Distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network

Abstract

The invention discloses a distributed optical fiber sensing multi-point positioning method based on weight redistribution and a convolutional neural network, and adopts different training and test position selection schemes and data acquisition quantities for different numbers of disturbance points. The number of samples at each training position is reduced along with the increase of disturbance points, and the number of samples at each test position is the same. Conventional preprocessing and weight redistribution are carried out on the collected time domain sensing signals; randomly selecting a certain amount of processed sample data from each training position to respectively form a training set and a verification set, wherein the sample number proportion of the two data sets is reduced along with the increase of disturbance points; and inputting the training samples into the convolutional neural network regression model, completing training of different numbers of disturbance position prediction models, and performing parameter optimization by using the verification set. And testing a plurality of models on the test sample, and selecting the test result with the minimum error as the disturbance position of the sample. According to the invention, accurate positioning of multi-point disturbance can be realized by using a relatively small amount of data.

Description

Technical Field

[0001] The present invention relates to a multi-point positioning method for distributed optical fiber sensing, in particular to a multi-point positioning method for distributed optical fiber sensing based on weight redistribution and convolutional neural network. Background Art

[0002] Distributed optical fiber sensing systems use optical fibers as sensing media and utilize the changes in the optical signals propagating in the optical fibers to reflect the changes in external physical quantities, thus having been rapidly developed in the fields of pipeline leakage monitoring, large-scale structural health monitoring, etc. Among them, the Sagnac distributed optical fiber sensing system can use two beams of light with zero optical path difference for interference, has strong anti-interference ability and low requirements for light sources, and is one of the current research hotspots. Traditional zero-frequency method and second Fourier transform (FFT) have relatively high requirements for the signal-to-noise ratio of signals. If there is a large amount of noise in the signal, the acquisition of zero frequency and the calculation of second FFT will become difficult, resulting in a large positioning error and making it more difficult to achieve the positioning of multiple leakage points.

[0003] In recent years, with the rapid development of machine learning technology, researchers have begun to combine machine learning methods with traditional positioning methods to improve the accuracy and robustness of positioning and achieve multi-point positioning using machine learning algorithms. Using classification algorithms, the positioning problem can be transformed into a multi-classification problem of sensing signals caused by disturbances at different positions of the sensing optical fiber. The positioning resolution depends on the length of the selected optical fiber segment. Due to the difficulty of data acquisition, the length of the optical fiber segment is generally taken to be relatively long, resulting in low positioning accuracy. The regression algorithm can achieve accurate prediction of any disturbance position, ensuring the continuous monitoring of the distributed optical fiber sensing system, but there are problems such as the need to collect a large amount of data for training the model, which is relatively difficult in practical applications. Summary of the Invention

[0004] To solve the problems of the prior art, the present invention provides a multi-point positioning method for distributed optical fiber sensing based on weight redistribution and convolutional neural network. By adopting different training and test position selection and data acquisition schemes for different numbers of disturbance points, the workload of data acquisition is reduced; using the time-domain signal as the input, the weight redistribution network is used to redistribute the weights of the signal to extract data features. The training samples are input into the convolutional neural network regression model to complete the training of the prediction models for different numbers of disturbance positions, and the validation set is used for parameter optimization. The test samples are tested with multiple models, and the test result with the smallest error is selected as the disturbance position of the sample, which can achieve accurate positioning of multi-point disturbances with relatively less data volume.

[0005] To achieve the above object of the invention-creation, the present invention adopts the following technical solutions:

[0006] Distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network, comprising the following steps:

[0007] 1) Select the positions for data acquisition, and take single-point, two-point, and more than three-point disturbances respectively.

[0008] 2) Collect the time-domain signals output by the sensing system multiple times at each selected position. The number of samples at each training position decreases with the increase in the number of disturbance points, while the sampling rate is correspondingly increased to ensure the positioning accuracy while reducing the data acquisition volume. The number of samples at each test position is the same.

[0009] 3) Perform conventional preprocessing on the collected data.

[0010] 4) Re-distribute the weights of the preprocessed time-domain input signals through the weight redistribution network to extract the features of the sensing signals, so as to improve the positioning accuracy.

[0011] 5) Randomly select a certain number of processed sample data from each training position to form a training set and a validation set respectively. The ratio of the number of samples in the training set and the validation set decreases with the increase in the number of disturbance points. The processed sample data at the test position is used as the test set.

[0012] 6) Input the training set samples of single-point, two-point, and more than three-point disturbances into the convolutional neural network regression model respectively to complete the training of multiple disturbance position prediction models; use the validation set for parameter optimization to obtain the best training effect.

[0013] 7) Input the test samples into the trained disturbance position prediction models respectively to obtain multiple prediction results, and select the result with the smallest error as the disturbance position corresponding to the test sample to achieve multi-point disturbance positioning.

[0014] Preferably, in the step 1), for single-point disturbance, select the training positions at a fixed interval with the maximum acceptable positioning error, and randomly select the same number of other sensing positions as the test positions; for two-point disturbance, appropriately increase the fixed interval to d meters, and use all possible positions of the two-point disturbance as the training positions, and use two new positions with intervals smaller than the minimum interval of the training positions, the same as the training positions, and random position intervals as the test positions, and the number is appropriately reduced compared with the training positions. For disturbances of more than three points, evenly divide the sensing optical fiber into several segments, select part of the optical fiber in each segment as the data acquisition interval, further divide it at an interval of d meters, and use all possible positions of multiple disturbances as the training positions; use the new positions within the position interval involved in the training positions as the test positions. Similarly, the number of test positions is appropriately reduced compared with the training positions.

[0015] Preferably, the method for selecting the data acquisition interval in the step 1) is: uniformly select half of the optical fiber section far from or close to the sensing optical fiber end as the data acquisition interval.

[0016] Preferably, in the step 3), at least one of low-pass filtering, downsampling, and normalization methods is used for conventional preprocessing.

[0017] Preferably, the weight redistribution network structure in the step 4) is one global average pooling layer and two fully connected layers.

[0018] Preferably, in the step 4), first, global average pooling is used to extract the global features of the input signal, then two fully connected networks are used to generate normalized attention weights, and the weights are multiplied by the input signal to obtain the sensing signal with redistributed weights.

[0019] Preferably, in the step 6), during the regression training process, mean squared error is selected as the loss function; the adaptive moment estimation Adam optimization algorithm is used to optimize the model parameters to minimize the loss function value.

[0020] Preferably, in the step 7), the test data of single-point, two-point, and three-point perturbations are respectively predicted by three trained perturbation position prediction models to obtain multiple prediction results. The magnitudes of the prediction error values of each model are compared, and the prediction result with a smaller error is selected as the final multi-point perturbation positioning result.

[0021] Preferably, in the step 7), when using different models for prediction, there may be a situation where null values appear in the prediction results due to the inconsistency between the number of perturbation positions of the test samples and the number of model outputs, that is, the model mismatch, resulting in the inability to perform position prediction. In this case, these prediction results are directly discarded, and only the prediction results of the models without null values, that is, the errors of the positioning results where the prediction model can work normally, are compared.

[0022] Compared with the prior art, the present invention has the following obvious prominent substantial features and remarkable advantages:

[0023] 1. Using the time-domain signal as the input, without the need for other transformations, the signal processing is simple and fast;

[0024] 2. By adding a weight redistribution network to realize the extraction of the time-domain waveform features of the sensing signal, the positioning result is more accurate compared to only using a convolutional neural network;

[0025] 3. Different training and test position selection and data acquisition schemes are adopted for different numbers of disturbance points. Especially when there are multiple points of disturbance, after segmenting the sensing optical fiber, some position intervals are selected for data acquisition at large intervals and high sampling rates, reducing the amount of data acquisition and still being able to maintain a high positioning accuracy;

[0026] 4. It has a wide range of applications and can be used for disturbance positioning in distributed optical fiber sensing systems with ring or linear Sagnac or other reflective structures. As long as there is more than one waveform change in the time-domain waveform output by the system and the time difference of the waveform changes is related to the disturbance position, this method can be used for positioning in distributed optical fiber sensing systems. Brief Description of the Drawings

[0027] Figure 1 It is a flow chart of the positioning method of the present invention.

[0028] Figure 2 It is a schematic diagram of the system structure adopted in the embodiment.

[0029] Figure 3 It is the time-domain waveform of the sensing signal collected at different leakage positions.

[0030] Figure 4 It is the weight redistribution network structure used in the embodiment.

[0031] Figure 5 It is the convolutional neural network structure used in the embodiment. Detailed Description of the Invention

[0032] The above scheme is further described below in conjunction with specific implementation examples. The preferred implementation examples of the present invention are described in detail as follows:

[0033] Example 1:

[0034] In this embodiment, referring to Figure 1 , a distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network includes the following steps:

[0035] 1) Select the positions for data acquisition, and select single-point, two-point, and more than three-point disturbances respectively for data acquisition;

[0036] For single-point disturbance, select the training positions at a fixed interval with the maximum positioning error that can be accepted, and randomly select the same number of other sensing positions as the test positions;

[0037] For two-point disturbance, appropriately increase the fixed interval to d meters, use all possible positions where two-point disturbance may occur as the training positions, and use two new positions with intervals smaller than the minimum interval of the training positions, the same as the training positions, and random position intervals as the test positions, and the number is appropriately reduced compared to the training positions;

[0038] For perturbations of more than three points, the sensing optical fiber is evenly divided into several segments, and a part of the optical fiber in each segment is selected as the data acquisition interval, which is further divided at intervals of d meters. All positions where multiple perturbations may occur are used as training positions;

[0039] New positions within the position intervals involved in the training positions are used as test positions. Similarly, the number of test positions is appropriately reduced compared to the training positions;

[0040] 2) The time-domain signals output by the sensing system are collected multiple times at each selected position. The number of samples at each training position decreases as the number of perturbation points increases, while the sampling rate is correspondingly increased to ensure the positioning accuracy while reducing the data acquisition volume. The number of samples at each test position is the same;

[0041] 3) Perform conventional preprocessing on the collected data;

[0042] 4) Re-distribute the weights of the preprocessed time-domain input signals through a weight re-distribution network to extract the characteristics of the sensing signals and improve the positioning accuracy;

[0043] 5) Randomly select a certain number of processed sample data from each training position to form a training set and a validation set respectively. The ratio of the number of samples in the training set and the validation set decreases as the number of perturbation points increases; The processed sample data at the test positions are used as the test set;

[0044] 6) Input the training set samples of single-point, two-point, and more than three-point perturbations into the convolutional neural network regression model respectively to complete the training of multiple perturbation position prediction models; Use the validation set to optimize the parameters to obtain the best training effect;

[0045] 7) Input the test samples into the trained various perturbation position prediction models respectively to obtain multiple prediction results, and select the result with the smallest error as the perturbation position corresponding to the test sample to achieve multi-point perturbation positioning.

[0046] The distributed optical fiber sensing and positioning method based on weight re-distribution and convolutional neural network in this embodiment uses the time-domain signal as the input without the need for other transformations, and the signal processing is simple and fast; Extract features through the weight re-distribution network, and different training and test position selection and data acquisition schemes are adopted for different numbers of perturbation points. Especially in the case of multi-point perturbations, the sensing optical fiber is segmented, and part of the position intervals are selected for data acquisition. The positioning can still be accurate and stable with relatively small data volume; It can be used for perturbation positioning of distributed optical fiber sensing systems with ring or linear Sagnac or other reflective structures. As long as there is more than one waveform change in the time-domain waveform output by the system, and the time difference of the waveform changes is related to the perturbation position, this method can be used for positioning.

[0047] Example 2:

[0048] This example is basically the same as Example 1, with the special feature being that:

[0049] In this example, the method for selecting the data acquisition interval in step 1) is: uniformly select half of the optical fiber segment far from or close to the sensing fiber end as the data acquisition interval.

[0050] In this example, in step 3), at least one of low-pass filtering, downsampling, and normalization methods is used for the conventional preprocessing.

[0051] In this example, in step 4), the weight redistribution network structure is one layer of global average pooling layer and two layers of fully connected layers.

[0052] This example can locate multiple-point disturbances at any position on the sensing fiber, and the location process is simple, easy to implement, and efficient.

[0053] Example 3:

[0054] This example is basically the same as the above examples, with the special feature being that:

[0055] In this example, in step 4), first, global average pooling is used to extract the global features of the input signal, then two layers of fully connected networks are used to generate normalized attention weights, and the weights are multiplied by the input signal to obtain the sensing signal with redistributed weights.

[0056] In step 6), during the regression training process, mean squared error is selected as the loss function; the Adam optimization algorithm for adaptive moment estimation is used to optimize the model parameters to minimize the loss function value.

[0057] In step 7), the test data of single-point, two-point, and three-point disturbances are respectively predicted through three trained disturbance position prediction models to obtain multiple prediction results. Compare the magnitudes of the prediction error values of each model, and select the prediction result with a smaller error as the final multiple-point disturbance location result.

[0058] In step 7), when using different models for prediction, there will be a situation where null values appear in the prediction results due to the inconsistency between the number of disturbance positions in the test samples and the number of model outputs, that is, the model does not match, resulting in the inability to perform position prediction. For this situation, directly discard these prediction results and only compare the prediction results of the models without null values, that is, the errors of the location results where the prediction model can work properly.

[0059] This embodiment is based on the distributed fiber optic sensing multi-point positioning method of weight redistribution and convolutional neural network. Different training and test position selection and data collection schemes are adopted for different numbers of disturbance points to reduce the workload of data collection. Time domain signals are used as input, and the weight redistribution network is used to redistribute the weights of the signals to extract data features. The training samples are input into the convolutional neural network regression model to complete the training of the prediction model of different numbers of disturbance positions, and the validation set is used for parameter optimization. Multiple models are tested on the test samples, and the test result with the smallest error is selected as the disturbance position of the sample. This can achieve accurate positioning of multi-point disturbances with relatively small amounts of data.

[0060] Embodiment 4:

[0061] This embodiment is basically the same as the above embodiment, with the following features:

[0062] In this embodiment, a linear Sagnac distributed optical fiber sensing system is selected and used for pipeline leakage monitoring. Theoretically, leakage at more than three points may occur, but the probability is extremely low. Therefore, this embodiment only realizes the positioning of single-point, two-point, and three-point leakage. The sensing system is simulated using OptiSystem software to verify the feasibility of the distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network in this embodiment.

[0063] like Figure 2 As shown, the simulated linear Sagnac distributed optical fiber sensing system includes a continuous laser 1, a 2×2 bidirectional 3dB optical coupler 2, a delay fiber 3, a 2×2 bidirectional 3dB optical coupler 4, a first sensing fiber 5, a first phase modulator 6, a second sensing fiber 7, a second phase modulator 8, a third sensing fiber 9, a third phase modulator 10, a fourth sensing fiber 11, a Faraday rotation mirror 12, and a photodetector 13. The sum of the lengths of the first sensing fiber 5, the second sensing fiber 7, the third sensing fiber 9, and the fourth sensing fiber 11 is equal to the total length of the sensing fiber. Since the bandwidth of the pipeline leakage signal is about 60kHz, and the energy is mainly concentrated at the moment when the leakage occurs, the phase modulator is driven by a Sinc function with a bandwidth of 62kHz and a duration of 20μs to simulate the disturbance caused by different numbers of leaks at different positions in the pipeline to the sensing fiber laid thereon.

[0064] In this embodiment, the length of the delay optical fiber 3 is 1 km. On the premise of keeping the total length of the sensing optical fiber unchanged at 40 km, by changing the lengths of different segments of the sensing optical fiber, different leakage positions are simulated. The phase change amounts of the first phase modulator 6 and the second phase modulator 8 are set to 0°, and the phase change amount of the third phase modulator 10 is set to 5° to simulate the action of a single leakage on the optical fiber. Similarly, in the case of two-point leakage, the phase change amount of the first phase modulator 6 is set to 0°, and the phase change amounts of the second phase modulator 8 and the third phase modulator 10 are set to 5°. For three-point leakage, the change amounts of the three phase modulators are all 5°.

[0065] According to the process as Figure 1 shown to locate the leakage.

[0066] For the case of single-point leakage, assuming that the maximum acceptable error is 100 m, the 40-km sensing optical fiber is divided at this interval to obtain 401 training positions. In addition, 400 new positions at non-training positions are randomly selected on the sensing optical fiber as the leakage points to be located, that is, the test positions.

[0067] For the case of two-point leakage, for two-point perturbations, the fixed interval is appropriately increased to 400 m to obtain 101 positions. The training positions are all the positions where two-point leakage may occur among these 101 positions. The specific number is the combination number of selecting 2 different positions from 101 positions, that is, 5050. 100 groups of new positions are respectively selected at intervals of 200 - 400 m, 400 - 450 m, and random position intervals, totaling 300 groups, as the test positions.

[0068] For the case of three-point leakage, according to the distances from the leakage points to the mirror, the positions of the three leakage points are called P1, P2, and P3 in order from far to near. First, the sensing optical fiber is divided into 4 segments at an interval of 10 km, and then the part 4.8 km away from the Faraday rotator mirror at the end of the sensing optical fiber in each segment is selected as the position interval for collecting leakage signals. Each interval is divided at an interval of 400 m to obtain 13 position points. All the positions where the three leakage points may appear are used as the training positions. Consider three different combination cases:

[0069] (1) The three points are in the same interval. It may be any one of the 4 intervals. The number of positions in each interval is the combination number of selecting 3 different positions from 13 positions, totaling

[0070] (2) P1 and P2 are in the same interval farther from the Faraday rotator mirror, and P3 is in another closer interval. At this time, there are 6 different combinations of selectable intervals. The number of positions where two leakage points appear simultaneously in the same interval is the combination number of selecting 2 different positions from 13 positions, and the number of positions where the other leakage point appears is 13. Therefore, the total number of leakage positions is ;

[0071] (3) The three points are in different intervals respectively. There are 4 possible combinations of intervals where leakage may occur. The number of positions where a single leakage may occur in each interval is 13. In total, there are 13 * 13 * 13 * 4 = 8788.

[0072] Actually, in the second combination, there is another possibility, that is, P1 appears in the farther interval, and P2 and P3 are in the same closer interval. The position change rule of this combination is the same as that of the selected acquisition position combination, only the difference in position distance. Therefore, in this embodiment, only the positioning result of one of the combinations is considered for verification. A number of new positions within the position intervals involved in the training positions are used as test positions. 30 are selected for each position interval combination. The total number of test positions for the three cases is 120, 180, and 120.

[0073] The time-domain signals output by the sensing system are collected multiple times at each selected position. The number of samples of the training positions with single-point, two-point, and three-point leaks are 21, 11, and 4 respectively, and the number of samples of each test position is 3. The number of test positions is appropriately reduced compared to the training positions. In order to maintain a high positioning accuracy when the leakage point interval increases, for the two-point and three-point leakage cases, the sampling rate is increased during data collection, and downsampling is performed during subsequent preprocessing. Finally, the number of sampling points of the single-point, two-point, and three-point leakage signals is 4096.

[0074] The collected data is subjected to conventional preprocessing, mainly including at least one of removing the baseline, low-pass filtering, and normalization. Limited by the simulation software, there is always an unwanted noise baseline signal in the collected sensing signals. Therefore, the collected signals need to be subtracted from the output signals of the system without leakage first. After removing the baseline signal, there is still a large amount of high-frequency noise in the signal, and it needs to be low-pass filtered to obtain a pure sensing signal. Figure 3It is the preprocessed result obtained when three leakage signals act on different positions of the sensing optical fiber. It can be seen that a single leakage will cause waveform changes at two places at different times. The first place corresponds to the change in the interference light intensity generated after the two beams of light first pass through the leakage position, and the second place corresponds to the second change in the interference light intensity generated after the two beams of light are reflected and then pass through the leakage point again. Therefore, the time difference between the two waveform changes is the time for the light to travel back and forth from the leakage point to the Faraday rotator mirror. Therefore, a total of 6 waveform changes are generated by the three leakages. The earliest and latest waveform changes correspond to the farthest leakage points, and the closer the leakage point is to the Faraday rotator mirror, the smaller the time difference between the corresponding two waveform changes. Therefore, based on the corresponding relationship between this waveform change and the leakage position, the time-domain signal can be used for model training.

[0075] Since the data collected by the software simulation system is the same every time, although it is collected multiple times, in fact, there is only one pure noise-free data sample for each leakage position. In order to make the data samples collected at each position different and verify the noise tolerance of the model, noise is added to the preprocessed data samples at each position. For the data collected at the training positions, except for one noise-free sample, Gaussian white noise with different intensities is added to each training position sample of single-point leakage, so that the signal-to-noise ratio of the sample becomes 1 - 20 dB, and the change step is 1 dB, so as to obtain 21 sample data with different signal-to-noise ratios;

[0076] For two-point leakage, noise is added at ten different signal-to-noise ratios of 1 dB, 2 dB, 3 dB, 9 dB, 10 dB, 11 dB, 17 dB, 18 dB, 19 dB, and 20 dB respectively, so that 11 sample data with different signal-to-noise ratios are obtained for each training position;

[0077] For the case of three-point leakage, the signal-to-noise ratios of 5 dB, 15 dB, and 25 dB are selected, so that 4 sample data with different signal-to-noise ratios are obtained for each training position.

[0078] For the samples at the test positions, three signal-to-noise ratios of 5.5 dB, 15.5 dB, and 25.5 dB are selected to add noise, so that 3 noisy sample data with different signal-to-noise ratios are obtained for each test position.

[0079] The preprocessed data samples are input into a weight reallocation network to reallocate the weights of the signals, which plays a role in feature extraction. As Figure 4 shown, first, global average pooling is used to extract the global features of the input signal, and then a two-layer fully connected network is used to generate normalized attention weights. Multiply the weights by the input signal to obtain the sensing signal with reallocated weights.

[0080] For each training location with single-point, two-point, and three-point leaks, 16, 8, and 3 sample data after weight redistribution are randomly selected to form the training set, and the remaining data forms the validation set. The sample data after weight redistribution at the test location is used as the test set.

[0081] The training set samples of single-point, two-point, and three-point leaks are respectively input into the convolutional neural network regression model to complete the training of the single-point, two-point, and three-point leak location prediction models. The convolutional neural network regression model consists of three convolutional layers. As Figure 5 shown, each convolutional layer is followed by a pooling layer, and finally three fully connected layers are connected. The convolutional kernel size of the 3 convolutional layers is 3, and the stride is 1. During the single-point and multi-point leakage location training processes, the number of convolutional kernels is respectively selected as 1, 8, 16 and 1, 16, 32. The "ReLU" activation function is used in the convolutional layers, and the padding mode is "SAME". The window size and stride of the three pooling layers are both 2. In the three fully connected layers, the first two layers use the ReLU activation function to perform the regression task, and the last layer uses the linear activation function. For the single-point, two-point, and three-point location tasks, the number of neurons in the three fully connected layers is respectively selected as 1024, 128, 1, 4096, 1024, 2, and 4096, 1024, 3. During the regression training process, the mean squared error (MSE) is usually selected as the loss function. The value of MSE determines whether the model training converges and the optimization direction of the network. The expression of MSE is

[0082]

[0083] where n is the number of samples, y i and y p are the predicted location and the true location of the test sample, and i is the sample number.

[0084] The training of the location model is based on the TensorFlow neural network framework, with TensorFlow version 2.6.2 and Python version 3.7.0. The training is performed using the NVIDIA GeForce RTX 3060 GPU. During the training phase, the adaptive moment estimation Adam optimization algorithm is used to optimize the model parameters to minimize the loss function value.

[0085] According to the test results of the network model on the validation set, the hyperparameters are fine-tuned to optimize the model. The prediction results of the model location are evaluated by the mean absolute error (MAE). MAE represents the average value of the difference between the predicted location and the true location. The expression of MAE is:

[0086]

[0087] Finally, a learning rate of 0.0001 and a batch size of 32 were selected as the hyperparameters of the convolutional neural network model.

[0088] The test data of single-point, two-point, and three-point leaks were respectively predicted through three trained leak location prediction models to obtain multiple prediction results. The magnitudes of the prediction error values of each model were compared, and the prediction result with a smaller error was selected as the final multi-point leak location result. When using different models for prediction, there may be cases where null values appear in the prediction results due to the inconsistency between the number of leak positions in the test samples and the number of model outputs, that is, model mismatch, which means that position prediction cannot be performed. For such cases, these prediction results are directly discarded, and only the errors of the location results without null values, that is, the prediction models can work normally, are compared.

[0089] Finally, the MAEs of the test data of single-point, two-point, and three-point leaks were 57.12m, 105.30m, and 102.14m respectively, all within the acceptable range.

[0090] In summary, the above-described embodiment of the distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network adopts different selection schemes for training and test positions and data acquisition amounts for different numbers of disturbance points. The number of samples at each training position decreases with the increase in the number of disturbance points, and the number of samples at each test position is the same. The collected time-domain sensing signals are respectively subjected to conventional preprocessing and weight redistribution. A certain number of sample data processed as described above are randomly selected from each training position to form a training set and a validation set respectively, and the ratio of the number of samples in the two data sets decreases with the increase in the number of disturbance points. The training samples are input into the convolutional neural network regression model to complete the training of prediction models for different numbers of disturbance positions, and the validation set is used for parameter optimization. The test samples are tested with multiple models, and the test result with the smallest error is selected as the disturbance position of the sample. The present invention can achieve accurate positioning of multi-point disturbances with a relatively small amount of data.

[0091] The above has described the embodiments of the present invention in conjunction with the drawings. However, the present invention is not limited to the above embodiments, and various changes can be made according to the purpose of the invention of the present invention. Any changes, modifications, substitutions, combinations, or simplifications made based on the spirit and principle of the technical solution of the present invention shall be equivalent replacement methods, as long as they conform to the invention purpose of the present invention and do not depart from the technical principle and inventive concept of the present invention, they all belong to the protection scope of the present invention.

Claims

1. A distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network, characterized in that It includes the following steps: 1) Select the positions for data acquisition, and take single-point, two-point, and more than three-point perturbations respectively; 2) Collect the time-domain signals output by the sensing system multiple times at each selected position. The number of samples at each training position decreases with the increase in the number of perturbation points, while the sampling rate is correspondingly increased to ensure the positioning accuracy while reducing the data acquisition volume. The number of samples at each test position is the same; 3) Perform conventional preprocessing on the collected data; 4) Re-distribute the weights of the preprocessed time-domain input signals through a weight re-distribution network to extract the features of the sensing signals and improve the positioning accuracy; 5) Randomly select a certain number of processed sample data from each training position to form a training set and a validation set respectively. The ratio of the number of samples in the training set and the validation set decreases with the increase in the number of perturbation points; the processed sample data at the test position is used as the test set; 6) Input the training set samples with single-point, two-point, and more than three-point perturbations into the convolutional neural network regression model respectively to complete the training of multiple perturbation position prediction models; use the validation set for parameter optimization to obtain the best training effect; 7) Input the test samples into the trained perturbation position prediction models respectively to obtain multiple prediction results, and select the result with the smallest error as the perturbation position corresponding to the test sample to achieve multi-point perturbation positioning.

2. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, wherein In the step 1), for single-point perturbation, select the training positions at a fixed interval with the maximum acceptable positioning error, and randomly select the same number of other sensing positions as the test positions; for two-point perturbation, appropriately increase the fixed interval to d meters, take all the positions where the two-point perturbation appears as the training positions, and take two new positions with an interval smaller than the minimum interval of the training positions, the same as the training positions, and random position intervals as the test positions, and the number is appropriately reduced compared to the training positions; for more than three-point perturbations, evenly divide the sensing optical fiber into several segments, select part of the optical fiber in each segment as the data acquisition interval, further divide it at an interval of d meters, and take all the positions where multiple perturbations appear as the training positions; take the new positions within the position interval involved in the training positions as the test positions. Similarly, the number of test positions is appropriately reduced compared to the training positions.

3. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 2, characterized in that In the step 1), the method for selecting the data acquisition interval is: uniformly select half of the optical fiber segment far from or close to the end of the sensing optical fiber as the data acquisition interval.

4. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, wherein In the step 3), the conventional preprocessing uses at least one of low-pass filtering, downsampling, and normalization methods.

5. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, characterized in that In the step 4), the weight re-distribution network structure is one layer of global average pooling layer and two layers of fully connected layers.

6. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 5, characterized in that, In the step 4), first extract the global features of the input signal through global average pooling, then use two layers of fully connected networks to generate normalized attention weights, multiply the weights by the input signal, and obtain the sensing signal with re-distributed weights.

7. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, characterized in that, In the step 6), during the regression training process, select the mean square error as the loss function; use the adaptive moment estimation Adam optimization algorithm to optimize the model parameters to minimize the loss function value.

8. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, characterized in that, In step 7), the test data with single-point, two-point, and three-point perturbations are respectively predicted by three trained perturbation position prediction models to obtain multiple prediction results. Compare the magnitudes of the prediction error values of each model, and select the prediction result with a smaller error as the final multi-point perturbation positioning result.

9. The distributed optical fiber sensing multi-point positioning method based on weight redistribution and convolutional neural network according to claim 1, characterized in that In step 7), when using different models for prediction, there may be a situation where null values appear in the prediction results due to the inconsistency between the number of perturbed positions of the test samples and the number of model outputs, that is, the model is not matched, resulting in the inability to perform position prediction. In this case, directly discard these prediction results and only compare the prediction results of the models without null values, that is, the errors of the positioning results where the prediction model can work properly.

Images