CN112985574B - High-precision classification identification method for optical fiber distributed acoustic sensing signals based on model fusion - Google Patents

Abstract

The invention discloses a high-precision classification and identification method for optical fiber distributed acoustic sensing signals based on model fusion, and belongs to the technical field of optical fiber distributed sensing and machine learning. According to the method, a stacking strategy is introduced into DAS signal identification, a decision tree model, a random forest model, a support vector machine model and an extreme value gradient lifting algorithm model in machine learning are integrated, original data obtained by a DAS system are preprocessed, characteristics of multiple analysis domains are extracted, an artificial neural network is used for further extracting and classifying the characteristics, a prediction result of the model is relearned by using logistic regression, and a final prediction of an environment state where a current input signal is located is obtained. Compared with the traditional method, the identification method has higher identification rate and shorter discrimination time, and has great significance in pursuing real-time optical fiber distributed sensing detection.

Description

High-precision classification identification method for optical fiber distributed acoustic sensing signals based on model fusion

Technical Field

The invention relates to the technical field of optical fiber distributed sensing and machine learning, in particular to an optical fiber distributed acoustic wave sensing signal classification and identification method based on model fusion.

Background

Optical fiber Distributed Acoustic Sensing (DAS) systems have been widely used in oil and gas exploration and other fields. The detection of the external disturbance signal is realized by demodulating the phase of the coherent Rayleigh scattering light, so that the external disturbance can be qualitatively detected, and the magnitude of the external disturbance signal can be reflected by the magnitude of the phase amplitude, namely, the quantitative measurement is realized.

Signal processing of DAS data is largely divided into three phases. The first stage is mainly aimed at detection and positioning, and improves the signal-to-noise ratio of a detection positioning signal by preprocessing means such as empirical mode decomposition, wavelet transformation, Hilbert-Huang transformation, signal denoising and separation and the like, reduces the influence of the noise of the system and environmental noise on a detection result, and can improve the accuracy of the system to a certain degree. However, in practical application, it is found that the noise environment on site is complex and variable, and the false alarm rate of the system is mainly caused by insufficient recognition of the noise source for the evolution and change of the target time state.

The second stage begins exploring multi-domain feature analysis and recognition model exploration. Focusing on the research on a target classification and identification method in a specific application environment, the method extracts features including signal amplitude in a time domain, a horizontal zero crossing rate, FFT spectral energy distribution features of a frequency spectrum, a short-time Fourier transform spectrum, a Mel frequency cepstrum (MFCC) and the like, and then combines classifiers such as a neural network, a Gaussian Mixture Model (GMM), a support vector machine, a random forest, a Hidden Markov Model (HMM) and the like. However, the feature extraction method mainly performs professional analysis and design according to a specific environment or limited signal samples, has poor generalization capability in the face of actual complex application environments, is long in development period, wastes time and labor, and is difficult to realize high-precision classification and identification of signals in a complex background noise environment.

With the rapid development of technologies such as artificial intelligence, machine learning and the like, the automatic extraction of features can be realized on different target event signals by using a deep learning method, so that the feature extraction mode of the DAS signal obtains breakthrough progress, and the identification accuracy of events is further improved. The deep learning method in the signal processing field improves the accuracy of the target classification recognition system, but the network is complex and has many parameters, a large amount of data is required for model training, and the training time is long. In practical application, a large number of samples are difficult to collect, and training of a deep learning high-precision large model is difficult to support.

Disclosure of Invention

The invention aims to overcome the problems that in the prior art, a large number of samples are needed to support a large high-precision model for identifying different events by using a deep learning method and the problem that the signal characteristics are extracted by using a traditional signal processing method and deep knowledge in the field of digital signal processing is needed in a fiber Distributed Acoustic Sensing (DAS) system, and provides a model fusion-based high-precision classification and identification method for fiber distributed acoustic sensing signals.

The purpose of the invention is realized by the following technical scheme:

step 1: a data processing unit of the DAS receives an external sound wave signal and constructs a database of original signals;

step 2: carrying out data standardization processing on the acquired external sound wave signals by adopting maximum-minimum standardization;

and step 3: extracting amplitude-frequency characteristics of a time domain, a frequency domain and a transform domain from the external sound wave signals subjected to the standardization in the step 2, and combining to obtain a feature vector a;

and 4, step 4: automatically extracting features of the feature vector a obtained in the step 3 by using an ANN based on a BP algorithm to further mine the amplitude-frequency characteristics of the external sound wave signal, so as to obtain a feature vector b and a prediction result y _ hat of the ANN; specifically, when the feature vector a obtained after the process of step 3 is propagated in the forward direction by using the BP algorithm, the input sample is transmitted from the input layer, and is transmitted to the output layer after being processed layer by each hidden layer. If the actual output of the output layer does not match the expected output, the error is propagated back to the error stage. The error back transmission is to back transmit the output error to the input layer by layer through the hidden layer in a certain form, and distribute the error to all units of each layer, thereby obtaining the error signal of each layer of units, and the error signal is used as the basis for correcting the weight of each unit. The weight adjustment processes of each layer of signal forward propagation and error backward propagation are carried out until a set iteration number is reached or the loss is within an acceptable range. In the training process, the weight of the model is continuously adjusted, and the model is continuously learned in the direction which accords with the reality of the sample. And further feature mining is carried out on the feature vector a by adopting an ANN, the vectors after the full connection layer are output to obtain a description data feature distribution vector b, and further feature mining is carried out on the feature vector a by adopting the ANN to obtain a description data feature distribution vector b and a prediction y _ hat.

And 5: performing feature fusion on the feature vector a obtained in the step 3 and the feature vector b obtained by further mining in the step 4 to obtain a feature vector W, and using the feature vector W as a training set for training a first-layer base model to obtain a prediction result matrix y _ com of the first-layer base model;

step 6: and (4) merging the ANN prediction result y _ hat obtained in the step (4) and the prediction result matrix y _ com obtained in the step (5) to be used as a training set of the second-layer base model for training, so as to obtain the final prediction result of the final ensemble learning model.

Specifically, the DAS system in step 1 includes a light source, a modulation unit, an amplification unit, an optical fiber sensing unit, and a data processing unit.

Specifically, the signal database obtained after normalization in step 2 is divided into a training set and a test set according to a ratio of 7: 3.

Specifically, the step 3 of extracting the time domain, frequency domain and transform domain features of the normalized signal is specifically as follows:

the time domain features include: the pulse intensity, average amplitude, short-time energy autocorrelation coefficient and signal energy of the signal;

the frequency domain features include: variance, skewness, kurtosis, information entropy and mean value of signal amplitude;

the transform domain features include: mel-frequency cepstrum coefficients.

Specifically, the step 5 specifically includes: performing feature fusion on the obtained feature vector a and the obtained feature vector b to serve as input of a first layer base model, utilizing a stacking strategy to enhance the characteristics of the weak classifiers, performing cross validation learning on the first layer base model respectively to obtain prediction of the first layer base model, and combining the prediction y _ hat of the ANN, and combining the prediction result of the ANN and the prediction result of the first layer base model to obtain prediction output y _ com; the weak classifier can be enhanced mainly by using a stacking strategy, and the first layer base models of the weak classifier are DT, RF, SVM and XGB respectively. And (3) taking the feature vector W as an input signal of the first-layer base model, respectively carrying out 4-fold cross validation on the four weak classifiers to obtain prediction results of a training set and a test set, combining the judgment result y _ hat of the ANN obtained in the step 4 to obtain prediction outputs y _ com of five models in total, and respectively obtaining the training set and the test set of the second-layer base model by adopting a mode of stacking and combining in rows. Wherein the training set is used to train a second level basis model (LR). And training the prediction outputs of the five models by using a logistic regression model, and further obtaining the weight vector of the model so as to obtain the final discrimination result.

The invention brings the advantages that:

1. and a stacking strategy is introduced, so that the accuracy of the traditional machine learning model is improved.

2. The ANN is used for further feature extraction of the data, so that the problems of single feature, omission and the like caused by manual feature selection can be solved.

3. Compared with a method for extracting features by using a Convolutional Neural Network (CNN), the ANN reduces the operation of convolutional pooling compared with the CNN, greatly reduces the calculated amount and the parameter amount, can avoid the problems that deep learning needs a large number of samples for fitting and the training time is long, and has great significance in the field of optical fiber sensing pursuing real-time performance.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the invention without limiting the invention. In the figure:

FIG. 1 is a block diagram of a system according to embodiment 1 of the present invention;

FIG. 2 is a flowchart of a method of example 1 of the present invention;

FIG. 3 is a block diagram of a network model according to embodiment 1 of the present invention;

FIG. 4 is a confusion matrix diagram of the model used in example 1 of the present invention.

Detailed Description

The technical solutions of the present invention will be described clearly and completely with reference to the accompanying drawings, and it should be understood that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In the description of the present invention, it should be noted that directions or positional relationships indicated by "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", and the like are directions or positional relationships based on the drawings, and are only for convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be configured and operated in a specific orientation, and thus, cannot be construed as limiting the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

In the description of the present invention, it should be noted that, unless otherwise explicitly stated or limited, the terms "mounted," "connected," and "connected" are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.

The data processing unit receives an external sound wave signal and extracts the amplitude-frequency characteristics of a time domain, a frequency domain and a transform domain of the external sound wave signal; the amplitude-frequency characteristics of the signals are further mined by utilizing the characteristic of automatic feature extraction of an Artificial Neural Network (ANN) based on a BP algorithm; the amplitude-frequency characteristics of the signals and the further mined data characteristics are subjected to feature fusion, a stacking method in machine learning is introduced into DAS signal identification for the first time, and the method is utilized to improve the identification accuracy of a traditional machine learning classifier, so that the high-precision classification identification method of the optical fiber distributed acoustic sensing signals is realized. The method comprises the steps of collecting external sound wave signals by using a data processing unit, extracting characteristics of each sample data from a time domain, a frequency domain and a transformation domain by using a traditional digital signal processing method, and constructing a characteristic vector a.

When the BP algorithm is used for forward propagation, input samples are transmitted from an input layer, processed layer by each hidden layer and transmitted to an output layer. If the actual output of the output layer does not match the expected output, the error is propagated back to the error stage. The error back transmission is to back transmit the output error to the input layer by layer through the hidden layer in a certain form, and distribute the error to all units of each layer, thereby obtaining the error signal of each layer of units, and the error signal is used as the basis for correcting the weight of each unit. The weight adjustment process of each layer of signal forward propagation and error backward propagation is performed in cycles. In the training process, the weight of the model is continuously adjusted, and the model is continuously learned in the direction which accords with the reality of the sample. And taking the feature vector a as the input of the ANN, and performing further feature mining to obtain a description data feature distribution vector b. And splicing the obtained feature vector a and the feature vector b to obtain a feature vector c as the input of the model.

Specifically, the weak classifiers can be enhanced by using a stacking strategy, and the four weak classifiers are DT, RF, SVM and XGB respectively. And taking the obtained feature vector c as an input signal of the first-layer base model, and respectively carrying out cross validation learning on the four weak classifiers to obtain the prediction of the four weak classifiers. And (4) collecting the judgment results of the ANN obtained in the second step to obtain the prediction outputs of the five models in total, training the prediction outputs of the five models by using a Logistic Regression (LR) model, and further obtaining the weight vector of the model so as to obtain the final judgment result.

A stacking strategy is introduced into DAS signal processing for the first time, and an integrated learning method is utilized to enhance a weak classifier.

Example one

The optical fiber distributed acoustic wave sensing communication optical cable safety monitoring system, the system structure and the working principle thereof are shown in fig. 1. The system hardware consists of three parts, namely a sensing optical cable, optical signal demodulation equipment and a signal processing host. The sensing optical cable generally utilizes a common single-mode communication optical cable, and the internal components of the optical signal demodulation equipment mainly comprise an optical device, a collection card and other equipment. A narrow linewidth laser generates a path of coherent light signal which is modulated into an optical pulse signal by an acousto-optic modulator, the optical pulse signal is intensively amplified by an erbium-doped fiber amplifier (EDFA), and the amplified optical pulse signal is injected into a sensing optical cable through an isolator and ports 1 and 2 of a circulator; the optical pulse signals generate Rayleigh scattering along the transmission process of the optical cable, backward Rayleigh scattering optical signals return along the optical cable and are received by ports 2 and 3 of the circulator, the optical signals are demodulated by a demodulation system and converted into electric signals by a photoelectric detector, and the electric signals under the action of the sound waves and the vibration signals on the optical fiber are acquired by an acquisition system and are transmitted to a signal processing host in real time through interfaces such as a network. The signal processing host is a common computer host or an AI chip and is used for analyzing and processing the optical fiber detection signal, event information around the communication optical cable can be obtained through a specific signal processing algorithm, and the sensing time can be intelligently analyzed, processed and classified and identified.

Based on the basis, in the application of safety monitoring of the communication optical cable, sample databases of different types of time are constructed, the spatial resolution of the embedded communication optical cable is 5.16m, the embedded communication optical cable is divided into 100 sensing units, the acquisition time length of each signal is 30s, and the acquisition time length is also the original input data of the model.

In the monitoring of the transmission optical cable, four main events frequently encountered include road breaker excavation, excavator excavation, manual excavation, background noise and the like. Label of the event is respectively marked as 0 to 3, in addition, an original data set is randomly disordered and then divided into a training set and a testing set according to the proportion of 7:3, and the constructed data set of various events is shown in table 1.

TABLE 1

Event type	Training/test set size	Event label
			Road breaking machine excavation	58/25	0
Digging machine	72/31	1
			Manual touch cable	63/27	2
Background noise	58/25	3

Sniping is carried out on the collected time, each data sample in the data set is preprocessed through normalization operation, and then ten characteristics shown in the table 2 are extracted from the data in a time domain, a frequency domain and a transformation domain;

TABLE 2

Analytical domains	Feature(s)
		Time domain	Pulse intensity, average amplitude, short-time energy autocorrelation coefficient and energy
Frequency domain	Variance, skewness, kurtosis, information entropy, mean
		Transform domain	Mel frequency cepstrum coefficients

And simultaneously inputting the preprocessed data into an ANN for feature extraction, wherein 16 neurons in a full connecting layer are total, performing feature fusion on the 16-dimensional ANN feature vector and the 10-dimensional features to be used as 26-dimensional feature vector input of a stacking model, and training the judgment result of the ANN as the second-layer input of the stacking model.

And inputting the judgment result of the ANN after the softmax layer and the result of the stacking first layer basis model into the LR layer together for next fitting. And finally, performing stacking training in a 4-fold cross validation mode, wherein a prediction output obtained by combining the prediction of the four base learners of the first layer and the prediction of the ANN is used as the input of the LR of the second layer, and the final output of the LR is the final signal identification result. The whole process is shown in fig. 3.

The data set is tested by using the high-precision classification and identification method of the optical fiber distributed acoustic wave sensing signals based on model fusion, the confusion matrix is shown in fig. 4, the established high-precision classification and identification method of the optical fiber distributed acoustic wave sensing signals based on model fusion has the advantage of high identification rate, the only misjudgment is from judging background noise as artificial contact cable, and the situation of misjudgment of high threat events does not exist in practical application.

In the embodiment, the identification result of DAS signal safety monitoring based on the traditional machine learning method is compared with the identification result of the method provided by the invention, and the advantages and the effectiveness of the method provided by the invention are verified. The specific method comprises the following steps:

(1) and inputting the feature set into a traditional machine learning model for classification and identification, wherein the traditional machine learning model selects 4 representative models, which are also the basic models in the method provided by the invention. Respectively as follows: a support vector machine SVM, a random forest RF, an XGboost model XGB and a decision tree model DT.

(2) In general, four standard performance metrics commonly used in classification (e.g., Accuracy, Precision, Recall, and F-measure) are used to evaluate the recognition results, as follows:

Accuracy＝(TP+TN)/(TP+FP+TN+FN)

Precision＝(TP+TN)/(TP+FP)

Recall＝TP/(TP+FN)

F-measure＝2×Precision×Recall/(Precision+Recall)

the constructed 26-dimensional feature vectors are input into different models for training, and the recognition results are shown in table 3.

TABLE 3

It can be seen that of the remaining four models, the highest accuracy is the random forest model, which has an accuracy of 94.44%, whereas the accuracy of the model proposed herein is up to 99.07%. This demonstrates the effectiveness of the stacking model proposed herein for DAS signal identification, far beyond the commonly used machine learning classification methods.

For different event types, the identification accuracy of the stacking model is higher than that of the other five models, and the classification accuracy of the event 2 which is easier to make mistakes by the other four methods is higher, and the superiority of the stacking model provided by the text is shown through comparison. This is because the machine learning based stacking model utilizes multi-fold cross validation, making the results more robust. The effectiveness and the outstanding advantages of the method provided by the invention in solving the problem of safety monitoring of the communication optical cable are demonstrated.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims (4)

1. The method for identifying the high-precision classification of the optical fiber distributed acoustic sensing signals based on model fusion is characterized by comprising the following steps:

step 1: a data processing unit of the DAS receives an external sound wave signal and constructs a database of original signals;

step 2: carrying out data standardization processing on external sound wave signals in a database by adopting maximum-minimum standardization;

and step 3: extracting amplitude-frequency characteristics of a time domain, a frequency domain and a transform domain from the external sound wave signals subjected to the standardization in the step 2, and combining to obtain a feature vector a;

and 4, step 4: automatically extracting features of the feature vector a obtained in the step 3 by using an ANN based on a BP algorithm to further mine the amplitude-frequency characteristics of the external sound wave signal, so as to obtain a feature vector b and a prediction result y _ hat of the ANN;

and 5: performing feature fusion on the feature vector a obtained in the step 3 and the feature vector b obtained by further mining in the step 4 to obtain a feature vector W, and using the feature vector W as a training set for training a first-layer base model to obtain a prediction result matrix y _ com of the first-layer base model;

step 6: merging the ANN prediction result y _ hat obtained in the step 4 and the prediction result matrix y _ com obtained in the step 5, and training the merged prediction result as a training set of a second-layer base model to obtain a final prediction result of the final ensemble learning model;

the step 5 specifically comprises the following steps: performing feature fusion on the obtained feature vector a and the obtained feature vector b to serve as input of a first layer base model, utilizing a stacking strategy to enhance the characteristics of the weak classifiers, performing cross validation learning on the first layer base model respectively to obtain prediction of the first layer base model, and combining the prediction y _ hat of the ANN, and combining the prediction result of the ANN and the prediction result of the first layer base model to obtain prediction output y _ com;

the first-layer base models are respectively a decision tree, a random forest, a support vector machine and a gradient lifting algorithm.

2. The model fusion-based optical fiber distributed acoustic wave sensing signal high-precision classification and identification method according to claim 1, characterized in that: the DAS system in the step 1 comprises a light source, a modulation unit, an amplification unit, an optical fiber sensing unit and a data processing unit.

3. The model fusion-based optical fiber distributed acoustic wave sensing signal high-precision classification and identification method according to claim 1, characterized in that: in the step 2, the external sound wave signals in the database after the standardization processing are divided into a training set and a test set according to the ratio of 7: 3.

4. The model fusion-based optical fiber distributed acoustic wave sensing signal high-precision classification and identification method according to claim 1, characterized in that: the step 3 of extracting time domain, frequency domain and transform domain features of the normalized signal is specifically as follows:

the time domain features include: the pulse intensity, average amplitude, short-time energy autocorrelation coefficient and signal energy of the signal;

the frequency domain features include: variance, skewness, kurtosis, information entropy and mean value of signal amplitude;

the transform domain features include: mel-frequency cepstrum coefficients.

Images