Low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network

Abstract

The invention relates to a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network, and the method comprises the following steps: S1, adjusting the time domain and frequency domain sampling rates of a Brillouin optical time domain sensor in a model training stage, obtaining a low-resolution Brillouin frequency spectrum, and taking the low-resolution Brillouin frequency spectrum as a sample of a data training set; and meanwhile, changing an external physical quantity applying mode of the optical fiber to be measured, such as a heating or strain position and a temperature or strain amplitude, establishing quasi-continuous temperature or strain field distribution information under different parameter conditions, and taking the quasi-continuous temperature or strain field distribution information as a label of a data training set. The method breaks through the limitation of the time domain sampling rate on the spatial resolution, saves the data storage space, reduces the calculation time consumption of the temperature or strain field, and improves the demodulation precision of the physical quantity distribution information.

Description

Technical field:

[0001] The invention belongs to the technical field of optical fiber sensing, and in particular relates to a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network. Background technology:

[0002] Optical fiber has great practical value in fields such as communication and sensing due to its low loss, corrosion resistance and electromagnetic interference resistance. Distributed optical fiber sensing technology can use optical effects to measure the distribution of various physical quantities along the optical fiber, such as temperature, strain and vibration. At present, distributed optical fiber sensing systems mainly use three scattering effects in optical fibers to achieve distributed physical quantity measurement. Among them, distributed optical fiber sensing systems based on Rayleigh scattering are used for short-distance temperature measurement and optical fiber attenuation monitoring, and distributed optical fiber sensing systems based on Raman scattering have also been widely used due to their high sensitivity to temperature measurement. Finally, distributed optical fiber sensing systems based on Brillouin scattering have attracted widespread attention from researchers because of their high sensitivity to a variety of physical quantities.

[0003] When the incident light propagates in the optical fiber, the excited phonons interact with the photons, generating scattered light with a certain frequency difference from the incident light frequency, including Stokes light with a frequency downshift and Anti-Stokes light with a frequency upshift. Its power spectrum is approximately a Lorentz line shape, and the offset between the center frequency and the center frequency of the incident light is called the Brillouin frequency shift (BFS). When the physical quantity outside the optical fiber changes, the Brillouin frequency shift at each location will also change accordingly. By establishing a mapping relationship between the two, the distributed measurement of the physical quantity can be achieved. The mainstream methods are two methods based on time domain signals, one is Brillouin Optical Time-Domain Analysis (BOTDA), and the other is Brillouin Optical Time-Domain Reflectometry (BOTDR). Taking Brillouin optical time-domain reflectometry technology as an example, by scanning the power of scattered light at different frequency differences, a set of Brillouin scattering spectra can be obtained. After the Brillouin frequency shift of the Brillouin scattering spectrum at each optical fiber position is extracted and matched with the physical quantity, a single physical quantity measurement is completed.

[0004] Brillouin Optical Time-Domain Sensor (BOTDS) has excellent sensing accuracy, which enables it to accurately locate temperature / strain events along the optical fiber. This feature is critical in many application scenarios, such as manufacturing process monitoring and aircraft structure monitoring, where damaged modules must be accurately located and repaired to prevent safety accidents. For these systems, BOTDS needs to measure the edge, length and amplitude of the event with extremely high accuracy and speed.

[0005] Traditional Brillouin optical time-domain sensing systems need to perform frequency scanning to form a two-dimensional Brillouin frequency spectrum with high frequency resolution, so they need to perform repeated measurements of different scattering frequencies, which causes the measurement time to increase exponentially; at the same time, the high-precision demodulation requirements of the physical quantity field also require the sensing system to have an extremely high time-domain sampling rate, which will reduce the data processing speed, damage the real-time nature of data processing, and increase the cost of hardware equipment. Existing post-processing methods, such as deconvolution-based recovery technology, are used to improve spatial resolution and avoid performance degradation caused by pulse width. However, these traditional methods, including those based on convolutional neural network technology, cannot overcome the accuracy limitations caused by the system sampling rate. For example, when the transition zone length of the monitoring event is 1 cm, and the distance between adjacent points in the sampled data is 50 cm, the highest spatial resolution that can be achieved by the traditional method is limited to 50 cm. This limitation leads to a significant decrease in the spatial accuracy of physical quantity demodulation, which is caused by the inherent discreteness of the sampled data and is difficult to overcome by existing traditional methods. Summary of the invention:

[0006] The technical problem to be solved by the present invention is to provide a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network, which obtains the Brillouin frequency spectrum through a lower frequency domain sampling rate, thereby avoiding the increase in measurement time caused by a large number of frequency sweeps, and adopts a lower time domain sampling rate to reduce the amount of data to be processed and alleviate storage pressure. Subsequently, an artificial neural network (ANN) is used to establish a mapping relationship between a discrete low-resolution Brillouin frequency spectrum and a quasi-continuous distribution field of temperature or strain. Finally, in the measurement stage, the downsampled sensor data is input, and the quasi-continuous distribution information of the physical quantity field to be measured is output.

[0007] The technical solution of the present invention is to provide a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network, comprising the following steps:

[0008] S1. In the model training stage, the time domain and frequency domain sampling rates of the Brillouin optical time domain sensor are adjusted to obtain a low-resolution Brillouin frequency spectrum, which is used as a sample of the data training set; at the same time, the external physical quantity application method of the optical fiber to be tested is changed, such as the heating or strain position and the temperature or strain amplitude, to establish quasi-continuous temperature or strain field distribution information under different parameter conditions, which is used as the label of the data training set, and these distribution information are associated with the corresponding low-resolution Brillouin frequency spectrum to construct a low-resolution Brillouin frequency spectrum-temperature or strain quasi-continuous distribution field sample data set;

[0009] S2. Build an artificial neural network model and use a low-resolution Brillouin frequency spectrum-temperature or strain quasi-continuous distribution field sample data set to train the parameters of the artificial neural network model; by continuously adjusting the model parameters, it can accurately predict the corresponding temperature or strain distribution from the low-resolution Brillouin frequency spectrum;

[0010] S3. In the measurement phase, the time domain and frequency domain sampling rates of the Brillouin optical time domain sensor are reduced to obtain the Brillouin time domain traces at each sensing position at different scattering frequencies of the optical fiber to be measured, and form a low-resolution Brillouin frequency spectrum from them; wherein the frequency domain sampling rate reduction can be achieved by sampling at lower equal frequency intervals or unequal frequency intervals; the time domain sampling rate is achieved by sampling at lower equal time intervals;

[0011] S4. Input the low-resolution sensor data into the trained artificial neural network model to obtain a high-precision quasi-continuous distribution temperature or strain field.

[0012] The present invention only adopts a Brillouin frequency spectrum with a relatively low sampling rate in the Brillouin optical time domain sensor, and utilizes a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network to build an artificial neural network model without introducing additional hardware equipment or increasing the system sampling rate. In addition, the present invention utilizes a "low-resolution Brillouin frequency spectrum-temperature or strain quasi-continuous distribution field" sample data set to break through the limitation of the time domain sampling rate on the spatial resolution, save data storage space, reduce the calculation time of the temperature or strain field, and improve the demodulation accuracy of the physical quantity distribution information.

[0013] Preferably, in step S1, the Brillouin optical time domain sensor may be a Brillouin optical time domain analyzer or a Brillouin optical time domain reflectometer.

[0014] Preferably, in step S1, the parameter conditions include the heating or strain position of the optical fiber to be tested, the temperature or strain amplitude, the line width of the Brillouin spectrum, the detection pulse width, the detector bandwidth and the amplitude signal-to-noise ratio of the data.

[0015] Preferably, in step S1, the sample is low-resolution data corresponding to the Brillouin spectrum.

[0016] Preferably, in step S1, the label is data corresponding to a quasi-continuous distribution field of temperature or strain.

[0017] Preferably, in step S2, the artificial neural network model includes an input layer, a hidden layer and an output layer, the input layer has nodes corresponding to low-resolution Brillouin frequency spectrum data, and the output layer has nodes corresponding to high-precision temperature or strain field distribution; wherein the hidden layer can be any type of network layer, including but not limited to a long and short-term memory network (Long and Short-Term Memory, LSTM) layer, a fully connected layer and a convolutional layer; during training, the sample data set is divided into a training set and a test set, the model is trained using the training set and the model performance is tested using the test set.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] The present invention only adopts a Brillouin frequency spectrum with a relatively low sampling rate in the Brillouin optical time domain sensor, and utilizes a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network. Without introducing additional hardware equipment or increasing the system sampling rate, the present invention breaks through the limitation of the time domain sampling rate on the spatial resolution, saves data storage space, reduces the calculation time of the temperature or strain field, and improves the demodulation accuracy of the physical quantity distribution information. Description of the drawings:

[0020] Figure 1 It is a schematic diagram of the process of the present invention;

[0021] Figure 2 is the Brillouin frequency spectrum obtained using a lower sampling rate;

[0022] Figure 3 This is the starting position, length and amplitude information of the event segment after extraction in the embodiment of the present invention. Specific implementation method:

[0023] The present invention will be further described below with reference to the accompanying drawings:

[0024] The present invention provides a low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network, such as Figure 1 As shown, the present invention performs low-resolution Brillouin frequency spectrum demodulation through the following steps:

[0025] 1) Set the sampling range of the Brillouin optical time domain sensor to 0-200MHz, use a larger 20MHz interval in the scanning frequency dimension, and include a total of 11 frequency scanning points; set the time domain sampling rate to 200MHz, set the local optical fiber length to be measured to 25 meters, and include a total of 51 sampling points, where the interval between points is 0.5 meters; the starting point position of the temperature event is set at intervals of 0.01 meters, from 7.5 meters to 12.5 meters, a total of 501 groups; the event length is set at intervals of 0.01 meters, ranging from 0.5 meters to the maximum value that can be set for each group of data. And use the obtained discrete low-resolution Brillouin frequency spectrum data as samples of the data training set. The temperature value is set at intervals of 10℃, with a total of 7 groups from 20℃ to 80℃; the detection pulse width is set at intervals of 10ns, with a total of 4 groups from 20ns to 50ns; the amplitude signal-to-noise ratio is set at intervals of 5dBm, with a total of 5 groups from 10dBm to 30dBm; and the detector bandwidth value is set at intervals of 10MHz, with a total of 18 groups from 30MHz to 200MHz. The low-resolution Brillouin frequency spectrum under different parameter settings is matched with the position and amplitude distribution of the temperature field to construct a training data set.

[0026] 2) The training data set was divided into training set, test set and validation set in a ratio of 8:1:1, and an artificial neural network model was established, including 1 input layer, 1 LSTM layer, 2 fully connected layers and 1 output layer. The number of nodes in each layer was 50, 128, 256, 512 and 501 respectively. The training of the neural network model was completed through the back propagation algorithm.

[0027] 3) Using a traditional Brillouin optical time-domain reflectometer, an average Brillouin frequency spectrum of 4096 times was obtained on a 1000-meter-long bend-insensitive (G.657) optical fiber. The incident pulse width was set to 20ns, corresponding to a theoretical spatial resolution of 2m. The photodetector measured 11 time domain traces at intervals of 20MHz within a sweep frequency range of 200MHz with a time domain sampling rate of 200MSample / s, and collected 2001 sampling points at each frequency point to be measured. A fiber section heated to 40°C was introduced, and its heating length was 1.2 meters. The Brillouin frequency spectrum obtained by low-resolution sampling is shown below. Figure 2 shown.

[0028] 4) Through the artificial neural network trained in step 2), the quasi-continuous distribution result of the temperature field can be obtained from the Brillouin frequency spectrum data obtained in step 3), and the demodulation result is as follows: Figure 3As shown. After using the low-resolution Brillouin frequency spectrum demodulation method of the present invention, the recognition accuracy of the temperature event segment is extremely high. When the pulse width is 20ns, the measurement uncertainty of the starting point of the temperature event segment is 5.6 cm, the measurement uncertainty of the length of the temperature event segment is 4.1 cm, the measured length of the temperature event segment is 115.7 cm, and the temperature measurement error is 0.3°C, which shows that the present invention can demodulate the quasi-continuous distribution information of the temperature event segment with extremely high accuracy. Compared with the high-resolution Brillouin frequency spectrum obtained with a frequency sampling interval of 1MHz and a time domain sampling rate of 2GHz, the present invention can reduce the measurement time by 20 times, reduce the storage space to 10%, and greatly reduce the hardware cost of the acquisition equipment.

[0029] The present invention can obtain the Brillouin frequency spectrum through a relatively low sampling rate, thereby avoiding the increase in measurement time caused by a large number of frequency sweeps, and obtains high-precision quasi-continuous physical quantity field distribution through a low-resolution Brillouin frequency spectrum demodulation method based on an artificial neural network. Without introducing additional hardware equipment or increasing the system sampling rate, the limitation of the time domain sampling rate on the spatial resolution is broken through, data storage space is saved, the calculation time of the temperature or strain field is reduced, and the demodulation accuracy of the distribution information of the temperature or strain field is improved.

[0030] The above is only an explanation of the preferred embodiments of the present invention, which should not be construed as a limitation on the claims. Any equivalent process changes made using the present invention specification are included in the patent protection scope of the present invention.

Claims

1. A low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network, characterized in that: The following steps are included: S1. In the model training stage, the time domain and frequency domain sampling rates of the Brillouin optical time domain sensor are adjusted to obtain a low-resolution Brillouin frequency spectrum, which is used as a sample of the data training set; at the same time, the heating or strain position and the temperature or strain amplitude of the optical fiber to be tested are changed to establish quasi-continuous temperature or strain field distribution information under different parameter conditions, which is used as a label of the data training set, and these distribution information are associated with the corresponding low-resolution Brillouin frequency spectrum to construct a low-resolution Brillouin frequency spectrum-temperature or strain quasi-continuous distribution field sample data set; S2. Build an artificial neural network model and use a low-resolution Brillouin frequency spectrum-temperature or strain quasi-continuous distribution field sample data set to train the parameters of the artificial neural network model; S3. In the measurement phase, the time domain and frequency domain sampling rates of the Brillouin optical time domain sensor are reduced to obtain the Brillouin time domain traces at each sensing position at different scattering frequencies of the optical fiber to be measured, and form a low-resolution Brillouin frequency spectrum; S4. Input the low-resolution sensor data into the trained artificial neural network model to obtain a high-precision quasi-continuous distribution temperature or strain field.

2. The low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network according to claim 1, characterized in that: In step S1, the Brillouin optical time domain sensor may be a Brillouin optical time domain analyzer or a Brillouin optical time domain reflectometer.

3. The low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network according to claim 1, characterized in that: In step S1, the parameter conditions include the heating or strain position of the optical fiber to be tested, the temperature or strain amplitude, the line width of the Brillouin spectrum, the detection pulse width, the detector bandwidth and the amplitude signal-to-noise ratio of the data.

4. The low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network according to claim 1, characterized in that: In step S1, the sample is low-resolution data corresponding to the Brillouin spectrum.

5. The low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network according to claim 1, characterized in that: In step S1, the label is data corresponding to the quasi-continuous distribution field of temperature or strain.

6. The low-resolution Brillouin frequency spectrum demodulation method based on artificial neural network according to claim 1, characterized in that: In step S2, the artificial neural network model includes an input layer, a hidden layer and an output layer, the input layer has nodes corresponding to low-resolution Brillouin frequency spectrum data, and the output layer has nodes corresponding to high-precision temperature or strain field distribution; wherein the hidden layer can be any type of network layer, including but not limited to long short-term memory network layer, fully connected layer and convolutional layer; during training, the sample data set is divided into a training set and a test set, the model is trained through the training set and the model performance is tested using the test set.

Images