm vector is also different,which is determined by the location of the node and line devices in the OTN.Therefore,there is a correlation between different elements in the alarm vector and the fault time and location,which is consistent with the characteristics of LSTM neural network.

4.4 Measurement

The simulation index and experimental results shown in this paper are two commonly indexes in the field of optical transport network fault location,which are: fault location time and mean square error of fault location.At the same time,the fault location model of optical transport network is a two-classification model.In order to measure the accuracy of the model more comprehensively,F1 score is introduced.

In measuring the effect of the neural network model on the fault location of the optical transport network,only the accuracy of the positioning is analysed,and sometimes some results is not good enough.Due to unequal data distribution,99% accuracy is likely to occur,ignoring the location of a small number of possible failures.Therefore,relying on the accuracy rate to evaluate the algorithm model is far from scientific and comprehensive.

Therefore,the concept of statistics inF-Measure was introduced.F-Measure is an indicator used to measure the accuracy of a binary classification model.It also takes into account the accuracy and recall of the classification model,which can be regarded as a weighted average of model accuracy and recall.Accuracy rate is the ratio of the number of positive samples that are classified correctly to the number of positive samples divided by the classifier and recall rate is the ratio of the number of positive samples that are correctly classified to the number of positive samples.P and R represent the accuracy rate and recall rate respectively in the paper.

When the parameter α=1,F-Measure isF1-Measure,also denoted asF1score.The maximum value is 1 and the minimum valueis 0.The confusion matrix is shown in Table.1.

Table I.Confusion matrix.

Fig.5.MSE of BP neural network and LSTM.

In this article,True Positive (TP)represents that there is a fault in the location where the fault originally occurred in the test fault sequence; True Negative (TN)represents that the location of the original fault in the test fault sequence is judged as no fault,and it is a leak judgment behavior; False Positive (FP)indicates that the original fault-free position in the test fault sequence is judged to be faulty and belongs to a misjudgment behavior; False Negative (FN)indicates that the original fault-free position in the test fault sequence is judged to be fault-free.

Formula (19)is derived from formula (17)and (18).It can be seen from (19)that the larger F1 is,the better the training effect of this neural network model is.

V.RESULTS AND DISCUSSION

Using expert experience to obtain 10000 sets of data,9000 sets of data were set as training set,and 1000 groups of data were selected as test set to verify the results of neural network model.The trained BP neural network and LSTM neural network were trained and processed by Keras of Python.After 150 iterations,the mean square error and F1 score were calculated with the test sample.From the positioning time,the positioning time of the test sample by the BP neural network is 0.3063 seconds,and the positioning time of the LSTM test sample is 0.2151 seconds.

Figure5 shows the mean squared error curves of the BP neural network and LSTM.The red curve is the mean square error curve of LSTM,and the blue curve is the mean square error curve of the BP neural network.It can be seen from the figure that when the number of iterations is 45 times,the decreasing trend of the mean square error of the two began to become slow.When iterates 150 times,the mean square error decreases to the lowest point.The mean square error of the LSTM model and BP neural network model is not very different.However,LSTM converges faster than BP neural network.

Figure 6 shows the F1 score curve of the BP neural network and LSTM.After training,the F1 score of the BP neural network is generally 0.932346723044397 and theF1score of the LSTM can be set to 0.961566888396156.Under the same number of iterations,the F1 score of LSTM tends to be stable compared to the BP neural network and the F1 score of LSTM can achieve a peak value of 1 more times.The optimal number of iterations for the LSTM is 45.

As for BP neural network,there are peaks in F1 score,but the curve shows that the F1 score of BP neural network is not very stable.The number of iterations of the BP neural network is the 55th.Table 2 shows the results of BP and LSTM in this training.

Therefore,for the optical transport network channel model of this paper,LSTM performs better in fault location than BP neural network.As can be seen from the positioning time,LSTM is slightly superior to BP neural network and has a shorter time.In order to ensure its positioning effect,the number of iterations set in this experiment is more.As far as MSE is concerned,BP neural network is not much different from LSTM,but BP neural network is likely to face the problem of overfitting.At the same time,from the curve,it can be seen that BP neural network does not converge as fast as LSTM.From the F1 score curve,when it is 0,it means that this fault location is all invalid; but when it is equal to 1,it means that this fault location is all located correctly.Both LSTM and BP neural networks perform well on this parameter,but as can be seen from the curves,LSTM has more advantages and is more stable and stable.

At the same time,the structure of LSTM neural network is complex,the function between Input Gate,Output Gate,and Forget Gate may be repeated,and the simple structure can be improved.However,the feasibility of the improved neural network needs to be further analyzed.

Fig.6.F1 score of BP neural network and LSTM.

Table II.Results of BP and LSTM.

At present,the application of this method in OTN is still aimed at small-scale networks,that is,the number of channels in the network is less,corresponding to the number of traffic wavelengths carried on the link in OTN is less,and the number of traffic requests is lower.

This method is only suitable for static network,once the size of the network changed,such as the network topology and the wavelength variability of the optical signal on the link,it is necessary to re-select the data set for the training of the model.

The data input in this method is the set of alarm signals in the network,that is,the OTN network management unit sends out the alarm information.The alarm information is derived from the hard fault in OTN or soft fault which has exceeded the prescribed standard.Therefore,it is necessary to re-select the input data set to train for the deterioration of internal optical performance parameters in OTN.

In this paper,due to the introduction of the concepts of membership and F1 score,the leakage alarms and false alarms are fully considered in the tested test data sets.The application of neural network can accurately and quickly locate faults in the optical transport network.

ACKNOWLEDGEMENT

We gratefully acknowledge anonymous reviewers who read drafts and made many helpful suggestions.