Method for identifying small-scale hot zone length for distributed optical fiber temperature measurement system

Abstract

The present invention discloses a method for identifying the length of a small-scale hot zone for a distributed optical fiber temperature measurement system, including determining the cycle length of the small-scale hot zone response pattern according to the minimum sampling interval of the acquisition card. L , the cycle length L is equally divided into n parts, and the length of each part is L / n ; determine m the lengths of the small-scale hot zones to be identified, which are respectively marked as l 1 ‑l m ; construct training samples; train a fully connected neural network to obtain a small-scale hot zone length identification model; use the small-scale hot zone length identification model to identify the small-scale hot zones to be identified. The present invention solves the technical problem that the traditional signal processing method of the distributed optical fiber temperature measurement system cannot accurately determine the length of the small-scale hot zone. This method, combined with the total variation deconvolution algorithm, can greatly improve the spatial resolution of the distributed optical fiber temperature measurement system without supervision, enabling it to be better applied to small-scale monitoring scenarios and expanding the application range of the distributed optical fiber temperature measurement system.

Description

Technical Field

[0001] The present invention relates to a method for identifying the length of a hot zone of a distributed optical fiber, and particularly to a method for identifying the length of a small-scale hot zone for a distributed optical fiber temperature measurement system. Background Art

[0002] Distributed optical fiber temperature sensors based on the Raman scattering temperature sensing principle have become important safety monitoring devices in industries such as electric power, coal mines, and transportation due to their large sensing range, low deployment cost, and strong anti-interference ability.

[0003] Distributed optical fiber Raman temperature sensors mainly measure the temperature at each point along the optical fiber based on the optical time domain reflectance of the optical fiber and the backscattering Raman scattering temperature effect of the optical fiber. The main performance indicators include temperature resolution, spatial resolution, minimum sampling interval, maximum sensing distance, etc. Among them, the spatial resolution is one of the most important indicators in the practical application of distributed optical fiber Raman temperature sensors, representing the minimum distance unit that the system can resolve, that is, the shortest optical fiber length required for the system to demodulate the correct temperature and length information.

[0004] We refer to the hot zone with a length greater than the spatial resolution in the measurement scenario as a large-scale hot zone, and the hot zone with a length less than the spatial resolution as a small-scale hot zone. When a large-scale hot zone appears in the measurement scenario, its system response can accurately reflect the length and temperature of the hot zone. However, when a small-scale hot zone appears, its system response cannot reflect the true length and temperature of the hot zone. The present invention refers to this as the phenomenon of false expansion and under-response of the distributed optical fiber Raman temperature sensor to small-scale hot zones. For example, see Figures 1(a) and 1(b) in the accompanying drawings of the specification. 0.4 m hot zone and 0.6 m hot zones both correspond to 3 response points, and the minimum sampling interval of the system is 0.4 m , and the direct calculation result is 1.2 m , which is significantly inconsistent with the actual hot zone length. In the method of reconstructing small-scale hot zone signals by total variation deconvolution to improve the spatial resolution, different hot zone lengths correspond to different regularization parameters and coefficients of different weighted finite difference matrices. If the parameters do not match the hot zone length, inaccurate reconstruction will occur.

[0005] The spatial resolution of existing distributed optical fiber temperature measurement systems (RDTS) is mostly in the order of meters due to the influence of the pulse width, which severely restricts their application in small-scale scenarios. In order to enhance the applicability of distributed optical fiber Raman temperature sensors in small-scale scenarios, the width and amplitude of the system response curve of the small-scale hot zone should be consistent with the actual situation. Among them, the width of the system response curve of the small-scale hot zone represents the hot zone length, and the amplitude represents the hot zone temperature.

[0006] For the problem of insufficient amplitude response in small-scale hot regions, it can be solved by the method of reconstructing small-scale hot region signals through total variation deconvolution. However, for the problem of excessive response width in small-scale hot regions, there is currently no corresponding research and method, and the process of reconstructing small-scale hot region signals through total variation deconvolution requires knowing the length of the hot region to be carried out correctly. Therefore, currently, it is first necessary to enable the distributed fiber Raman temperature sensor to have the ability to identify the length of small-scale hot regions in order to break through the bottleneck of low spatial resolution and thus have a broader application scenario. Summary of the Invention

[0007] The object of the present invention is to provide a method for identifying the length of small-scale hot regions in a distributed optical fiber temperature measurement system, which can solve the problem of being unable to accurately identify the length when there are small-scale hot regions with a length less than the spatial resolution in the measurement scenario.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows: A method for identifying the length of small-scale hot regions in a distributed optical fiber temperature measurement system, the distributed optical fiber temperature measurement system includes a distributed optical fiber and a data acquisition card, and includes the following steps;

[0009] (1) Determine the cycle length of the small-scale hot region response mode according to the minimum sampling interval of the data acquisition card L , the cycle length L is equally divided into n parts, and the length of each part L / n ;

[0010] (2) Determine the lengths of m small-scale hot regions to be identified, and mark them as l 1 -l m ;

[0011] (3) Construct training samples, including steps (31)-(35);

[0012] (31) Randomly set m small-scale hot regions on the distributed optical fiber, with lengths of l 1 -l m , heat the small-scale hot regions on a constant temperature heating table, and use the distributed optical fiber temperature measurement system to perform k times of sampling, and obtain a set of system responses each time of sampling, and a total of k sets of system responses are obtained. The system response contains Q response data of response points, Q = fiber length / minimum sampling interval;

[0013] (32) Move the distributed optical fiber in the direction away from its head end L / n by a distance, and perform k sub-sampling again to obtain k sets of system responses;

[0014] (33) Repeat step (32) until moving n-1 times, and a total of n×k sets of system responses are obtained;

[0015] (34) For a set of system responses, which altogether contain m response data of small-scale hot spots, for each small-scale hot spot, with the response point having the highest amplitude as the center, take its first a points and the following b points to determine an interval containing a + b + 1 points. The response data within this interval forms a training sample, and use the length of this small-scale hot spot as the label of this training sample. Altogether, m training samples are obtained;

[0016] (35) Operate on all system responses according to step (34), and a total of n×k×m training samples are obtained;

[0017] (4) Feed the training samples into the fully connected neural network in sequence, and use the label of this training sample as the expected output to train the fully connected neural network to obtain a small-scale hot spot length recognition model;

[0018] (5) Obtain the response data of the small-scale hot spot to be recognized, determine an interval with the response point having the highest amplitude as the center, form the response data within this interval into a test sample, and feed it into the small-scale hot spot length recognition model to output the length of this small-scale hot spot.

[0019] As a preference: In step (1), the minimum sampling interval L S =v / 2 f , v represents the propagation speed of light in the optical fiber, f is the sampling frequency of the acquisition card, and L = L S .

[0020] As a preference: In step (34), a is 4 - 9, b is 5 - 10.

[0021] As a preference: Step (4) trains the fully connected neural network as;

[0022] The training samples are sequentially input into the fully-connected neural network. After calculation by the hidden layer, the loss function is updated, and the updated loss function is backpropagated to the hidden layer to continuously adjust the parameters in the model. After the number of cycles reaches the preset number of training rounds, the update is stopped and a small-scale hot zone length recognition model is formed. The loss function uses the cross-entropy loss function.

[0023] Compared with the prior art, the advantages of the present invention are as follows:

[0024] (1) It solves the technical problem that in the monitoring scenario, there are small-scale hot zones with lengths less than the device spatial resolution, and the traditional signal processing method of the distributed optical fiber temperature measurement system cannot accurately determine the length of the hot zone. By utilizing the phenomenon that the distributed optical fiber temperature measurement system has a periodic cyclic response pattern to small-scale hot zones, special training samples are constructed, and the response patterns included in the training set are expanded, greatly improving the generalization ability of the model, so as to accurately identify the length of the small-scale hot zone.

[0025] (2) Since the response pattern is only related to the length of the small-scale hot zone and has nothing to do with the temperature of the hot zone, the model trained with the training samples constructed based on the change law of the response pattern still has strong generalization ability at different temperatures, and when constructing the training samples, there is no need to consider too much the influence brought by the temperature accuracy. It only requires that the optical fiber of the hot zone is in the heating state, further reducing the operation difficulty.

[0026] (3) Based on the ability of this method to accurately identify the length of the small-scale hot zone, combined with the total variation deconvolution algorithm, the spatial resolution of the distributed optical fiber temperature measurement system can be effectively improved, enabling it to be better applied to small-scale monitoring scenarios and expanding the application range of the distributed optical fiber temperature measurement system. Description of the Drawings

[0027] Figure 1 a For the prior art 0.4 m Hot zone length RDTS response;

[0028] Figure 1 b For the prior art 0.6 m Hot zone length RDTS response;

[0029] Figure 2 Is the flow schematic diagram of the present invention;

[0030] Figure 3 For 5 groups with a length of 40 cm Schematic diagram of the difference in the response patterns of small-scale hot zones;

[0031] Figure 4 a Is the schematic diagram of the change law of the first-round cycle of the hot zone response pattern when the system sampling interval is 40 cm ;

[0032] Figure 4 b The system sampling interval is 40 cm Schematic diagram of the second-round cyclic change law of the hot zone response mode

[0033] Figure 5 The curve of the hot zone temperature changing with time Specific implementation mode

[0034] The present invention will be further described below in conjunction with the accompanying drawings

[0035] Example 1, see Figure 1a and Figure 1b The spatial resolution of the distributed optical fiber temperature measurement system is jointly determined by the light source, APD, and acquisition card. In this example, Zhang Hong, we use the spatial resolution of the distributed optical fiber temperature measurement system as 1.6 m The system performance indicators are as shown in Table 1 below

[0036] Table 1 System performance indicators of the distributed optical fiber temperature measurement system

[0037]

[0038] The distributed optical fiber temperature measurement system is also called RDTS. Based on this distributed optical fiber temperature measurement system, we obtain the RDTS response of the 0.4 m hot zone length and the RDTS response of the 0.6 m hot zone length. As shown in Figure 1a and Figure 1b The 0.4 m hot zone and the 0.6 m hot zone both correspond to 3 response points, and the minimum sampling interval of the system is 0.4 m The direct calculation result is 1.2 m which is significantly inconsistent with the actual hot zone length

[0039] Example 2: See Figure 2 -Figure 4, a method for identifying the length of a small-scale hot zone for a distributed optical fiber temperature measurement system, the distributed optical fiber temperature measurement system includes a distributed optical fiber and an acquisition card, and includes the following steps

[0040] (1) Determine the cycle length of the small-scale hot zone response mode according to the minimum sampling interval of the acquisition card L The cycle length L is equally divided into n parts, and the length of each part L / n ;

[0041] (2) Determine the lengths of m small-scale hot zones to be identified, and mark them respectively as l 1-l m ;

[0042] (3) Construct training samples, including steps (31)-(35);

[0043] (31) Randomly set m small-scale hot zones on the distributed optical fiber, with lengths of l 1 -l m , place the small-scale hot zones on a constant-temperature heating table for heating, and use the distributed optical fiber temperature measurement system to perform k samplings. Each sampling obtains a set of system responses, and a total of k sets of system responses are obtained. The system responses contain Q response data of response points, Q = fiber length / minimum sampling interval;

[0044] (32) Move the distributed optical fiber L / n distance in the direction away from its head end, and perform k samplings again to obtain k sets of system responses;

[0045] (33) Repeat step (32) until it is moved n-1 times, and a total of n×k sets of system responses are obtained;

[0046] (34) For a set of system responses, which contains m response data of small-scale hot zones, for each small-scale hot zone, with the response point with the highest amplitude as the center, take its first a points and the last b points to determine an interval containing a + b + 1 points. The response data within this interval constitutes the training sample, and use the length of this small-scale hot zone as the label of this training sample. A total of m training samples are obtained;

[0047] (35) Perform the operation according to step (34) on all system responses, and a total of n×k×m training samples are obtained;

[0048] (4) Feed the training samples into the fully connected neural network in sequence, and use the label of this training sample as the expected output to train the fully connected neural network to obtain a small-scale hot zone length recognition model;

[0049] (5) Obtain the response data of the small-scale hot zone to be recognized, determine an interval with the response point with the highest amplitude as the center, and the response data within this interval constitutes the test sample. Feed it into the small-scale hot zone length recognition model to output the length of this small-scale hot zone.

[0050] Among them, the minimum sampling interval in step (1) L S =v / 2 f , v represents the propagation speed of light in the optical fiber, f is the sampling frequency of the acquisition card, and L = L S .

[0051] In step (34), a is 4 to 9, b is 5 to 10.

[0052] Step (4) trains the fully connected neural network as;

[0053] The training samples are sequentially input into the fully connected neural network. After calculation by the hidden layer, the loss function is updated, and the updated loss function is backpropagated to the hidden layer to continuously adjust the parameters in the model; after the number of cycles reaches the preset number of training rounds, the update is stopped and a small-scale hot zone length recognition model is formed; the loss function uses the cross-entropy loss function.

[0054] This embodiment is based on the fully connected neural network classification model. The principle of this model is to input the amplitude arrangement of the response signal into the network, and use the difference in the morphology between the response signals of different hot zone lengths as features to classify the hot zone length.

[0055] The basis for the model to classify the hot zone length is based on the morphology of the hot zone response signal. Therefore, when constructing the training set, the differences in the morphology of the hot zone response signals of different lengths should be considered emphatically. Observing a large number of experimental results, it is found that the response curves of the distributed fiber Raman temperature sensor to small-scale hot zones of the same length will show different morphologies, which the present invention calls different response modes. See Figure 3 , five groups of 40 cm long hot zones are simultaneously placed in the heating device to heat up, but their positions on the optical fiber are randomly arranged. After the heating device is heated to 70 ℃ , 30,000 groups of data are collected. After averaging, the data of the five groups of hot zones are extracted, and the abscissas are aligned to obtain Figure 3 . From Figure 3 it can be seen that; the response modes of the 5 groups are different, and the fully connected neural network model will not be able to correctly identify the length of the prediction object.

[0056] In the present invention, the construction of the training samples will fully consider the influence brought by the response mode, and study the cyclic law of the response mode. See Figure 4a and Figure 4b, by changing the position of the small-scale hot zone of the same length on the optical fiber multiple times, so that its response can include various response modes that may appear. Starting from Figure 4a As can be seen, randomly select a location on the optical fiber to arrange a small-scale hot zone with a length of 40 cm . The measured response curve is used as the initial response curve. After moving the hot zone backward by 40 cm , the distribution of the response curve obtained is completely consistent with the initial response, which is recorded as the first round of cycle here. The above phenomenon indicates that the response mode of the small-scale hot zone returns to the initial state after one round of movement of the hot zone. To further verify the law of response mode change, continue to move the hot zone backward to obtain Figure 4b . After moving the optical fiber backward by 50 cm , its response mode is the same as that after moving backward by 10 cm in the first round of cycle, and so on. After completing the second 40 cm movement, it returns to the initial corresponding mode. Figure 4a And Figure 4b Prove that there is a regularity in the change of the response mode of the distributed optical fiber temperature measurement system to the small-scale hot zone. In the training set of the fully connected neural network, the shape of each sample curve, that is, the response mode, must be fully considered as an important feature for its classification. The training set constructed in this way solves the problem of insufficient generalization ability of the small-scale hot zone length recognition model.

[0057] Example 3: Refer to Figures 2 to 5 . Based on Example 2, we give a specific value.

[0058] (1) According to the minimum sampling interval of the acquisition card, which is 40 cm , determine the cycle length L of the response mode of the small-scale hot zone as 40 cm , and divide 40 cm equally into n = 4 parts, with each part having a length of 10 cm ;

[0059] In step (1), the minimum sampling interval L S =v / 2 f , v represents the propagation speed of light in the optical fiber. In the optical fiber with SiO 2 as the base material, usually take 2×10 8 m / s , f is the sampling frequency of the acquisition card = 250 MHz , and L = L S .

[0060] (2) Specifically, it is divided into steps (21)-(22);

[0061] (21) Determine m= the lengths of 5 small-scale hot spots to be identified, respectively marked as l 1 -l 5 , where l 1 = 40 cm 、 l 2 = 50 cm 、 l 3 = 60 cm 、 l 4 = 80c m 、 l 5 = 100 cm ;

[0062] (22) In order to obtain more system responses, lay 5 copies of small-scale hot spots with a length of l 1 -l 5 on the distributed optical fiber. It can be regarded that there are 25 segments of small-scale hot spots on the distributed optical fiber, that is, 25 groups of system responses of small-scale hot spots can be obtained each time of acquisition.

[0063] (3) Specifically, it is divided into steps (31)-(35);

[0064] (31) The same as step (31) of Embodiment 2, the optical fiber is 10 km , and the minimum sampling interval is 0.4 m , then the number of effective response points for one acquisition is Q = 100000 / 0.4 = 250000.

[0065] (32) Move the distributed optical fiber 10 cm in the direction away from its head end, and perform k times of sampling again to obtain k groups of system responses;

[0066] (33) Repeat step (32), move a total of 3 times, and a total of 4 ×k groups of system responses are obtained;

[0067] (34)Same as step (34) of Embodiment 2. In this embodiment, in a set of system responses, a total of 25 small-scale hot spots are included. For each small-scale hot spot, with the response point having the highest amplitude as the center, the first 4 response points and the next 5 response points are selected to form an interval with a length of 10 response points. The response data within this interval constitutes a training sample, and the length of this small-scale hot spot is used as the label of this training sample. A total of 25 training samples are obtained;

[0068] (35)Operate on all system responses according to step (34), and a total of n×k×m = 4 × 33568 × 25 = 3356800 training samples are obtained. Here, k= 33568 is generally set according to the actual required data volume.

[0069] (4)Same as step (4) of Embodiment (2), a small-scale hot spot length recognition model is obtained. The specific training method includes steps (41)-(42):

[0070] (41)Training set segmentation: 97% of all training samples are used as the training set, and 3% are used as the test set. The training set is the data sample for model fitting. During training, the gradient descent is performed on the training error to adjust the weight parameters of each neuron. The validation set is used to adjust the hyperparameters of the model and conduct a preliminary evaluation of the model's capabilities.

[0071] (42)Hyperparameter setting: Through continuous attempts, the following various hyperparameters of this network are finally determined. The batch size is 1024, the number of training epochs Epoch is 200, the learning rate a is 0.001, the optimizer is selected as "Adam", the loss function is selected as the cross-entropy loss function, the activation function of the hidden layer is "Relu", and the activation function of the output layer is "Softmax".

[0072] (5)Same as step (5) of Embodiment 2, except that when constructing the sample to be measured, the sample to be measured also includes the corresponding data of 10 response points.

[0073] Finally, in this embodiment, the number of training set samples is 3356800 groups, the hot spot responses of the test set are randomly distributed, and the number of samples is 150000 groups. We call the k groups of system responses obtained for the first time in step (32) of Embodiment 3 the first large group, and the cm obtained after moving the hot spot by 10 kThe group of system responses is called the second largest group, and so on. After moving three times, a total of four large groups of system responses are obtained. The training sets of each of these four large groups of system responses are extracted according to step (34), and in addition, the training sets extracted from the four large groups are combined as the training set containing all the response patterns of this embodiment. The above various types of training sets are respectively input into the training layer of the neural network for training to verify the influence of the response pattern on the model generalization. The model performance is shown in Table 2.

[0074] Table 2: Comparison table of model prediction accuracy rates obtained from different training sets

[0075]

[0076] The prediction results of each hot zone length recognition model for the same group of test sets vary. The key reason is whether the response pattern existing in the test set is included in the model. The recognition model obtained from the training set containing all the data has the best comprehensive performance, and the total correct rate reaches 87.66%. This result also proves that the response pattern problem of the distributed optical fiber temperature measurement system for small-scale hot zones is a point that cannot be ignored in the process of studying small-scale temperature events.

[0077] The above results show that the model trained by this method can effectively identify the length of small-scale hot zones in practical applications. The functions of the length of small-scale hot zones include but are not limited to: being used to assign the best parameters in signal reconstruction calculation, being used to understand the detailed distribution of small-scale hot zones, etc. Since this method does not require any hardware modification, has no additional cost and has a high correct rate, it can be widely applied to commercial distributed optical fiber temperature measurement systems to expand the application range and improve the measurement accuracy.

[0078] In the temperature setting of the training set construction experiment, there is no need to precisely control the temperature of the hot zone. During the experiment, it is only necessary to ensure that the optical fiber in the hot zone is heated. However, in applications, it is recommended to control the temperature of the hot zone to fluctuate within a certain range to further enhance the temperature generalization of the model. The specific implementation method is to place the constant temperature heating table where the optical fiber in the hot zone is located on the smart socket and execute the cyclic switch task, then the temperature can be controlled to fluctuate within a certain range. The temperature fluctuation curve of the hot zone over time is as Figure 5 shown.

[0079] The above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for identifying the length of a small-scale hot zone in a distributed optical fiber temperature measurement system, the distributed optical fiber temperature measurement system including a distributed optical fiber and a data acquisition card, Characterized in that: It includes the following steps; (1) Determine the cycle length of the small-scale hot zone response pattern according to the minimum sampling interval of the acquisition card L , the cycle length L is equally divided into n parts, and the length of each part L / n , where the minimum sampling interval L S =v / 2 f , v represents the propagation speed of light in the optical fiber, f is the sampling frequency of the acquisition card, and L = L S ; (2)Determine m the lengths of the small-scale thermal regions to be recognized, and mark them respectively as l 1 -l m ; (3) Construct training samples, including steps (31)-(35); Randomly set m small-scale hot zones on the distributed optical fiber, with lengths of l 1 -l m , place the small-scale hot zones on a constant-temperature heating table for heating, and use a distributed optical fiber temperature measurement system to perform k times of sampling. Each sampling obtains a set of system responses, and a total of k sets of system responses are obtained. The system response contains Q response data of response points, Q = optical fiber length / minimum sampling interval; Move the distributed optical fiber in a direction away from its head end L / n by a distance, and perform k subsampling again to obtain k groups of system responses; Repeat step (32) until the movement n- is done once, and a total of n×k groups of system responses are obtained; (34) For a set of system responses, which altogether contain m response data of small-scale hot zones, for each small-scale hot zone, taking the response point with the highest amplitude as the center, taking its first a points and the subsequent b points, determining an interval containing a + b + 1 points, forming the response data within this interval into a training sample, using the length of this small-scale hot zone as the label of this training sample, and altogether obtaining m training samples; (35)Operate on all system responses according to step (34), and a total of n×k×m training samples are obtained; (4) Send the training samples into the fully connected neural network in sequence, and train the fully connected neural network with the label of the training sample as the expected output to obtain a small-scale hot zone length recognition model; (5) Obtain the response data of the small-scale hot zone to be recognized, determine an interval with the response point having the highest amplitude as the center, and form a test sample with the response data within this interval, and send it into the small-scale hot zone length recognition model to output the length of the small-scale hot zone.

2. The method for identifying the length of a small-scale hot zone in a distributed optical fiber temperature measurement system according to claim 1, Characterized in that: In step (34), a is from 4 to 9, b is from 5 to 10.

3. The method for identifying the length of a small-scale hot zone in a distributed optical fiber temperature measurement system according to claim 1, Characterized in that: Step (4) trains the fully connected neural network as; Input the training samples into the fully connected neural network in sequence, update the loss function after calculation by the hidden layer, and backpropagate the updated loss function to the hidden layer to continuously adjust the parameters in the model; after the number of cycles reaches the preset number of training rounds, stop updating and form a small-scale hot zone length recognition model; the loss function uses the cross-entropy loss function.

Images