A fast training identification algorithm for phi-otdr signal based on space-time domain combination, mfcc and pre-extraction technology

Abstract

The application discloses a kind of based on space-time domain combination, MFCC and pre-extraction technology Φ-OTDR signal fast training identification algorithm.The algorithm includes four steps: ① first, Φ-OTDR signal data set is combined with time domain and space domain each other;②Second, Mel Frequency Cepstrum Coefficient (MelFrequencyCepstrumCoefficient, MFCC) is solved;③Then, Mel Frequency Cepstrum Coefficient is extracted in advance, and stored;④Finally, import Mel Frequency Cepstrum Coefficient stored in advance for training when training.Overall, compared with prior art, the following beneficial effects can be achieved by the above technical solutions of the application: ①Compared with the method of detecting single optical fiber node for identification and classification, the method of mutual combination of time domain and space domain in the present solution can avoid excessive interference caused by single node error, has stronger anti-interference ability, and introduces the information of space domain simultaneously;②Compared with the method of directly training and identifying optical fiber node data, the method of converting into Mel Frequency Cepstrum Coefficient in the present solution improves information feature density and reduces data volume during training;③Compared with the method of extracting features and training simultaneously, the pre-extraction technology in the present solution can reduce training time;④Compared with the method of using complex network structure, the simplified network structure in the present solution not only trains fast, but also identifies fast, and is more practical.

Description

Technical Field

[0001] The present invention relates to the technical field of back-end signal processing of a phase-sensitive optical time domain reflectometer (φ-OTDR), and in particular to a φ-OTDR signal fast training and recognition algorithm based on a combination of time and space domains, MFCC, and pre-extraction technology. Background Art

[0002] Distributed acoustic sensors (DAS) based on φ-OTDR technology are currently being applied in many everyday scenarios due to their high spatial resolution, high sensitivity, and wide measurement range. DAS are also widely used to identify various intrusion signals. However, due to their high sampling frequency and the large number of fiber nodes, the amount of data generated per second by a single fiber is enormous.

[0003] Many researchers are currently working on signal classification and recognition using distributed acoustic sensors. Most focus on improving classification accuracy by pre-collecting data and then performing offline processing. While some current algorithms offer high offline recognition accuracy, they suffer from complex and bloated models, long training times, and slow recognition speeds. Summary of the Invention

[0004] To address the issues described in the background, this paper proposes a rapid training and recognition algorithm for Φ-OTDR signals based on a combination of time and space domains, MFCC, and pre-extraction techniques. This algorithm incorporates three preprocessing methods—a combination of time and space domains, Mel-frequency cepstral coefficients, and pre-extraction—and a streamlined network architecture. While ensuring high recognition accuracy, this algorithm exhibits strong anti-interference capabilities and rapid model training, making it more practical. In experiments, 50 rounds of training on a single graphics card took only 1.64 minutes, achieving an accuracy of 99.79%, providing advantages for subsequent large-scale data training.

[0005] 1. The technical solution of the present invention comprises the following steps:

[0006] S1: First, the Φ-OTDR signal dataset is combined with the time domain and the spatial domain.

[0007] S2: Secondly, calculate the Mel Frequency Cepstrum Coefficient (MFCC).

[0008] S3: Then, the Mel-frequency cepstral coefficients are pre-extracted and stored.

[0009] S4: Finally, at the time of training, the pre-stored Mel frequency cepstral coefficients are imported for training.

[0010] 2. In some optional embodiments, step S1 comprises:

[0011] S11: The data set we adopted has 6 categories, namely background sound (background), digging sound (dig), knocking sound (knock), shaking sound (shake), walking sound (walk), and the ratio of training set and validation set is 4:1. The data set has a total of 15612 files in.mat format. After parsing, each file is a matrix with a shape of 10000x12, where 10000 is the sampling frequency representing the information in the time domain, and 12 is the 12 adjacent optical fiber nodes representing the information in the spatial domain.

[0012] S12: When the event signal occurs, multiple nodes in the optical fiber may all be able to sense it, with the nodes close to the signal being strong and the nodes far away being weak. Therefore, for event signals such as digging sound (dig), knocking sound (knock), shaking sound (shake), and walking sound (walk), we add and average the signals of the 12 adjacent nodes, which can avoid excessive interference caused by a single node error and has stronger anti-interference ability, while introducing information in the spatial domain. For background signal (background), since the background signal intensity of different nodes is inconsistent, averaging will interfere with recognition, so we directly split the 10000x12 matrix into 12 10000 sequences.

[0013] 3. In some optional embodiments, step S2 comprises:

[0014] S21: Mel frequency cepstral coefficients have been widely used in the field of speech recognition because of their good effect. Distributed acoustic sensors are essentially sound recognition, and the signals recognized are similar to human speech signals, so Mel frequency cepstral coefficients are introduced. The main steps for calculating Mel frequency cepstral coefficients are as follows: ① pre-emphasis; ② frame division; ③ windowing; ④ Fast Fourier Transform (FFT); ⑤ Mel filter bank; ⑥ logarithmic operation; ⑦ Discrete Cosine Transform (DCT).

[0015] S22: Solve the Mel frequency cepstrum coefficient, we use the Librosa toolkit under the feature.mfcc function. The code is: mfcc = librosa.feature.mfcc (y = k, sr = 10000), where k is the sequence information, and 10000 is the sampling rate. After solving, 10000 data becomes a 20x20 matrix data containing rich feature signals. This reduces the data amount during neural network training and improves the information feature density.

[0016] 4. In some optional embodiments, step S3 comprises:

[0017] S31: During training, if Mel frequency cepstrum coefficients are extracted at the same time, the training time will be increased. Extracting all files before training and saving them in npy format can reduce the training time. Of course, saving them in other formats is also possible, as long as they can be directly imported during training.

[0018] S32: Use the os.walk() function to traverse the files under the path, then operate them one by one to become Mel frequency cepstrum coefficients, and finally save them to a specific folder using the save function under the numpy toolkit. During training and testing, the data can be directly imported using the load function under the numpy toolkit.

[0019] 5. In some optional embodiments, step S4 comprises:

[0020] S41: Considering the rich feature information of Mel frequency cepstrum coefficients, it is decided to use a simplified convolutional neural network, which makes the model not only faster to train, but also faster to recognize, and more practical.

[0021] S42: The convolutional neural network mainly consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The network structure is generally as follows: after the input layer, first pass through 2 convolutional layers and 1 pooling layer, then pass through 2 convolutional layers and 1 pooling layer, then use the flatten function to flatten, then enter the fully connected layer, and finally enter the output layer. Also use ReLu and Dropout functions.

[0022] S43: Import the dataset using the load function under the numpy toolkit, send it to the network structure, define the number of iterations, optimization method, loss function, batch size, learning rate, and num_workers parameters, and then you can start training.

[0023] 6. In general, the above technical solutions of the present invention can achieve the following beneficial effects compared with the prior art:

[0024] (1) Compared with the method of detecting a single optical fiber node for identification and classification, the method of combining the time domain and the spatial domain in this scheme can avoid excessive interference caused by a single node error, has stronger anti-interference ability, and introduces spatial domain information.

[0025] (2) Compared with the method of directly training and identifying fiber node data, the method of converting it into Mel frequency cepstral coefficients in this scheme improves the information feature density and reduces the amount of data during training.

[0026] (3) Compared with the method of extracting features and performing training at the same time, the pre-extraction technology in this scheme can reduce the training time.

[0027] (4) Compared with the method using a complex network structure, the simplified network structure used in this scheme not only makes the model training faster, but also the recognition faster, making it more practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is an implementation flow chart of the present invention.

[0029] Figure 2 It is a schematic diagram of optical fiber detection of mining signals of the present invention.

[0030] Figure 3 It is a diagram combining the time domain and the space domain of the mining signal of the present invention.

[0031] Figure 4 It is a background signal diagram of 12 adjacent optical fiber nodes of the present invention.

[0032] Figure 5 This is a comparison chart of the accuracy of two different background signal processing methods during training according to the present invention.

[0033] Figure 6 This is a comparison chart of the six-category accuracy after training using two different background signal processing methods of the present invention.

[0034] Figure 7 It is a schematic diagram of the conversion into Mel frequency cepstral coefficients of the present invention.

[0035] Figure 8 It is a neural network structure diagram of the present invention.

[0036] Figure 9 This is a comparison chart of training time under different parameter settings of the present invention. DETAILED DESCRIPTION

[0037] The technical solutions of the present application will be further described below with reference to the drawings.

[0038] A Φ-OTDR signal fast training recognition algorithm based on space-time domain combination, MFCC and pre-extraction technology, the technical solutions include the following steps:

[0039] S1: First, combine the Φ-OTDR signal dataset in time domain and space domain.

[0040] In this embodiment, step S1 can be realized by the following way:

[0041] S11: The dataset we use has 6 categories, which are background sound (background), digging sound (dig), knocking sound (knock), shaking sound (shake), walking sound (walk), and the ratio of training set and validation set is 4:1. The dataset has a total of 15612 files, and the file format is.mat format. After parsing, each file is a matrix with a shape of 10000x12, where 10000 is the sampling frequency, representing the information in the time domain, and 12 is the number of adjacent optical fiber nodes, which are represented by the dashed line in Figure 2 , representing the information in the spatial domain.

[0042] S12: When an event signal occurs, multiple nodes in the optical fiber may be able to sense it, with the nodes closer to the signal being stronger and the nodes farther away being weaker. Therefore, for event signals such as digging sound (dig), knocking sound (knock), shaking sound (shake), and walking sound (walk), we add and average the signals of the 12 adjacent nodes, as shown in Figure 3 . This can avoid excessive interference caused by a single node error and has stronger anti-interference ability, while introducing information in the spatial domain. For background signals (background), as shown in Figure 4 , the sum and average of different node background signals are inconsistent, and the sum and average will interfere with the recognition. At this time, we directly split the 10000x12 matrix into 12 10000 sequences. We also tested the accuracy of background signals in the sum and average and split two ways in the experiment. In Figure 5 , the straight line represents splitting, and the dashed line represents sum and average, and the accuracy of the splitting method is higher during training. In Figure 6 , the shaded rectangle represents sum and average, and the blank rectangle represents splitting, 0-5 represents 6 event categories, and overall, the splitting method has higher accuracy, and the recognition accuracy of the background signal (category 0) reaches 100%.

[0043] S2: Secondly, calculate the Mel Frequency Cepstrum Coefficient (MFCC).

[0044] In this embodiment, step S2 can be implemented in the following manner:

[0045] S21: Mel-frequency cepstral coefficients have been widely used in speech recognition due to their effectiveness. Distributed acoustic sensors are essentially sound recognition, and the signals they recognize are similar to human speech signals, which is why they are used. The main steps for calculating Mel-frequency cepstral coefficients are as follows: ① Pre-emphasis; ② Framing; ③ Windowing; ④ Fast Fourier Transform (FFT); ⑤ Mel filter bank; ⑥ Logarithmic operation; ⑦ Discrete Cosine Transform (DCT).

[0046] S22: To find the Mel frequency cepstral coefficient, we use the feature.mfcc function in the Librosa toolkit. The code is: mfcc = librosa.feature.mfcc(y = k, sr = 10000), where k is the sequence information and 10000 is the sampling rate. After solving, the 10000 data is transformed into a 20×20 matrix data containing rich feature signals, as shown in the figure. Figure 7 This reduces the amount of data required for neural network training and increases the information feature density.

[0047] S3: Then, the Mel-frequency cepstral coefficients are pre-extracted and stored.

[0048] In this embodiment, step S3 can be implemented in the following manner:

[0049] S31: Extracting Mel-frequency cepstral coefficients while training will increase training time. Extracting all files and saving them in npy format before training can reduce training time. Saving them in other formats is also acceptable as long as they can be imported directly during training.

[0050] S32: Use the os.walk() function to traverse the files in the path, convert them into Mel-frequency cepstral coefficients one by one, and finally use the save function in the numpy toolkit to save them to a specific folder. Later, during training and testing, the data can be directly imported using the load function in the numpy toolkit.

[0051] S4: Finally, during training, the pre-stored Mel frequency cepstral coefficients are imported for training.

[0052] In this embodiment, step S4 can be implemented in the following manner:

[0053] S41: Considering the rich feature information of Mel-frequency cepstral coefficients, we decided to adopt a streamlined convolutional neural network, which not only makes the model faster to train, but also faster to recognize, making it more practical.

[0054] S42: The convolutional neural network mainly consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The network structure is generally as follows: in order, after the input layer, it first passes through 2 convolutional layers and 1 pooling layer, then passes through 2 convolutional layers and 1 pooling layer, then uses the flatten function to flatten, then enters the fully connected layer, and finally enters the output layer. Functions such as ReLu and Dropout are also used. The specific structure of the convolutional neural network is as follows Figure 8 shown.

[0055] S43: Import the dataset through the load function under the numpy toolkit, send it to the network structure, define the number of iterations, optimization method, loss function, batch size, learning rate, num_workers and other parameters, and then you can start training. If the computer CPU and GPU performance is good, the training batch size and num_workers can be set larger. After testing, if the setting is larger, the training speed will be faster when the accuracy does not change much. Of course, this is when the computer performance is acceptable. Too large a value will also cause the program to report an error. Figure 9 As shown in the figure, we set the training batch size from 64 to 128 and num_workers from 4 to 8. The training time for 50 rounds was shortened from 1.64 minutes to 1.36 minutes, and the training speed was increased by about 1.2 times.

[0056] It should be noted that, depending on the needs of the method implementation, the various steps described in this application can be split into more steps, or two or more steps or parts of steps can be combined into new steps to achieve the objectives of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A Φ-OTDR signal fast training and recognition method based on the combination of time and space domains, MFCC and pre-extraction technology, characterized in that: include: S1: Combine the time domain and space domain of the Φ-OTDR signal dataset; S2: Find the Mel frequency cepstral coefficient; S3: pre-extract the Mel frequency cepstral coefficients and store them; S4: During training, import the pre-stored Mel frequency cepstral coefficients for training; Said S1 comprises: S11: The dataset used is classified into background sound, digging sound, knocking sound, vibration sound, and walking sound, with a training set and validation set ratio of 4:

1. The dataset contains 15,612 files in .mat format. After parsing, each file is a 10,000 × 12 matrix, where 10,000 is the sampling frequency, representing information in the time domain, and 12 is 12 adjacent fiber nodes, representing information in the spatial domain. S12: When an event signal occurs, for event signals such as digging, knocking, vibration, and walking, the signals of the 12 adjacent nodes are summed and averaged. For background signals, since the background signal strength of different nodes is inconsistent, the 10,000 × 12 matrix is ​​directly split into 12 sequences of 10,000. The S2 includes: S21: The steps for obtaining the Mel frequency cepstrum coefficient are as follows: ① pre-emphasis; ② framing; ③ windowing; ④ fast Fourier transform; ⑤ Mel filter bank; ⑥ logarithmic operation; ⑦ discrete cosine transform; S22: Calculate the Mel frequency cepstral coefficient using the feature.mfcc function in the Librosa toolkit. The code is: mfcc = librosa.feature.mfcc, where k is the sequence information and 10,000 is the sampling rate. After solving, the 10,000 data points are converted into a 20 × 20 matrix containing rich feature signals. The S4 includes: S41: Considering the rich feature information of Mel-frequency cepstral coefficients, we decided to adopt a streamlined convolutional neural network, which not only makes the model faster to train but also faster to recognize, making it more practical; S42: This convolutional neural network mainly consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The network structure is as follows: after the input layer, it first passes through two convolutional layers and one pooling layer, then passes through two more convolutional layers and one pooling layer, then flattens using the flatten function, then enters the fully connected layer, and finally enters the output layer. ReLu and Dropout functions are also used. S43: Import the dataset through the load function under the numpy toolkit, send it to the network structure, define the number of iterations, optimization method, loss function, batch size, learning rate and num_workers parameters, and perform training.

2. The method according to claim 1, characterized in that The S3 includes: S31: During training, if you extract the Mel-frequency cepstral coefficients while training, it will increase the training time. Extract all the files directly before training and save them in npy format; S32: Use the os.walk() function to traverse the files in the path, and then operate them one by one to convert them into Mel-frequency cepstral coefficients. Finally, use the save function under the numpy toolkit to save them to a specific folder. During subsequent training and testing, the load function under the numpy toolkit can be used to directly import data.

Images